Self-Supervised Learning Collaborative Sequence Recommendation Method, Device and Medium Based on Time Interval

Abstract

The present invention proposes a self-supervised learning collaborative sequence recommendation method, device, and medium based on time intervals. The present invention uses technologies such as graph neural networks, attention mechanisms, and self-supervision to construct a self-supervised learning collaborative sequence recommendation model based on time intervals. By using user personal information and interaction data, it completes the attention weight change mechanism and self-supervised learning. During the training process, the model continuously changes relevant parameters therein, fully considering the periodic interests of users, and effectively fusing collaborative information and sequence information, greatly improving the personalized recommendation performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of personalized recommendation, and particularly relates to a self-supervised learning collaborative sequence recommendation method, device, and medium based on time intervals. Background Art

[0002] With the continuous development of information technology, the Internet plays an increasingly important role in today's society. People perform numerous activities on the Internet every day, such as watching videos, reading novels, shopping, browsing news, etc. However, with the continuous increase of information, people also need to spend a lot of time selecting the content they are truly interested in from the network, which brings a lot of inconvenience to people. Therefore, in order to solve the many problems brought by "information overload" in people's daily lives, a recommendation system has emerged. It will judge and predict the needs of users and recommend the content that users are most likely to be interested in to users, which brings great convenience to people and alleviates the trouble of people selecting from a vast amount of information.

[0003] The core of the recommendation system lies in using information such as user attributes and historical interaction sequences to model user preferences, obtaining a better user representation, and inferring the items that users may be interested in based on this. Therefore, the recommendation system must have the ability to identify user preferences and predict whether users are interested in items. For personalized recommendation, different users need to obtain different user preferences and representations. According to different methods of realizing the above capabilities, existing recommendation algorithms can be roughly divided into those based on content, social information, etc. Among them, content-based recommendation is a recommendation made based on the information of items. It does not require the evaluation opinions of users on items. Instead, it mainly uses the items that users have interacted with and uses machine learning methods to obtain the interests and representations required by users from the description of content features.

[0004] Recently, with the continuous deepening of research, content-based recommendations have been further refined, and sequential recommendations have emerged as a result. As a type of content-based recommendation, sequential recommendations have received extensive attention because they can further fit the interest transfer of users. Therefore, graph-based recommendations, deep learning-based recommendations, matrix factorization-based recommendations, etc. have also emerged in sequential recommendations. Early research on sequential recommendations used Markov chains to mine sequential patterns in historical data. FMPC models users' long-term preferences by combining first-order Markov chains and matrix factorization for next-basket recommendations. In the era of deep learning, researchers use different neural networks for sequential recommendations. GRU4Rec is the first model to apply recurrent neural networks (RNNs) to session-based recommendations. Caser uses a convolutional neural network (CNN) as the backbone network to embed the most recent items into an "image" and then uses convolutional filters to learn sequential patterns as local features of the image. SASRec captures self-attention-based sequential dynamics. CTA proposes using multiple kernel functions to learn temporal dynamics and then deploying different weighted kernels in the context information. IMfOU models user representations using relative recommendation time intervals. By introducing a knowledge graph, Chrous constructs the relationships between two different items and considers the distance between the interaction time and the recommendation time for recommendations. There are also other recent results based on self-attention. The above models generally assume that the next item to be recommended depends heavily on the most recent interaction, which is usually not true in practice. SLRC uses a Hawkes process to model the temporal dynamics of repeated consumption, but it only realizes that there are different time periods for different items and ignores that for the same item, its life cycle is also different among different users, which will cause the system to fail to learn user representations and thus affect the final recommendation performance. Summary of the Invention

[0005] The object of the present invention is to address the problem of personalized recommendation, overcome the deficiencies of the prior art, and propose a self-supervised learning collaborative sequential recommendation method, device, and medium based on time intervals by using techniques such as attention mechanisms and graph neural networks, fully considering and mining various relationships among users.

[0006] The present invention is implemented through the following technical solutions. The present invention proposes a self-supervised learning collaborative sequential recommendation method based on time intervals, and the method specifically includes the following steps:

[0007] Step 1: Obtain user personal information and a user interaction sequence dataset, preprocess the dataset, and divide it into a training set and a test set;

[0008] Step 2: Construct a self-supervised time interval-aware sequential recommendation model;

[0009] Step 3: Train the self-supervised time interval-aware sequential recommendation model constructed in Step 2;

[0010] Step 4: Input the user interaction sequence to be recommended into the self-supervised time interval-aware sequential recommendation model trained in Step 3, calculate the recommendation score of the item to be recommended for this user, and recommend the item to the user according to the recommendation score.

[0011] Furthermore, Step 1 includes the following steps:

[0012] Step 1.1: Extract the interaction sequence of the user from the user interaction sequence dataset;

[0013] Step 1.2: Divide the preprocessed dataset into a training set and a test set.

[0014] Furthermore, Step 2 includes the following steps:

[0015] Step 2.1: Obtain the embedded representation of the user's personal information and the embedded representation of the user interaction sequence;

[0016] Step 2.2: Construct a user-item bipartite graph and learn better user and item representations;

[0017] Step 2.3: Take the user's historical interaction records as the input of the gated recurrent unit and input them into the gated recurrent unit to learn the user representation with sequence information;

[0018] Step 2.4: Perform an attention mechanism between the user representation and the items in the sequence to calculate the attention score of the user for each item;

[0019] Step 2.5: Through the attention weight change mechanism, change the attention weight value of the user for each item, and re-aggregate the items according to the changed attention weights to generate a better user representation with sequence information;

[0020] Step 2.6: Through self-supervised technology, complete the aggregation of the user representation with sequence information and collaborative information;

[0021] Step 2.7: Effectively fuse the user representation optimized in Step 2.6 with the original user representation generated by the gated recurrent unit to obtain the final user representation and use it for subsequent recommendations.

[0022] Furthermore, Step 3 includes the following steps:

[0023] Step 3.1: Input the data of the training set in Step 1 into the model in Step 2 to obtain the final vector representation of the final user;

[0024] Step 3.2: Input the user's final vector representation into the prediction module to obtain the predicted score of the user for the item;

[0025] Step 3.3: Update the parameters of the model by calculating the difference between the predicted score value and the true value to optimize the loss function, and train to obtain the optimal time interval-aware collaborative sequence recommendation model with self-supervised learning.

[0026] The present invention proposes an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the time interval-based self-supervised learning collaborative sequence recommendation method are implemented.

[0027] The present invention proposes a computer-readable storage medium for storing computer instructions, and when the computer instructions are executed by a processor, the steps of the time interval-based self-supervised learning collaborative sequence recommendation method are implemented.

[0028] Compared with the prior art, the beneficial effect of the present invention is that the score of attention is dynamically adjusted based on the time interval, so that the recommendation process is no longer that the closer to the user interaction time, the greater the proportion, as in the traditional method. At the same time, through the self-supervised technology, the user representation with sequence information and the user representation with collaborative information are effectively fused to obtain a better user representation and improve the recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of the time interval-aware collaborative sequence recommendation method with self-supervised learning;

[0030] Figure 2 It is a model diagram of the time interval-aware collaborative sequence recommendation method with self-supervised learning;

[0031] Figure 3 It is an example diagram of the change of attention weights. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0033] As Figures 1 to 2 shown, the present invention provides a time interval-aware sequence recommendation method with self-supervision, specifically including the following steps:

[0034] Step 1: Obtain user personal information and user interaction sequence data sets, divide the data sets into training sets and test sets, and perform preprocessing;

[0035] Step 1 specifically includes the following steps:

[0036] Step 1.1: Extract user interaction sequences from the user interaction sequence dataset, such as the user's serial line historical interaction records for items such as commodities, microblogs, and advertisements;

[0037] Step 1.2: Preprocess the data set to filter out user interaction records from multiple terminals and overly unpopular items in the data set, for example, the user interaction sequence is less than 5, and the number of item interactions is less than 5;

[0038] Step 1.3: According to the leave-one-out method, take the last interacted item as the test set and the rest as the training set;

[0039] Step 2: Build a time interval-aware sequence recommendation model with self-supervision.

[0040] The step 2 specifically comprises the following steps:

[0041] Step 2.1: Obtain the embedding representation of user personal information and the embedding representation of user interaction sequence. First, use independent hot encoding to obtain the ID encoding of each item v and each user u.

[0042] In step 2.1, let U, I, and T be the user set, item set, and time set in the dataset respectively. For each user u belonging to U, set the user's interaction sequence to: [(s 1 ,t 1 ),(s 2 ,t 2 ),···,(s n ,t n )], here s i Belongs to the project set, t i Belonging to the time set, each pair (s i , t i ) indicates that the user is at timestamp t i Departments and Projects i Interaction. + ={y ui |u∈U,i∈I}.

[0043] Step 2.2: Build a user-item bipartite graph to learn better user and item representations;

[0044] Construct a bipartite graph G = {V, E}, where the node set V = U∪I includes all users and items, and the edge set E = O +Represents all historical interactions. Based on the graph G, the adjacency matrix between users and items is defined as R ∈ R m×n , where m is the number of users and n is the number of items. For each interaction pair (u, i) in R, if user u has not interacted with item i, then r ui = 0; otherwise, r ui is the number of interactions between user u and item i. Let E I ∈ R n×d be the item embedding matrix generated by the item embedding layer, and e i I be the embedding vector of item i, E U ∈ R m×d be the user embedding matrix generated by the user embedding layer, and e j U be the vector of embedded user j, where e i I ∈ E I , e j U ∈ E U , and d is the embedding size.

[0045] Use a graph convolutional network to obtain the collaborative information of users. First, construct an adjacency matrix from user-item interactions. Take a user as an example. For any interacting user-item pair (u, i), the information propagation process at the l-th layer is defined as:

[0046]

[0047]

[0048] where W 1 , W 2 ∈ R d×d are trainable weight matrices used to extract useful information during the propagation process, and p ui is the coefficient controlling the decay factor. e u (l -1) is the user representation generated from the previous layer, which remembers the messages from its (l - 1)-hop neighbors. At the same time, add a user self-loop to ensure the retention of the original information. Then, the representation of user u at the l-th layer can be obtained. Note that e (0) u and e (0) i are respectively initialized as e u U and e i I . A user (or item) can receive messages propagated from its l-hop neighbors. At the l-th layer, the representation of user u is updated as follows:

[0049]

[0050] where N u represents the l-hop neighbors of user u. The same operation is performed on the items. After l-layer propagation, multiple representations of user u and item i are obtained. Since the representations obtained at different layers of the graph convolutional network emphasize the information transmitted by different connections, they can reflect different features of users (or items). Therefore, they are concatenated to form the final embedding of each user (or each item), and the final user embedding and item embedding are obtained as follows:

[0051]

[0052] Here || is the concatenation operation. By doing so, on the one hand, the initial embedding can be enriched through the embedding propagation layer; on the other hand, the propagation range can be controlled by adjusting l.

[0053] Step 2.3: Take the user representation learned in Step 2.2 as the input of the gated recurrent unit and input it into the gated recurrent unit to learn the user representation with sequential information;

[0054] In Step 2.3, in order to capture the sequentiality between items in the sequence, a learnable positional embedding is injected into each item in the sequence:

[0055]

[0056] where e i * is the item representation obtained in the previous step. In particular, define Xu = [e 1 s , e 2 s , ···, e L s as the user sequence without timestamps, aiming to understand the influence of the user's historical interaction sequence. Select the gated recurrent unit to capture the sequential information of the user:

[0057]

[0058] Step 2.4: Perform an attention mechanism between the user representation and the items in the sequence to calculate the attention scores of the user for each item;

[0059]

[0060] where e i s is the i-th item of X u .

[0061] Then, the intensity function is used to construct the influence of the time interval on the attention value. Generally speaking, most existing methods assume that the closer an item is to the recommended time, the greater its impact on the recommendation result. However, according to observations, the influence of certain items on the final recommendation takes some time to manifest. To address this issue, after obtaining the user representation from the gated recurrent unit based on sequence information, the attention score of each item in the sequence is first calculated as:

[0062]

[0063] where △t is the time interval between the item interaction time and the recommended time, and N(△t|μ i u ,σ i u ) is a Gaussian distribution with mean μ i u and standard deviation σ i u , and π is used to control the change ratio. Here, μ and σ are learned based on the items purchased by the user. Therefore, the corresponding attention scores can be adjusted according to the variation of the learned intensity function for different items by different users. By introducing the kernel function, items with more recent interactions may have lower attention scores than previous items.

[0064] Step 2.5: Through the attention weight change mechanism, change the attention weight value of the user for each item, and re - complete the aggregation of items according to the changed attention weights to generate a better user representation with sequence information;

[0065] Combined with Figure 3 , in Step 2.5, depict the user representation by combining the weights learned from the above - adjusted attention weights:

[0066]

[0067] Step 2.6: Through self - supervised techniques, complete the aggregation of user representations with sequence information and collaborative information;

[0068] In Step 2.6, inspired by the success of self - supervised learning on graphs, integrate self - supervised learning into the proposed network to combine user representations from sequential information and collaborative information to further enhance the user representation. In particular, an auxiliary task is designed in the following two steps to benefit the recommendation task.

[0069] 1: Create self-supervised signals. In the method of the present invention, user representations are generated in two ways, namely, gated recurrent units and graph convolutional networks. Since these two methods only use sequential information or collaborative information, the two types of user embeddings can effectively complement each other, thus generating a richer user representation. For each batch containing n users during the training process, there is a bijective mapping between the two types of user embeddings. Naturally, these two types can be used as the ground truth for self-supervised learning for each other. If two user embeddings both represent the same user, mark this pair as a positive ground truth, otherwise mark it as negative.

[0070] 2: Self-supervised learning. Compare the two types of user embeddings and use InfoNCE as the learning objective with the standard binary cross-entropy loss between samples from real samples (positive) and corrupted samples (negative), which is defined as follows:

[0071]

[0072] where, e u g is the user representation obtained by the graph convolutional network, and e u S is the user representation obtained by the gated recurrent unit, and f D (·): R d ×R d →R is the discriminator function, which takes two vectors as input and then scores the consistency between them. The discriminator is implemented as the dot product between two vectors. This learning objective is used to maximize the mutual information between user embeddings learned in different ways. By doing so, they can obtain information from each other to improve their performance.

[0073] Step 2.7: Effectively fuse the user representation optimized in Step 2.6 with the original user representation generated by the gated recurrent unit to obtain the final user representation and use it for subsequent recommendations.

[0074] where, in order to obtain the final user representation, use a linear strategy to adjust the weights of the user representation (e u adj ) and the user representation obtained by the gated recurrent unit (e u S ) for fusion:

[0075]

[0076] Step 3: Train the time interval-aware collaborative sequential recommendation model with self-supervised learning in Step 2, which specifically includes the following steps:

[0077] Step 3.1: Input the data of the training set in Step 1 into the model in Step 2 to obtain the final vector representation of the end user;

[0078] Step 3.2: Input the final representation of the user into the prediction module to obtain the predicted score of the user for the item;

[0079] As Figure 2 shown, for user a, the prediction module in Step 3.2 performs an inner product operation on the final representation vector e a of user a and the vector representation e z of candidate item z, and obtains the recommendation score of user a with respect to candidate item z through the softmax function operation;

[0080] Step 3.3: Update the parameters of the model by calculating the difference between the predicted score value and the true value to optimize the loss function, and train to obtain the optimal time interval-aware collaborative sequential recommendation model with self-supervised learning.

[0081] Among them, the calculation process of the loss function in Step 3.3 is as follows:

[0082]

[0083] where Θ = {E I , E U , E P , E μ , E σ} is a subset of the model parameters, including the item embedding matrix, the position embedding matrix, the user embedding matrix, the mu matrix, and the sigma matrix. ||·|| represents the Frobenius norm, λ is the regularization parameter, and σ(·) is the sigmoid activation function. D is the training data set, which includes the true interaction pairs (u, t, i) and the negative sample item j that u did not interact with at time t.

[0084] Finally, unify the recommendation task and the self-supervised task in Step 2.6 into a master-slave learning framework, with the recommendation task as the main task and the self-supervised task as the auxiliary task. The joint learning objective is formalized as:

[0085]

[0086] where β controls the influence of the self-supervised task.

[0087] Step 4: Input the personal information and interaction sequence of the user to be recommended into the self-supervised time interval-aware sequential recommendation model after training, updating, and optimizing in Step 3, calculate the score of the item to be recommended for this user, and recommend the item to the user according to the recommendation score.

[0088] By calculating the embedding e of the user u and the embedding e of each candidate item o to generate a recommendation list, where r o = e u e o T , and select the top K candidate items with the highest similarity scores.

[0089] The present invention provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the self-supervised learning collaborative sequence recommendation method based on time intervals are implemented.

[0090] The present invention provides a computer-readable storage medium for storing computer instructions, and when the computer instructions are executed by a processor, the steps of the self-supervised learning collaborative sequence recommendation method based on time intervals are implemented.

[0091] The above has introduced in detail the self-supervised learning collaborative sequence recommendation method, device, and medium based on time intervals proposed by the present invention. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. Self-supervised learning collaborative sequence recommendation method based on time interval, characterized in that, the method specifically includes the following steps: Step 1: Obtain user personal information and user interaction sequence datasets, preprocess the datasets and divide them into training sets and test sets; Step 2: Construct a time interval-aware sequence recommendation model with self-supervision; Step 3: Train the time interval-aware sequence recommendation model constructed with self-supervision in Step 2; Step 4: Input the user interaction sequence to be recommended into the time interval-aware sequence recommendation model trained in Step 3, calculate the recommendation score of the item to be recommended for this user, and recommend the item to the user according to the recommendation score; The Step 2 includes the following steps: Step 2.1: Obtain the embedded representation of user personal information and the embedded representation of the user interaction sequence; In step 2.1, let U, I, and T be the user set, item set, and time set in the dataset respectively; for each user u belonging to U, set the interaction sequence of the user as: [(i 1 , t 1 ), (i 2 , t 2 ), ···, (i n , t n )], where i k belongs to the item set, t k belongs to the time set, and each pair (i k , t k ) represents that the user interacts with item i k at timestamp t k ; O + = {r ui |u ∈ U, i ∈ I}; Step 2.2: Construct a user-item bipartite graph and learn better user and item representations; Construct a bipartite graph \(G = \{V, E\}\), where the node set \(V = U\cup I\) includes all users and items, and the edge set \(E = O\). + represents all historical interactions; Based on the graph \(G\), the adjacency matrix between users and items is defined as \(R\in\mathbb{R}\). m×n , where \(m\) is the number of users and \(n\) is the number of items; For each interaction pair \((u, i)\) in \(R\), if user \(u\) has not interacted with item \(i\), then \(r\). ui \(= 0\); Otherwise, \(r\). ui is the number of interactions between user \(u\) and item \(i\); Let \(E\). I \(\in\mathbb{R}\). n×d be the item embedding matrix generated by the item embedding layer, and \(e\). i I be the embedding vector of item \(i\), \(E\). U \(\in\mathbb{R}\). m×d be the user embedding matrix generated by the user embedding layer, and \(e\). u U be the vector of the embedded user \(u\), where \(e\). i I \(\in E\). I , \(e\). u U \(\in E\). U , and \(d\) is the embedding size. Use a graph convolutional network to obtain the collaborative information of users; first construct an adjacency matrix from user-item interactions; for any interacting user-item pair (u, i), define the information transfer process of the l-th layer as: where W 1 , W 2 ∈R d×d is a trainable weight matrix for extracting useful information during propagation, while p ui is the coefficient controlling the decay factor; e u (l-1) is the user representation generated from the previous layer, which remembers the messages from its (l - 1)-hop neighbors; meanwhile, a user self-loop is added to ensure the retention of the original information; e u (0) and e i (0) are respectively initialized as e u U and e i I ; then, the representation of user u at the l-th layer can be obtained; a user or an item can receive the messages propagated from its l-hop neighbors; at the l-th layer, the representation of user u is updated as follows: where N u represents the l-hop neighbors of user u; perform the same operation on items; after l-layer propagation, multiple representations of user u and item i are obtained; connect the representations obtained from different layers that emphasize the information transmitted by different connections to form the final embedding of each user or each item, and obtain the final user representation and item representation as follows: where || is the concatenation operation; Step 2.3: Use the learned user representation as the input of the gated recurrent unit and input it into the gated recurrent unit to learn the user representation with sequence information; in Step 2.3, inject a learnable position embedding into each item in the sequence: where e i * is the item representation obtained in the previous step; define Xu = [e 1 s , e 2 s , ···, e L s as the user interaction sequence without timestamps, and select a gated recurrent unit to capture the sequence information of the user: Step 2.4: Perform an attention mechanism between the user representation and the items in the sequence to calculate the attention score of the user for each item; In Step 2.4, after obtaining the user representation from the gated recurrent unit based on sequence information, first calculate the attention score of each item in the sequence as: where △t is the time interval between the item interaction time and the recommendation time, and N(△t|μ i u ,σ i u ) is a Gaussian distribution with mean μ i u and standard deviation σ i u . The change ratio is controlled by π, where μ and σ are learned based on the items purchased by the user; Step 2.5: Through the attention weight change mechanism, change the value of the attention weight of the user for each item, and re-complete the aggregation of items according to the changed attention weight to generate a better user representation with sequence information; in Step 2.5, depict the user representation by combining the weights learned from the adjusted attention weights above: Step 2.6: Through self-supervised technology, complete the aggregation of user representations with sequence information and collaborative information; adopt the standard binary cross-entropy loss between samples from real samples and corrupted samples in InforNCE as the learning objective, which is defined as follows: Among them, e u * is the user representation obtained by the graph convolutional network, and e u S is the user representation obtained by the gated recurrent unit, and f D (·): R d ×R d →R is a discriminator function that takes two vectors as inputs and then scores the consistency between them; the discriminator is implemented as the dot product between two vectors; this learning objective is used to maximize the mutual information between user representations learned in different ways; Step 2.7: Effectively fuse the user representation optimized in Step 2.6 with the original user representation generated by the gated recurrent unit to obtain the final user representation for subsequent recommendations; specifically: use a linear strategy to adjust the weights of the user representation e u adj and the user representation e u S obtained through the gated recurrent unit for fusion: Among them, e u is represented for the end user.

2. The self-supervised learning collaborative sequence recommendation method based on time interval according to claim 1, characterized in that, the Step 1 includes the following steps: Step 1.1: Extract the interaction sequences of users from the user interaction sequence dataset; Step 1.2: Divide the preprocessed dataset into a training set and a test set.

3. The self-supervised learning collaborative sequence recommendation method based on time interval according to claim 1, characterized in that, the Step 3 includes the following steps: Step 3.1: Input the data of the training set in Step 1 into the model in Step 2 to obtain the final vector representation of the final user; Step 3.2: Input the final vector representation of the user into the prediction module to obtain the prediction score of the user for the item. Step 3.3: Update the parameters of the model by calculating the difference between the predicted score value and the true value to optimize the loss function, and train to obtain an optimal time interval-aware collaborative sequence recommendation model with self-supervised learning.

4. An electronic device, comprising a memory and a processor, the memory storing a computer program, wherein, when the processor executes the computer program, the steps of the method according to any one of claims 1-3 are implemented.

5. A computer-readable storage medium for storing computer instructions, wherein, when the computer instructions are executed by the processor, the steps of the method according to any one of claims 1-3 are implemented.

Images