WO2020108605A1 - 一种推荐方法、装置及存储介质 - Google Patents

一种推荐方法、装置及存储介质 Download PDF

Info

Publication number
WO2020108605A1
WO2020108605A1 PCT/CN2019/121919 CN2019121919W WO2020108605A1 WO 2020108605 A1 WO2020108605 A1 WO 2020108605A1 CN 2019121919 W CN2019121919 W CN 2019121919W WO 2020108605 A1 WO2020108605 A1 WO 2020108605A1
Authority
WO
WIPO (PCT)
Prior art keywords
attention
social
group
candidate item
candidate
Prior art date
Application number
PCT/CN2019/121919
Other languages
English (en)
French (fr)
Inventor
张旭
林乐宇
葛凯凯
唐琳瑶
刘雨丹
陈鑫
闫肃
庄凯
王伟
张晶
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Priority to JP2021527278A priority Critical patent/JP7299320B2/ja
Publication of WO2020108605A1 publication Critical patent/WO2020108605A1/zh
Priority to US17/183,251 priority patent/US11709902B2/en
Priority to US18/309,300 priority patent/US20230259566A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/01—Social networking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9536—Search customisation based on social or collaborative filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition

Definitions

  • This application relates to the field of Internet technology, and in particular, to a recommendation method, device, and storage medium.
  • AI Artificial Intelligence
  • Artificial Intelligence is a theory, method, technology, and application system that uses digital computers or digital computer-controlled machines to simulate, extend, and expand human intelligence, sense the environment, acquire knowledge, and use knowledge to obtain the best results.
  • artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence.
  • Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machine has the functions of perception, reasoning and decision-making.
  • Artificial intelligence technology is a comprehensive subject, covering a wide range of fields, both hardware-level technology and software-level technology.
  • Basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, and mechatronics.
  • Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning/deep learning.
  • Machine learning is a multidisciplinary interdisciplinary subject, involving multiple subjects such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and so on. Specially study how the computer simulates or realizes human learning behavior to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve its performance.
  • Machine learning is the core of artificial intelligence, and is the fundamental way to make computers intelligent, and its applications are in various fields of artificial intelligence.
  • Machine learning and deep learning usually include technologies such as artificial neural networks, confidence networks, reinforcement learning, transfer learning, inductive learning, and pedagogical learning.
  • Projects on the Internet refer to data information used for consumer consumption, participation, and behavioral interaction in the Internet. For example, products, articles, advertisements, and virtual information.
  • different users have different needs for the project. In this way, different items need to be recommended to different users. For example, in some recommendation methods, the user's preferences can be mined based on the user's friend information.
  • Embodiments of the present application provide a recommendation method, device, and storage medium for improving the accuracy of recommending items that the user is interested in to the user, avoiding recommending items that the user is not interested in to the user multiple times, improving resource utilization, and thus improving the user Experience.
  • a recommendation method provided by an embodiment of the present application executed by a computing device, includes:
  • an embodiment of the present application provides a recommendation device, including:
  • the obtaining module is used to obtain candidate items to be recommended to users of the social platform;
  • the first determining module is configured to determine, for at least one target social object of each type of social relationship, of the at least two different types of social relationships of the social platform user in the social platform, respectively Candidate's single attention degree;
  • a second determination module configured to determine the comprehensive attention degree of the target social object of different types to the candidate item according to the single attention degree of each target social object to the candidate item;
  • the third determining module is configured to determine whether to recommend the candidate item to the social platform user according to the comprehensive attention degree.
  • an embodiment of the present application provides a computing device, including at least one processor and at least one memory, wherein the memory stores a computer program, and when the program is executed by the processor, the The processor executes the steps of the recommendation method in the embodiment of the present application.
  • an embodiment of the present application provides a storage medium that stores computer instructions.
  • the computer instructions run on a computer, the computer is caused to perform the steps of the recommended method in the embodiments of the present application.
  • the recommendation method in the embodiment of the present application when obtaining candidate items to be recommended to a user of a social platform, at least one goal in each type of social relationship among at least two different types of social relationships of the user in the social platform is targeted to the user Social objects, respectively determine the single attention degree of the target social object to the candidate item, that is, the social object that fully utilizes the user's multiple social relationships in the social platform to pay attention to the candidate item, and then according to the individual target social object's single item of the candidate item Attention, determine the comprehensive attention of different types of target social objects to the candidate item, and then determine whether to recommend the candidate item to the social platform user according to the comprehensive attention, so it can improve the accuracy of recommending items to the user Degree, try to avoid sending users items that are not of interest to users, thereby improving resource utilization and improving user experience.
  • FIG. 1 is a schematic diagram of an application scenario provided by an embodiment of the present application
  • FIG. 3 is a schematic structural diagram of a project recommendation model provided by an embodiment of this application.
  • FIG. 4 is a schematic diagram of interaction between a WeChat user and a friend, group, or candidate item to be recommended provided by an embodiment of the present application;
  • FIG. 5 is a schematic diagram of a second attention model architecture provided by an embodiment of this application.
  • FIG. 6 is a schematic diagram of obtaining the importance weight of each group member of a group provided by an embodiment of the present application.
  • FIG. 7 is a schematic structural diagram of a first attention model provided by an embodiment of this application.
  • FIG. 8 is a schematic diagram of an accuracy evaluation result of a recommended item provided by an embodiment of this application.
  • FIG. 9 is a schematic diagram of a recommendation device provided by an embodiment of this application.
  • FIG. 10 is a schematic diagram of a computing device provided by an embodiment of the present application.
  • artificial intelligence technology has been researched and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, autonomous driving, drones , Robots, intelligent medical care, intelligent customer service, voice recognition, etc., I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important value.
  • the user’s preferences can be determined based on the user’s friends’ information, but these methods only take into account information about some of the user’s friends.
  • the user does not necessarily have the same or similar preferences as the part of the friend. Mining user preferences from the perspective of friends cannot obtain accurate user preference information. Therefore, the above recommendation method has a technical problem that the accuracy of item recommendation is low and affects the user experience.
  • SACF Social Attentional Collaborating Filtering
  • Deep learning It is a new field in machine learning research. Its motivation is to establish and simulate the human brain to analyze and learn neural networks. It mimics the mechanism of the human brain to interpret data, such as images, sounds, and text.
  • Attention Mechanism a branch of deep learning, can also be called neural attention mechanism, derived from the study of human vision.
  • cognitive science due to the bottleneck of information processing, humans will selectively pay attention to a part of all information while ignoring other visible information.
  • the above mechanism is usually called the attention mechanism.
  • Different parts of the human retina have different levels of information processing capabilities, only the foveal part of the retina has the strongest acuity.
  • humans need to select a specific part in the visual area and then focus on it. For example, when people read, usually only a small number of words to be read will be paid attention to and processed.
  • the attention mechanism has two main aspects: deciding which part of the input needs attention; allocating limited information processing resources to important parts.
  • Attention refers to the ability to ignore some information while paying attention to some information.
  • the inventor of the present application considers that social objects with different social relationships in the social platform have different influences on user behavior, for example, when a user needs to buy basketball shoes, they may follow the advice of a basketball friend when it comes to travel At this time, the user may turn to the suggestions of friends who like to travel in the social platform, and the user who plays basketball or the friend who likes to travel may be a friend with a friend relationship in the user's social platform, or may have a group relationship for the user to join
  • the members of the group of therefore, when recommending items to users, can combine the attention information of the social objects with different social relationships in the user's social platform on the items to improve the accuracy of recommending items to users.
  • the attention mechanism can be introduced into the project recommendation field to learn the "usefulness" of various social relationships in the user's social platform for project reviews
  • this application takes into account the different impacts of social objects in different social relationships on the social platform on user behavior, targeting different Social objects of social relationships, based on the attention mechanism, designed an attention model suitable for learning social objects to treat recommended items, so as to improve the learning effect of social objects of different social relationships on recommended items.
  • the attention mechanism considering the different activities of different group members in the group, the more active group members have a greater impact on the group, the higher the importance of the group.
  • the attention information of the group members with higher category importance on the candidate items can affect the user's interest in the candidate items, so the importance mechanism weight of each group member in the group is combined in the introduced attention mechanism to
  • the interest of the learning group on the item is to improve the accuracy of the group's attention to the item obtained by learning, and thus to improve the accuracy of recommending the item to the user.
  • an embodiment of the present application provides an item recommendation method.
  • the method When obtaining candidate items to be recommended to a social platform user, the method will target the user in at least two different types of social relationships in the social platform. At least one target social object in a type of social relationship, respectively determining the single attention degree of the target social object to the candidate item, that is, making full use of the user's attention information on the candidate item by the social objects of multiple social relationships in the social platform, and then The single attention degree of each target social object to the candidate item to determine the comprehensive attention degree of different types of target social objects to the candidate item, wherein the single attention degree of each target social object to the candidate item is attention based on the attention mechanism
  • the model training obtained that is, the learning advantage of the attention mechanism was fully utilized to accurately learn the single attention degree of each target social object for the candidate item, and then determine whether to recommend the candidate item to the social platform user according to the comprehensive attention degree Therefore, it can improve the accuracy of recommending items to users, avoid recommending items to users that are not of interest to users, improve resource
  • the item recommendation method in the embodiment of the present application may be applied to the application scenario shown in FIG. 1, which includes the user terminal 10 and the item recommendation computing device.
  • the item recommendation computing device The server device 11 may be recommended for the project, and the project recommendation server device 11 may be a server device, or a server device cluster or cloud computing center composed of several server devices.
  • the item recommendation server device 11 described in FIG. 1 may include the computing device shown in FIG. 10.
  • the user terminal 10 may be any smart terminal device that can run according to a program and automatically process large amounts of data at high speed.
  • Such terminal devices such as computers, ipads, mobile phones, etc., are installed on the user terminal 10 corresponding to the social platforms in the embodiments of the present application.
  • Social applications Application, APP
  • the social APP consists of a background server device that supports its operation, and the background server device can be a server device, or It may be a server device cluster or a cloud computing center composed of several server devices.
  • the backend server device supporting the operation of social APP and the project recommendation server device 11 may be an integrated server device cluster integrated together, or may be each shown in FIG. 1 An independent server device, the background server device that supports the operation of social APP in FIG. 1 is labeled 12.
  • the user terminal 10 is connected to the background server device 12 and the project recommendation server device 11 supporting social APP operation through a network, respectively, and the background server device 12 supporting the social APP operation and the project recommendation server device 11 are connected through a network, so that the user terminal 10 supports
  • the background server device 12 running the social APP and the project recommendation server device 11 can communicate with each other.
  • the network may be any one of communication networks such as a local area network, a wide area network, or a mobile Internet.
  • the item recommendation method can be applied to the item recommendation server device.
  • the item recommendation server device can target at least two different types of social platform users in the social platform In the social relationship of, at least one target social object in each type of social relationship determines the single attention of the target social object to the candidate item, and then determines the different types of target social objects according to the single attention of the target social object to the candidate item.
  • the target social object's comprehensive attention to the candidate item, and according to the comprehensive attention degree determine whether to recommend the candidate item to the social platform user, and when determining to recommend the candidate item to the social platform user, interact with the backend server device supporting the social platform,
  • the recommended candidate items are presented to the users of the social platform through the social platform in the user terminal.
  • the item recommendation method provided by the embodiment of the present application may be executed by the item recommendation server device shown in FIG. 1, or may be executed by the computing device shown in FIG. 10.
  • the item recommendation method includes:
  • Step 201 Obtain candidate items to be recommended to social platform users.
  • the social platform in step 201 refers to software for socializing on the Internet, such software as WeChat, QQ, etc.
  • candidate items refer to those on the Internet for user consumption, participation, or behavioral interaction
  • Data information such as commodities, articles, advertisements, virtual information, or points of interest (POI), in the embodiments of the present application does not specifically limit the type of candidate items, and can be determined according to actual conditions.
  • step 201 obtains the number of candidate items to be recommended to the social platform user, which may be one or multiple.
  • the number of obtained candidate items to be recommended to the social platform user is multiple,
  • the types of multiple candidate items may be the same or different, and are not specifically limited herein.
  • Step 202 Determine that the social platform user has at least one target social object in each type of social relationship among at least two different types of social relationships in the social platform.
  • social platform users for convenience of narrative, the social platform users referred to below are simply referred to as users.
  • social objects with different social relationships have different influences on their behavior, therefore,
  • the user's attention information on the candidate item in multiple social objects of different social types in the social platform can be comprehensively utilized to improve the accuracy of recommending the candidate item to the user.
  • the user may first determine at least two different types of social relationships in the social platform, and determine at least one target social object from each type of social relationship, for example, when the social platform is specifically a WeChat platform ,
  • the target social object can be a WeChat friend added by the user and have a friend relationship, or a WeChat group that has a group relationship added by the user, or an enterprise WeChat connection that has a business relationship between the user and the user People and so on.
  • the at least one target social object of at least two different types of social relationships determined in step 202 includes at least at least one of the user's joining in the social platform and having a group relationship (also referred to as group social relationship).
  • group social relationship also referred to as group social relationship
  • the user and the user considering that the user’s friend in the social platform has a friend relationship, the user and the user also share common needs or common characteristics in certain aspects. Therefore, the user added in the social platform and having The attention information of the friends of the friend relationship on the candidate items further improves the accuracy of recommending the candidate items to the user. Therefore, the at least one target social object of at least two different types of social relationships determined in step 202 also includes at least one friend added by the user in the social platform and having a friend relationship.
  • At least one social target social object of other types of social relationships may also be included.
  • Target social objects include, for example, at least one group that the user joins in the social platform and has a group relationship, and at least one friend that is added in the social platform and has a friend relationship.
  • Step 203 Determine the individual attention degree of the target social object to the candidate item respectively.
  • step 203 when the target social object is at least one group that the user joins and has a group owner relationship with the user, the specific execution process of step 203 includes:
  • the first attention model determines the attention parameters of each group for the candidate item according to the input vector representation of each group and the vector representation of the candidate item, and generates each group pair candidate according to the determined attention parameter The single item attention degree of the item, thereby obtaining the single item attention degree of the candidate item in each group of the at least one group output by the first attention model.
  • the first attention model Each group is trained with the importance weight of each group member in each group in the group.
  • the first attention model can determine the attention parameters of each group for the candidate item, and can also group members of each group in each group.
  • the importance weight in the group is weighted with each group's attention parameter of the candidate item, and then the single attention degree of the candidate item of the group is generated and output to enhance the obtained single item of the candidate item of the group Attention, and then improve the accuracy of recommending candidate items to users.
  • step 203 when the target social object is at least one friend added by the user in the social platform and has a friend relationship, the specific execution process of step 203 further includes:
  • the vector representation of at least one friend and the vector representation of the candidate item are input into a pre-trained second attention model, and the second attention model is pre-trained with the attention parameters of each friend of the social platform on the candidate item.
  • the second attention model determines the attention parameters of each friend to the candidate item according to the input vector representation of each friend and the vector representation of the candidate item, and generates and outputs a single item of each friend to the candidate item according to the determined attention parameter The attention degree, so as to obtain the single attention degree of the candidate item for each friend in the at least one friend output by the second attention model.
  • Step 204 Normalize each target social object's single attention degree of the candidate item.
  • Step 205 The target social objects of the same type after the normalization process are summed to the single attention of the candidate item to obtain the comprehensive attention degree of the social object of each type of social object to the candidate item.
  • the individual attention degree of each target social object to the candidate item may be normalized to facilitate subsequent calculation, and then At least one group in the article sums up the individual attention of each group to the candidate item after normalization processing to obtain the comprehensive attention degree of the candidate group by the user in the social platform; for the above At least one of the friends sums the single attention items of the normalized processing of the candidate items by each friend to obtain the comprehensive attention degree of the user's friends on the social platform for the candidate items.
  • Step 206 Perform feature fusion on the comprehensive attention degree of the candidate items of various types of social objects to obtain the recommendation index of the candidate items to be recommended.
  • the existing feature fusion technology may be used to perform feature fusion on the obtained comprehensive attention degree of each type of target social object to the candidate item. For example, if the additive fusion technology is selected, then the The comprehensive attention of the user's friends on the social platform to the candidate item and the comprehensive attention of the user's group on the social platform to the candidate item are used as the input parameters of the additive fusion algorithm, and then the additive fusion algorithm is obtained The output recommendation index of the recommended candidate item.
  • the recommendation index can be expressed as a star rating, rating score, and so on.
  • Step 207 Determine whether the recommendation index is greater than the threshold, if yes, go to step 208, otherwise go to step 209.
  • the recommendation index of the candidate item to be recommended after obtaining the recommendation index of the candidate item to be recommended, it can also be determined whether to recommend the candidate item to the social platform user according to the recommendation index. Specifically, it can be determined whether the obtained recommendation index of the candidate item to be recommended is greater than Threshold value, for example, when using the score score to represent the recommendation index, if the score score of the candidate item to be recommended is greater than the threshold value, then perform step 208: determine to recommend the candidate item to the user, otherwise perform step 209: determine not to recommend the candidate item to the user.
  • the recommendation index corresponding to each candidate item can be obtained according to the method described above, and then according to the recommendation index, when determining whether to recommend the candidate item to a social platform user, you can directly According to the recommendation index corresponding to each candidate item, it is determined that the candidate item recommended to the user is the candidate item with the highest recommendation index; it may also be sorted in order from the recommendation index from high to low, and the ranked candidate items are recommended to the user.
  • At least one target social object in each type of social relationship is targeted for the user in at least two different types of social relationships in the social platform , Separately determine the target social object's single attention to the candidate item, that is, the social object that fully uses the user's multiple social relationships in the social platform to pay attention to the candidate item, and then according to the target social object's single item of attention to the candidate item
  • the target social object's single attention to the candidate item that is, the social object that fully uses the user's multiple social relationships in the social platform to pay attention to the candidate item
  • the target social object's single item of attention to the candidate item To determine the comprehensive attention degree of different types of target social objects on the candidate items, wherein the single attention degree of each target social object on the candidate items is obtained by training an attention model based on the attention mechanism, that is, full use of attention
  • the learning advantage of the mechanism to accurately learn the individual attention degree of each target social object to the candidate item, and then determine whether to recommend the candidate item to the social platform user according to the comprehensive attention degree, so it
  • step 202 taking into account the actual application, social objects that frequently socialize with the user have a greater influence on the behavior of the user than objects occasionally socialized with the user.
  • the attention information of frequent social objects on candidate items can not only improve the accuracy of recommending candidate items to users, but also improve the speed of data processing. Therefore, at least one target social object determined in each type of social relationship in step 202 , Can be an object that socializes frequently with the user in the corresponding type of social relationship.
  • the method of determining the objects frequently socialized with the user may be determined by socializing with the user within a preset time period, for example, if socializing with the user within a week (including sending and receiving messages and sharing based on the social platform If the number of social activities reaches a threshold, the object is considered to be an object that interacts frequently with the user.
  • At least one target social object determined in each type of social relationship in step 202 may also be an object associated with a candidate item in the corresponding type of social relationship.
  • related means that the social object has interacted with the candidate item, for example, has used or paid attention to the candidate item.
  • the at least one target social object determined in each type of social relationship in step 202 may also be an object that frequently socializes with the user and is associated with the candidate item.
  • at least one target social object determined in each type of social relationship in step 202 may also be all social objects in the corresponding type of social relationship, or a part of social objects selected at random .
  • the above item recommendation method may be completed based on the item recommendation model, which includes the pre-trained first attention model and second attention model described above, specifically,
  • the social platform user who needs to recommend a project and the candidate items to be recommended to the social platform user can be input into the project recommendation model, and after performing the project recommendation method in the above embodiment through the project recommendation model, whether to recommend the candidate item to the social platform user is output
  • the output of whether to recommend the candidate items to the social platform user can be the top candidate item with the recommendation index, or it can be sorted according to the recommendation index from high to low And recommend each sorted candidate item to the user.
  • the input item recommendation model has one candidate item
  • the result of outputting whether to recommend the candidate item to the social platform user can be the output to recommend the candidate item to the social platform user or the output to not socialize Platform users recommend the candidate item.
  • the above steps 201 to 209 in the embodiment of the present application can be implemented by the item recommendation model, which can be supported by the item recommendation server device in the application scenario shown in FIG. 1
  • the algorithm of the project recommendation model in detail.
  • FIG. 3 it is a schematic structural diagram of a project recommendation model in an embodiment of the present application, in which at least one group that is added by a user in a social platform and has a group relationship and the one added in a social platform are specifically used And at least one friend who has a friend relationship will be described as an example.
  • the project recommendation model can be supported by the project recommendation server device.
  • the input of the project recommendation model includes a user in a social platform that needs to recommend a project, at least one group that the user joins in the social platform and has a group relationship, and At least one friend added in the social platform and having a friend relationship.
  • each user of the item to be recommended includes at least one group that he has joined and has a group relationship, and at least one friend he has added that has a friend relationship, that is,
  • Each user of the item to be recommended has three observable interactions between the user and the item, group, and friend on the social platform.
  • the above interaction can be stored in the background server device of the social platform. It can communicate with the background server device of the social platform.
  • the item recommendation service calls the above interaction stored in the background server device of the social platform to obtain the corresponding user of the item to be recommended Interaction information with friends, groups, and projects, and then, based on the acquired interaction information, the user, the user's friends, and the user's group are used as input for the project recommendation model, so as to pass the training of the project recommendation model, Obtain candidate items recommended to the user.
  • User i represents social platform user i
  • Friends of User i represents at least one friend of social platform user i
  • Groups of User i represents at least one group of social platform user i
  • Attentional Pooling represents attention pool It is used to obtain the group's comprehensive attention to candidate items and friends' comprehensive attention to candidate items.
  • Feature Fusion indicates feature fusion
  • Prediction indicates output results.
  • a sequence of binary 1s and 0s is used to represent the social platform user u i who needs to recommend the project.
  • the sequence includes a binary 1 in the binary.
  • the 1 represents the social platform user u i in the project recommendation model.
  • Location which means that you need to recommend items to the social platform user u i at that location.
  • a sequence of binary 1s and 0s represents at least one friend of u i in the social platform f (i,l)
  • a 1 in the sequence of f (i,l) represents a friend belonging to user u i
  • the position, that is, the user at the position of each 1 in the f (i,l) sequence is a friend of u i .
  • a sequence of binary 1s and 0s indicates that at least u i in the social platform A group g (i, s) , the 1 in the g (i, s) sequence represents the position of a group that the user u i joins, that is, the position of each 1 in the g (i, s) sequence
  • the groups of are all groups joined by user u i , where s and l are positive integers greater than or equal to 1.
  • Step 1 Preprocess the binary sequences corresponding to u i and f (i, l) respectively.
  • the binary sequences corresponding to u i and f (i, l) are processed through the embedding layer to obtain dense vectors corresponding to u i and f (i, l) respectively.
  • f (i, 1) , F (i, 2) , f (i, 3) ... means that f (i, l) includes dense vectors corresponding to each friend.
  • the embedding layer converts positive integers (subscripts) into dense vectors with a fixed size in order to perform operations.
  • Step 2 Perform concentration training on the dense vectors f (i,1) , f (i,2) , f (i,3) corresponding to each friend to obtain the friends' comprehensive attention to the candidate items.
  • the project recommendation model can use the user's friends in the social platform for attention training to Learning the attention degree of each friend on the candidate item, so as to obtain the comprehensive attention degree of the candidate item on the friend who has the social relationship of the friend, so as to improve the accuracy of recommending the candidate item to the user.
  • the process of obtaining the comprehensive attention degree of the candidate items in the second step specifically includes: (a) The dense vectors corresponding to each friend f (i,1) , f (i,2) , f (i,3) ... ...Input the pre-trained second attention model to obtain the single attention degree of each item of the friend output by each friend output by the second attention model.
  • the second attention model belongs to a sub-model project recommended model can be regarded as a functional module project recommended model, the second model of attention in pre-trained social platform in the u i friends attention each parameter candidate projects As shown in FIG. 5, after inputting the dense vector corresponding to each friend into the second attention model, the second attention model can determine the attention parameters corresponding to each friend according to the dense vector corresponding to each friend, and then The determined attention parameter is used as a parameter, and according to the following formula (1), the single attention degree of each friend to the candidate item is calculated and obtained:
  • f (i,l) represents the dense vectors f (i,1) , f (i,2) , f (i,3) corresponding to each friend ...
  • v j represents the candidate item
  • W f1 ⁇ R d*k , W f2 ⁇ R d*k , b f ⁇ R k , h f ⁇ R k are the attention parameters corresponding to the friend f (i, l) , these attention parameters are pre-trained
  • k means attention
  • ReLU is a non-linear activation function
  • T represents transpose
  • f1 in FIG. 5 represents formula (1).
  • Each friend's individual attention to the candidate item Perform normalization processing according to formula (2) to obtain the single item attention ⁇ (i,l) of each normalized friend to the candidate item:
  • j represents all the values that need to be normalized
  • f2 in FIG. 5 represents the formula (2).
  • each friend obtained by the formula (2) may perform the summing operation as shown in the formula (3) on the single attention degree of the candidate item to obtain the comprehensive attention degree F i of each friend on the candidate item:
  • ⁇ (i,l) is the normalized single attention of each friend to the candidate item
  • f (i,l) is the vector representation of each friend (dense vector)
  • f3 in Figure 5 represents the formula (3) .
  • Step 3 Pass g (i, s) to the pre-trained first attention model for attention training to obtain the group's comprehensive attention to the candidate items.
  • the process of obtaining the group's comprehensive attention to candidate items in step 3 specifically includes: (a) preprocessing the binary sequence corresponding to g (i, s) to obtain the group dense vector of each group, and each A dense vector of members of each group in the group.
  • the binary sequence g s (Group s in FIG. 6 ) corresponding to each group is obtained, that is, for each group, the position 1 of the group is reserved, and the Set the position 1 of the other group in g (i,s) to zero, and then process each group g s through the embedding layer to obtain the group dense vector of each group s is a positive integer that sequentially takes values from 1, and the maximum value is the total number of at least one group that the user joins in the social platform and has a group relationship.
  • the corresponding binary sequence representing the group members included in each group is processed through the embedding layer to obtain the dense vector f (s, k) of each group member in each group, where k is the value starting from 1 in sequence A positive integer, the maximum value of which is the total number of group members in the corresponding group.
  • f (s, k) the dense vector of each group member in each group, where k is the value starting from 1 in sequence A positive integer, the maximum value of which is the total number of group members in the corresponding group.
  • Users in Group s represents the group members included in each group.
  • the attention information of the group members on the candidate items can affect the user's interest in the candidate items, so when using the group for attention training, the importance weight of each group member in the group is combined to improve The accuracy of the obtained group's attention to the candidate item, thereby improving the accuracy of recommending items to the user.
  • the first attention model is trained with the importance weight of each group member in the group for the group.
  • the first attention model belongs to a sub-model in the item recommendation model and can be regarded as A functional module of the project recommendation model.
  • the first attention model can generate and output the dense vector f (s, k) of each group member in each group through the following formula (4) and formula (5) in turn
  • formula (5) is to normalize the processing result of formula (4), f (s, k) is the dense vector of each group member in the group, T represents transpose, h m , W m , b m As parameters, these parameters can be pre-trained.
  • f4 represents formula (4), and f5 represents formula (5).
  • the importance parameter of each group member in the group is obtained by the following formula (6) to obtain the importance weight g s of each group member in the group:
  • k is a positive integer that sequentially takes values from 1
  • the maximum value is the total number of group members in the corresponding group
  • f (s, k) is a vector representation of each group member in the group, that is, a group member dense vector
  • ⁇ (s, k) is the weight parameter of the corresponding group member
  • It is a vector representation of a group, that is, a group dense vector
  • f6 in FIG. 7 represents formula (6).
  • the first attention model the attention parameters of each group in the social platform for candidate items are pre-trained. Therefore, the first attention model can first determine the attention parameters of each group for candidate items, and then sequentially pass the formula ( 7) and formula (8), generate and output the single attention ⁇ (i, s) of each group to the candidate item:
  • formula (8) is to normalize the calculation result of formula (7), W g1 ⁇ R d*k , W g2 ⁇ R d*k , b g ⁇ R k , h g ⁇ R k are groups
  • the group's attention parameters for the candidate items, these parameters are pre-trained in the first attention model, g (i,1) , g (i,2) , g (i,3) in formula (7 )
  • ... g (i,s) represents the importance weight g s of each group member in the group obtained by formula (6), that is, g (i,1) means g 1 , g (i,2) Representing g 2 ... And so on.
  • f7 represents formula (7)
  • f8 represents formula (8).
  • each group of candidate items obtained by formula (8) can be summed as shown in formula (9) to obtain the group's comprehensive attention degree of candidate items G i :
  • ⁇ (i, s) represents the individual attention of each group to the candidate item
  • g (i, s) represents the vector representation of each group, that is, the group dense vector
  • s is a positive value that starts from 1 in sequence.
  • f9 in FIG. 7 represents formula (9).
  • Step 4 Integrate the features of the friend's comprehensive attention to the candidate item and the group's comprehensive attention to the candidate item to obtain the recommended candidate item score.
  • the feature fusion technology may be additive fusion technology or other types of fusion technology.
  • the comprehensive attention F i of the friend to the candidate item and the group to the candidate item The comprehensive attention degree G i is subjected to additive feature fusion through the following formula (10) to generate a score indicating the user's preference for the candidate item, which is used to characterize the recommendation index recommending the candidate item to the social platform user u i :
  • v j is a vector representation of the candidate item
  • F i represents the comprehensive attention degree of the candidate item by the friend
  • G i represents the comprehensive attention degree of the candidate item by the group
  • T represents the transposition
  • Step 5 According to the obtained scores of candidate items to be recommended, determine whether to recommend the candidate items to users of the social platform.
  • the candidate item with the highest recommended score can be output to the social platform user u i , or it can be selected according to the score from high
  • the candidate items are sorted in descending order and output according to the sorted candidate items for recommendation to the user.
  • the project recommendation method in the embodiment of the present application can be implemented by the project recommendation model, and the user of the recommended project and each of the items to be recommended are processed through the project recommendation model to obtain the items recommended to the user to be recommended.
  • the good model usually needs good training samples. Therefore, in this embodiment of the present application, the original data can also be processed to obtain more reliable training samples and based on different training samples. Training and optimizing the project recommendation model in multiple scenarios can improve the accuracy of the project recommendation model.
  • the training process is usually performed by the background server device. Since the training of each module of the model may be complicated and the calculation amount is large, the background server device implements the training process, so that the trained model and results can be applied To each intelligent terminal, to achieve the purpose of accurately recommending items to users to be recommended.
  • the project recommendation model can be optimized according to the following formula (11):
  • the project recommendation method in the embodiment of the present application can accurately learn the single attention degree of each target social object to the candidate item when it is executed through the project recommendation model, and then achieve the purpose of accurately recommending the item to the user to be recommended according to the comprehensive attention degree. Therefore, the accuracy of recommending items to users can be improved, thereby improving user experience.
  • the existing project recommendation accuracy evaluation method is adopted, and for the same social platform user to be recommended, the items recommended to the user based on the project recommendation method in the embodiments of the present application and the existing multiple recommendation methods are respectively recommended Perform a recommendation accuracy evaluation, in which the existing project recommendation accuracy evaluation method can use the Recall evaluation method, which is shown in formula (12):
  • k represents the total number of items recommended to users by each algorithm
  • relj takes the value 0 or 1, indicating whether the j-th item is in the test set's push list.
  • the existing project recommendation accuracy evaluation method can also use the Normalized Discounted Cumulative Gain evaluation method (NDCG evaluation method), which is shown in formula (13) and formula (14):
  • NDCG evaluation method Normalized Discounted Cumulative Gain evaluation method
  • k represents the total number of items recommended to users by each algorithm, and relj has a value of 0 or 1, indicating whether the j-th item is in the push list of the test set. If so, relj has a value of 1.
  • both the Recall evaluation method and the NDCG evaluation method are used.
  • the MP (Most Popular) method in FIG. 8 is used.
  • ItemKNN method, BPR method, neural collaborative filtering NCF (Neural Collaborative Filtering) method, SBPR method and SAMN method to recommend the recommendation accuracy of the items recommended to users.
  • SACF Social Attentional Collaborative Filtering
  • the effect of the MP method is not ideal. This also illustrates the importance of modeling user preferences, not just recommending popular items to users.
  • SACF the item recommendation method SACF in the embodiment of the present application, the accuracy of the items recommended to the user is higher than the existing multiple recommendation methods.
  • the Recall@10 evaluation method is used to implement this application.
  • the performance of the SACF method provided in the example on two data sets (WeChat-10k, WeChat-100k) is improved by about 3.67% and 5.01% respectively compared to the SAMN method.
  • the NDCG@10 evaluation method is adopted.
  • the SACF method provided in the examples of this application The performance on the two data sets is improved by 4.17% and 5.49% respectively compared to the SAMN method. Therefore, the item recommendation method in the embodiment of the present application achieves the purpose of accurately recommending items to users to be recommended, avoids recommending items that are not of interest to users to be recommended, and improves resource utilization and user experience.
  • an embodiment of this application provides an item recommendation device, as shown in FIG. 9, including:
  • the obtaining module 90 is used to obtain candidate items to be recommended to users of the social platform;
  • the first determining module 91 is configured to determine, for at least one target social object of each type of social relationship, at least one target social object in each type of social relationship among the at least two different types of social relationships of the social platform user in the social platform Describe the individual attention of candidate projects;
  • the second determination module 92 is configured to determine the comprehensive attention degree of the target social object of different types to the candidate item according to the single attention degree of each target social object to the candidate item;
  • the third determination module 93 is configured to determine whether to recommend the candidate item to the social platform user according to the comprehensive attention degree.
  • the different types of social relationships include group relationships, and at least one target social object in the social relationships includes at least one group that the social platform user joins in the social platform.
  • the first determining module is used to:
  • the single attention degree is the vector representation of the first attention model according to each group and the vector of the candidate item Means that the attention parameters of the candidate items for each group are determined, and the single attention degree of the candidate item for the group generated and output according to the determined attention parameters is determined.
  • the importance weight of each group member in the group is also trained for each group.
  • the individual attention of each group output by the first attention model to the candidate item is to further weight the importance of each group member in each group in each group, and compare each group with The attention parameter of the candidate item is obtained after weighting processing.
  • the different types of social relationships include friend relationships, and at least one target social object in the social relationship includes at least one friend added by the social platform user in the social platform.
  • the first determining module is also used to:
  • the single attention degree of each friend output by the second attention model to the candidate item where the single attention degree is the second attention model according to the vector representation of each friend and the vector representation of the candidate item.
  • the attention parameters of each friend to the candidate item are determined, and the single attention degree of each friend to the candidate item generated by each friend according to the determined attention parameter is determined.
  • the second determining module is used to:
  • the target social objects of the same type sum the single attention degrees of the candidate items to obtain the comprehensive attention degrees of the social objects of various types to the candidate items;
  • the third determining module is used to:
  • an embodiment of the present application provides a computing device, as shown in FIG. 10, including at least one processor 101 and at least one memory 102, wherein the memory 102 stores a computer program, when the program When executed by the processor 101, the processor 101 is caused to execute the steps of the item recommendation method described above.
  • an embodiment of the present application provides a storage medium that stores computer instructions, and when the computer instructions run on a computer, the computer is allowed to execute the steps of the item recommendation method described above.
  • the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer usable program code.
  • a computer usable storage media including but not limited to disk storage and optical storage, etc.
  • each flow and/or block in the flowchart and/or block diagram and a combination of the flow and/or block in the flowchart and/or block diagram may be implemented by computer program instructions.
  • These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processing machine, or other programmable data processing device to produce a machine that enables the generation of instructions executed by the processor of the computer or other programmable data processing device
  • These computer program instructions may also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, the instructions
  • the device implements the functions specified in one block or multiple blocks of the flowchart one flow or multiple flows and/or block diagrams.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operating steps are performed on the computer or other programmable device to produce computer-implemented processing, which is executed on the computer or other programmable device
  • the instructions provide steps for implementing the functions specified in one block or multiple blocks of the flowchart one flow or multiple flows and/or block diagrams.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Economics (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请涉及人工智能的机器学***台用户的候选项目；针对所述社交平台用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定目标社交对象对所述候选项目的单项关注度；根据各个目标社交对象对所述候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度；并根据所述综合关注度确定是否向所述社交平台用户推荐所述候选项目。

Description

一种推荐方法、装置及存储介质
本申请要求于2018年11月29日提交中国专利局、申请号为201811445266.1、发明名称为“一种推荐方法、装置及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及互联网技术领域，尤其涉及一种推荐方法、装置及存储介质。
背景技术
人工智能(Artificial Intelligence,AI)是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法、技术及应用***。换句话说，人工智能是计算机科学的一个综合技术，它企图了解智能的实质，并生产出一种新的能以人类智能相似的方式做出反应的智能机器。人工智能也就是研究各种智能机器的设计原理与实现方法，使机器具有感知、推理与决策的功能。
人工智能技术是一门综合学科，涉及领域广泛，既有硬件层面的技术也有软件层面的技术。人工智能基础技术一般包括如传感器、专用人工智能芯片、云计算、分布式存储、大数据处理技术、操作/交互***、机电一体化等技术。人工智能软件技术主要包括计算机视觉技术、语音处理技术、自然语言处理技术以及机器学习/深度学习等几大方向。
机器学习(Machine Learning,ML)是一门多领域交叉学科，涉及概率论、统计学、逼近论、凸分析、算法复杂度理论等多门学科。专门研究计算机怎样模拟或实现人类的学习行为，以获取新的知识或技能，重新组织已有的知识结构使之不断改善自身的性能。机器学习是人工智能的核心，是使计算机具有智能的根本途径，其应用遍及人工智能的各个领域。机器学习和深度学习通常包括人工神经网络、置信网络、强化学习、迁移学习、归纳学习、式教学习等技术。
随着互联网技术的不断发展，互联网上的多类项目迅速增长，以满足用户在信息时代对信息的需求，互联网上的项目是指互联网中用于用户消费、参与以及进行行为交互的数据信息，比如商品、文章、广告以及虚拟信息等，然而，不同的用户对项目的需求也不同。这样，需要向不同的用户推荐不同的项目。比如，在一些推荐方法中，可以根据用户的好友的信息去挖掘用户的偏好。
发明内容
本申请实施例提供一种推荐方法、装置及存储介质，用于提升向用户推荐用户感兴趣的项目的准确度，避免多次向用户推荐用户不感兴趣的项目，提高资源利用率，进而提升用户体验。
一方面，本申请实施例提供的一种推荐方法，由计算装置执行，包括：
获得待推荐给社交平台用户的候选项目；
针对所述社交平台用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定各个目标社交对象对所述候选项目的单项关注度；
根据各个目标社交对象对所述候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度；
并根据所述综合关注度确定是否向所述社交平台用户推荐所述候选项目。
另一方面，本申请实施例提供了一种推荐装置，包括：
获得模块，用于获得待推荐给社交平台用户的候选项目；
第一确定模块，用于针对所述社交平台用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定目标社交对象对所述候选项目的单项关注度；
第二确定模块，用于根据各个目标社交对象对所述候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度；
第三确定模块，用于根据所述综合关注度确定是否向所述社交平台用户推荐所述候选项目。
另一方面，本申请实施例提供了一种计算装置，包括至少一个处理器、 以及至少一个存储器，其中，所述存储器存储有计算机程序，当所述程序被所述处理器执行时，使得所述处理器执行本申请实施例中的推荐方法的步骤。
另一方面，本申请实施例提供了一种存储介质，所述存储介质存储有计算机指令，当所述计算机指令在计算机上运行时，使得计算机执行本申请实施例中的推荐方法的步骤。
本申请实施例中的推荐方法，获得待推荐给社交平台用户的候选项目时，会针对用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定目标社交对象对候选项目的单项关注度，即充分利用用户在社交平台中的多种社交关系的社交对象对候选项目的关注信息，然后根据各个目标社交对象对候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度，然后再根据综合关注度确定是否向所述社交平台用户推荐所述候选项目，所以，能够提升向用户推荐项目的准确度，尽量避免向用户发送用户不感兴趣的项目，从而提高资源利用率，提升用户体验度。
附图简要说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例。
图1为本申请实施例提供的一种应用场景示意图；
图2为本申请实施例提供的一种推荐方法流程图；
图3为本申请实施例提供的项目推荐模型结构示意图；
图4为本申请实施例提供的微信用户与好友、群组、待推荐候选项目的交互示意图；
图5为本申请实施例提供的第二注意力模型架构示意图；
图6为本申请实施例提供的群组各群成员重要度权重获取示意图；
图7为本申请实施例提供的第一注意力模型架构示意图；
图8为本申请实施例提供的对推荐的项目的准确度评估结果示意图；
图9为本申请实施例提供的一种推荐装置示意图；
图10为本申请实施例提供的一种计算装置示意图。
实施方式
为使本申请实施例的目的、技术方案和优点更加清楚，下面将结合本申请实施例中的附图，对本申请的技术方案进行清楚、完整地描述，显然，所描述的实施例是本申请技术方案的一部分实施例，而不是全部的实施例。基于本申请文件中记载的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请技术方案保护的范围。
随着人工智能技术研究和进步，人工智能技术在多个领域展开研究和应用，例如常见的智能家居、智能穿戴设备、虚拟助理、智能音箱、智能营销、无人驾驶、自动驾驶、无人机、机器人、智能医疗、智能客服、语音识别等，相信随着技术的发展，人工智能技术将在更多的领域得到应用，并发挥越来越重要的价值。
在一些推荐方案中，可以根据用户的好友的信息来确定用户的偏好，但是这些方法仅考虑到用户的部分好友的信息，然而，用户并不一定与该部分好友具有相同或相似的偏好，仅从好友角度去挖掘用户的偏好，不能获得准确的用户偏好信息，所以，上述推荐方法存在项目推荐准确度低，影响用户体验的技术问题。
随着社交网络的发展，一些社交平台和应用程序允许用户加入他们感兴趣的群组并与群组成员通信。群组活动可以反映用户的兴趣，并且是对用户的朋友信息的补充。
在本申请实施例提供的项目推荐方法中，除了考虑好友信息，还进一步引入了与用户偏好高度相关的群组信息。在本申请一些实施例中，提出了一种社交注意力协同过滤(SACF，Social Attentional Collaborative Filtering)模型，共同利用用户的好友信息和群组信息，因为并非所有社交关系对确定用户的偏好都同样有用，该模型中引入了注意力机制来区分用户不同朋友和群体的影响力。此外，为了更好的利用群组信息，本申请实施例还提出了一种注意力模型，以提升不同社交关系的社交对象对待推荐项目的学习效果。
下面对本申请实施例中涉及的部分概念进行介绍。
深度学习：是机器学习研究中的一个新的领域，其动机在于建立、模拟人脑进行分析学习的神经网络，它模仿人脑的机制来解释数据，例如图像，声音和文本。
注意力机制：即Attention Mechanism，为深度学习的一个分支，又可称为神经注意力机制，源于对人类视觉的研究。在认知科学中，由于信息处理的瓶颈，人类会选择性地关注所有信息的一部分，同时忽略其他可见的信息，上述机制通常被称为注意力机制。人类视网膜不同的部位具有不同程度的信息处理能力，只有视网膜中央凹部位具有最强的敏锐度。为了合理利用有限的视觉信息处理资源，人类需要选择视觉区域中的特定部分，然后集中关注它，例如，人们在阅读时，通常只有少量要被读取的词会被关注和处理。综上，注意力机制主要有两个方面：决定需要关注输入的哪部分；分配有限的信息处理资源给重要的部分。
注意力：指可以在关注一些信息的同时忽略另一些信息的选择能力。
本申请的发明人考虑到，社交平台中具有不同社交关系的社交对象对用户行为具有不同的影响力，例如，当用户需要购买篮球鞋时，可能会遵循打篮球朋友的建议，当涉及到旅行时，用户可能会转向社交平台中喜欢旅行的朋友的建议，而用户打篮球的朋友或喜欢旅行的朋友，可能是用户社交平台中具有好友关系的好友，也可能为用户加入的具有群组关系的群组中的成员，因此，在向用户推荐项目时，可结合用户社交平台中具有不同社交关系的社交对象对项目的关注信息，以提升向用户推荐项目的准确度。
进一步的，考虑到注意力机制已经在各种机器学***台中各类社交关系对项目评论的“有用性”，从而进一步提升向用户推荐项目的准确度，进一步的，本申请在将注意力机制引入项目推荐方法中时，考虑到社交平台中不同社交关系中的社交对象对用户行为的不同影响，针对不同社交关系的社交对象，基于注意力机制，设计了适合学习各社交关系的社交对象对待推荐项目的注意力模型，以提升不同社交关系的社交对象对待推荐项目的学习效果，例如，在社交关系为群组关系时，考虑到群组中不同群成员在群组中的活跃度不同，越活跃的群成员对群组影响越大，其对群组的重要度也就越 高，群组中这类重要度越高的群成员对候选项目的关注信息，更能影响用户对候选项目的兴趣，因此在引入的注意力机制中结合群组中各群成员在群组中的重要度权重，以学习群组对项目的关注，以提升学习获得的群组对项目的关注度的准确度，进而提升向用户推荐项目的准确度。
基于此，本申请实施例提供了一种项目推荐方法，该方法在获得待推荐给社交平台用户的候选项目时，会针对用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定目标社交对象对候选项目的单项关注度，即充分利用用户在社交平台中的多种社交关系的社交对象对候选项目的关注信息，然后根据各个目标社交对象对候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度，其中，各个目标社交对象对候选项目的单项关注度是基于注意力机制的注意力模型训练获得的，即充分利用了注意力机制的学***台用户推荐所述候选项目，所以，能够提升向用户推荐项目的准确度，避免向用户推荐用户不感兴趣的项目，提高资源利用率，从而提升用户体验度。
本申请实施例中的项目推荐方法可以应用于如图1所示的应用场景，该应用场景中包括用户终端10、以及项目推荐计算装置，图1中所示的实施例中，项目推荐计算装置可以为项目推荐服务器设备11，项目推荐服务器设备11可以是一台服务器设备，也可以是若干台服务器设备组成的服务器设备集群或云计算中心。在一种实施方式中，图1中所述的项目推荐服务器设备11可以包含图10所示的计算装置。
用户终端10可以为任何能够按照程序运行，自动、高速处理大量数据的智能终端设备，这样的终端设备如电脑，ipad，手机等，用户终端10上安装有本申请实施例中的社交平台对应的社交应用程序(Application，APP)，例如微信、QQ等，用户终端10中还可以安装其他类型的APP，社交APP由支持其运行的后台服务器设备，其后台服务器设备可以是一台服务器设备，也可以是若干台服务器设备组成的服务器设备集群或云计算中心，支持社交APP运行的后台服务器设备与项目推荐服务器设备11可以是集成在一起的集 成服务器设备集群，也可以为图1所示的各自独立的服务器设备，图1中支持社交APP运行的后台服务器设备的标号为12。
用户终端10分别与支持社交APP运行的后台服务器设备12、项目推荐服务器设备11通过网络连接，支持社交APP运行的后台服务器设备12与项目推荐服务器设备11通过网络连接，以使得用户终端10、支持社交APP运行的后台服务器设备12、项目推荐服务器设备11之间可以通信。其中，网络可以为局域网、广域网或移动互联网等通信网络中的任意一种。
本申请实施例中，项目推荐方法可应用于项目推荐服务器设备中，项目推荐服务器设备在获得待推荐给社交平台用户的候选项目时，可针对社交平台用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定目标社交对象对候选项目的单项关注度，然后根据各个目标社交对象对候选项目的单项关注度，确定不同类型的目标社交对象对候选项目的综合关注度，并根据综合关注度确定是否向社交平台用户推荐所述候选项目，并在确定向社交平台用户推荐候选项目时，与支持社交平台的后台服务器设备交互，进而通过用户终端中的社交平台向社交平台用户呈现推荐的候选项目。
需要注意的是，上文提及的应用场景仅是为了便于理解本申请的精神和原理而示出，本申请实施例在此方面不受任何限制。相反，本申请实施例可以应用于适用的任何场景。
下面结合图1所示的应用场景，对本申请实施例提供的项目推荐方法进行说明。
如图2所示，本申请实施例提供的项目推荐方法，可由图1所示的项目推荐服务器设备执行，也可由图10所示的计算装置执行，该项目推荐方法包括：
步骤201：获得待推荐给社交平台用户的候选项目。
在本申请实施例中，步骤201中的社交平台是指互联网中用于进行社交的软件，这样的软件如微信，QQ等，候选项目是指互联网上用于用户消费、参与或进行行为交互的数据信息，比如商品、文章、广告、虚拟信息或者兴趣点(Point of Interest，简称POI)，本申请实施例中对候选项目的类型不作具 体限定，可根据实际情况确定。
本申请实施例中，步骤201获得待推荐给社交平台用户的候选项目的数目，可以为一个，也可以为多个，在获得的待推荐给社交平台用户的候选项目的数目为多个时，多个候选项目的类型可以相同，也可以不同，在此不作具体限定。
步骤202：确定社交平台用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象。
在本申请实施例中，考虑到社交平台用户(为叙述方便，将下文涉及的社交平台用户简称为用户)在社交平台中，不同社交关系的社交对象对其行为具有不同的影响力，因此，在向用户推荐候选项目时，可综合利用用户在社交平台中多个不同社交类型的社交对象对候选项目的关注信息，以提升向该用户推荐候选项目的准确度。
因此在步骤202中，可先确定用户在社交平台中的至少两种不同类型的社交关系，并从每种类型的社交关系中确定至少一个目标社交对象，例如，在社交平台具体为微信平台时，目标社交对象可以为用户添加的、且具有好友关系的微信好友，也可以为用户加入的、具有群组关系的微信群组，还可以为用户加入的、具有企业业务来往关系的企业微信联系人等等。
在本申请实施例中，考虑到社交平台中的群组活动变得越来越普遍，用户在社交平台中加入的群组中的群成员通常与用户的偏好高度相关，如在某些方面具有共同需求或共同特征，那么，使用用户在社交平台中加入的、且具有群组关系的群组对候选项目的关注信息，能够进一步提升向用户推荐候选项目的准确度。
例如，当用户需要购买化妆品时，在用户加入的化妆品主题相关的群组中与群成员讨论购买化妆品的话题，如哪些化妆品更适合用户，用户根据群成员的建议购买化妆品，对于与化妆品相关的候选项目，使用用户加入的化妆品主题相关的群组，能够进一步提升向用户推荐与化妆品相关的候选项目的准确度。
因此，在步骤202中确定的至少两种不同类型的社交关系的至少一个目标社交对象，至少包括用户在社交平台中加入的、且具有群组关系(又可称 为群组社交关系)的至少一个群组。
在本申请实施例中，考虑到用户在社交平台中的具有好友关系的好友，与用户在某些方面也具有共同需求或共同特征，因此，也可以使用用户在社交平台中添加的、且具有好友关系的好友对候选项目的关注信息，进一步提升向用户推荐候选项目的准确度。因此，在步骤202中确定的至少两种不同类型的社交关系的至少一个目标社交对象，至少还包括用户在社交平台中添加的、且具有好友关系的至少一个好友。
当然，本申请实施例中，还可以包括其他类型的社交关系的至少一个社交目标社交对象，为叙述方便，在下文中，具体以步骤202中确定的至少两种不同类型的社交关系中的至少一个目标社交对象，包括用户在社交平台中加入的、且具有群组关系的至少一个群组以及在社交平台中添加的、且具有好友关系的至少一个好友为例。
步骤203：分别确定目标社交对象对候选项目的单项关注度。
在本申请实施例中，目标社交对象为用户加入的、且与用户具有群主关系的至少一个群组时，步骤203的具体执行过程包括：
首先，将至少一个群组的向量表示和候选项目的向量表示，输入预先训练的第一注意力模型，其中，第一注意力模型中预先训练有社交平台的各群组对候选项目的注意力参数；
然后，第一注意力模型根据输入的各群组的向量表示以及候选项目的向量表示，确定出各群组分别对候选项目的注意力参数，以及根据确定的注意力参数生成各群组对候选项目的单项关注度，进而获得第一注意力模型输出的至少一个群组中每个群组对候选项目的单项关注度。
进一步的，在本申请实施例中，考虑到群组中不同群成员在群组中的活跃度不同，越活跃的群成员对群组影响越大，其对群组的重要度也就越高，群组中这类重要度越高的群成员对候选项目的关注信息，更能影响用户对候选项目的兴趣，因此，在本申请实施例中，第一注意力模型中预先还可以针对每个群组，训练有各群组中各群成员在群组中的重要度权重。
对应的，第一注意力模型在根据输入的群组中各群成员的向量表示，可确定出各群组分别对候选项目的注意力参数，还可将各群组中各群成员在各 群组中的重要度权重，与各群组分别对候选项目的注意力参数进行加权处理后，生成并输出该群组对候选项目的单项关注度，以提升获得的该群组对候选项目的单项关注度，进而提升向用户推荐候选项目的准确度。
在本申请实施例中，目标社交对象为用户在社交平台中添加的、且具有好友关系的至少一个好友时，步骤203的具体执行过程还包括：
首先，将至少一个好友的向量表示与候选项目的向量表示，输入预先训练的第二注意力模型，第二注意力模型中预先训练有社交平台的各好友对候选项目的注意力参数。
然后，第二注意力模型根据输入的各好友的向量表示与候选项目的向量表示，确定各好友对候选项目的注意力参数，根据确定的注意力参数，生成并输出各好友对候选项目的单项关注度，从而获得第二注意力模型输出的至少一个好友中每个好友对候选项目的单项关注度。
步骤204：将各个目标社交对象对候选项目的单项关注度进行归一化处理。
步骤205：将归一化处理后相同类型的目标社交对象对候选项目的单项关注进行求和处理，获得各类型的社交对象对候选项目的综合关注度。
在本申请实施例中，在分别确定各目标社交对象对候选项目的单项关注度之后，可将各个目标社交对象对候选项目的单项关注度进行归一化处理，以便于后续计算，然后针对上文中的至少一个群组，将各个群组对候选项目的归一化处理后的单项关注度，进行求和运算，获得用户在社交平台中的群组对候选项目的综合关注度；针对上文中的至少一个好友，将各个好友对候选项目的归一化处理后的单项关注度，进行求和运算，获得用户在社交平台中的好友对候选项目的综合关注度。
步骤206：将各类型的社交对象对候选项目的综合关注度进行特征融合，获得待推荐候选项目的推荐指数。
在本申请实施例中，可以采用现有的特征融合技术，对获得的各类型的目标社交对象对候选项目的综合关注度进行特征融合，例如，选择加成融合技术，那么，就可将上文获得的用户在社交平台中的好友对候选项目的综合关注度，以及用户在社交平台中的群组对候选项目的综合关注度，作为加成融合算法的输入参数，进而获得加成融合算法输出的对待推荐候选项目的推 荐指数。
其中，推荐指数的表示方式可以为星级等级、评分分数等。
步骤207：确定推荐指数是否大于阈值，若是，则执行步骤208，否则执行步骤209。
在本申请实施例中，在获得待推荐候选项目的推荐指数之后，还可以根据推荐指数，确定是否向社交平台用户推荐候选项目，具体的，可以确定获得的待推荐候选项目的推荐指数是否大于阈值，例如，在使用评分分数表示推荐指数时，若待推荐候选项目的评分分数大于阈值，则执行步骤208：确定向用户推荐候选项目，否则执行步骤209：确定不向用户推荐候选项目。
在本申请实施例中，若候选项目包括多个，则可以根据上文叙述的方法获得每个候选项目对应的推荐指数，再根据推荐指数，确定是否向社交平台用户推荐候选项目时，可以直接根据各候选项目对应的推荐指数，确定向用户推荐的候选项目为推荐指数最靠前的候选项目；也可以按照推荐指数从高到低依次排序，并向用户推荐排序后的各候选项目。
所以，通过上述方法，在获得待推荐给社交平台用户的候选项目时，会针对用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定目标社交对象对候选项目的单项关注度，即充分使用用户在社交平台中的多种社交关系的社交对象对候选项目的关注信息，然后根据各个目标社交对象对候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度，其中，各个目标社交对象对候选项目的单项关注度是基于注意力机制的注意力模型训练获得的，即充分利用了注意力机制的学***台用户推荐所述候选项目，所以，能够提升向用户推荐项目的准确度，从而提升用户体验度。
作为一种方式，在步骤202中，考虑到实际应用中，通常与用户社交频繁的社交对象对用户的行为影响力，比偶尔与用户社交的对象对用户的行为影响力大，若是使用与用户社交频繁的社交对象对候选项目的关注信息，不仅能提升向用户推荐候选项目的准确度，还可以提升数据处理速度，因此， 步骤202中每种类型的社交关系中确定的至少一个目标社交对象，可以为对应类型社交关系中与用户社交频繁的对象。
在本申请实施例中，与用户社交频繁的对象的确定方式，可以为在预设时间段内与用户社交次数确定，例如，若在一周时间与用户社交(包括基于社交平台的收发消息、分享等社交活动)次数达到阈值的社交对象，则认为该对象为与用户社交频繁的对象。
进一步的，考虑到使用社交对象中与候选项目有关联的对象对候选项目的关注信息，比使用社交对象中与候选项目无关联的对象对候选项目的关注信息，更利于提升向用户推荐候选项目的准确度，因此，步骤202中每种类型的社交关系中确定的至少一个目标社交对象，还可以为对应类型社交关系中与候选项目有关联的对象。其中，有关联是指，社交对象与候选项目有过交互，例如使用、或关注过候选项目等。
作为一种方案，本申请实施例中，步骤202中每种类型的社交关系中确定的至少一个目标社交对象，也可以是同时与用户社交频繁的、且与候选项目有关联的对象。作为一种方案，本申请实施例中，步骤202中每种类型的社交关系中确定的至少一个目标社交对象，也可以是对应类型的社交关系中所有的社交对象，或随机选择的部分社交对象。
进一步的，在本申请实施例中，上述项目推荐方法可以基于项目推荐模型来完成，该项目推荐模型中包括上文叙述的预先训练的第一注意力模型以及第二注意力模型，具体地，可以将需要推荐项目的社交平台用户以及待推荐给社交平台用户的候选项目输入项目推荐模型，通过该项目推荐模型执行上述实施例中的项目推荐方法后，输出是否向社交平台用户推荐候选项目的结果，其中，输入项目推荐模型的候选项目为多个时，输出是否向社交平台用户推荐候选项目的结果可以为推荐指数最靠前的候选项目，也可以为按照推荐指数从高到低依次排序并向用户推荐排序后的各候选项目，在输入项目推荐模型的候选项目为一个时，输出是否向社交平台用户推荐候选项目的结果可以为输出向社交平台用户推荐该候选项目或输出不向社交平台用户推荐该候选项目。
也就是说，本申请实施例中的上述步骤201至步骤209，即项目推荐方法， 是可以通过项目推荐模型来实现的，项目推荐模型可由图1所示的应用场景中的项目推荐服务器设备支持，下文对项目推荐模型的算法进行具体说明。
如图3所示，为本申请实施例中的项目推荐模型结构示意图，其中，具体以使用用户在社交平台中加入的、且具有群组关系的至少一个群组，以及在社交平台中添加的、且具有好友关系的至少一个好友为例进行说明。
该项目推荐模型可由项目推荐服务器设备支持，该项目推荐模型的输入Input包括需要推荐项目的社交平台中的用户，该用户在社交平台中加入的、且具有群组关系的至少一个群组，以及在社交平台中添加的、且具有好友关系的至少一个好友。
如图4所示，在实际应用中，社交平台中包括多个用户，每一个用户都可以为需要推荐项目的用户，那么可以预先使用一个集合如U＝{u1，u2，...，un}表示社交平台中的多个待推荐项目的用户，可使用另一个集合如I＝{i1，i2，...，im}表示待推荐的m个项目，可使用另一个集合如G＝{g1，g2，...，gs}表示社交平台中的s个组群组，其中每一个群组包括多个群成员，每个群成员可为社交平台中的用户。那么，每一个待推荐项目的用户如图4所示，在社交平台中包括其加入的、且具有群组关系的至少一个群组，以及其添加的、且具有好友关系的至少一个好友，即每一个待推荐项目的用户在社交平台中都存在该用户与项目、群组、好友之间的三种可观察到的交互。
因此，可以使用一个矩阵如X＝[x ik] n*n表示各待推荐项目的用户与其好友的交互，可以使用一个矩阵如R＝[r ij] n*m表示各待推荐项目的用户与待推荐项目的交互，也可以使用一个矩阵如Y＝[y il] n*s表示各待推荐项目的用户与群组的交互，上述交互可以存储在社交平台的后台服务器设备中，项目推荐服务可与社交平台的后台服务器设备进行通信，针对社交平台的中的任一待推荐项目的用户，项目推荐服务调用社交平台的后台服务器设备中存储的上述交互，以获取待推荐项目的用户对应的与好友、群组、项目的交互信息，然后，基于获取的交互信息，将该用户，该用户的好友，以及该用户的群组作为项目推荐模型的输入，以便通过该项目推荐模型的训练，获得向该用户推荐的候选项目。
具体的，在图3中User i表示社交平台用户i，Friends of User i表示社交 平台用户i的至少一个好友，Groups of User i表示社交平台用户i的至少一个群组，Attentional Pooling表示注意力池化，用于获得群组对候选项目的综合关注度，以及好友对候选项目的综合关注度，Feature Fusion表示特征融合，Prediction表示输出结果。
在该项目推荐模型中，用二进制1和0组成的序列表示需要推荐项目的社交平台用户u i，该序列中包括二进制中的一个1，该1表示该社交平台用户u i在项目推荐模型的位置，即表示需要向该位置上的社交平台用户u i推荐项目。
在该项目推荐模型中，用二进制1和0组成的序列表示u i在社交平台中的至少一个好友f (i,l)，f (i,l)序列中的1表示属于用户u i的好友所在的位置，即f (i,l)序列中的每个1所在的位置上的用户均为u i的好友，同样的，用二进制1和0组成的序列表示u i在社交平台中的至少一个群组g (i,s)，g (i,s)序列中的1表示用户u i加入的一个群组所在的位置，即g (i,s)序列中的每个1所在的位置上的群组均为用户u i加入的群组，其中，s与l为大于等于1的正整数。
下文对结合图4对图3所示的项目推荐模型的处理步骤进行介绍：
第1步：分别对u i和f (i,l)对应的二进制序列进行预处理。
具体包括：分别将u i和f (i,l)对应的二进制序列通过嵌入层进行处理，获得u i和f (i,l)各自对应的密集向量，图4中，f (i,1)、f (i,2)、f (i,3)……表示f (i,l)包括各好友对应的密集向量。
其中，嵌入层是将正整数(下标)转换为具有固定大小的密集向量，以便进行运算。
第2步：将各好友对应的密集向量f (i,1)、f (i,2)、f (i,3)……进行注意力训练，获得好友对候选项目的综合关注度。
考虑到用户在社交平台中的具有好友关系的好友，与用户在某些方面也具有共同需求或共同特征，因此，项目推荐模型中可以使用用户在社交平台中的各好友进行注意力训练，以学习各好友对候选项目的关注度，从而获得具有好友社交关系的好友对候选项目的综合关注度，以提升向用户推荐候选项目的准确度。
第2步中获取好友对候选项目的综合关注度的处理过程具体包括：(a) 将各好友对应的密集向量f (i,1)、f (i,2)、f (i,3)……输入预先训练的第二注意力模型，获得第二注意力模型输出的各好友对候选项目的单项关注度。
第二注意力模型属于项目推荐模型中的一个子模型，可视为项目推荐模型的一个功能模块，第二注意力模型中预先训练有社交平台中的u i各好友对候选项目的注意力参数，如图5所示，在将各好友对应的密集向量输入第二注意力模型之后，第二注意力模型可根据各好友对应的密集向量，确定与各好友分别对应的注意力参数，然后将确定的注意力参数作为参数，根据以下公式(1)，计算获得各好友对候选项目的单项关注度：
Figure PCTCN2019121919-appb-000001
其中，f (i,l)表示各好友对应的密集向量f (i,1)、f (i,2)、f (i,3)…，v j表示候选项目，W f1∈R d*k、W f2∈R d*k、b f∈R k、h f∈R k是与好友f (i,l)对应的注意力参数，这些注意力参数都是预先训练出来的，k表示注意力网络的维度，ReLU是非线性激活函数，T表示转置，图5中f1表示公式(1)。
将获得的各好友对候选项目的单项关注度
Figure PCTCN2019121919-appb-000002
按照公式(2)进行归一化处理，获得归一化后的各好友对候选项目的单项关注度α (i,l)：
Figure PCTCN2019121919-appb-000003
其中，j表示所有需要归一化的值，图5中f2表示公式(2)。
(b)将各好友对候选项目的单项关注度进行求和运算，获得好友对候选项目的综合关注度。
具体的，可将公式(2)获得的各个好友对候选项目的单项关注度进行如公式(3)所示的求和运算，获得各个好友对候选项目的综合关注度F i：
Figure PCTCN2019121919-appb-000004
其中，α (i,l)为归一化后的各好友对候选项目的单项关注度，f (i,l)为各好友的向量表示(密集向量)，图5中f3表示公式(3)。
第3步：将g (i,s)传递给预先训练的第一注意力模型进行注意力训练，获得群组对候选项目的综合关注度。
考虑到社交平台中的群组活动变得越来越普遍，用户在社交平台中加入的群组中的群成员通常与用户的偏好高度相关，因此，项目推荐模型中可以使用用户在社交平台中的群组进行注意力训练，以学习各群组对候选项目的关注度，从而获得具有群组社交关系的群组对候选项目的综合关注度，以提升向用户推荐候选项目的准确度。
第3步中获取群组对候选项目的综合关注度的处理过程具体包括：(a)将g (i,s)对应的二进制序列进行预处理，获得各群组的群组密集向量，以及各群组中各群成员的密集向量。
例如图6所示，根据g (i,s)，获得各群组对应的二进制序列g s(图6中的Group s)，即针对各群组，保留该群组所在的位置的1，将g (i,s)中其他群组中的所在的位置的1置零，然后将各群组g s通过嵌入层进行处理，获得各群组的群组密集向量
Figure PCTCN2019121919-appb-000005
s为依次从1开始取值的正整数，其最大值为用户在社交平台中加入的、且具有群组关系的至少一个群组的总数。
将表示各群组中包括的群成员的对应的二进制序列通过嵌入层进行处理，获得各群组中各群成员的密集向量f (s,k)，其中，k为依次从1开始取值的正整数，其最大值为对应群组中群成员的总数，图6中Users in Group s表示各群组中包括的群成员。
(b)针对每一个群组，获取该群组中各群成员在该群组中的重要度权重。
考虑到群组中不同群成员在群组中的活跃度不同，越活跃的群成员对群组影响越大，其对群组的重要度也就越高，群组中这类重要度越高的群成员对候选项目的关注信息，更能影响用户对候选项目的兴趣，因此在使用群组进行注意力训练时，结合了群组中各群成员在群组中的重要度权重，以提升获得的群组对候选项目的关注度的准确度，进而提升向用户推荐项目的准确 度。
其中，第一注意力模型针对每一个群组，训练有该群组中各群成员在该群组中的重要度权重，第一注意力模型属于项目推荐模型中的一个子模型，可视为项目推荐模型的一个功能模块。
请结合图7所示，第一注意力模型可将各群组中各群成员的密集向量f (s,k)，依次通过以下公式(4)和公式(5)生成并输出该群组中各群成员在该群组中的重要度参数γ (s,k)：
Figure PCTCN2019121919-appb-000006
Figure PCTCN2019121919-appb-000007
其中，公式(5)是对公式(4)的处理结果进行归一化，f (s,k)为群组中各群成员的密集向量，T表示转置，h m、W m、b m为参数，这些参数可以预先训练，图7中f4表示公式(4)，f5表示公式(5)。
然后，将各群成员在该群组中的重要度参数通过以下公式(6)，获得该群组中各群成员在该群组中的重要度权重g s：
Figure PCTCN2019121919-appb-000008
其中，k为依次从1开始取值的正整数，其最大值为对应群组中群成员的总数，f (s,k)为群组中各群成员的向量表示，即群成员密集向量，γ (s,k)为对应群成员的权重参数，
Figure PCTCN2019121919-appb-000009
为群组的向量表示，即群组密集向量，图7中f6表示公式(6)。
(c)获取各群组对候选项目的注意力参数，并根据各群组中各群成员在该群组中的重要度权重，将群成员分别对候选项目的注意力参数进行加权运算，获得各群组对候选项目的单项关注度。
第一注意力模型中预先训练有社交平台中的各群组对候选项目的注意 力参数，因此，第一注意力模型可以先确定各群组对候选项目的注意力参数，然后依次通过公式(7)和公式(8)，生成并输出各群组对候选项目的单项关注度β (i,s)：
Figure PCTCN2019121919-appb-000010
Figure PCTCN2019121919-appb-000011
其中，公式(8)是对公式(7)的计算结果进行归一化处理，W g1∈R d*k、W g2∈R d*k、b g∈R k、h g∈R k是群组对候选项目的注意力参数，这些参数是第一注意力模型中预先训练的，公式(7)中的g (i,1)、g (i,2)、g (i,3)……g (i,s)分别表示通过公式(6)获得的群组中各群成员在该群组中的重要度权重g s，即g (i,1)表示g 1，g (i,2)表示g 2……以此类推，图7中f7表示公式(7)，f8表示公式(8)。
(d)将各群组对候选项目的单项关注度进行求和运算，获得群组对候选项目的综合关注度。
具体的，可将公式(8)获得的各群组对候选项目的单项关注度β (i,s)进行如公式(9)所示的求和运算，获得群组对候选项目的综合关注度G i：
Figure PCTCN2019121919-appb-000012
其中，β (i,s)表示各群组对候选项目的单项关注度，g (i,s)表示各群组的向量表示，即群组密集向量，s为依次从1开始取值的正整数，其最大值为用户在社交平台中加入的、且具有群组关系的至少一个群组的总数，图7中f9表示公式(9)。
第4步：将好友对候选项目的综合关注度以及群组对候选项目的综合关注度进行特征融合，获得推荐候选项目的分数。
其中，特征融合技术可以为加成融合技术，也可以为其他类型的融合技术，在此以加成融合技术为例，可将好友对候选项目的综合关注度F i以及群组对候选项目的综合关注度G i，通过以下公式(10)进行加成特征融合，生成表示用户对候选项目的偏好的分数，该分数用于表征向社交平台用户u i推荐候选项目的推荐指数：
Figure PCTCN2019121919-appb-000013
其中，v j为候选项目的向量表示，F i表示好友对候选项目的综合关注度，G i表示群组对候选项目的综合关注度，T表示转置。
第5步：根据获得的待推荐候选项目的分数，确定是否向社交平台用户推荐候选项目。
具体包括：在输入的候选项目为1个时，在根据上述第1步至第4步获得推荐该候选项目的分数之后，可以确定该分数是否大于推荐阈值，若是，则输出向社交平台用户u i推荐候选项目v j的结果，否则输出不向社交平台用户u i推荐该候选项目v j。
在输入的候选项目为多个时，在根据上述第1步至第4步获得推荐各候选项目的分数之后，可以输出向社交平台用户u i推荐分数最高的候选项目，也可以按照分数从高到低依次对各候选项目进行排序，并输出按照排序后的各候选项目向用户进行推荐。
基于上述实施例可知，本申请实施例中项目推荐方法是可以通过项目推荐模型执行的，通过项目推荐模型对待推荐项目用户和各待推荐项目进行处理，获得向待推荐用户推荐的项目，在应用之前需要先训练项目推荐模型，而通常好的模型需要有好的训练样本，因此，本申请实施例中还可针对原始数据进行处理，获得更为可靠的训练样本，并基于不同的训练样本从多个场景训练和优化项目推荐模型，可以提高项目推荐模型的准确性，下面对本申请实施例中项目推荐模型的训练过程进行具体说明：
需要说明的是，通常训练过程是由后台服务器设备执行，由于模型的各 个模块训练可能比较复杂，计算量较大，因此，由后台服务器设备实现训练过程，从而可以将训练好的模型和结果应用到各个智能终端，实现准确向待推荐用户推荐项目的目的。
需要说明的是，在基于不同的训练样本从多个场景训练和优化项目推荐模型，可以根据以下公式(11)，优化项目推荐模型：
Figure PCTCN2019121919-appb-000014
其中，j表示候选项目(即上文中的v j)，i表示用户(即上文中的u i)，k表示随机抽样负项，
Figure PCTCN2019121919-appb-000015
表示对推荐候选项目的推荐分数，L BPR表示排列损失，
Figure PCTCN2019121919-appb-000016
表示针对用户i的随机抽样负项，λ Θ(∥Θ∥ 2)表示正则项，用于防止过拟合，D表示样本集合。
所以，本申请实施例中项目推荐方法在通过项目推荐模型执行时，能够准确学习各个目标社交对象对候选项目的单项关注度，然后再根据综合关注度实现准确向待推荐用户推荐项目的目的，所以，能够提升向用户推荐项目的准确度，从而提升用户体验度。
下文中，采用现有的项目推荐准确度评估方法，针对相同的待推荐社交平台用户，分别对基于本申请实施例中的项目推荐方法、现有的多个推荐方法获得的向用户推荐的项目进行推荐准确度评估，其中，现有的项目推荐准确度评估方法可以采用Recall评估方法，该评估方法如公式(12)所示：
Figure PCTCN2019121919-appb-000017
其中，k表示各算法获得的向用户推荐的项目总数，relj取值0或1，表示第j个项目是否在测试集的推向列表中，若是，relj取值为1，否则取值为0，
Figure PCTCN2019121919-appb-000018
表示测试集中用户u所评定的候选项目的总数。
现有的项目推荐准确度评估方法也可以采用Normalized Discounted Cumulative Gain评估方法即NDCG评估方法，该评估方法如公式(13)和公式(14)所示：
Figure PCTCN2019121919-appb-000019
Figure PCTCN2019121919-appb-000020
其中，k表示各算法获得的向用户推荐的项目总数，relj取值0或1，表示第j个项目是否在测试集的推向列表中，若是，relj取值为1。
如图8所示，在本申请实施例中，同时采用了Recall评估方法以及NDCG评估方法，对本申请实施例中的项目推荐方法、现有的推荐方法如图8中的MP(Most Popular)方法、ItemKNN方法、BPR方法、神经协同过滤NCF(Neural Collaborative Filtering)方法、SBPR方法以及SAMN方法获得的向用户推荐的项目进行推荐准确度评估。
在图8，使用SACF(Social Attentional Collaborative Filtering)表示本申请实施例中的项目推荐方法，从图8中显示的各推荐方法的评估结果中可以看出，首先，MP方法的效果不太理想，这也说明了对用户的偏好建模的重要性，而不仅仅是推荐热门的项目给用户。其次，采用本申请实施例中的项目推荐方法SACF获得的向用户推荐的项目的准确度高于现有的多种推荐方法，从图8可以看出，采用Recall@10评估方法，本申请实施例提供的SACF方法在两个数据集(WeChat-10k，WeChat-100k)上的性能分别比SAMN方法提升了约3.67％和5.01％，采用NDCG@10评估方法，本申请实施例提供的SACF方法在两个数据集上的性能分别比SAMN方法提升了月4.17％和5.49％。所以，本申请实施例中的项目推荐方法实现了准确向待推荐用户推荐项目的目的，避免向待推荐用户推荐不感兴趣的项目，提升了资源利用率和用户体 验度。
基于同一构思，本申请实施例中提供了一种项目推荐装置，如图9所示，包括：
获得模块90，用于获得待推荐给社交平台用户的候选项目；
第一确定模块91，用于针对所述社交平台用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定目标社交对象对所述候选项目的单项关注度；
第二确定模块92，用于根据各个目标社交对象对所述候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度；
第三确定模块93，用于根据所述综合关注度确定是否向所述社交平台用户推荐所述候选项目。
所述不同类型的社交关系包括群组关系，所述社交关系中的至少一个目标社交对象，包括所述社交平台用户在所述社交平台中加入的至少一个群组。
所述第一确定模块，用于：
将所述至少一个群组的向量表示和所述候选项目的向量表示，输入预先训练的第一注意力模型，所述第一注意力模型中预先训练有所述社交平台的各群组对所述候选项目的注意力参数；
并获得所述第一注意力模型输出的各群组对所述候选项目的单项关注度，该单项关注度为所述第一注意力模型根据各群组的向量表示以及所述候选项目的向量表示，确定各群组分别对所述候选项目的注意力参数，以及根据确定的注意力参数生成并输出的群组对所述候选项目的单项关注度。
所述第一注意力模型中预先还针对每个群组，训练有该群组中各群成员在群组中的重要度权重；以及
所述第一注意力模型输出的各群组对所述候选项目的单项关注度，是进一步将各群组中各群成员在各群组中的重要度权重，与各群组分别对所述候选项目的注意力参数进行加权处理后获得的。
所述不同类型的社交关系包括好友关系，所述社交关系中的至少一个目 标社交对象，包括所述社交平台用户在所述社交平台中添加的至少一个好友。
所述第一确定模块，还用于：
将所述至少一个好友的向量表示与所述候选项目的向量表示，输入预先训练的第二注意力模型，所述第二注意力模型中预先训练有所述社交平台的各好友对所述候选项目的注意力参数；
并获得所述第二注意力模型输出的各好友对所述候选项目的单项关注度，该单项关注度为所述第二注意力模型根据各好友的向量表示和所述候选项目的向量表示，确定各好友对所述候选项目的注意力参数，并根据确定的注意力参数生成的各好友对所述候选项目的单项关注度。
所述第二确定模块，用于：
将各个目标社交对象对所述候选项目的单项关注度进行归一化处理；
将归一化处理后相同类型的目标社交对象对所述候选项目的单项关注度进行求和处理，获得各类型的社交对象对所述候选项目的综合关注度；以及
所述第三确定模块，用于：
将各类型的社交对象对所述候选项目的综合关注度进行特征融合，获得推荐所述候选项目的推荐指数；
并根据所述推荐指数，确定是否向所述社交平台用户推荐所述候选项目。
基于同一构思，本申请实施例中提供了一种计算装置，如图10所示，包括至少一个处理器101、以及至少一个存储器102，其中，所述存储器102存储有计算机程序，当所述程序被所述处理器101执行时，使得所述处理器101执行上文所述的项目推荐方法的步骤。
基于同一构思，本申请实施例中提供了一种存储介质，所述存储介质存储有计算机指令，当所述计算机指令在计算机上运行时，使得计算机执行上文所述的项目推荐方法的步骤。
本领域内的技术人员应明白，本申请的实施例可提供为方法、***、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个 其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器和光学存储器等)上实施的计算机程序产品的形式。
本申请是根据本申请实施例的方法、设备(***)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
显然，本领域的技术人员可以对本申请进行各种改动和变型而不脱离本申请的精神和范围。这样，倘若本申请的这些修改和变型属于本申请权利要求及其等同技术的范围之内，则本申请也意图包含这些改动和变型在内。

Claims (17)

  1. 一种推荐方法，由计算装置执行，包括：
    获得待推荐给社交平台用户的候选项目；
    针对所述社交平台用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分别确定各个目标社交对象对所述候选项目的单项关注度；
    根据各个目标社交对象对所述候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度；
    并根据所述综合关注度确定是否向所述社交平台用户推荐所述候选项目。
  2. 如权利要求1所述的方法，其中，所述不同类型的社交关系包括群组关系，所述社交关系中的至少一个目标社交对象，包括所述社交平台用户在所述社交平台中加入的至少一个群组。
  3. 如权利要求2所述的方法，其中，所述分别确定各个目标社交对象对所述候选项目的单项关注度，具体包括：
    将所述至少一个群组的向量表示和所述候选项目的向量表示，输入预先训练的第一注意力模型，所述第一注意力模型中预先训练有所述社交平台的各群组对所述候选项目的注意力参数；
    并获得所述第一注意力模型输出的各群组对所述候选项目的单项关注度，该单项关注度为所述第一注意力模型根据各群组的向量表示以及所述候选项目的向量表示，确定各群组分别对所述候选项目的注意力参数，以及根据确定的所述注意力参数生成并输出各群组对所述候选项目的单项关注度。
  4. 如权利要求3所述的方法，其中，所述第一注意力模型中预先还针对每个群组，训练有该群组中各群成员在群组中的重要度权重；以及
    所述第一注意力模型输出的各群组对所述候选项目的单项关注度，是进一步将各群组中各群成员在各群组中的重要度权重，与各群组分别对所述候选项目的注意力参数进行加权处理后获得的。
  5. 如权利要求2-4中任一项所述的方法，其中，所述不同类型的社交关系还包括好友关系，所述社交关系中的至少一个目标社交对象，包括所述社交平台用户在所述社交平台中添加的至少一个好友。
  6. 如权利要求5所述的方法，其中，所述分别确定各个目标社交对象对所述候选项目的单项关注度，还包括：
    将所述至少一个好友的向量表示与所述候选项目的向量表示，输入预先训练的第二注意力模型，所述第二注意力模型中预先训练有所述社交平台的各好友对所述候选项目的注意力参数；
    并获得所述第二注意力模型输出的各好友对所述候选项目的单项关注度，该单项关注度为所述第二注意力模型根据各好友的向量表示和所述候选项目的向量表示，确定各好友对所述候选项目的注意力参数，并根据确定的注意力参数生成的各好友对所述候选项目的单项关注度。
  7. 如权利要求1-4、6中任一项所述的方法，其中：
    所述根据各个目标社交对象对所述候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度，具体包括：
    将各个目标社交对象对所述候选项目的单项关注度进行归一化处理；
    将归一化处理后相同类型的目标社交对象对所述候选项目的单项关注度进行求和处理，获得各类型的社交对象对所述候选项目的综合关注度；以及
    所述根据所述综合关注度确定是否向所述社交平台用户推荐所述候选项目，具体包括：
    将各类型的社交对象对所述候选项目的综合关注度进行特征融合，获得推荐所述候选项目的推荐指数；
    并根据所述推荐指数，确定是否向所述社交平台用户推荐所述候选项目。
  8. 一种推荐装置，包括：
    获得模块，用于获得待推荐给社交平台用户的候选项目；
    第一确定模块，用于针对所述社交平台用户在社交平台中的至少两种不同类型的社交关系中，每种类型的社交关系中的至少一个目标社交对象，分 别确定各个目标社交对象对所述候选项目的单项关注度；
    第二确定模块，用于根据各个目标社交对象对所述候选项目的单项关注度，确定不同类型的目标社交对象对所述候选项目的综合关注度；
    第三确定模块，用于根据所述综合关注度确定是否向所述社交平台用户推荐所述候选项目。
  9. 如权利要求8所述的装置，其中，所述不同类型的社交关系包括群组关系，所述社交关系中的至少一个目标社交对象，包括所述社交平台用户在所述社交平台中加入的至少一个群组。
  10. 如权利要求9所述的装置，其中，所述第一确定模块，还用于：
    将所述至少一个群组的向量表示和所述候选项目的向量表示，输入预先训练的第一注意力模型，所述第一注意力模型中预先训练有所述社交平台的各群组对所述候选项目的注意力参数；
    并获得所述第一注意力模型输出的各群组对所述候选项目的单项关注度，该单项关注度为所述第一注意力模型根据各群组的向量表示以及所述候选项目的向量表示，确定各群组分别对所述候选项目的注意力参数，以及根据确定的所述注意力参数生成并输出各群组对所述候选项目的单项关注度。
  11. 如权利要求10所述的装置，其中，所述第一注意力模型中预先还针对每个群组，训练有该群组中各群成员在群组中的重要度权重；以及
    所述第一注意力模型输出的各群组对所述候选项目的单项关注度，是进一步将各群组中各群成员在各群组中的重要度权重，与各群组分别对所述候选项目的注意力参数进行加权处理后获得的。
  12. 如权利要求8-10中任一项所述的装置，其中，所述不同类型的社交关系还包括好友关系，所述社交关系中的至少一个目标社交对象，包括所述社交平台用户在所述社交平台中添加的至少一个好友。
  13. 如权利要求12所述的装置，其中，所述第一确定模块，还用于：
    将所述至少一个好友的向量表示与所述候选项目的向量表示，输入预先训练的第二注意力模型，所述第二注意力模型中预先训练有所述社交平台的 各好友对所述候选项目的注意力参数；
    并获得所述第二注意力模型输出的各好友对所述候选项目的单项关注度，该单项关注度为所述第二注意力模型根据各好友的向量表示和所述候选项目的向量表示，确定各好友对所述候选项目的注意力参数，并根据确定的所述注意力参数生成各好友对所述候选项目的单项关注度。
  14. 如权利要求12所述的装置，其中，所述第二确定模块，还用于：
    将各个目标社交对象对所述候选项目的单项关注度进行归一化处理；
    将归一化处理后相同类型的目标社交对象对所述候选项目的单项关注度进行求和处理，获得各类型的社交对象对所述候选项目的综合关注度。
  15. 如权利要求14所述的装置，其中，所述第三确定模块，还用于：
    将各类型的社交对象对所述候选项目的综合关注度进行特征融合，获得推荐所述候选项目的推荐指数；
    并根据所述推荐指数，确定是否向所述社交平台用户推荐所述候选项目。
  16. 一种计算装置，包括至少一个处理器、以及至少一个存储器，其中，所述存储器存储有计算机程序，当所述程序被所述处理器执行时，使得所述处理器执行权利要求1～7任一项所述方法的步骤。
  17. 一种非易失性计算机可读存储介质，所述存储介质存储有计算机指令，当所述计算机指令在计算机上运行时，使得计算机执行如权利要求1-7任一项所述的方法的步骤。
PCT/CN2019/121919 2018-11-29 2019-11-29 一种推荐方法、装置及存储介质 WO2020108605A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
JP2021527278A JP7299320B2 (ja) 2018-11-29 2019-11-29 推薦方法と装置、コンピューティング装置、及びプログラム
US17/183,251 US11709902B2 (en) 2018-11-29 2021-02-23 Recommendation method and apparatus, and storage medium
US18/309,300 US20230259566A1 (en) 2018-11-29 2023-04-28 Recommending an item based on attention of user group to item

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811445266.1A CN110245301A (zh) 2018-11-29 2018-11-29 一种推荐方法、装置及存储介质
CN201811445266.1 2018-11-29

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/183,251 Continuation US11709902B2 (en) 2018-11-29 2021-02-23 Recommendation method and apparatus, and storage medium

Publications (1)

Publication Number Publication Date
WO2020108605A1 true WO2020108605A1 (zh) 2020-06-04

Family

ID=67882442

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/121919 WO2020108605A1 (zh) 2018-11-29 2019-11-29 一种推荐方法、装置及存储介质

Country Status (4)

Country Link
US (2) US11709902B2 (zh)
JP (1) JP7299320B2 (zh)
CN (1) CN110245301A (zh)
WO (1) WO2020108605A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112084407A (zh) * 2020-09-08 2020-12-15 辽宁工程技术大学 一种融合图神经网络和注意力机制的协同过滤推荐方法
CN112328893A (zh) * 2020-11-25 2021-02-05 重庆理工大学 一种基于记忆网络和协同注意力的推荐方法
CN113434762A (zh) * 2021-06-28 2021-09-24 平安银行股份有限公司 基于用户信息的关联推送方法、装置、设备及存储介质
CN113761378A (zh) * 2021-09-14 2021-12-07 上海任意门科技有限公司 内容排序方法、计算设备和计算机可读存储介质
CN116738034A (zh) * 2022-10-10 2023-09-12 荣耀终端有限公司 一种信息推送方法及***

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2014180792A1 (en) * 2013-05-05 2014-11-13 Lantiq Deutschland Gmbh Training optimization of multiple lines in a vectored system using a prepared-to-join group
CN110245301A (zh) * 2018-11-29 2019-09-17 腾讯科技（深圳）有限公司 一种推荐方法、装置及存储介质
US11157702B2 (en) * 2019-03-06 2021-10-26 International Business Machines Corporation Utilizing varying coordinates related to a target event to provide contextual outputs
US11170789B2 (en) * 2019-04-16 2021-11-09 Microsoft Technology Licensing, Llc Attentive adversarial domain-invariant training
CN110781389A (zh) * 2019-10-18 2020-02-11 支付宝(杭州)信息技术有限公司 用于为用户生成推荐的方法和***
CN113032685B (zh) * 2019-12-06 2023-11-28 北京达佳互联信息技术有限公司 基于社交关系的对象推送方法、装置、设备及存储介质
CN111523045A (zh) * 2020-03-09 2020-08-11 浙江工业大学 一种协同过滤推荐方法
CN112231567A (zh) * 2020-10-20 2021-01-15 王明烨 一种精确推送信息的方法及装置
CN112364259A (zh) * 2020-11-24 2021-02-12 深圳市元征科技股份有限公司 一种信息推荐方法、装置、设备及介质
CN113761338A (zh) * 2021-01-12 2021-12-07 北京沃东天骏信息技术有限公司 信息推送方法、装置和计算机可读存储介质
CN112883268B (zh) * 2021-02-22 2022-02-01 中国计量大学 一种考虑用户多兴趣以及社交影响的会话推荐方法
CN113326440B (zh) * 2021-08-03 2021-11-02 腾讯科技（深圳）有限公司 基于人工智能的推荐方法、装置及电子设备
CN113609398B (zh) * 2021-08-17 2023-09-19 石家庄铁道大学 一种基于异构图神经网络的社交推荐方法
CN113868542B (zh) * 2021-11-25 2022-03-11 平安科技（深圳）有限公司 基于注意力模型的推送数据获取方法、装置、设备及介质
CN114510653B (zh) * 2022-04-21 2022-07-15 中国科学技术大学 社交群体推荐方法、***、设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130325560A1 (en) * 2012-06-05 2013-12-05 Rank Miner, Inc. System, Method and Apparatus for Voice Analytics of Recorded Audio
CN107767279A (zh) * 2017-10-12 2018-03-06 燕山大学 一种基于lda的加权平均的个性化好友推荐方法
CN108763314A (zh) * 2018-04-26 2018-11-06 深圳市腾讯计算机***有限公司 一种兴趣推荐方法、装置、服务器及存储介质
CN110245301A (zh) * 2018-11-29 2019-09-17 腾讯科技（深圳）有限公司 一种推荐方法、装置及存储介质
CN110351318A (zh) * 2018-04-04 2019-10-18 腾讯科技（深圳）有限公司 应用推荐的方法、终端和计算机存储介质

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8489515B2 (en) * 2009-05-08 2013-07-16 Comcast Interactive Media, LLC. Social network based recommendation method and system
JP6217228B2 (ja) * 2013-08-12 2017-10-25 大日本印刷株式会社 情報提供装置、プログラム及び情報提供システム
US10944707B2 (en) * 2014-09-26 2021-03-09 Line Corporation Method, system and recording medium for providing video contents in social platform and file distribution system
CN105574045B (zh) * 2014-10-17 2019-12-10 深圳市腾讯计算机***有限公司 一种视频推荐方法和服务器
CN104573109A (zh) * 2015-01-30 2015-04-29 深圳市中兴移动通信有限公司 一种基于群组关系的自动推荐方法、终端及***
CN106294497B (zh) * 2015-06-09 2020-05-12 深圳市腾讯计算机***有限公司 信息推荐方法和装置
US10558714B2 (en) * 2016-12-28 2020-02-11 Facebook, Inc. Topic ranking of content items for topic-based content feeds
CN106649842A (zh) 2016-12-30 2017-05-10 上海博泰悦臻电子设备制造有限公司 一种基于融合数据的交叉推荐方法、***及一种车机
JP6744353B2 (ja) * 2017-04-06 2020-08-19 ネイバー コーポレーションＮＡＶＥＲ Ｃｏｒｐｏｒａｔｉｏｎ ディープラーニングを活用した個人化商品推薦

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130325560A1 (en) * 2012-06-05 2013-12-05 Rank Miner, Inc. System, Method and Apparatus for Voice Analytics of Recorded Audio
CN107767279A (zh) * 2017-10-12 2018-03-06 燕山大学 一种基于lda的加权平均的个性化好友推荐方法
CN110351318A (zh) * 2018-04-04 2019-10-18 腾讯科技（深圳）有限公司 应用推荐的方法、终端和计算机存储介质
CN108763314A (zh) * 2018-04-26 2018-11-06 深圳市腾讯计算机***有限公司 一种兴趣推荐方法、装置、服务器及存储介质
CN110245301A (zh) * 2018-11-29 2019-09-17 腾讯科技（深圳）有限公司 一种推荐方法、装置及存储介质

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112084407A (zh) * 2020-09-08 2020-12-15 辽宁工程技术大学 一种融合图神经网络和注意力机制的协同过滤推荐方法
CN112084407B (zh) * 2020-09-08 2024-03-12 辽宁工程技术大学 一种融合图神经网络和注意力机制的协同过滤推荐方法
CN112328893A (zh) * 2020-11-25 2021-02-05 重庆理工大学 一种基于记忆网络和协同注意力的推荐方法
CN112328893B (zh) * 2020-11-25 2022-08-02 重庆理工大学 一种基于记忆网络和协同注意力的推荐方法
CN113434762A (zh) * 2021-06-28 2021-09-24 平安银行股份有限公司 基于用户信息的关联推送方法、装置、设备及存储介质
CN113761378A (zh) * 2021-09-14 2021-12-07 上海任意门科技有限公司 内容排序方法、计算设备和计算机可读存储介质
CN116738034A (zh) * 2022-10-10 2023-09-12 荣耀终端有限公司 一种信息推送方法及***

Also Published As

Publication number Publication date
US11709902B2 (en) 2023-07-25
US20230259566A1 (en) 2023-08-17
JP7299320B2 (ja) 2023-06-27
US20210173884A1 (en) 2021-06-10
CN110245301A (zh) 2019-09-17
JP2021532515A (ja) 2021-11-25

Similar Documents

Publication Publication Date Title
WO2020108605A1 (zh) 一种推荐方法、装置及存储介质
US11735163B2 (en) Human-machine dialogue method and electronic device
WO2020107806A1 (zh) 一种推荐方法及装置
US10354184B1 (en) Joint modeling of user behavior
WO2022016522A1 (zh) 推荐模型的训练方法、推荐方法、装置及计算机可读介质
JP2018190396A (ja) ネットワークレーティング予測エンジン
US8868471B1 (en) Evaluation of task judging results
US20160162837A1 (en) Collaboration pattern creation by crowdsourcing participants
Kokkodis Dynamic, multidimensional, and skillset-specific reputation systems for online work
Alshmrany Adaptive learning style prediction in e-learning environment using levy flight distribution based CNN model
US20210182976A1 (en) Methods and systems for social media recommendations engines
Shen et al. A voice of the customer real-time strategy: An integrated quality function deployment approach
CN109977979B (zh) 定位种子用户的方法、装置、电子设备和存储介质
CN115631008B (zh) 商品推荐方法、装置、设备及介质
Mogavi et al. HRCR: hidden Markov-based reinforcement to reduce churn in question answering forums
KR102601446B1 (ko) 인공지능 모델을 활용한 인플루언서 기반 판매 제품매칭 플랫폼 서비스 제공 방법, 장치 및 시스템
KR102592129B1 (ko) 인공지능에 기반한 사용자 맞춤형 의류 추천시스템
Pandita et al. Acceptance of Ride-sharing in India: Empirical Evidence from the UTAUT Model.
KR20220075623A (ko) 구매상품 리뷰 큐레이션 및 라이브커머스를 통한 상품 구매 방법, 장치 및 프로그램
Änäkkälä Exploring value in eCommerce artificial intelligence and recommendation systems
CN112561000B (zh) 基于组合模型的分类方法、装置、设备及存储介质
CN113010772B (zh) 一种数据处理方法、相关设备及计算机可读存储介质
KR102537900B1 (ko) 사용자 특징을 지시하는 태그를 생성하기 위한 장치, 방법 및 명령을 기록한 기록 매체
Sripituksakul Education about Digital Marketing to Students through Entrepreneurship Programs
WO2022262561A1 (zh) 多媒体资源的处理方法、装置、设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19889547

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2021527278

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19889547

Country of ref document: EP

Kind code of ref document: A1