CN114528490A - Self-supervision sequence recommendation method based on long-term and short-term interests of user - Google Patents

Self-supervision sequence recommendation method based on long-term and short-term interests of user Download PDF

Info

Publication number
CN114528490A
CN114528490A CN202210151706.2A CN202210151706A CN114528490A CN 114528490 A CN114528490 A CN 114528490A CN 202210151706 A CN202210151706 A CN 202210151706A CN 114528490 A CN114528490 A CN 114528490A
Authority
CN
China
Prior art keywords
term
user
sequence
short
long
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202210151706.2A
Other languages
Chinese (zh)
Inventor
王也
阎震
韩启龙
宋洪涛
李丽洁
王宇华
马志强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Engineering University
Original Assignee
Harbin Engineering University
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 Harbin Engineering University filed Critical Harbin Engineering University
Priority to CN202210151706.2A priority Critical patent/CN114528490A/en
Publication of CN114528490A publication Critical patent/CN114528490A/en
Pending legal-status Critical Current

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
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • G06N3/04—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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a self-supervision sequence recommendation method based on long-term and short-term interests of a user. Step 1: acquiring a sequence data set of user information, project information and user behaviors, preprocessing the data set, and dividing the data set into a training set and a test set; and 2, step: constructing a self-supervision sequence recommendation model based on long-term and short-term interests of a user; and step 3: training the self-supervision sequence recommendation model based on the long-term and short-term interests of the user in the step 2 by using a training set; and 4, step 4: inputting the personal information and the interaction sequence of the user to be recommended into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user trained in the step 3, calculating the recommendation score of the item to be recommended relative to the user, and recommending the item to the user according to the recommendation score. The invention is used for solving the problem of mutual constraint relation between long-term interest and short-term interest of the user in the prior art, and realizing more accurate recommendation of the long-term interest and the short-term interest of the user.

Description

Self-supervision sequence recommendation method based on long-term and short-term interests of user
Technical Field
The invention belongs to the field of sequence recommendation, and particularly relates to a self-supervision sequence recommendation method based on long-term and short-term interests of a user.
Background
The behavior data of people in the internet can indicate important information, such as user preferences and behavior patterns, and a recommendation system can use the information to provide personalized services for users and improve the experience of the users.
The modeling of user interaction with an item by conventional sequence recommendation systems can be generalized into two main approaches. The first approach is Collaborative Filtering (CF) based on matrix factorization to obtain user preferences, focusing on mining the static associations of users from their interactions with items, which are represented by a traditional collaborative filtering model. However, these works only take into account the specific user-item relationships from the static view, neglecting the evolution of user preferences implicit in the serialized interactions, and do not take into account the impact of the evolution of user preferences on future purchases. The second method is to mine the relationship between the user and the item based on the sequence pattern to make personalized recommendation. Among them, the user's steady long-term interest is a preference caused by personal habits for a long time; short-term interest is a preference determined by items recently purchased by the user. This type of work includes: and modeling an interaction sequence of the user and the commodity according to the Markov chain model. The long-term interest and the short-term interest of the user play an important role in commodity selection of the user, so that the accuracy can be effectively improved by combining the long-term interest and the short-term interest into recommendation of the user.
Deep neural networks are increasingly used by many to construct sequence recommendation systems by taking advantage of the natural nature of the integrated relationships between different entities (e.g., users, items, interactions) in building and capturing sequences.
For sequence recommendation systems, recurrent neural networks were originally proposed and valued by researchers because of their structural advantages in modeling sequence data, but they also suffer from the drawback that higher order complex relationships cannot be modeled. Then, the convolutional neural network and the graph neural network are also applied to a sequence recommendation system for modeling complex interaction relation, so that the defects in the cyclic neural network are overcome.
In recent years, people pay more and more attention to improving the sequence recommendation performance by capturing the long-term and short-term interests of users, but the prior art only considers the complementary relation of the long-term and short-term interests of the users and does not consider the mutual constraint relation between the long-term and short-term interests of the users. In addition, the existing recommendation technology using the hypergraph can better learn high-order context information so as to better model the short-term interest of the user, but does not fully consider the sequence information of user interaction items, and particularly does not consider the utilization of the sequence information when constructing the hypergraph.
In summary, the current research work mainly provides a more accurate modeling method for the long-term interest and the short-term interest of the user, and can fully utilize the complementary and constrained relationships between the long-term interest and the short-term interest of the user to achieve a better recommendation effect.
Disclosure of Invention
The invention provides a self-supervision sequence recommendation method based on long-term and short-term interests of a user, which is used for solving the problem of mutual constraint relation between the long-term and short-term interests of the user in the prior art and realizing more accurate recommendation of the long-term and short-term interests of the user.
The invention is realized by the following technical scheme:
a self-supervision sequence recommendation method based on long-term and short-term interests of a user comprises the following steps:
step 1: acquiring a sequence data set of user information, project information and user behaviors, preprocessing the data set, and dividing the data set into a training set and a test set;
step 2: constructing a self-supervision sequence recommendation model based on long-term and short-term interests of a user;
and step 3: training the self-supervision sequence recommendation model based on the long-term and short-term interests of the user in the step 2 by using a training set;
and 4, step 4: inputting the personal information and the interaction sequence of the user to be recommended into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user trained in the step 3, calculating the recommendation score of the item to be recommended relative to the user, and recommending the item to the user according to the recommendation score.
Further, the step 1 specifically includes the following steps:
step 1.1: dividing a user interaction sequence into a long-term sequence L and a short-term sequence S;
step 1.2: constructing a hypergraph G according to the short-term sequence S divided in the step 1.1;
let G ═ V, E denote a hypergraph, where the set of V contains N vertices, each vertex representing an item, E contains M hyperedges, each hyperedge epsilon E represents a session and contains five vertices; each super-edge is assigned a forward weight WεεRepresenting the weight occupied by each item in the super edge, and forming all weights into a diagonal matrix W epsilon RM×M(ii) a The hypergraph is represented by a correlation matrix H ∈ RN×MIf the hyper-edge ε E contains a vertex viE is V then H iε1, otherwise 0; for each vertex and hyper-edge, their degree is defined as
Figure BDA0003510855050000021
Definition DhAnd B are both diagonal matrices; to utilize the sequence relation between the item interactions in the conversation, a degree matrix D in the sequence sense is definedp∈RN×NRegarding the former as influencing the latter through the context of the interactive items in each session;
standardizing for each item:
Figure BDA0003510855050000031
obtaining a final vertex degree matrix:
D＝2×(Dh+Dp)
further, the step 2 of constructing an automatic supervision sequence recommendation model based on long-term and short-term interests of the user specifically comprises the following steps:
step 2.1: obtaining an embedded representation of an item;
step 2.2: inputting the user long-term sequence L marked out in the step 1.1 and the project embedding expression into a GRU layer, and capturing the long-term interest theta of the user through a feedforward neural network layerl；
Step 2.3: convolving the hypergraph G constructed in the step 1.2, and obtaining a short-term interest expression theta of the user by the convolution result and the short-term interaction sequence S of the user divided in the step 1.1 through a soft attention layers；
Step 2.4: the long-term interest theta of the user obtained in the step 2.2lAnd short term interest θ from step 2.3sObtaining an auto-supervised loss L by an auto-supervised learning layerS；
Step 2.5: the long-term interest theta of the user obtained in the step 2.2lAnd short term interest θ from step 2.3sObtaining a final user representation theta through the fusion layer;
step 2.6: and calculating the scores of the candidate items according to the final representation of the user.
Further, the step 2.2 is specifically to model the long-term preference of the user according to the long-term sequence partitioned in the step 1.1, capture the evolution in the long-term sequence by using a GRU network, and represent θ by taking the last hidden unit state as the long-term preference of the userl。
Further, the step 2.3 performs hypergraph convolution, and the hypergraph convolution update item represents:
Figure BDA0003510855050000032
the hypergraph convolution takes a formula from right to left as a convolution process from an item to a hyperedge to an item;
Figure BDA0003510855050000033
representing slave verticesInformation aggregation to the super edge, and then left multiplication H is carried out to aggregate the information from the super edge to the top point; after L-layer convolution, the average of all layer results is taken as the representation of the final item
Figure BDA0003510855050000034
For one session s ═ is,1,is,2,...,is,m]Aggregating the items in the conversation to obtain a representation θ of the end user's short-term interests：
αt＝fTσ(W1xs+W2xt+b
Figure BDA0003510855050000041
Wherein xsIndicating that the current session is averaged from items within the session, xtRepresenting the tth interactive item in the conversation, wherein the short-term interest representation is obtained by the items in the conversation through a soft attention mechanism; f is an element of Rd，W1∈Rd×dAnd W2∈Rd×dAre the attention parameters used to learn the weights.
Further, the step 2.4 is specifically,
the method of comparative learning is used for fully playing the constraint relationship between the positive sample and the negative sample and calculating the loss by adopting a standard binary system cross entropy loss function between the positive sample and the negative sample
Figure BDA0003510855050000042
Figure BDA0003510855050000043
Wherein
Figure BDA0003510855050000044
Is to
Figure BDA0003510855050000045
Is rearranged in rows and columns to obtain negative samples, fD(·):Rd×d→ R is a discriminator function that takes two vectors as input and scores the agreement of the two.
Further, said step 2.5 is specific in that, since step 2.4 effectively facilitates obtaining each other's information more accurately, the resulting representation of the user is obtained in an additive manner:
Figure BDA0003510855050000046
further, the step 2.6 obtains the scores of the items by multiplying the final user representation obtained in the step 2.5 by the item set points
Figure BDA0003510855050000047
Calculating the probability of each item appearing through softmax
Figure BDA0003510855050000048
Further, the step 3 specifically includes the following steps:
step 3.1: inputting data in the training set into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user in the step 2 to obtain a final expression vector of the user;
step 3.2: inputting the final expression vector of the user into a prediction module to obtain a recommendation score of the user relative to the item;
step 3.3: updating the parameters of the model by calculating the error between the predicted score value and the true score value to optimize the recommendation loss function, combining the self-supervision loss in step 2.4 to obtain the final loss
Figure BDA0003510855050000049
Wherein β is a weight set to 0.01; and repeatedly training to obtain an optimal self-supervision sequence recommendation model based on long-term and short-term interests of the user.
Further, the step 4: inputting the personal information and the interaction sequence of the user to be recommended into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user trained in the step 3, calculating the recommendation score of the item to be recommended relative to the user, and recommending the item to the user according to the recommendation score.
The invention has the beneficial effects that:
the invention fully utilizes the constraint relation between the long-term interest and the short-term interest of the user and considers the sequence relation in the short-term conversation during the construction of the hypergraph so as to achieve better recommendation effect.
According to the method and the device, the constraint relation between the long-term interest and the short-term interest of the user is fully exerted through self-supervision learning, and the recommendation accuracy rate is further improved.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention.
FIG. 2 is a diagram of a user long-short term interest self-supervised recommendation model of the present invention.
Fig. 3 is a short-term session hypergraph diagram of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
A sequence recommendation system based on three deep neural networks.
1) A recurrent neural network-based sequence recommendation system. Given a historical sequence of user-commodity interactions, a recurrent neural network-based sequence recommendation system attempts to build a sequence dependency through a given interaction, thereby predicting the next possible interaction. The GRU4Rec model introduces a GRU network into sequence recommendation, models a user interaction sequence through the GRU network, learns the evolution of user interest, recommends the user according to the evolution of the user interest, and obtains a very good recommendation effect by utilizing the excellent performance of the recurrent neural network on the modeling of a sequence relation, but does not recommend the user interest by considering the long-term interest and the short-term interest together.
2) A convolutional neural network based sequence recommendation system. Convolutional Neural Networks (CNNs) can be used to extract features from text, audio, and pictures. Unlike the recurrent neural network, given the order of a user-commodity interaction, the convolutional neural network first embeds all the interactions into a matrix, and then treats this matrix as a picture in time and in potential space. Finally, the convolutional neural network learns this sequence pattern as a local feature of this picture, using convolutional filtering for subsequent recommendations. The Caser model utilizes a convolutional neural network to model the short-term interest of the user, takes the latest L items of the user interaction as the short-term sequences of the user interaction, respectively carries out horizontal convolution and vertical convolution on the embedding of the short-term sequences of the user, carries out splicing operation on the convolution results, takes the spliced embedding as the short-term interest of the user, also adopts a splicing mode when the long-term interest and the short-term interest of the user are fused, and takes the spliced embedding as the final interest representation of the user so as to recommend the user the articles meeting the interest of the user. Compared with a cyclic neural network, the convolutional neural network has the advantage that higher-dimensionality information can be captured by the convolutional neural network, and compared with a sequence recommendation model which is recommended by the cyclic neural network, the sequence recommendation model is relatively good at the moment, the experimental effect on a plurality of data sets is not weaker than that of other sequence recommendation models, but only the complementary relation among long-term and short-term interests of a user is considered, the constraint relation is not considered, and meanwhile, the short-term interest modeling of the user is too simple.
3) A sequence recommendation system based on a graph neural network. With the rapid development of graph neural networks, sequence recommendation systems based on graph neural networks have been designed that utilize graph neural networks to model and capture the transitions of more complex user-commodity interactions in sequences. When each sequence is mapped onto a path, a directed graph is first built over the sequence data, and each interaction is treated as a node in the graph. The embedding of the user or goods is then learned on the graph in order to embed more complex relationships throughout the graph. The SRGNN model constructs an interaction sequence of a user into a directed graph, and takes the sequence relation in the interaction sequence of the user into consideration as the directional relation in the directed graph during composition, the constructed directed graph can be used for clearly knowing which other items can influence the item representation in the graph, so that the item representation in the graph is updated through a graph neural network, the updated item representation is made to obtain an interest representation of the user through an attention mechanism, and recommendation is made according to the interest representation. The graph neural network is applied to the session recommendation to model the user interests, and compared with the sequence recommendation model which is represented in good terms at that time, the graph neural network has a better recommendation effect compared with other models, so that the graph neural network has strong capability of capturing the complex relationships among items in the user interaction sequence and has a large development space. As researchers have explored the use of graph neural networks in sequence recommendations, in recent years, hypergraph neural networks have been applied to sequence recommendations. The hypergraph neural network belongs to a graph neural network, but is different from a standard graph neural network in that the hypergraph neural network does not construct the relationship between the interaction sequences of the users into a standard graph structure any more, but into a hypergraph structure. The hypergraph differs from the normal graph in that there are two vertices on one edge of the normal graph, and there may be more than two vertices on one edge of the hypergraph, such an edge being called a hyperedge. By using the hypergraph, it is no longer possible to restrict to the paired relationship of two points on one side, and more complex relationships can be captured. The HyperRec model divides a complete interaction sequence of a user into a plurality of short-term sequences, constructs each short-term sequence into a hypergraph, utilizes hypergraph convolution to capture the correlation of articles in the short-term sequences, models the user interest of each short-term sequence by obtaining the dynamic representation of the articles, finally obtains the dynamic representation of the user interest of all the short-term sequences at the current moment by a self-attention mechanism, and performs next recommendation for the user by using the dynamic representation to obtain good recommendation effect, but sequence information in the user interaction sequence is not considered when constructing the hypergraph.
A self-supervision sequence recommendation method based on long-term and short-term interests of a user comprises the following steps:
step 1: acquiring a sequence data set of user information, project information and user behaviors, preprocessing the data set, and dividing the data set into a training set and a test set; dividing the data set into training set and testing set according to 8:2 ratio
Step 1.1, the complete interactive sequence of the user is divided into a sequence capable of representing long-term preference and a session capable of representing current short-term interest, and because the long-term preference of the user is modeled, the complete interactive sequence can more accurately represent the long-term preference of the user, all the interactive sequences are used as sequences for capturing the long-term preference of the user, and aiming at the short-term interest of the modeled user, nearly five item interactions are used as short-term session sequences for constructing the hypergraph in order to prevent the influence effect of the data sparsity problem of the hypergraph session.
Step 2: constructing a self-supervision sequence recommendation model based on long-term and short-term interests of a user;
and step 3: training the self-supervision sequence recommendation model based on the long-term and short-term interests of the user in the step 2 by using a training set;
and 4, step 4: inputting the personal information and the interaction sequence of the user to be recommended into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user trained in the step 3, calculating the recommendation score of the item to be recommended relative to the user, and recommending the item to the user according to the recommendation score.
Further, the step 1 specifically includes the following steps:
step 1.1: dividing a user interaction sequence into a long-term sequence L and a short-term sequence S;
step 1.2: constructing a hypergraph G according to the short-term sequence S divided in the step 1.1;
constructing the hypergraph in the manner shown in fig. 3, and constructing the hypergraph by using the short-term session part of the whole sequence, wherein the step considers the mutual influence of similar sessions on the items in the respective sessions and also considers the sequence relation among the interactive items in the same session, namely the last interactive item has influence on the current interactive item.
Let G ═ V, E denote a hypergraph, where the set of V contains N vertices, each vertex representing an item, E contains M hyperedges, each hyperedge epsilon E represents a session and contains five vertices; each super-edge is assigned a forward weightWεεRepresenting the weight occupied by each item in the super edge, and forming all weights into a diagonal matrix W epsilon RM×M(ii) a The hypergraph is represented by a correlation matrix H ∈ RN×MIf the hyper-edge ε E contains a vertex viE is then H iε1, otherwise 0; for each vertex and hyper-edge, their degree is defined as
Figure BDA0003510855050000081
Definition DhAnd B are both diagonal matrices; to utilize the sequence relation between the item interactions in the conversation, a degree matrix D in the sequence sense is definedp∈RN×NRegarding the former as influencing the latter through the context of the interactive items in each session; such as i1、i2Of (2), consider i1Will influence i2Then will be
Figure BDA0003510855050000082
Setting as 1;
standardizing for each item:
Figure BDA0003510855050000083
obtaining a final vertex degree matrix:
D＝2×(Dh+Dp)
further, the step 2 of constructing an automatic supervision sequence recommendation model based on long-term and short-term interests of the user specifically comprises the following steps:
step 2.1: obtaining an embedded representation of an item; generating an embedded representation of all items through an embedding matrix;
step 2.2: inputting the user long-term sequence L marked out in the step 1.1 and the project embedded representation into a GRU layer, and capturing the long-term interest theta of the user through a feedforward neural network layerl；
Step 2.3: convolving the hypergraph G constructed in the step 1.2, and obtaining the short-term interest expression of the user by the convolution result and the short-term interaction sequence S of the user divided in the step 1.1 through a soft attention layerθs；
Step 2.4: the long-term interest theta of the user obtained in the step 2.2lAnd short term interest θ from step 2.3sObtaining an auto-supervised loss L by an auto-supervised learning layerS；
Step 2.5: the long-term interest theta of the user obtained in the step 2.2lAnd short term interest θ from step 2.3sObtaining a final user representation theta through the fusion layer;
step 2.6: and calculating the scores of the candidate items according to the final representation of the user.
Further, the step 2.2 is specifically to model the long-term preference of the user according to the long-term sequence partitioned in the step 1.1, capture the evolution in the long-term sequence by using a GRU network, and represent θ by taking the last hidden unit state as the long-term preference of the userl。
Further, step 2.3 performs hypergraph convolution, and the hypergraph convolution update item represents:
Figure BDA0003510855050000084
the hypergraph convolution takes a formula from right to left as a convolution process from an item to a hyperedge to an item;
Figure BDA0003510855050000085
representing information aggregation from the vertex to the super edge, and then left multiplying H to aggregate information from the super edge to the vertex; after L-layer convolution, the average of all layer results is taken as the representation of the final item
Figure BDA0003510855050000091
For one session s ═ is,1,is,2,...,is,m]Aggregating the items in the conversation to obtain a representation θ of the end user's short-term interests：
αt＝fTσ(W1xs+W2xt+b)
Figure BDA0003510855050000092
Wherein xsIndicating that the current session is averaged from items within the session, xtRepresenting the tth interactive item in the conversation, wherein the short-term interest representation is obtained by the items in the conversation through a soft attention mechanism; f is an element of Rd，W1∈Rd×dAnd W2∈Rd×dAre the attention parameters used to learn the weights.
Further, the step 2.4 is specifically,
the method of comparative learning is used for fully playing the constraint relationship between the positive sample and the negative sample and calculating the loss by adopting a standard binary system cross entropy loss function between the positive sample and the negative sample
Figure BDA0003510855050000093
Figure BDA0003510855050000094
Wherein
Figure BDA0003510855050000095
Is to
Figure BDA0003510855050000096
Is rearranged in rows and columns to obtain negative samples, fD(·):Rd×d→ R is a discriminator function that takes two vectors as input and scores the agreement of the two. The consistency is scored in a point-and-multiply mode; the learning goal of this part may be to maximize the mutual information between the user's long-term preferences modeled with different angles and the embedding of short-term interests.
Further, the step 2.5 is specific, since the step 2.4 effectively facilitates obtaining each other's information more accurately, but the complementary relationship between the two still needs to be fully developed; the way of obtaining the addition yields the final representation of the user:
Figure BDA0003510855050000097
further, the step 2.6 obtains the scores of the items by multiplying the final user representation obtained in the step 2.5 by the item set points
Figure BDA0003510855050000098
Calculating the probability of each item appearing through softmax
Figure BDA0003510855050000099
Further, the step 3 specifically includes the following steps:
step 3.1: inputting data in the training set into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user in the step 2 to obtain a final expression vector of the user;
step 3.2: inputting the final expression vector of the user into a prediction module to obtain a recommendation score of the user relative to the item;
step 3.3: updating the parameters of the model by calculating the error between the predicted score value and the true score value to optimize the recommendation loss function, combining the self-supervision loss in step 2.4 to obtain the final loss
Figure BDA0003510855050000101
Wherein β is a weight set to 0.01; and repeatedly training to obtain an optimal self-supervision sequence recommendation model based on long-term and short-term interests of the user.
Preferably, the recommended loss function described in step 3.3 is specifically calculated as follows:
Figure BDA0003510855050000102
further, step 4 inputs the personal information and the interaction sequence of the user to be recommended into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user trained in step 3, calculates the recommendation score of the item to be recommended relative to the user, and recommends the item to the user according to the recommendation score.
And 4, sequencing the item recommendation scores, and recommending the top k items with the highest scores to the user.

Claims (10)

1. A self-supervision sequence recommendation method based on long-term and short-term interests of a user is characterized by comprising the following steps:
step 1: acquiring a sequence data set of user information, project information and user behaviors, preprocessing the data set, and dividing the data set into a training set and a test set;
and 2, step: constructing a self-supervision sequence recommendation model based on long-term and short-term interests of a user;
and 3, step 3: training the self-supervision sequence recommendation model based on the long-term and short-term interests of the user in the step 2 by using a training set;
and 4, step 4: inputting the personal information and the interaction sequence of the user to be recommended into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user trained in the step 3, calculating the recommendation score of the item to be recommended relative to the user, and recommending the item to the user according to the recommendation score.
2. The self-supervision sequence recommendation method based on user long-term and short-term interests according to claim 1, wherein the step 1 specifically comprises the following steps:
step 1.1: dividing a user interaction sequence into a long-term sequence L and a short-term sequence S;
step 1.2: constructing a hypergraph G according to the short-term sequence S divided in the step 1.1;
let G ═ V, E denote a hypergraph, where the set of V contains N vertices, each vertex representing an item, E contains M hyperedges, each hyperedge epsilon E represents a session and contains five vertices; each super-edge is assigned a forward weight WεεRepresenting the weight occupied by each item in the super edge, and forming all weights into a diagonal matrix W epsilon RM×M(ii) a The hypergraph is represented by a correlation matrix H ∈ RN×MIf the hyper-edge ε E contains a vertex viE is then Hiε1, otherwise 0;for each vertex and hyper-edge, their degree is defined as
Figure FDA0003510855040000011
Definition DhAnd B are both diagonal matrices; to utilize the sequence relation between the item interactions in the conversation, a degree matrix D in the sequence sense is definedp∈RN×NRegarding the former as influencing the latter through the context of the interactive items in each session;
standardizing for each item:
Figure FDA0003510855040000012
obtaining a final vertex degree matrix:
D＝2×(Dh+Dp)。
3. the method for recommending an unsupervised sequence based on long-term and short-term interests of a user according to claim 2, wherein the step 2 of constructing an unsupervised sequence recommendation model based on long-term and short-term interests of a user specifically comprises the following steps:
step 2.1: obtaining an embedded representation of an item;
step 2.2: inputting the user long-term sequence L marked out in the step 1.1 and the project embedded representation into a GRU layer, and capturing the long-term interest theta of the user through a feedforward neural network layerl；
Step 2.3: convolving the hypergraph G constructed in the step 1.2, and obtaining a short-term interest expression theta of the user by the convolution result and the short-term interaction sequence S of the user divided in the step 1.1 through a soft attention layers；
Step 2.4: the long-term interest theta of the user obtained in the step 2.2lAnd short term interest θ from step 2.3sObtaining an auto-supervised loss L by an auto-supervised learning layerS；
Step 2.5: the long-term interest theta of the user obtained in the step 2.2lAnd short term interest θ from step 2.3sThrough the fusion layerTo the final user representation θ;
step 2.6: and calculating the scores of the candidate items according to the final representation of the user.
4. The method according to claim 1, wherein the step 2.2 is specifically to model the long-term preference of the user according to the long-term sequence divided in the step 1.1, capture the evolution in the long-term sequence by using a GRU network, and represent the last hidden unit state as the long-term preference of the user by θl。
5. The method for recommending an unsupervised sequence based on long-term and short-term interests of a user according to claim 2, wherein said step 2.3 is a hypergraph convolution, which updates the item representation:
Figure FDA0003510855040000021
the hypergraph convolution takes a formula from right to left as a convolution process from an item to a hyperedge to an item;
Figure FDA0003510855040000022
representing the information aggregation from the vertex to the super edge, and then left multiplying H to aggregate the information from the super edge to the vertex; after L-layer convolution, the average of all layer results is taken as the representation of the final item
Figure FDA0003510855040000023
For one session s ═ is,1,is,2,...,is,m]Aggregating the items in the conversation to obtain a representation θ of the end user's short-term interests：
αt＝fTσ(W1xs+W2xt+b)
Figure FDA0003510855040000024
Wherein xsIndicating that the current session is averaged from items within the session, xtRepresenting the tth interactive item in the conversation, wherein the short-term interest representation is obtained by the items in the conversation through a soft attention mechanism; f is an element of Rd，W1∈Rd×dAnd W2∈Rd×dAre the attention parameters used to learn the weights.
6. The method for recommending an unsupervised sequence based on long-term and short-term interests of a user according to claim 2, wherein said step 2.4 is specifically,
the method of comparative learning is used for fully playing the constraint relationship between the positive sample and the negative sample and calculating the loss by adopting a standard binary system cross entropy loss function between the positive sample and the negative sample
Figure FDA0003510855040000031
Figure FDA0003510855040000032
Wherein
Figure FDA0003510855040000033
Is to
Figure FDA0003510855040000034
Is rearranged in rows and columns to obtain negative samples, fD(·):Rd×d→ R is a discriminator function that takes two vectors as input and scores the agreement of the two.
7. The method as claimed in claim 2, wherein the step 2.5 is to obtain the final list of the user in an additive manner, since the step 2.4 effectively facilitates obtaining each other's information more accuratelyThe following steps:
Figure FDA0003510855040000035
8. the method as claimed in claim 2, wherein the step 2.6 is to determine the score of each item by multiplying the final user representation obtained in the step 2.5 by the item set point
Figure FDA0003510855040000036
Calculating the probability of each item appearing through softmax
Figure FDA0003510855040000037
9. The method as claimed in claim 2, wherein the step 3 specifically comprises the following steps:
step 3.1: inputting data in the training set into the self-supervision sequence recommendation model based on the long-term and short-term interests of the user in the step 2 to obtain a final expression vector of the user;
step 3.2: inputting the final expression vector of the user into a prediction module to obtain a recommendation score of the user relative to the item;
step 3.3: updating the parameters of the model by calculating the error between the predicted score value and the true score value to optimize the recommendation loss function, combining the self-supervision loss in step 2.4 to obtain the final loss
Figure FDA0003510855040000038
Wherein β is a weight set to 0.01; and repeatedly training to obtain an optimal self-supervision sequence recommendation model based on long-term and short-term interests of the user.
10. The method as claimed in claim 2, wherein the step 4 inputs the personal information and the interaction sequence of the user to be recommended into the self-supervised sequence recommendation model based on the long-term and short-term interests of the user trained in the step 3, calculates the recommendation score of the item to be recommended relative to the user, and recommends the item to the user according to the recommendation score.
CN202210151706.2A 2022-02-18 2022-02-18 Self-supervision sequence recommendation method based on long-term and short-term interests of user Pending CN114528490A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210151706.2A CN114528490A (en) 2022-02-18 2022-02-18 Self-supervision sequence recommendation method based on long-term and short-term interests of user

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210151706.2A CN114528490A (en) 2022-02-18 2022-02-18 Self-supervision sequence recommendation method based on long-term and short-term interests of user

Publications (1)

Publication Number Publication Date
CN114528490A true CN114528490A (en) 2022-05-24

Family

ID=81622053

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210151706.2A Pending CN114528490A (en) 2022-02-18 2022-02-18 Self-supervision sequence recommendation method based on long-term and short-term interests of user

Country Status (1)

Country Link
CN (1) CN114528490A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115099886A (en) * 2022-05-25 2022-09-23 华南理工大学 Long and short interest sequence recommendation method and device and storage medium

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115099886A (en) * 2022-05-25 2022-09-23 华南理工大学 Long and short interest sequence recommendation method and device and storage medium
CN115099886B (en) * 2022-05-25 2024-04-19 华南理工大学 Long-short interest sequence recommendation method, device and storage medium

Similar Documents

Publication Publication Date Title
Chen et al. Deep reinforcement learning in recommender systems: A survey and new perspectives
CN111949865A (en) Interest point recommendation method based on graph neural network and user long-term and short-term preference
CN109062962B (en) Weather information fused gated cyclic neural network interest point recommendation method
CN111222054A (en) Session social contact recommendation method based on context neighbor relation modeling
CN110851760B (en) Human-computer interaction system for integrating visual question answering in web3D environment
CN111241394B (en) Data processing method, data processing device, computer readable storage medium and electronic equipment
CN114817663B (en) Service modeling and recommendation method based on class perception graph neural network
CN116664719B (en) Image redrawing model training method, image redrawing method and device
CN109464803A (en) Virtual objects controlled, model training method, device, storage medium and equipment
CN114519145A (en) Sequence recommendation method for mining long-term and short-term interests of users based on graph neural network
CN110738314B (en) Click rate prediction method and device based on deep migration network
CN113297487A (en) Attention mechanism-based sequence recommendation system and method for enhancing gated cyclic unit
CN112116589B (en) Method, device, equipment and computer readable storage medium for evaluating virtual image
CN114637911A (en) Next interest point recommendation method of attention fusion perception network
CN112819575A (en) Session recommendation method considering repeated purchasing behavior
CN115588122A (en) News classification method based on multi-mode feature fusion
CN114896515A (en) Time interval-based self-supervision learning collaborative sequence recommendation method, equipment and medium
CN114528490A (en) Self-supervision sequence recommendation method based on long-term and short-term interests of user
CN114529077A (en) Click rate prediction method based on intra-session heterogeneous behaviors
Zheng et al. Kernelized deep learning for matrix factorization recommendation system using explicit and implicit information
CN113590976A (en) Recommendation method of space self-adaptive graph convolution network
CN112559904A (en) Conversational social recommendation method based on door mechanism and multi-modal graph network
CN117056595A (en) Interactive project recommendation method and device and computer readable storage medium
CN116776003A (en) Knowledge graph recommendation method, system and equipment based on comparison learning and collaborative signals
CN114936890A (en) Counter-fact fairness recommendation method based on inverse tendency weighting method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination