A method for user preference prediction with sensitive behavior data protection

Abstract

The application discloses a user preference prediction method for sensitive behavior data protection and relates to the technical field of big data analysis. The application introduces the localized differential privacy of random response in the construction of dynamic characteristics of preferences, protects the security of potential feature data of the preferences, fuses a time decay factor in the process of learning dynamic characteristics, optimizes the dependency relationship between features in different periods, and realizes user dynamic preference prediction under dynamic change and sensitive behavior protection, so that the application realizes the construction of a privacy protection mechanism for sensitive behavior of users and retains the availability of data, and improves the accuracy of future preference prediction.

Description

Technical Field

[0001] This invention relates to a user preference prediction method for protecting sensitive behavioral data, and relates to the field of big data analytics. Background Technology

[0002] User selection behavior reveals their preferences for specific item categories. The aggregation of multiple preferences can create a user profile, and a good user profile has a decisive impact on the construction of subsequent recommendation models and the improvement of recommendation performance. User profiles are sensitive to users, and the user behavior data used to construct these profiles contains a large amount of sensitive behavioral data. Once attackers obtain this sensitive behavioral data, combined with certain background knowledge, they can use inference attacks to uncover potential user preferences and thus target specific users. Therefore, research on user preference perception under the protection of sensitive behaviors is of great significance.

[0003] Current research on user preference models, while incorporating temporal context into preference prediction, largely relies on predictions based on time series or time-series events. Time-series and time-series event-based prediction algorithms typically follow a two-step strategy: first, projecting the time set into a latent representation; second, learning a model with the latent representation. However, this approach often leads to information loss or suboptimal prediction performance. Some researchers have considered the dynamic and static aspects of user preferences separately. For example, Yu et al. designed an attention-based adaptive model to learn the temporal dependencies of elements and sets, providing an open update mechanism that allows it to discover hidden shared patterns across different sequences, fusing dynamic and static information to improve prediction performance. While this method has advantages in prediction performance compared to current state-of-the-art methods, it only utilizes element label embeddings, neglecting the impact of element features on prediction and the security issues of sensitive behaviors behind user preferences. Furthermore, studies treating each user's historical interaction item as an independent entity ignore the possibility of multiple interaction items occurring at a given moment.

[0004] In summary, existing research on privacy-preserving recommendations based on user preferences still has the following problems: (1) it ignores the situation where there are multiple interactive items at a certain time; (2) it ignores the problem that long-term preferences gradually decrease in their impact on future preferences over time; and (3) it ignores the privacy and security issues of sensitive behaviors that play a decisive role in user preference models.

[0005] Temporal set forecasting differs from time series forecasting or temporal event forecasting, which only considers the relationships between individual items in a time series. A temporal set is a collection of interactions that occur simultaneously at a given moment. For example, when shopping at a supermarket, purchasing multiple items at once creates a set of items with the same timestamp, which is considered a temporal set. Figure 1The diagram illustrates a comparison between time series prediction, time event prediction, and temporal set prediction. Temporal sets are more complex than time series or temporal events, containing more information and thus possessing greater research value. Mining temporal sets can better capture potential user preference characteristics and the dynamics of user preferences. Therefore, it is essential to provide a temporal set-based dynamic preference perception method based on localized differential privacy. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a user preference prediction method for protecting sensitive behavioral data.

[0007] The technical solution of the present invention: One objective of this invention is to provide a method for predicting user preferences while protecting sensitive behavioral data. This method includes the following steps: Step 1: Treat the categories of user's historical interactions as sensitive user behavior data, and add a perturbation mechanism locally to ensure that it avoids the risk of privacy leakage during uploading and use; Step 2: Using the similarity of subgraph nodes with time decay factors in the form of temporal sets, learn the data processed in Step 1 to achieve dynamic preference representation learning and obtain dynamic representation; Step 3: Integrate dynamic and static representations to achieve user preference prediction.

[0008] Further specifying, step 1 is: given a random response algorithm and its domain Sum range With probability Preserve the true value, in terms of probability Randomly replace other values; perform random response algorithms locally on each user's sensitive behavior data. deal with; like In any adjacent sensitive behavior dataset and ( The output results obtained from ) satisfy ,but Satisfy localized differential privacy; In the formula It is the privacy budget, which indicates the level of privacy protection.

[0009] Further specifying, step 2 is: S1: Embedding learning is performed on the nodes in the temporal subset graph to obtain node feature embeddings. Then, the cross product is used to obtain the similarity score between all temporal subset graphs in the temporal set sequence. The similarity between temporal subset graphs is optimized by adding a time decay factor to obtain the dynamic similarity score between nodes. The similarity scores of all nodes are aggregated to obtain the user's score. Relationship matrix of historical interactive items ; S2: Based on the obtained first A temporal set node relationship matrix Combined with the first a set of time states Graph convolutional networks are used to learn the node feature embeddings for each temporal subset of the graph. ; in Indicates the first The node feature matrix of the layer, User The The node relationship matrix of each temporal set, It is a degree matrix. It is the first The weight matrix, Indicates the activation function; S3: Perform self-attention processing on each temporal set in the user temporal set subgraph sequence to obtain the weight values ​​between nodes, and update the node embedding matrix in the temporal set. The node embedding vectors in the _ temporal set are stacked together to obtain the _ ... Embedment matrix of all nodes in a temporal set; S4: Construct an embedding matrix sequence based on the embedding matrix of each temporal set. Attention weights between embedding matrices are obtained through a self-attention mechanism; in These represent different time lengths; , , These are obtained from the embedding matrix, specifically the query matrix, key matrix, and value matrix. It is a learnable weight matrix; These are the attention weights between two different temporal sets; S5: Based on the node relationship matrix The first and self-attention mechanism obtained The neighbor information of each temporal set is used, considering the dependencies between temporal sets, and the attention weights and temporal decay of each temporal set are combined to learn the information in the first temporal set. Global dynamic preference representation of a set of time states ; in It is the first Node feature embedding at time step Represent a function used to compute a given i-th... Embedding matrix of each temporal set and the Nonlinear transformation results of the temporal set relation matrix; Indicates the first The weighted sum of the feature embeddings of all previous temporal sets is obtained based on the attention weights; It is the decay coefficient of the weighted sum of historical states, which aims to emphasize the importance of the current temporal set by reducing the weight of historical states; It is the first Attention weights for each temporal set, representing the embedding matrix. The importance of.

[0010] To further specify, S1 is as follows: First, the original features of the nodes are mapped to a high-dimensional embedding space through a multi-layer fully connected network; in It is the mapped feature embedding; and They are respectively of 3D attribute features and 1D category features; Then, the cross product is used to obtain the similarity score between all temporal set subgraphs in the temporal set sequence; in express Middle node Feature embedding, express Middle node Feature embedding, There are two nodes. and Similarity results between them; The similarity scores of all node pairs are combined into the first... The and the first Similarity matrix of temporal sets To ensure that the similarity scores are on the same order of magnitude, they are normalized. The similarity between temporal subset graphs is optimized by adding a time decay factor to obtain the dynamic similarity score between nodes; in It is a node and nodes In The similarity of the temporal sets at any given moment. Is and Time-of-flight similarity matrix Middle node and nodes Similarity score between them It is a node and nodes At any moment The number of co-occurrences of edges. It is the time decay coefficient; in It is the time difference between the current time and the reference time. This is the decay rate, representing the magnitude of decay per unit time. When The larger the value, the faster the influence of historical data will diminish. Finally, the similarity scores of all nodes are aggregated to obtain the user's score. Relationship matrix of historical interactive items .

[0011] To further specify, S2 is as follows: Let the user temporal set graph sequence be... ,in Indicates user No. A set of tenses, Include Nodes Using the formula Learn the feature vector of each node in each temporal set subgraph ; In users Historical Interaction Item Relationship Matrix Extract the temporal subset graph for each user. The similarity scores between nodes constitute relation matrix ,Right now ; According to the obtained number A temporal set node relationship matrix Combined with the first a set of time states Graph convolutional networks are used to learn the node feature embeddings for each temporal subset of the graph. .

[0012] To further specify, S3 is as follows: Let the input be the user Temporal set graph sequence The corresponding feature vector sequence ,in, , User The set of nodes in the t-th temporal set. It is each temporal set Feature matrix, Indicates the number of nodes; yes The node relationship matrix; Initialize node In the Feature embeddings of a temporal set: in It is a node In the Eigenvectors of a set of time states These are trainable parameter weights; Calculate neighbor nodes Query, key, and value matrix: in , , These are the three matrix parameters in the self-attention mechanism. , , It is to make nodes Feature vector mapping to The result obtained in 3D space; Calculate neighbor nodes Attention weights: Aggregate neighbor information and update nodes Feature vectors: in It is the first The number of nodes in a temporal set subgraph. It is the first Temporal central nodes The weighted sum of the feature vectors of the neighboring nodes; The first The node embedding vectors in the _ temporal set are stacked together to obtain the _ ... The embedding matrix of all nodes in a temporal set.

[0013] Further specifying, step 3 is: The embedding matrix obtained by all users is treated as a static representation , will users The dynamic embedding matrix is ​​a dynamic representation These two representations are used as inputs and integrated into the gating information component to learn the user. Hidden state Through formula Learning through iterative updates: in It is a node Mapped to user Static representation The corresponding index function; It is an indicator vector consisting of 0s and 1s, where a value of 1 indicates that the corresponding node is in the user's history. middle; These are the trainable parameters of the update gate that control the static and dynamic representations; Further calculate the probability of all elements appearing in the next temporal subset; in , These are trainable parameters; Since the prediction of the next temporal set can be viewed as a multi-label learning problem, the loss function using L2 regularization is as follows: in These are the trainable parameters of the model. It is the number of training samples. It is a hyperparameter that controls L2 regularization. and It is the first of the reference truth values The first of the peacekeeping prediction values Each element in the user The probability of the next set appearing is determined by optimizing the proposed model through minimizing the equation until convergence; Finally, the top K items are selected as the first K items based on the Top-K criteria. A set of temporal parameters is recommended to the user to obtain the final result. .

[0014] Beneficial effects: This invention protects the security of potential preference feature data by introducing localized differential privacy of random responses into the construction of dynamic preference features. Furthermore, it integrates a time decay factor during the learning of dynamic features to optimize the dependencies between features at different times. Under dynamic changes and with sensitive behavior protection, this invention enables the prediction of dynamic user preferences, achieving a privacy protection mechanism for sensitive user behaviors while preserving data availability and improving the accuracy of future preference predictions. Attached Figure Description

[0015] Figure 1 This is a comparative diagram of time series forecasting, time event forecasting, and temporal set forecasting; Figure 2 The framework diagram of the temporal set dynamic preference awareness method based on localized differential privacy provided by the present invention is shown below. Figure 3 This is a schematic diagram illustrating the learning process of the similarity matrix between two temporal sets. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Those skilled in the art will understand that, unless otherwise stated, the singular forms “a,” “an,” and “,” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0018] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 2 As shown, a method for predicting user preferences for the protection of sensitive behavioral data includes the following steps: Step 1: Treat the categories of user's historical interactions as sensitive user behavior data, and add a perturbation mechanism locally to ensure that it avoids the risk of privacy leakage during uploading and use; Step 2: Using the similarity of subgraph nodes with time decay factors in the form of temporal sets, learn the user history processed in Step 1 to achieve dynamic preference representation learning and obtain dynamic representation; Step 3: Integrate dynamic and static representations to achieve user preference prediction.

[0019] Specifically, based on the definition of localized differential privacy and the description of sensitive behaviors, the definition of localized differential privacy for sensitive user behaviors is given below: Temporal set and temporal set sequence Given a dataset of user history interactions , ,in It is a collection of users. It is a collection of items based on the user's historical interactions. It is a set of historical interaction timestamps. The set of any number of items at a given timestamp is called a temporal set. , by temporal set The sequence formed is called a temporal set sequence. .user The temporal set sequence is represented as .

[0020] Temporal set diagram With temporal set graph Given the first Temporal set of timestamps The graph formed by this set of temporal states is called the temporal set graph of the t-th timestamp. Among them, users Temporal set at the t-th timestamp The resulting graph is called the temporal set subgraph of the t-th timestamp. .

[0021] Sensitive behaviors Given a user graph ,in Indicates user A collection of historical interactive projects, for any one project , and They are respectively of 3D attribute features and 1D category features. (Referring to the user...) about Categories of historical interactive items For users Sensitive behaviors.

[0022] Sensitive behaviors here refer to the project attributes in a user's historical interaction behavior that can directly indicate the user's preference for a certain type of project. User The history records all project attribute sets, including attribute features and category features.

[0023] Sensitive Behavior Localization Differential Privacy: Given Each user has a corresponding privacy algorithm. and its domain Sum range If the algorithm In any user-sensitive behavior and The same output result was obtained. If the following inequalities are satisfied, then satisfy -Localized differential privacy.

[0024] in It is the privacy budget, which indicates the level of privacy protection. The smaller the value, the more noise is needed to ensure the differential privacy property, thus resulting in a higher level of privacy protection.

[0025] User preference prediction with sensitive behavior protection: given a user set Given any user Sensitive behaviors and Temporal set sequence of step size For users Perceive user preferences for protecting sensitive behaviors and predict user behavior. In the The temporal set of steps is formally defined as: in Represents a preference-aware recommendation function The set of parameters, It is a budget for privacy protection of sensitive behaviors. satisfy -Localized differential privacy.

[0026] Predicting dynamic user preferences while protecting sensitive behaviors is a highly challenging research task, especially under dynamic changes. Protecting and predicting sensitive behaviors within user preferences becomes even more difficult. Specifically, one challenge is how to construct a privacy protection mechanism for sensitive user behaviors while maintaining data availability. Consider a user's potential item category preferences hidden behind their purchasing or browsing behavior as sensitive behaviors. When the data volume is large enough, attackers can use this information to build user profiles and launch inference attacks. However, like user preferences, these sensitive behaviors are dynamic. This sensitive behavior data also plays a crucial role in constructing user preferences. Therefore, constructing a privacy protection mechanism with data availability is particularly important.

[0027] Secondly, obtaining dynamic user preference representations presents another challenge. User preferences are variable and dependent on time factors. Furthermore, there are correlations between different items. Effectively utilizing this information to construct dynamic user preferences and integrating static representations to predict future temporal sets is crucial for achieving more accurate predictions of user preferences under dynamic changes and with sensitive behavior protection.

[0028] The detailed implementation process of the dynamic preference-aware model with sensitive behavior protection mechanisms is as follows: Step 1: Randomization of Local Sensitive Behaviors: Given Each user contains [number] users, and each user has [number] users. Each sensitive behavior has a record. Each value can be selected. Given a random response algorithm. and its domain Sum range With probability Preserve the true value, in terms of probability Randomly replace other values; perform random response algorithms locally on each user's sensitive behavior data. Processing. If In any adjacent sensitive behavior dataset and ( () and The output results obtained from only one user-sensitive behavior dataset) satisfy ,but Satisfy localized differential privacy.

[0029] The proof is as follows: The probability that a record retains its true value is: Each user has There are 100 records, therefore the probability of retaining a user's actual sensitive behavior value is 100%. The probability of not retaining the true value is Furthermore, because each user's sensitive behavior records are independent, therefore... The probability of retaining the true sensitive behavior value from a user's sensitive behavior records is: , Let be the probability of not retaining the true value. Therefore, we have: For adjacent datasets Because there are still If a user's sensitive behavior records remain unchanged, what is the probability of the true sensitive behavior value? The probability that other values ​​are randomly selected is Therefore, .

[0030] Substituting the two formulas above into the inequality, we can obtain... Take the logarithm of both sides. After sorting, Simplified to, The proof is complete, indicating that the condition is satisfied. -Localized differential privacy.

[0031] Step 2: Dynamic Preference Representation Learning S1: Node relationship matrix with time decay To quickly obtain similarity results between temporal subgraphs, we first perform embedding learning on the nodes in the temporal subgraphs to obtain node feature embeddings. The original node features are then mapped to a high-dimensional embedding space through a multi-layer fully connected network.

[0032] in It is the feature embedding after mapping. and They are respectively of 1D attribute features and 1D category features.

[0033] The cross product is used to obtain the similarity score between all temporal subgraphs in the temporal set sequence.

[0034] in express Middle node Feature embedding, express Middle node Feature embedding. There are two nodes. and The similarity results between them. Figure 3 This is a schematic diagram of the learning process of the similarity matrix between two temporal sets.

[0035] The similarity scores of all node pairs are combined into the first... The and the first Similarity matrix of temporal sets To ensure that the similarity scores are on the same order of magnitude, they are normalized.

[0036] Finally, the similarity between temporal subgraphs is optimized by adding a time decay factor to obtain the dynamic similarity score between nodes.

[0037] in It is a node and nodes In The similarity of the temporal sets at any given moment. Is and Time-of-flight similarity matrix Middle node and nodes Similarity score between them It is a node and nodes At any moment The number of co-occurrences of edges. It is the time decay coefficient.

[0038] in It is the time difference between the current time and the reference time; This is the decay rate, representing the magnitude of decay per unit time. When The larger the value, the faster the influence of historical data will diminish.

[0039] Aggregate the similarity scores of all nodes to obtain the user's score. Relationship matrix of historical interactive items .

[0040] S2: Temporal set relation matrix and node embedding learning Let the user temporal set graph sequence be... ,in Indicates user No. A set of tenses, Include Nodes Using the formula Learn the feature vector of each node in each temporal set subgraph The user obtained in the previous step Historical Interaction Item Relationship Matrix Extract the temporal subset graph for each user. The similarity scores between nodes constitute relation matrix ,Right now .

[0041] According to the obtained number A temporal set node relationship matrix Combined with the first a set of time states Graph convolutional networks are used to learn the node feature embeddings for each temporal subset of the graph. .

[0042] in Indicates the first The node feature matrix of the layer, User The The node relationship matrix of each temporal set, It is a degree matrix. It is the first The weight matrix, This represents the activation function.

[0043] S3: Temporal Set Self-Attention Mechanism Learning The input is user Temporal set graph sequence The corresponding feature vector sequence ,in, , User The set of nodes in the t-th temporal set. It is each temporal set Feature matrix, Indicates the number of nodes. yes The node relationship matrix.

[0044] A self-attention mechanism is applied to each temporal set in the user temporal set subgraph sequence to obtain the weight values ​​between nodes and update the node embedding matrix in the temporal set.

[0045] Initialize node In the Feature embeddings of a temporal set: in It is a node In the Eigenvectors of a set of time states These are trainable parameter weights.

[0046] Calculate neighbor nodes Query, key, and value matrix: in , , These are the three matrix parameters in the self-attention mechanism. , , It is to make nodes Feature vector mapping to The result obtained in 3D space.

[0047] Calculate neighbor nodes Attention weights: Aggregate neighbor information and update nodes Feature vectors: in It is the first The number of nodes in a temporal set subgraph. It is the first Temporal central nodes The weighted sum of the feature vectors of the neighboring nodes. The first... The node embedding vectors in the _ temporal set are stacked together to obtain the _ ... Embedding matrix of all nodes in a time-state set .

[0048] S4: Global Dynamic Node Representation Learning In the temporal set subgraph sequence, in order to learn the dynamics and evolution patterns of the dependencies between temporal sets, an embedding matrix sequence is constructed based on the embedding matrix of each temporal set. Attention weights between embedding matrices are obtained through a self-attention mechanism.

[0049] in These represent different time lengths. , , The query matrix, key matrix, and value matrix are obtained from the embedding matrix, respectively. It is a learnable weight matrix. It is the attention weight between two different temporal sets.

[0050] Based on the node relationship matrix The first and self-attention mechanism obtained The neighbor information of each temporal set is used, considering the dependencies between temporal sets, and the attention weights and temporal decay of each temporal set are combined to learn the information in the first temporal set. Global dynamic preference representation of a set of time states .

[0051] in It is the first Node feature embedding at time step Represent a function used to compute a given i-th... Embedding matrix of each temporal set and the The nonlinear transformation result of the relation matrix of each temporal set. Indicates the first The weighted sum of the feature embeddings of all previous temporal sets is calculated using attention weights. It is the decay coefficient of the weighted sum of historical states, and its purpose is to emphasize the importance of the current temporal set by reducing the weight of historical states. It is the first Attention weights for each temporal set, representing the embedding matrix. The importance of.

[0052] Step 3: Fusion of dynamic and static preferences Fusion of dynamic and static preference representations; regarded as user In the The dynamic feature embedding obtained from the temporal set learns all the information about the user's previous actions with time decay. However, it ignores other similar user-to-user interactions. The impact of personalization. Therefore, adding information from other similar users can reduce data sparsity and improve the robustness of the model.

[0053] The embedding matrix obtained by all users is treated as a static representation , will users The dynamic embedding matrix is ​​a dynamic representation These two representations are used as input and fused into the gating information component to learn the user. Hidden state Through formula Learning through iterative updates: in It is a node Mapped to user Static representation The corresponding index function. It is an indicator vector consisting of 0s and 1s, where a value of 1 indicates that the corresponding node is in the user's history. middle. It controls the updating of gating trainable parameters for both static and dynamic representations.

[0054] Prediction layer: Based on the user's current state obtained earlier, calculate the probability of all elements appearing in the next temporal subset.

[0055] in , These are trainable parameters.

[0056] Finally, the top K items are selected as the first K items based on the Top-K criteria. A set of temporal parameters is recommended to the user to obtain the final result. .

[0057] During training, predicting the next time-series can be viewed as a multi-label learning problem. Therefore, the loss function using L2 regularization is as follows: in These are the trainable parameters of the model. It is the number of training samples. It is a hyperparameter that controls L2 regularization. and It is the first of the reference truth values The first of the peacekeeping prediction values Each element in the user The probability of the next set appearing is determined by optimizing the proposed model through minimizing the equation until convergence.

[0058] Example 1: The algorithm flow is as follows: Input: User graph , , All user embedding matrices (static representation), user temporal set sequence Privacy Budget Other parameter sets Output: User exist Prediction results of temporal sets 1) for i=1 to do 2) Local sensitive behavior randomization 3) endfor 4) for i=1 to t do 5) Historical interactive project feature embedding processing 6) Obtain node similarity in subgraphs of different temporal states 7) No. Similarity between nodes of each temporal subset 8) Acquiring users Relationship matrix of historical interactive items 9) Node feature embedding for each temporal subset graph 10) No. The weighted sum of the feature vectors of the neighboring nodes of node i in each temporal set 11) Update # Embedding matrix of time-state set nodes 12) Get the Dynamic preference representation of a time set 13) endfor 14) for j=1 to do 15) user Preference representation fusion 16) endfor 17) Predict the next temporal set 18) Return Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make various modifications and alterations without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. A method for predicting user preferences while protecting sensitive behavioral data, characterized in that, The method includes the following steps: Step 1: Treat the categories of user's historical interactions as sensitive user behavior data, and add a perturbation mechanism locally to ensure that it avoids the risk of privacy leakage during uploading and use; Step 2: Using the similarity of subgraph nodes with time decay factors in the form of temporal sets, learn the user history processed in Step 1 to achieve dynamic preference representation learning and obtain dynamic representation; Step 3: Integrate dynamic and static representations to achieve user preference prediction.

2. The prediction method according to claim 1, characterized in that, Step 1 is: Given a random response algorithm and its domain Sum range With probability Preserve the true value, in terms of probability Randomly replace other values; perform random response algorithms locally on each user's sensitive behavior data. deal with; like In any adjacent sensitive behavior dataset and ( The output results obtained from ) satisfy ,but Satisfy localized differential privacy; In the formula It is the privacy budget, which indicates the level of privacy protection.

3. The prediction method according to claim 1, characterized in that, Step 2 is: S1: Embedding learning is performed on the nodes in the temporal subset graph to obtain node feature embeddings. Then, the cross product is used to obtain the similarity score between all temporal subset graphs in the temporal set sequence. The similarity between temporal subset graphs is optimized by adding a time decay factor to obtain the dynamic similarity score between nodes. The similarity scores of all nodes are aggregated to obtain the user's score. Relationship matrix of historical interactive items ; S2: Based on the obtained first A temporal set node relationship matrix Combined with the first a set of time states Graph convolutional networks are used to learn the node feature embeddings for each temporal subset of the graph. ; in Indicates the first The node feature matrix of the layer, User The The node relationship matrix of each temporal set, It is a degree matrix. It is the first The weight matrix, Indicates the activation function; S3: Perform self-attention processing on each temporal set in the user temporal set subgraph sequence to obtain the weight values ​​between nodes, and update the node embedding matrix in the temporal set. The node embedding vectors in the _ temporal set are stacked together to obtain the _ ... Embedment matrix of all nodes in a temporal set; S4: Construct an embedding matrix sequence based on the embedding matrix of each temporal set. Attention weights between embedding matrices are obtained through a self-attention mechanism; in These represent different time lengths; , , These are obtained from the embedding matrix, specifically the query matrix, key matrix, and value matrix. It is a learnable weight matrix; These are the attention weights between two different temporal sets; Based on the node relationship matrix The first and self-attention mechanism obtained The neighbor information of each temporal set is used, considering the dependencies between temporal sets, and the attention weights and temporal decay of each temporal set are combined to learn the information in the first temporal set. Global dynamic preference representation of a set of time states ; in It is the first Node feature embedding at time step Represent a function used to compute a given i-th... Embedding matrix of each temporal set and the Nonlinear transformation results of the temporal set relation matrix; Indicates the first The weighted sum of the feature embeddings of all previous temporal sets is obtained based on the attention weights; It is the decay coefficient of the weighted sum of historical states, which aims to emphasize the importance of the current temporal set by reducing the weight of historical states; It is the first Attention weights for each temporal set, representing the embedding matrix. The importance of.

4. The prediction method according to claim 3, characterized in that, S1 specifically refers to: First, the original features of the nodes are mapped to a high-dimensional embedding space through a multi-layer fully connected network; in It is the mapped feature embedding; and They are respectively of 3D attribute features and 1D category features; Then, the cross product is used to obtain the similarity score between all temporal set subgraphs in the temporal set sequence; in express Middle node Feature embedding, express Middle node Feature embedding, There are two nodes. and Similarity results between them; The similarity scores of all node pairs are combined into the first... The and the first Similarity matrix of temporal sets To ensure that the similarity scores are on the same order of magnitude, they are normalized. The similarity between temporal subset graphs is optimized by adding a time decay factor to obtain the dynamic similarity score between nodes; in It is a node and nodes In The similarity of the temporal sets at any given moment. Is and Time-of-flight similarity matrix Middle node and nodes Similarity score between them It is a node and nodes At any moment The number of co-occurrences of edges. It is the time decay coefficient; in It is the time difference between the current time and the reference time. This is the decay rate, representing the magnitude of decay per unit time. When The larger the value, the faster the influence of historical data will diminish. Finally, the similarity scores of all nodes are aggregated to obtain the user's score. Relationship matrix of historical interactive items .

5. The prediction method according to claim 3, characterized in that, S2 specifically refers to: Let the user temporal set graph sequence be... ,in Indicates user No. A set of tenses, Include Nodes Using the formula Learn the feature vector of each node in each temporal set subgraph ; In users Historical Interaction Item Relationship Matrix Extract the temporal subset graph for each user. The similarity scores between nodes constitute relation matrix ,Right now ; According to the obtained number A temporal set node relationship matrix Combined with the first a set of time states Graph convolutional networks are used to learn the node feature embeddings for each temporal subset of the graph. .

6. The prediction method according to claim 3, characterized in that, S3 specifically refers to: Let the input be the user Temporal set graph sequence The corresponding feature vector sequence ,in, , User The set of nodes in the t-th temporal set. It is each temporal set Feature matrix, Indicates the number of nodes; yes The node relationship matrix; Initialize node In the Feature embeddings of a temporal set: in It is a node In the Eigenvectors of a set of time states These are trainable parameter weights; Calculate neighbor nodes Query, key, and value matrix: in , , These are the three matrix parameters in the self-attention mechanism. , , It is to make nodes Feature vector mapping to The result obtained in 3D space; Calculate neighbor nodes Attention weights: Aggregate neighbor information and update nodes Feature vectors: in It is the first The number of nodes in a temporal set subgraph. It is the first Temporal central nodes The weighted sum of the feature vectors of the neighboring nodes; The first The node embedding vectors in the _ temporal set are stacked together to obtain the _ ... The embedding matrix of all nodes in a temporal set.

7. The prediction method according to claim 1, characterized in that, Step 3 is: The embedding matrix obtained by all users is treated as a static representation , will users The dynamic embedding matrix is ​​a dynamic representation These two representations are used as inputs and integrated into the gating information component to learn the user. Hidden state Through formula Learning through iterative updates: in It is a node Mapped to user Static representation The corresponding index function; It is an indicator vector consisting of 0s and 1s, where a value of 1 indicates that the corresponding node is in the user's history. middle; These are the trainable parameters of the update gate that control the static and dynamic representations; Further calculate the probability of all elements appearing in the next temporal subset; in , These are trainable parameters; Since predicting the next timescale can be viewed as a multi-label learning problem, the loss function using L2 regularization is: in These are the trainable parameters of the model. It is the number of training samples. It is a hyperparameter that controls L2 regularization. and It is the first of the reference truth values The first of the peacekeeping prediction values Each element in the user The probability of the next set appearing is determined by optimizing the proposed model through minimizing the equation until convergence; Finally, the top K items are selected as the first K items based on the Top-K criteria. A set of temporal parameters is recommended to the user to obtain the final result. .

Images