CN117743675A

CN117743675A - Resource recall method

Info

Publication number: CN117743675A
Application number: CN202310177168.9A
Authority: CN
Inventors: 沈慧; 苏睿龙; 高波; 欧宝源
Original assignee: Xiaohongshu Technology Co ltd
Current assignee: Xiaohongshu Technology Co ltd
Priority date: 2023-02-25
Filing date: 2023-02-25
Publication date: 2024-03-22

Abstract

The embodiment of the application discloses a resource recall method. The method comprises the following steps: acquiring a time length behavior sequence of a target object, and generating a time length interest representation of the time length behavior sequence, wherein the time length behavior sequence comprises at least one time length behavior feature of the target object; acquiring an interactive behavior sequence of a target object, and generating an interactive interest representation of the interactive behavior sequence, wherein the interactive behavior sequence comprises at least one interactive behavior feature of the target object; fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the target object to obtain a multi-interest characterization vector of the target object; and determining recall resources of the target object based on the multi-interest characterization vector of the target object. By adopting the method and the device, the interest preference of the object can be accurately mined, so that the accuracy of resource recall is improved.

Description

Resource recall method

Technical Field

The application relates to the technical field of computer application, in particular to a resource recall method.

Background

With the rapid development of computing technology, recommending resources (such as resource recall) for users in various social applications and search scenarios is a hotspot of great research. Currently, recall of resources is performed for a target user, and the recall resources are generally determined by calculating the matching degree between the target user and the resources through the interaction behavior of the target user. However, when the conventional recommendation recall algorithm models the behavior of the user, the user is usually characterized as a single vector, all interests of the user are mixed in a 64-bit vector, interest preference of the user cannot be accurately mined, and therefore the accuracy of resource recall is low.

Disclosure of Invention

The embodiment of the application provides a resource recall method which can accurately mine interest preference of an object, thereby improving the accuracy of resource recall.

In one aspect, an embodiment of the present application provides a resource recall method, where the method includes:

acquiring a time length behavior sequence of a target object, and generating a time length interest characterization of the time length behavior sequence; wherein the sequence of time duration behaviors includes at least one time duration behavior feature of the target object;

acquiring an interactive behavior sequence of a target object, and generating an interactive interest representation of the interactive behavior sequence; wherein the interactive behavior sequence comprises at least one interactive behavior feature of the target object;

fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the target object to obtain a multi-interest characterization vector of the target object;

and determining recall resources of the target object based on the multi-interest characterization vector of the target object.

In one embodiment, the generating the duration interest representation of the duration behavior sequence includes:

acquiring the weight of each duration behavior feature;

and carrying out multi-layer perceptron operation on the at least one time length behavior feature based on the weight of each time length behavior feature to obtain the time length interest characterization of the time length behavior sequence.

acquiring the weight of each duration behavior feature;

In one embodiment, the generating the interactive interest representation of the interactive behavior sequence includes:

acquiring the weight of each interactive behavior feature;

and carrying out multi-layer perceptron operation on the at least one interactive behavior feature based on the weight of each interactive behavior feature to obtain the interactive interest representation of the interactive behavior sequence.

In one embodiment, the performing a multi-layer perceptron operation on the at least one interactive behavior feature based on the weights of the interactive behavior features to obtain an interactive interest representation of the interactive behavior sequence includes:

splicing the identity characteristic of the target object and the at least one interactive behavior characteristic to obtain at least one spliced interactive behavior characteristic;

and carrying out multi-layer perceptron operation on at least one spliced interactive behavior feature based on the weight of each interactive behavior feature to obtain the interactive interest characterization of the interactive behavior sequence.

In one embodiment, the fusing the identity token, the duration interest token and the interaction interest token of the target object to obtain a multi-interest token vector of the target object includes:

generating weights of identity characterization, duration interest characterization and interaction interest characterization through a gating network;

and carrying out weighted operation on the identity characterization, the duration interest characterization and the interaction interest characterization based on the weights of the identity characterization, the duration interest characterization and the interaction interest characterization to obtain a multi-interest characterization vector of the target object.

In one embodiment, the multiple interest characterization vector includes multiple interest vectors; the determining recall resources of the target object based on the multiple interest characterization vectors of the target object includes:

acquiring resources corresponding to each interest vector in the interest vectors based on the corresponding relation between the interest vector and the resources;

and taking the obtained resource as a recall resource of the target object.

In one embodiment, the method further comprises:

obtaining a candidate set, the candidate set comprising a plurality of resources;

extracting the characteristics of each resource in the candidate set to obtain a resource characteristic vector of each resource;

For any resource of the plurality of resources, determining an interest vector with the smallest spatial distance with the any resource based on the spatial distances of the respective interest vector and the resource feature vector of the any resource;

and establishing a corresponding relation between the determined interest vector and any resource.

In one embodiment, the multiple interest characterization vector is derived based on an interest characterization model, the method further comprising:

obtaining a training sample; wherein the training samples comprise positive samples and negative samples, the positive samples comprise first training resources pushed to a training object in a history time period and clicked by the training object, and the negative samples comprise second training resources not pushed to the training object in the history time period;

acquiring a duration behavior sequence and an interaction behavior sequence of the training object about the first training resource;

invoking an initial interest characterization model, and generating a duration interest characterization of the duration behavior sequence of the training object and an interaction interest characterization of the interaction behavior sequence of the training object;

fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the training object to obtain a multi-interest characterization vector of the training object;

Determining a first interest vector with the smallest space distance to the first training resource and a second interest vector with the smallest space distance to the second training resource based on each interest vector in the multi-interest characterization vectors of the training object;

and training the initial interest characterization model by taking the space distance between the first interest vector and the first training resource as a training target and increasing the space distance between the second interest vector and the second training resource to obtain the interest characterization model.

In one embodiment, the method further comprises:

optimizing the dense parameters and the sparse parameters of the interest characterization model obtained by the last optimization every first preset time period to obtain the interest characterization model;

optimizing sparse parameters of the interest characterization model obtained by the last optimization every second preset time period to obtain the interest characterization model, wherein the duration indicated by the first preset time period is longer than that indicated by the second preset time period.

On the other hand, the embodiment of the application provides a resource recall device, which comprises:

the first generation unit is used for acquiring a time length behavior sequence of the target object and generating a time length interest representation of the time length behavior sequence; wherein the sequence of time duration behaviors includes at least one time duration behavior feature of the target object;

The second generation unit is used for acquiring the interactive behavior sequence of the target object and generating interactive interest characterization of the interactive behavior sequence; wherein the interactive behavior sequence comprises at least one interactive behavior feature of the target object;

the fusion unit is used for fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the target object to obtain a multi-interest characterization vector of the target object;

and the determining unit is used for determining recall resources of the target object based on the multi-interest characterization vector of the target object.

In another aspect, an embodiment of the present application provides a computer device, including a processor, a storage device, and a communication interface, where the processor, the storage device, and the communication interface are connected to each other, where the storage device is configured to store a computer program that supports the computer device to perform the method, the computer program includes program instructions, and the processor is configured to invoke the program instructions to perform the following steps:

In another aspect, embodiments of the present application provide a computer-readable storage medium storing a computer program comprising program instructions that, when executed by a processor, cause the processor to perform the above-described resource recall method.

In another aspect, embodiments of the present application provide a computer program product comprising a computer program adapted to be loaded by a processor and to perform the above-described resource recall method.

In the embodiment of the application, the duration behavior sequence of the target object is acquired, the duration interest representation of the duration behavior sequence is generated, in addition, the interactive behavior sequence of the target object is acquired, the interactive interest representation of the interactive behavior sequence is generated, then the identity representation, the duration interest representation and the interactive interest representation of the target object are fused to obtain the multi-interest representation vector of the target object, and the interest representation vector of the target object on different interests can be extracted by splitting different behavior sequences of the target object, so that the learning of the difference of different behaviors of the target object is facilitated, and the interest preference of the target object is accurately mined in different behaviors. Furthermore, based on the multi-interest characterization vector of the target object obtained in the mode, recall resources of the target object are determined, and accuracy of resource recall can be improved.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow chart of a resource recall method according to an embodiment of the present application;

FIG. 2 is a schematic diagram of an architecture of a characterization model of interest provided in an embodiment of the present application;

FIG. 3 is a schematic diagram of a user attention mechanism according to an embodiment of the present application;

fig. 4 is a schematic structural diagram of a gating network according to an embodiment of the present application;

FIG. 5 is a flowchart of another resource recall method according to an embodiment of the present application;

FIG. 6 is a schematic structural diagram of a resource recall device according to an embodiment of the present application;

fig. 7 is a schematic structural diagram of a computer device according to an embodiment of the present application.

Detailed Description

The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without making any inventive effort, are intended to be within the scope of the present application.

In the whole link system of the recommendation system, recall is the most basic module, and has the significance of narrowing the calculation range of candidate sets, and is responsible for selecting a set possibly interested by a user from a large number of candidate sets, wherein the closer the recall resource is to the interest of the user, the higher the accuracy of the whole recommendation system is. However, when the traditional resource recall algorithm models the behavior of the user, the user is characterized as a single vector, all interests of the user are mixed in the 64-bit vector, the expression of interest diversity of the user is not facilitated, and interest preference of the user cannot be accurately mined, so that the accuracy of resource recall is low.

Based on the method, the resource recall method based on multi-sequence multi-interest modeling is provided, different behavior sequences of the target object are respectively described, the multi-interest characterization vector of the target object is extracted in a multi-vector mode, interest preference of the target object can be accurately mined, the accuracy of resource recall is improved, resources of interest of more target objects are supplemented, and interactive willingness of the target object is enhanced. Meanwhile, as the behavior habits of different objects can be learned individually, the real interest preference of the objects can be observed in the different behavior habits, and the interest expression of the object masses can be better mined and reserved, so that the method has a positive driving effect on the diversity of recall resources. In addition, multiple interest characterization of the object under different behaviors is extracted, and compared with single interest characterization, sparse long-tail interest of the object can be recorded better, for example, a certain object browses a note related to a vehicle and collects the note, but the object browses more notes related to a through-put, if the object is characterized as a one-way quantity by adopting a traditional resource recall algorithm, the note related to the through-put of the object can be submerged, so that the object cannot be identified to be interested in the through-put and the vehicle, and the embodiment of the application can play a positive driving role in long-tail distribution.

The target object may refer to a user, in particular at least one user, or at least one type of user.

A resource refers to a resource released by a content release platform, such as an electronic resource or an entity resource. The assets may include advertisements, video, audio, text, images, merchandise, or the like. By way of example, the resources may include notes or short videos, among others.

In the specific embodiments of the present application, data related to an object, such as seed resources for generating interaction behavior by a target object within a preset period of time, when the embodiments of the present application are applied to specific products or technologies, permission or consent of the object needs to be obtained, and collection, use and processing of related data need to comply with local laws and regulations and standards.

The resource recall method provided by the embodiment of the application can be applied to a resource recall device, the resource recall device can be installed or integrated in a content release platform, the content release platform can be operated in computer equipment, the computer equipment can comprise terminal equipment or a server and the like, and the computer equipment comprises but is not limited to smart phones, vehicle-mounted equipment, wearable equipment or computers and the like.

Referring to fig. 1, fig. 1 is a schematic flow chart of a resource recall method provided in an embodiment of the present application, where the resource recall method may be executed by a resource recall device or a computer device; the resource recall scheme shown in fig. 1 includes, but is not limited to, steps S101 to S104, wherein:

s101, acquiring a time length behavior sequence of a target object, and generating a time length interest characterization of the time length behavior sequence.

Wherein the sequence of time duration behavior may comprise at least one time duration behavior feature of the target object. The at least one time duration behavior feature may include one or more of the following: the browsing time length of each resource, the running time length of the client corresponding to the target object, the browsing time length of the resources with different resource types, and the like.

For example, after the target object submits the access request to the content publishing platform, the content publishing platform may push the determined resource to the target object, and the target object may browse some or all of the resources, for example, if a click operation of the target object on the target resource is detected, the content publishing platform may play the target resource, and further may count the browsing duration of the target object on the target resource, that is, the playing duration of the target resource. For another example, the target object may log in the content publishing platform through the account number, and then the running duration of the client corresponding to the target object in the unit time may be counted, for example, the duration that the target object browses the content publishing platform every day. For another example, if the click operation of the target object on the target resource is detected, the resource type of the target resource and the browsing duration of the target object on the target resource can be counted, so that the browsing duration of the target object on the resources with different resource types in the preset time period can be obtained. For example, if the target object browses three notes related to the punch-through in a week, where the browsing duration of the target object for the three notes related to the punch-through is 2 minutes (min), 1min, and 5min, respectively, it may be determined that the browsing duration of the target object for the punch-through type notes in a week is 2+1+5=8 min. Further, the counted at least one time length behavior feature may be ranked according to a time sequence to obtain a time length behavior sequence of the target object, for example, the target object browses the first resource in 2023, 1 month, 5 days, 9:00, the browsing time length of the first resource is 1min, and browses the second resource in 2023, 1 month, 5 days, 9:01, and the browsing time length of the second resource is 2min, and then the browsing time lengths of the resources may be ranked based on the browsing time points of the resources to obtain the time length behavior sequence of the target object.

In one implementation, the manner of generating the duration interest representation of the duration behavior sequence may include: the method comprises the steps of obtaining weights of all duration behavior features, and carrying out weighting operation on at least one duration behavior feature based on the weights of all duration behavior features to obtain duration interest characterization of a duration behavior sequence.

The weights of the behavior characteristics of each duration can be set by a developer based on experience, or can be learned through a neural network, and the weights are not limited by the embodiment of the application. For example, the at least one duration behavior feature includes three dimensions of duration behavior features of browsing duration of each resource, running duration of a client corresponding to the target object, and browsing duration of resources of different resource types, and then weights of the duration behavior features of each dimension may be set.

The multi-layer perceptron operation is performed on at least one time length behavior feature based on the weight of each time length behavior feature, namely the multi-interest characterization of the time length behavior sequence of the target object is extracted, and the time length interest characterization of the time length behavior sequence can be obtained through calculation according to the following formula (1).

Wherein V is _u May represent a duration interest characterization, H may represent a duration behavior sequence, And W is ₂ Is a preset parameter. The softmax function, also called as normalized exponential function, is a popularization of a two-class function sigmoid on multiple classes, and the softmax function has the function of displaying the results of the multiple classes in the form of probability.

In one implementation, based on the weights of the duration behavior features, performing a multi-layer perceptron operation on at least one duration behavior feature to obtain a duration interest representation mode of the duration behavior sequence, which may include: and performing attention calculation on at least one time length behavior feature by using the identity feature of the target object to obtain the processed at least one time length behavior feature, and performing multi-layer perceptron calculation on the processed at least one time length behavior feature based on the weight of each time length behavior feature to obtain the time length interest representation of the time length behavior sequence.

Wherein the identity of the target object may be used to identify identity information of the target object, e.g., the identity may include one or more of the following: age, gender, occupation, geographical area in which the user is located, hobbies, interests, etc.

Illustratively, the attention calculation may be performed on at least one duration behavior feature through formula (2), to obtain at least one processed duration behavior feature.

H＝sigmoid(W ₁ [e _u-static ，e _u-pt ]+b ₁ ) Formula (2)

Wherein H may represent at least one time-long behavioral characteristic after processing, e _u-static Identity (user static feature), e, which may represent a target object _u-pt Can represent at least one time-long behavior feature of the target object, i.e. a time-long behavior sequence (long term pagetime sequence), W of the target object ₁ And b ₁ Is a preset parameter. The Sigmoid function is a Sigmoid function common in biology, also known as an Sigmoid growth curve.

In this embodiment, for a duration behavior sequence, the interest signal of the target object is implicit, for example, the browsing time of the target object on the first resource is longer than the browsing time of the target object on the second resource, but it cannot be directly indicated that the target object is interested in the first resource, or the interest degree of the target object on the first resource is higher than the interest degree of the target object on the second resource. Therefore, by using the identity characteristics of the target object to perform attention calculation on at least one long-duration behavior characteristic, important information can be extracted from one long sequence, namely unimportant information can be filtered, and compared with the method for directly acquiring the interest characterization of the target object based on all the information in the long sequence, the interest preference of the target object can be accurately mined.

S102, acquiring an interaction behavior sequence of the target object, and generating an interaction interest representation of the interaction behavior sequence.

Wherein the interaction behavior sequence may comprise at least one interaction behavior feature of the target object. The at least one interactive behavior feature may include one or more of: the interaction behavior of each resource in the preset time period, the interaction behavior of a certain type of resource in the preset time period and the like, wherein the interaction behavior can comprise one or more of the following: praise, share, comment or collection, etc.

For example, after the target object submits the access request to the content publishing platform, the content publishing platform may push the determined resource to the target object, and the target object may perform interactive operations on some or all of the resources, such as praise, share, comment, collection, or the like. If the interactive operation of the target object on the target resource is detected, the interactive behavior characteristics of the target object on the target resource can be counted. For example, if the target object browses three notes related to the punch-through in a week, and the target object shares all the three notes related to the punch-through, the interaction behavior of the target object on the punch-through type notes in the week may be determined to include: and (5) three times of sharing. Further, the counted at least one interaction behavior feature may be ranked according to a time sequence to obtain an interaction behavior sequence of the target object, for example, the target object shares a first resource in 2023, 1 month, 5 days, 9:00, shares a second resource in 2023, 1 month, 5 days, 9:01, and shares a third resource in 2023, 1 month, 5 days, 9:05, and then the sharing behavior features of the resources may be ranked based on the sharing time points of the resources to obtain the interaction behavior sequence of the target object.

In one implementation, the manner in which the interactive interest representation of the interactive behavior sequence is generated may include: and acquiring the weight of each interactive behavior feature, and performing multi-layer perceptron operation on at least one interactive behavior feature based on the weight of each interactive behavior feature to obtain the interactive interest representation of the interactive behavior sequence.

The weights of the interaction behavior features may be set by a developer based on experience, or may be learned through a neural network, and are not limited in the embodiments of the present application. For example, the at least one interaction behavior feature includes interaction behavior features of two dimensions of interaction behavior of each resource in a preset time period and interaction behavior of a certain type of resource in the preset time period, and then the weight of the interaction behavior feature of each dimension can be set. For another example, the at least one interaction behavior feature includes interaction behavior features of four dimensions including praise behavior, sharing behavior, comment behavior and collection behavior of each resource in a preset time period, and then the weight of the interaction behavior feature of each dimension can be set.

The method includes the steps of carrying out multi-layer perceptron operation on at least one interactive behavior feature based on the weight of each interactive behavior feature, namely extracting multi-interest characterization of the interactive behavior sequence of the target object, and calculating to obtain the interactive interest characterization of the interactive behavior sequence through the formula (1). In calculating the interactive interest characterization, V in equation (1) _u The interactive interest characterization may be represented and H may represent an interactive behavior sequence. Used to calculate interactive interest characterizationAnd W is ₂ And calculating the +.>And W is ₂ May be the same or different, and is not specifically limited by the embodiments of the present application.

In one implementation, based on the weights of the interaction behavior features, performing a multi-layer perceptron operation on at least one interaction behavior feature to obtain an interaction interest representation mode of the interaction behavior sequence may include: the identity characteristics of the target object and at least one interaction behavior characteristic are spliced to obtain at least one spliced interaction behavior characteristic, and based on the weight of each interaction behavior characteristic, multi-layer perceptron operation is performed on the at least one spliced interaction behavior characteristic to obtain the interaction interest representation of the interaction behavior sequence.

In this embodiment of the present application, for the interactive behavior sequence, the interest signal of the target object is explicit, for example, the target object collects the first resource, which is very likely to indicate that the target object is interested in the first resource. Therefore, the identity characteristic and at least one interactive behavior characteristic of the target object can be directly spliced, and then the spliced at least one interactive behavior characteristic is extracted for interactive interest characterization.

The multi-interest characterization of the target object under different behaviors is extracted through a multi-head attention mechanism, and sparse long-tail interests of the target object can be recorded better than single interest characterization.

S103, fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the target object to obtain a multi-interest characterization vector of the target object.

In one implementation, weights of the identity token, the duration interest token and the interaction interest token can be generated through a gating network, and based on the weights of the identity token, the duration interest token and the interaction interest token, weighting operation is performed on the identity token, the duration interest token and the interaction interest token, so that a multi-interest token vector of the target object is obtained.

Specifically, interest characterization generated by each of two behavior sequences of the target object can be fused, and identity characterization, duration interest characterization and interaction interest characterization of the target object are fused by using a Gate control network (Gate). The gating network may produce a weight for each token, and the multiple token weights are combined into a multiple interest token vector for the target object. Illustratively, the multiple interest characterization vector of the target object may be calculated by the following equation (3).

O _u ＝G⊙V _u-pt +(1-G)⊙V _u-eg Formula (3)

G＝sigmoid(W _g [e _u-static ，V _u-pt ，V _u-eg ]+b _g ) Formula (4)

Wherein O is _u May represent multiple interest token vectors, G may represent weights for duration interest tokens, V _u-pt Can represent duration interest characterization, V _u-eg The interactive interest characterization may be represented. Wherein G can be calculated by the above formula (4), e _u-static Can represent identity representation of target object, W _g And b _g Can be preset parameters。

S104, determining recall resources of the target object based on the multi-interest characterization vector of the target object.

In one approach, recall resources for a target object may be determined in real-time based on multiple interest characterization vectors for the target object. For example, a candidate set is obtained, the candidate set comprises a plurality of resources, feature extraction is carried out on each resource in the candidate set, a resource feature vector of each resource is obtained, and for any resource in the plurality of resources, a recall resource of a target object is taken as a resource with a space distance smaller than a distance threshold based on a multi-interest feature vector of the target object and a space distance of the resource feature vector of any resource.

In one implementation, the multiple interest token vector may include multiple interest vectors, on the basis of which a manner of determining recall resources for the target object based on the multiple interest token vector for the target object may include: based on the corresponding relation between the interest vectors and the resources, acquiring the resources corresponding to each interest vector in the interest vectors, and taking the acquired resources as recall resources of the target object.

In one example, after the multi-interest characterization vector of the target object is obtained, the target object may submit an online access request for the resource, after the online access request of the target object is received, the resource corresponding to each interest vector in the plurality of interest vectors may be obtained based on the correspondence between the interest vector and the resource, and the obtained resource is used as a recall resource of the target object, so that the resource is pushed to the target object based on the recall resource of the target object.

In another example, the target object may submit an online access request to the resource, after receiving the online access request of the target object, the multi-interest characterization vector of the target object may be obtained in real time in a manner provided by the embodiment of the present application, then, based on a correspondence between the interest vector and the resource, the resource corresponding to each interest vector in the plurality of interest vectors is obtained, and the obtained resource is used as a recall resource of the target object, and further, the resource is pushed to the target object based on the recall resource of the target object.

Optionally, if the target object successfully logs in the client through the account, automatically submitting an online access request for the resource. Or if the target object is detected to perform the sliding operation in the display page, automatically submitting an online access request for the resource. Or if the switching operation of the target object to a certain page is detected, automatically submitting an online access request for the resource. Alternatively, the target object submits an online access request for the resource in a search field.

In the embodiment of the present application, since the correspondence between the interest vector and the resource is pre-established, after the multi-interest characterization vector of the target object is obtained, the resource corresponding to each interest vector in the plurality of interest vectors may be directly obtained based on the correspondence between the interest vector and the resource, and the obtained resource is used as the recall resource of the target object, so that the efficiency of resource recall may be improved.

In one implementation, a candidate set may be obtained, where the candidate set includes a plurality of resources, extracting features of each resource in the candidate set to obtain a resource feature vector of each resource, determining, for any resource in the plurality of resources, an interest vector with a minimum spatial distance from any resource based on a spatial distance between each interest vector and the resource feature vector of any resource, and establishing a correspondence between the determined interest vector and any resource.

Wherein if there is a minimum spatial distance of a certain interest vector from a certain resource, then an object that has an interest signal indicated by the interest vector is most likely to be interested in the resource.

When the recommendation system acts on line, a plurality of interest vectors of the target object are recalled, so that different behavior preferences of the target object can be fully learned, and finally the effects of improving the interaction rate of the system and the diversity of resource distribution are achieved.

In the embodiment of the application, the duration behavior sequence of the target object is acquired, the duration interest representation of the duration behavior sequence is generated, the interactive behavior sequence of the target object is acquired, the interactive interest representation of the interactive behavior sequence is generated, the identity representation, the duration interest representation and the interactive interest representation of the target object are fused to obtain the multi-interest representation vector of the target object, recall resources of the target object are determined based on the multi-interest representation vector of the target object, and interest preference of the object can be accurately mined, so that accuracy of resource recall is improved.

The multiple interest characterization vectors of the target object in the embodiment of the present application may be obtained based on an interest characterization model, and a schematic architecture of the interest characterization model is shown in fig. 2. The interest characterization model adopts a double-tower structure to learn the similarity of objects and resources, wherein an object side modeling structure is shown in fig. 2, a double-behavior sequence (an instant long-behavior sequence and an interactive behavior sequence) is received as input to perform interest modeling, a hierarchical attention module is adopted at the top layer, a plurality of interest vectors of the objects are extracted, and then matching modeling is performed with the vectors of target resources.

Specifically, the identity of the object (static feature) may be transmitted as an input to an embedding layer (embedding layer) in the interest characterization model, and then the output of the embedding layer is used as an input to a fully connected layer (Fully Connected layer, FC layer) to obtain the identity of the object. When the recommendation algorithm/language input modeling is carried out, the input text needs to be processed, the text is subjected to one-hot too sparse, then training is carried out through an embedding layer, and the text is mapped onto a shorter word vector so as to densify the matrix and extract the hidden characteristics of the matrix. In addition, when the output can be flattened and can be connected to the output layer, the addition of a fully connected layer generally enables these nonlinear combination features to be learned in a simple manner.

Meanwhile, the duration behavior sequence (long term pagetime sequence) of the object can be used as input and transmitted to another enabling layer in the interest characterization model, and then the duration behavior sequence is subjected to attention calculation by using the identity characteristics of the object before entering a Multi-interest extraction module (MIE, multi-Interest Extractor) corresponding to the duration behavior sequence (namely, the duration Multi-interest extraction module in FIG. 2) by using a user attention mechanism (UAT, user attention network) so as to obtain the duration interest characterization of the object.

In addition, the interaction behavior sequence (long term pagetime sequence) of the object can be used as input to be transmitted to another embedding layer in the interest characterization model, and then the identity features of the object are directly spliced to be used as input of a multi-interest extraction module (i.e. the interaction multi-interest extraction module in fig. 2) corresponding to the interaction behavior sequence, so as to obtain the interaction interest characterization of the object.

And extracting interest characterization vectors of the object on different interests by splitting different behavior sequences of the object and respectively adopting different model structures, thereby being beneficial to learning the difference of different behaviors of the object.

After the identity, duration and interaction interests of the object are obtained, the identity, duration and interaction interests of the object may be fused using a Gate network (Gate). And the gating network generates a weight for each characterization, and finally, the multiple characterizations are weighted and combined to obtain the multiple interest characterization vectors of the object. After obtaining the multi-interest token vector for the object, the argmax selection module may be used to calculate the interest vector that best matches a certain resource.

Through the interest fusion module, behavior habits of different objects are learned individually, real interest preference of the objects is obtained in the different behavior habits, and interest expression of the object masses is better mined and reserved.

Alternatively, the structure diagram of the user attention mechanism in the embodiment of the present application may be as shown in fig. 3, and the user attention mechanism may include a fully connected layer and a sigmoid module.

Alternatively, the structure of the gating network in the embodiment of the present application may be as shown in fig. 4, where the gating network may at least include a full connection layer and a sigmoid module.

Based on the above description, please refer to fig. 5, fig. 5 is a flow chart of another resource recall method provided in an embodiment of the present application, where the resource recall method may be executed by a resource recall device or a computer device; the resource recall scheme shown in fig. 5 includes, but is not limited to, steps S501 to S510, wherein:

s501, obtaining a training sample, wherein the training sample comprises a positive sample and a negative sample, the positive sample comprises a first training resource pushed to a training object in a historical time period and clicked by the training object, and the negative sample comprises a second training resource not pushed to the training object in the historical time period.

For example, an exposure sample may be obtained during a preset period of time, for example, the preset period of time is one month, then a first training resource pushed to the training object during the last month and clicked by the training object may be obtained, which indicates that the first training resource is a resource of interest to the training object, so the first training resource is taken as a positive sample, and a second training resource not pushed to the training object during the last month may be obtained, which indicates that the second training resource is not a resource of interest to the training object, so the second training resource is taken as a negative sample. For example, the negative samples may be random negative samples in a batch window.

S502, acquiring a duration behavior sequence and an interaction behavior sequence of a training object about a first training resource.

S503, calling an initial interest characterization model, and generating a duration interest characterization of a duration behavior sequence of the training object and an interaction interest characterization of an interaction behavior sequence of the training object.

S504, fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the training object to obtain a multi-interest characterization vector of the training object.

S505, based on each interest vector in the multi-interest characterization vectors of the training object, determining a first interest vector with the smallest space distance to the first training resource and a second interest vector with the smallest space distance to the second training resource.

S506, training the initial interest characterization model to obtain an interest characterization model by taking the space distance between the first interest vector and the first training resource as a training target and increasing the space distance between the second interest vector and the second training resource.

For example, argmax may be used to calculate the interest vector that best matches each training resource, i.e., the interest vector that has the smallest spatial distance to each training resource. Illustratively, the interest vector having the smallest spatial distance from any training resource can be calculated by the following formula (5).

Wherein,can represent the interest vector with the smallest space distance between the training object u and the training resource i, O _u Multiple interest token vector, which may represent training object u,/->Multiple interest token vector matrix transpose, e, that can represent training objects _i The token vector, which may represent training resource i, u may refer to any training object and i may refer to any training resource.

The loss value calculated during training of the initial interest characterization model may be obtained by likelihood function calculation, that is, may be obtained by the following formula (6).

Wherein p (x) _i I u) may represent the probability that the spatial distance from training resource i is smallest given training object u, x _i It is possible to represent a first training resource,an interest vector with the smallest spatial distance of the training object u from the second training resource j may be represented.

Illustratively, the maximized likelihood function is used to calculate a set of model parameters such that the probability of the first training resource occurring is maximized, which can be calculated by the following equation (7).

Wherein L can represent a loss value calculated when training the initial interest characterization model, U can represent all training objects, I _u The first training resource corresponding to any training object u may be represented.

In one implementation manner, the dense parameters and the sparse parameters of the interest characterization model obtained by the last optimization can be optimized every first preset time period to obtain the interest characterization model. For example, assuming that the first preset time period is one day, if the dense parameters and the sparse parameters of the interest characterization model are optimized at 2023, 1 and 6 days, then the dense parameters and the sparse parameters of the interest characterization model obtained by the last optimization may be optimized at 2023, 1 and 7 days to obtain the interest characterization model.

In one implementation manner, the sparse parameters of the interest characterization model obtained by the last optimization can be optimized every second preset time period to obtain the interest characterization model, and the duration indicated by the first preset time period is longer than the duration indicated by the second preset time period. For example, assuming that the second preset time period is one hour, if the sparse parameters of the interest characterization model are optimized at 2023, 1 month, 7 days, 9:00, then the sparse parameters of the interest characterization model obtained by the last optimization may be optimized at 2023, 1 month, 7 days, 10:00, to obtain the interest characterization model.

In order to enhance the distinguishing capability of the model on the negative samples, when the negative samples are calculated in the batch processing window, the interest vector matched with the negative samples, namely the interest vector with the smallest space distance with the negative samples, is found, so that the learning capability of the model on the difficult-to-separate samples can be enhanced.

S507, calling an interest characterization model, obtaining a duration behavior sequence of the target object, and generating a duration interest characterization of the duration behavior sequence.

S508, the interactive behavior sequence of the target object is obtained, and the interactive interest characterization of the interactive behavior sequence is generated.

S509, fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the target object to obtain a multi-interest characterization vector of the target object.

S510, determining recall resources of the target object based on the multi-interest characterization vector of the target object.

In the embodiment, the step S507 to the step S510 may be referred to as a specific description of the step S101 to the step S104, which is not repeated in the embodiment of the present application.

In the embodiment of the application, a duration behavior sequence and an interaction behavior sequence of a training object about a first training resource are acquired, an initial interest characterization model is called, a duration interest characterization of the duration behavior sequence of the training object and an interaction interest characterization of the interaction behavior sequence of the training object are generated, identity characterization, duration interest characterization and interaction interest characterization of the training object are fused to obtain multiple interest characterization vectors of the training object, a first interest vector with the minimum spatial distance to the first training resource and a second interest vector with the minimum spatial distance to the second training resource are determined based on each interest vector in the multiple interest characterization vectors of the training object, so that the spatial distance between the first interest vector and the first training resource is reduced, the spatial distance between the second interest vector and the second training resource is increased to serve as a training target, and the initial interest characterization model is trained to obtain the interest characterization model. Then, invoking an interest characterization model, acquiring a duration behavior sequence of the target object, generating a duration interest characterization of the duration behavior sequence, acquiring an interaction behavior sequence of the target object, generating an interaction interest characterization of the interaction behavior sequence, fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the target object to obtain a multi-interest characterization vector of the target object, determining recall resources of the target object based on the multi-interest characterization vector of the target object, and accurately mining interest preference of the target object, thereby improving accuracy of resource recall.

The present embodiment also provides a computer storage medium having stored therein program instructions for implementing the corresponding method described in the above embodiments when executed.

Referring to fig. 6 again, fig. 6 is a schematic structural diagram of a resource recall device according to an embodiment of the present application.

In one implementation manner of the resource recall device, the resource recall device comprises the following structure.

The first generation unit 601 is configured to obtain a duration behavior sequence of a target object, and generate a duration interest representation of the duration behavior sequence; wherein the sequence of time duration behaviors includes at least one time duration behavior feature of the target object;

a second generating unit 602, configured to obtain an interaction behavior sequence of a target object, and generate an interaction interest representation of the interaction behavior sequence; wherein the interactive behavior sequence comprises at least one interactive behavior feature of the target object;

a fusion unit 603, configured to fuse the identity token, the duration interest token, and the interaction interest token of the target object, so as to obtain a multiple interest token vector of the target object;

a determining unit 604, configured to determine recall resources of the target object based on the multiple interest token vector of the target object.

In one embodiment, the first generating unit 601 generates a duration interest representation of the duration behavior sequence, including:

acquiring the weight of each duration behavior feature;

In one embodiment, the first generating unit 601 performs a multi-layer perceptron operation on the at least one duration behavior feature based on the weights of the duration behavior features, to obtain a duration interest representation of the duration behavior sequence, including:

using the identity characteristic of the target object to perform attention calculation on the at least one duration behavior characteristic to obtain the processed at least one duration behavior characteristic;

and carrying out multi-layer perceptron operation on the at least one processed time length behavior feature based on the weight of each time length behavior feature to obtain the time length interest characterization of the time length behavior sequence.

In one embodiment, the second generating unit 602 generates the interactive interest representation of the interactive behavior sequence, including:

acquiring the weight of each interactive behavior feature;

In one embodiment, the second generating unit 602 performs a weighted operation on the at least one interaction behavior feature based on the weights of the interaction behavior features, to obtain the interaction interest representation of the interaction behavior sequence, including:

In one embodiment, the fusing unit 603 fuses the identity token, the duration interest token, and the interaction interest token of the target object to obtain a multiple interest token vector of the target object, including:

In one embodiment, the multiple interest characterization vector includes multiple interest vectors; the determining unit 604 determines recall resources for the target object based on the multiple interest token vector for the target object, including:

and taking the obtained resource as a recall resource of the target object.

In one embodiment, the resource recall device further comprises:

an obtaining unit 605 for obtaining a candidate set, the candidate set including a plurality of resources;

a feature extraction unit 606, configured to perform feature extraction on each resource in the candidate set, so as to obtain a resource feature vector of each resource;

the determining unit 604 is further configured to determine, for any resource of the plurality of resources, an interest vector with a minimum spatial distance to the any resource based on the spatial distances between the respective interest vectors and the resource feature vectors of the any resource;

a relationship establishing unit 607, configured to establish a correspondence between the determined interest vector and the any resource.

In one embodiment, the multiple interest token vector is derived based on an interest token model, and the resource recall device further comprises:

An acquisition unit 605 for acquiring a training sample; wherein the training samples comprise positive samples and negative samples, the positive samples comprise first training resources pushed to a training object in a history time period and clicked by the training object, and the negative samples comprise second training resources not pushed to the training object in the history time period;

the obtaining unit 605 is further configured to obtain a duration behavior sequence and an interaction behavior sequence of the training object with respect to the first training resource;

the first generating unit 601 is further configured to invoke an initial interest representation model to generate a duration interest representation of the duration behavior sequence of the training object;

the second generating unit 602 is further configured to invoke an initial interest representation model, generate a fusion unit 603, and use the fusion unit to represent the interactive interest of the interactive behavior sequence of the training object;

the fusion unit 603 is further configured to fuse the identity token, the duration interest token, and the interaction interest token of the training object, so as to obtain a multi-interest token vector of the training object;

a determining unit 604, configured to determine, based on each interest vector in the multiple interest token vectors of the training object, a first interest vector with a minimum spatial distance to the first training resource, and a second interest vector with a minimum spatial distance to the second training resource;

And the training unit 608 is configured to train the initial interest characterization model with the space distance between the first interest vector and the first training resource being reduced, and the space distance between the second interest vector and the second training resource being increased as a training target, so as to obtain the interest characterization model.

In one embodiment, the resource recall device further comprises:

the optimizing unit 609 is configured to optimize the dense parameter and the sparse parameter of the interest characterization model obtained by the previous optimization every first preset time period to obtain the interest characterization model;

the optimizing unit 609 is further configured to optimize the sparse parameter of the interest characterization model obtained by the previous optimization every second preset time period to obtain the interest characterization model, where the duration indicated by the first preset time period is greater than the duration indicated by the second preset time period.

In this embodiment of the present application, the first generating unit 601 obtains a duration behavior sequence of the target object, and generates a duration interest representation of the duration behavior sequence, the second generating unit 602 obtains an interaction behavior sequence of the target object, and generates an interaction interest representation of the interaction behavior sequence, and the fusion unit 603 fuses the identity representation, the duration interest representation and the interaction interest representation of the target object to obtain a multiple interest representation vector of the target object, and the determining unit 604 determines, based on the multiple interest representation vector of the target object, that recall resources of the target object can accurately mine interest preference of the object, thereby improving accuracy of resource recall.

Referring to fig. 7 again, fig. 7 is a schematic structural diagram of a computer device provided in an embodiment of the present application, where the computer device in the embodiment of the present application includes a power supply module and other structures, and includes a processor 701, a storage 702, and a communication interface 703. Data can be interacted among the processor 701, the storage device 702 and the communication interface 703, and a corresponding target detection method is realized by the processor 701.

The storage 702 may include volatile memory (RAM), such as random-access memory (RAM); the storage 702 may also include a non-volatile memory (non-volatile memory), such as a flash memory (flash memory), a Solid State Drive (SSD), etc.; the storage 702 may also include a combination of the types of memory described above.

The processor 701 may be a central processing unit (central processing unit, CPU). The processor 701 may also be a combination of a CPU and a GPU. In the server, a plurality of CPUs and GPUs can be included as required to perform corresponding data processing. In one embodiment, storage 702 is used to store program instructions. The processor 701 may invoke program instructions to implement the various methods as referred to above in embodiments of the present application.

In a first possible implementation manner, the processor 701 of the computer device invokes the program instructions stored in the storage 702, to obtain a duration behavior sequence of the target object, and generate a duration interest representation of the duration behavior sequence; wherein the sequence of time duration behaviors includes at least one time duration behavior feature of the target object; acquiring an interactive behavior sequence of a target object, and generating an interactive interest representation of the interactive behavior sequence; wherein the interactive behavior sequence comprises at least one interactive behavior feature of the target object; fusing the identity characterization, the duration interest characterization and the interaction interest characterization of the target object to obtain a multi-interest characterization vector of the target object; and determining recall resources of the target object based on the multi-interest characterization vector of the target object.

In one embodiment, the processor 701, when generating the duration interest representation of the duration behavior sequence, may perform the following operations:

acquiring the weight of each duration behavior feature;

In one embodiment, when the processor 701 performs a multi-layer perceptron operation on the at least one duration behavior feature based on the weights of the duration behavior features to obtain the duration interest representation of the duration behavior sequence, the following operations may be performed:

In one embodiment, the processor 701, when generating the interactive interest representation of the interactive behavior sequence, may perform the following operations:

acquiring the weight of each interactive behavior feature;

In one embodiment, when the processor 701 performs a multi-layer perceptron operation on the at least one interaction behavior feature based on the weights of the interaction behavior features to obtain the interaction interest representation of the interaction behavior sequence, the following operations may be performed:

In one embodiment, the processor 701 may perform the following operations when fusing the identity token, the duration interest token, and the interaction interest token of the target object to obtain a multiple interest token vector of the target object:

In one embodiment, the multiple interest characterization vector includes multiple interest vectors; the processor 701 may perform the following operations when determining recall resources for the target object based on the multiple interest token vector for the target object:

and taking the obtained resource as a recall resource of the target object.

In one embodiment, the processor 701 is further configured to perform the following operations:

In one embodiment, the multiple interest token vector is obtained based on an interest token model, and the processor 701 is further configured to perform the following operations:

invoking an initial interest characterization model to generate a duration interest characterization of the duration behavior sequence of the training object;

invoking an initial interest characterization model to generate a fusion unit 603, which is used for the interactive interest characterization of the interactive behavior sequence of the training object;

In the embodiment of the present application, the processor 701 obtains a duration behavior sequence of a target object, generates a duration interest representation of the duration behavior sequence, obtains an interaction behavior sequence of the target object, generates an interaction interest representation of the interaction behavior sequence, fuses the identity representation, the duration interest representation and the interaction interest representation of the target object to obtain a multi-interest representation vector of the target object, and determines, based on the multi-interest representation vector of the target object, that recall resources of the target object can accurately mine interest preference of the object, thereby improving accuracy of resource recall.

Those skilled in the art will appreciate that the processes implementing all or part of the methods of the above embodiments may be implemented by a computer program for instructing relevant hardware, and the program may be stored in a computer readable storage medium, and the program may include the processes of the embodiments of the methods as above when executed. The storage medium may be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a random-access Memory (Random Access Memory, RAM), or the like. The computer-readable storage medium of (a) may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function, and the like; the storage data area may store data created from the use of blockchain nodes, and the like.

The above disclosure is only a few examples of the present application, and it is not intended to limit the scope of the claims, and those skilled in the art will understand that all or a portion of the above-described embodiments may be implemented and equivalents may be substituted for elements thereof, which are included in the scope of the present invention.

Claims

1. A method of recall of a resource, comprising:

2. The method of claim 1, wherein generating the duration interest representation of the duration behavior sequence comprises:

Acquiring the weight of each duration behavior feature;

and calculating the at least one time length behavior feature through a multi-layer perceptron based on the weight of each time length behavior feature to obtain the time length interest representation of the time length behavior sequence.

3. The method according to claim 2, wherein the obtaining the duration interest representation of the duration behavior sequence based on the weights of the duration behavior features through multi-layer perceptron operation on the at least one duration behavior feature includes:

and carrying out weighted operation on at least one time length behavior feature after processing based on the weight of each time length behavior feature to obtain the time length interest representation of the time length behavior sequence.

4. The method of claim 1, wherein generating the interactive interest representation of the interactive behavior sequence comprises:

acquiring the weight of each interactive behavior feature;

and based on the weight of each interactive behavior feature, carrying out multi-layer perceptron operation on at least one interactive behavior feature to obtain the interactive interest representation of the interactive behavior sequence.

5. The method of claim 4, wherein the weighting the at least one interaction behavior feature based on the weights of the respective interaction behavior features results in an interaction interest characterization of the interaction behavior sequence, comprising:

and carrying out weighted operation on at least one spliced interactive behavior feature based on the weight of each interactive behavior feature to obtain the interactive interest representation of the interactive behavior sequence.

6. The method of claim 1, wherein the fusing the identity token, the duration interest token, and the interaction interest token of the target object to obtain a multi-interest token vector of the target object comprises:

7. The method of claim 1, wherein the multiple interest characterization vector comprises a plurality of interest vectors; the determining recall resources of the target object based on the multiple interest characterization vectors of the target object includes:

and taking the obtained resource as a recall resource of the target object.

8. The method of claim 7, wherein the method further comprises:

9. The method of claim 1, wherein the multiple interest characterization vector is derived based on an interest characterization model, the method further comprising:

10. The method according to claim 9, wherein the method further comprises: