Variant-based hypergraph convolutional network recommendation method

Abstract

The invention discloses a recommendation method of a hypergraph convolutional network based on variants. The recommendation method comprises the steps of collecting data and processing the data; dividing all project interaction data of a user into a plurality of time windows according to time, and constructing a hypergraph; all the hypergraphs serve as input and are sequentially input into a variant hypergraph convolutional network model according to a time sequence, and multi-layer convolutional operation is carried out on initial static embedding of user nodes obtained according to the input hypergraphs to obtain dynamic embedding of long and short term preferences of the user nodes; aggregating all user nodes in each hyperedge in the plurality of hypergraphs to obtain popularity dynamic embedding of the project; generating user project interactive embedding by combining the dynamic embedding of long and short term preferences of the user node and the popularity dynamic embedding of the project; the attention mechanism is adopted for attention aggregation, the prediction layer introduces dynamic embedding and static embedding for interactive prediction according to output after attention mechanism aggregation, a recommendation result is obtained, and the recommendation performance is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer intelligence technology, and in particular to a recommendation method based on a variant hypergraph convolutional network. Background Art

[0002] According to a statistical survey on the development of China's internet, internet penetration exceeds 76%, making the internet an integral part of people's daily lives. With the rapid growth of internet users, the amount of online information is also exploding, often leading to information overload when users search online. Recommendation systems emerged in response to the sheer volume and diversity of information. With the explosive growth in data volume and the diversification of data types, traditional recommendation methods face numerous challenges, such as the curse of dimensionality, data sparsity, and the inability to effectively capture deep data features.

[0003] The data involved in recommendation systems, such as users, items, and the relationships between users and items, are naturally suitable for representation using graph structures. However, the commonly used user-item bipartite graph cannot represent the temporal nature of interactions, and graph-based recommendation methods are inherently deficient in temporal representation. Temporal nature is crucial in the field of recommendation systems, as it depends on two aspects: the popularity of items changes over time, and the interests and preferences of users also change over time. Therefore, considering temporal nature in graph-based methods can significantly improve the effectiveness of recommendations. Existing technologies and research, such as TGAT, use timestamps and node attributes to dynamically generate time-aware embeddings, but this only captures local changes in temporal information and has difficulty modeling dynamics over long time spans. TGNN combines GCN and GRU modules to capture the dual dynamics of time series and graph structure, but this weakens the advantages of graph models. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides a recommendation method based on a variant hypergraph convolutional network, which adopts a variant hypergraph convolutional network to better model the popularity of items and the long-term and short-term preferences of users, thereby achieving better recommendation effects.

[0005] Technical solution: To achieve the above-mentioned purpose, the present invention provides a recommendation method based on a variant hypergraph convolutional network, comprising the following steps:

[0006] Step 1: Collect and process data. Divide all user-project interaction data into several time windows. Construct a hypergraph of all interaction data between projects and users within each time window. Construct a hyperedge for each interaction data between the same project and user in each time window.

[0007] Step 2: All hypergraphs are taken as input and fed into the variant hypergraph convolutional network model in chronological order. The static embedding layer obtains the initial static embedding of the user node based on the input hypergraph. The dynamic embedding layer of the user's long-term and short-term preferences obtains the dynamic embedding of the user node's long-term and short-term preferences through multi-layer convolution operations based on the initial static embedding.

[0008] Step 3: In the project popularity dynamic embedding layer, all user nodes on each hyperedge in several hypergraphs are aggregated to obtain the project popularity model, and then the project popularity dynamic embedding is obtained;

[0009] Step 4: The fusion layer combines the dynamic embedding of user node long-term and short-term preferences and the dynamic embedding of project popularity to generate user-project interaction embedding;

[0010] Step 5: Use the attention mechanism to aggregate the interaction sequences in the user-item interaction embedding; the prediction layer uses the output of the attention mechanism aggregation and introduces dynamic embedding and static embedding to perform interaction prediction to obtain the recommendation results.

[0011] Furthermore, in step 1, the user set is recorded as U = {u1, u2, ..., u N}, the item set is recorded as I = {i1, i2, ..., i P}; Divide all user project interaction data into several time windows T = {t1, t2, ..., t Q}, each time window builds a hypergraph Each hypergraph is expressed as:

[0012]

[0013] Where, is the set of user nodes in the hypergraph; is the set of hyperedges in the hypergraph; For size The diagonal matrix of represents the weights between hyperedges; For size The association matrix represents the connection relationship between user nodes and hyperedges.

[0014] Furthermore, when a user node In the time window t q Inside and a hyperedge When connected, On the contrary, when a user node In the time window t q Inside and a hyperedge When there is no connection, Set the number of hyperedges connected to user node u by matrix Indicates that the number of user nodes connected by hyperedges is represented by the matrix express;

[0015]

[0016] Where, is the number of hyperedges, is the number of user nodes.

[0017] Furthermore, all hypergraphs are taken as input and fed into the variant hypergraph convolutional network model in chronological order. First, the hyperedges connecting all items and user nodes are encoded through convolution operations, and then the hyperedge information connecting all items and user nodes is aggregated to obtain the dynamic embedding of each user node. The convolution operation of the first layer of convolution is represented by a matrix:

[0018]

[0019] Where, is the initial embedding of the first convolution user node u, for The transpose of P 0 is the trainable weight matrix between the initial static embedding and the first convolutional layer, τ(·) is the activation function;

[0020] Through the multi-layer hypergraph convolutional network, the information of high-order neighbors can be recursively aggregated. The output of the L-th layer convolution is expressed as:

[0021]

[0022] Where, is the initial embedding of the L-th convolution user node u, P (L-1) is the trainable weight matrix between the initial static embedding and the Lth convolutional layer;

[0023] At the same time, in order to prevent the numerical instability caused by the stacking of multiple convolutional layers, the L-th layer convolution is symmetrically normalized;

[0024]

[0025] Where, is a matrix to the power of -1 / 2, is a matrix to the power of -1.

[0026] Furthermore, since some preferences of the previous time period will be retained in the next time period, a gating unit is set to represent the initial embedding of the user node in each time window:

[0027]

[0028] Where W R and z R are respectively represented as the transformation matrix and vector of the gate control unit, g is the percentage value used to control the historical preference of the gate control unit, It is represented as the dynamic embedding of the user node in the previous hypergraph containing the user node u, e u is the static embedding of the user node, is the vector z R is the transpose of , and σ is the activation function.

[0029] Furthermore, if the user node u appears in the hypergraph for the first time, that is, there is no user node u in any previous hypergraph, then g = 0. e u is the corresponding user node u in the static embedding;

[0030] e u ∈E, E=[e1, e2, ..., e N ]

[0031] Where, e1, e2, ..., e N Represent user nodes u1, u2, ..., u N The corresponding initial static embedding.

[0032] Furthermore, all user nodes on each hyperedge in several hypergraphs are aggregated to obtain the popularity model of the project, and then the popularity dynamic embedding of the project is obtained. The calculation process of the popularity model of the project is as follows:

[0033]

[0034] Where, is a matrix to the power of -1 / 2, is a matrix to the power of -1 / 2, is the dynamic embedding of the output, P L is the trainable matrix of the Lth layer;

[0035] The resulting matrix In the time window t q The popularity of all items within is dynamically embedded.

[0036] Furthermore, the fusion layer combines the embedding of the long-term and short-term preferences of user nodes and the popularity embedding of items to generate the user-item interaction embedding; the calculation process is as follows:

[0037]

[0038] Where, e u and They are the static embedding of user nodes and the dynamic embedding of people, t q The popularity embedding of item i in the time window, W F and z represent the transformation matrix and vector of the fusion layer respectively.

[0039] Furthermore, the attention mechanism is used to aggregate the interaction sequences in the user-item interaction embedding. All the interaction data between a certain item i and the user node in each time window are represented by the interaction sequence L i express;

[0040]

[0041] Where, For project i and user node In the time window Interactions occurring within |L i | is the number of elements in the sequence;

[0042] For each interaction of item i, a positional encoding p is introduced k , where the position code p is introduced in the kth interaction k as follows:

[0043]

[0044] After introducing position encoding, the interaction sequence of item i is

[0045] Furthermore, based on the obtained interaction sequence, the attention score of the last interaction and each previous interaction is calculated; the calculation process is as follows:

[0046]

[0047] Where W Q and W K are all exchange matrices, d is the embedding size, For the last interaction, is the jth interaction in the interaction sequence, and att is the attention mechanism operation;

[0048] The result of attention aggregation represents the popularity embedding of item i after the last interaction. The calculation process of attention aggregation is as follows:

[0049]

[0050] Where W V is the interaction matrix of the attention mechanism.

[0051] Furthermore, the prediction layer is used to interactively predict the output of the attention mechanism after aggregation, and dynamic embedding and static embedding are introduced at the same time;

[0052]

[0053] Where, is the dynamic embedding of the user node in the previous hypergraph containing the user node u, e u is the initial static embedding, is the transpose of the attention-aggregated popularity embedding in the previous hypergraph containing item i.

[0054] Furthermore, the BPR loss function is used to optimize the variant-based hypergraph convolutional network model. The BPR loss function is calculated as follows:

[0055]

[0056] Where, ||θ|| 2 To use L2 regularization for the parameter θ, λ is used to control the degree of regularization, δ is the Sigmoid function, (i, t, u n1 ,u n2 ) is represented as the item i and user node u in the dataset C n1 An interaction occurs within time t, and item i interacts with user node u n2 No interaction occurs.

[0057] Beneficial effects: The present invention provides a recommendation method based on a variant hypergraph convolutional network, which divides all interactions between users and projects into several time windows, constructs a hypergraph in each time window according to the divided time windows, and embeds the hypergraph into the variant hypergraph convolutional network; this scheme can obtain deeper temporal information through the hypergraph of each time window and reduce the complexity of the model; a variant hypergraph convolutional network is designed and adopted, which takes into account the dual impact of time on users and projects, first dynamically embeds the long-term and short-term preferences of users, and then dynamically embeds the popularity of projects, while considering the impact of time on both users and projects, mining deeper temporal information, achieving better recommendation effects on data sets with more significant changes in project popularity, and improving recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A framework diagram of a proposed method based on a variant hypergraph convolutional network;

[0059] Figure 2 Schematic diagram of the dynamic embedding layer for users' long-term and short-term preferences. DETAILED DESCRIPTION

[0060] The present invention will be further described below with reference to the accompanying drawings.

[0061] like Figure 1 As shown, a proposed method based on a variant hypergraph convolutional network includes the following steps:

[0062] Step 1: Collect and process data. Divide all user-item interaction data into several time windows according to time. Construct a hypergraph from all interaction data between items and users in each time window. Construct the interaction data between the same item and user in each time window into a hyperedge. Therefore, a hypergraph contains several hyperedges, and the number of hyperedges is the number of items in the hypergraph.

[0063] Step 2: All hypergraphs are taken as input and input into the variant hypergraph convolutional network model in chronological order. The static embedding layer of the variant hypergraph convolutional network model obtains the initial static embedding of the user node based on the input hypergraph. The dynamic embedding layer of the user's long-term and short-term preferences of the variant hypergraph convolutional network model obtains the dynamic embedding of the user node's long-term and short-term preferences through multi-layer convolution operations based on the initial static embedding.

[0064] Step 3: In the dynamic embedding layer of the project popularity of the variant hypergraph convolutional network model, all user nodes on each hyperedge in several hypergraphs are aggregated to obtain project popularity modeling, and then the project popularity dynamic embedding is obtained. During the dynamic embedding, the user's long-term and short-term preferences are dynamically embedded first, and then the project popularity is dynamically embedded. This achieves better recommendation results on datasets with more significant changes in project popularity.

[0065] Step 4: The fusion layer of the variant hypergraph convolutional network model combines the dynamic embedding of user node long-term and short-term preferences and the dynamic embedding of project popularity to generate user-project interaction embedding;

[0066] Step 5: The variant hypergraph convolutional network model uses the attention mechanism to aggregate the interaction sequences in the user-item interaction embedding; the prediction layer of the variant hypergraph convolutional network model uses the output of the attention mechanism aggregation and introduces dynamic embedding and static embedding to perform interaction prediction to obtain the recommendation results.

[0067] In the step 1, the user set is recorded as U = {u1, u2, ..., u N}, N represents the total number of users; the item set is recorded as I = {i1, i2, ..., i P}, P represents the total number of projects; all user project interaction data are divided into several time windows T = {t1, t2, ..., t Q}, Q is the number of time windows, and there is no intersection between each time window; each time window constructs a hypergraph Each hypergraph is expressed as:

[0068]

[0069] Where, is the set of user nodes in the hypergraph, i.e., all users who interact with each other; is the set of hyperedges in the hypergraph, which approximates all items that interact; For size The diagonal matrix of represents the weights between hyperedges; For size The association matrix represents the connection relationship between user nodes and hyperedges.

[0070] When a user node In the time window t q Inside and a hyperedge When connected, That is On the contrary, when a user node In the time window t q Inside and a hyperedge When there is no connection, That is Set the number of hyperedges connected to user node u by matrix Indicates that the number of user nodes connected by hyperedges is represented by the matrix express;

[0071]

[0072] Where, is the number of hyperedges, is the number of user nodes.

[0073] like Figure 2 As shown in the figure, in step 2, all hypergraphs are taken as input and sequentially input into the variant hypergraph convolutional network model in chronological order. First, the hyperedges connecting all items and user nodes are encoded through convolution operations, and then the hyperedge information connecting all items and user nodes is aggregated to obtain the dynamic embedding of each user node. The convolution operation of the first layer of convolution is represented by a matrix:

[0074]

[0075] Where, is the initial embedding of the first convolution user node u, is a matrix The transpose of P 0 is the trainable weight matrix between the initial static embedding and the first convolutional layer, τ(·) is the activation function;

[0076] Through the multi-layer hypergraph convolutional network, the information of high-order neighbors can be recursively aggregated. The output of the L-th layer convolution is expressed as:

[0077]

[0078] Where, is the initial embedding of the L-th convolution user node u, P (L-1) is the trainable weight matrix between the initial static embedding and the Lth convolutional layer;

[0079] At the same time, in order to prevent the numerical instability caused by the stacking of multiple convolutional layers, the L-th layer convolution is symmetrically normalized;

[0080]

[0081] Where, is a matrix to the power of -1 / 2, is a matrix to the power of -1.

[0082] Although user preferences change over time, it is the short-term preferences that change dramatically over time. Relatively speaking, long-term preferences are not sensitive to time, so some preferences in the previous time period are retained in the next time period. Therefore, a gating unit is set to represent the initial embedding of user nodes in each time window:

[0083]

[0084] Where W R and z R are respectively represented as the transformation matrix and vector of the gate control unit, g is the percentage value used to control the historical preference of the gate control unit, It is represented as the dynamic embedding of the user node in the previous hypergraph containing the user node u, that is, the hypergraph containing the user node before the qth hypergraph but closest to it; e u is the static embedding of the user node, is the vector z R The transpose of , σ is the activation function; Denote as the input of the qth time window.

[0085] When the user node u appears in the hypergraph for the first time, that is, there is no user node u in any previous hypergraph, then g = 0, e u is the corresponding user node u in the static embedding;

[0086] e u ∈E, E=[e1, e2, ..., e N ]

[0087] Where, e1, e2, ..., e N Represent user nodes u1, u2, ..., u N The corresponding initial static embedding, E is the set of initial static embeddings of all user nodes.

[0088] The hypergraph is input into the variant hypergraph convolutional network model. In the first time window, the initial static embedding is input into the CHGCN module, and after l-layer convolution operation and normalization operation, the dynamic embedding of the user node is obtained. The dynamic embedding of the first user node is calculated with the initial static embedding to obtain the initial embedding of the first user node in the time window, and the initial embedding of the first user node is input into the second time window as the initial embedding. The CHGCN module in the second time window performs multi-layer convolution operation on the input first user node initial embedding to obtain the dynamic embedding of the second user node. The dynamic embedding of the second user node and the initial static embedding are calculated to obtain the initial embedding of the second user node in the time window as the initial embedding and input into the third time window; until the last time window Q, the dynamic embedding of the Q-1th user node obtained by the CHGCN module in the previous time window after multi-layer convolution operation is input. And dynamically embed the user node The initial embedding of the user node is calculated with the initial static embedding Input to the CHGCN module in the Qth time window and perform multi-layer convolution operations to obtain the dynamic embedding of the Qth user node The embedding of the user node in each time window is the dynamic embedding of the long-term and short-term preferences of the user node in each time window.

[0089] In step 3, the popularity of an item in a specific time period can be inferred from the users who interacted with it. All user nodes on each hyperedge in several hypergraphs are aggregated to obtain the item popularity model, and then the item popularity dynamic embedding is obtained. The calculation process of the item popularity model is as follows:

[0090]

[0091] Where, is a matrix to the power of -1 / 2, is a matrix to the power of -1 / 2, The dynamic embedding of the long-term and short-term preferences of user nodes calculated for each time window, P L is the trainable matrix of the Lth layer, and T represents the transpose operation of the matrix.

[0092] The resulting matrix In the time window tq The popularity of all items within is dynamically embedded.

[0093] In step 4, the fusion layer combines the embedding of the user node's long-term and short-term preferences with the item's popularity embedding to generate the user-item interaction embedding; the calculation process is as follows:

[0094]

[0095] Where, e u and are the static embedding of user nodes and the dynamic embedding of long-term and short-term preferences of user nodes, respectively. i and α d are the weights of the gated calculation, representing the weights of the user's long-term and short-term preferences and the popularity of the item respectively; t q The popularity embedding of item i in the time window, W F and z represent the transformation matrix and vector of the fusion layer respectively; z T The transpose of the vector representing the fusion layer.

[0096] In step 5, the attention mechanism is used to aggregate the interaction sequence in the user-item interaction embedding. All the interaction data between a certain item i and the user node in each time window are represented by the interaction sequence L i express;

[0097]

[0098] Where, For project i and user node In the time window Interactions occurring within |L i | is the number of elements in the sequence;

[0099] For each interaction of item i, a positional encoding p is introduced k , where the position code p is introduced in the kth interaction k as follows:

[0100]

[0101] Where, Represented as project i and user nodes In time interactive embedding, is the original sequence L i The embedding after mapping, where k represents the kth interaction;

[0102] After introducing position encoding, the interaction sequence of item i is

[0103] Based on the obtained interaction sequence, calculate the attention score of the last interaction and each previous interaction; the calculation process is as follows:

[0104]

[0105] Where W Q and W K are exchange matrices, d is the embedding size, For the last interaction, is the jth interaction in the interaction sequence, att is the attention mechanism operation, and the attention score is calculated; Expressed as The transpose of .

[0106] The result of attention aggregation represents the popularity embedding of item i after the last interaction. The calculation process of attention aggregation is as follows:

[0107]

[0108] Where, For project i at time Interaction sequence L i The popularity embedding of the user node after interaction, W V is the interaction matrix of the attention mechanism, |L i | is the number of elements in the interaction sequence.

[0109] In step 5, interactive prediction is performed on the output of the attention mechanism aggregation through the prediction layer, and dynamic embedding and static embedding are introduced at the same time;

[0110]

[0111] Where, is the dynamic embedding of the long-term and short-term preferences of the user node in the previous hypergraph containing the user node u, e u is the initial static embedding, is the transpose of the attention-aggregated popularity embedding in the previous hypergraph containing item i. The output of the attention mechanism is multiplied and fused with the sum of the dynamic embedding and the static embedding to obtain a final result, which is then input into the prediction layer for cross-prediction.

[0112] The variant-based hypergraph convolutional network model is optimized using the BPR loss function. The BPR loss function is calculated as follows:

[0113]

[0114] Where, ||θ|| 2L2 regularization is used for the parameter θ, θ is the set of all trainable parameters of the model; λ is used to control the degree of regularization parameter, δ is the Sigmoid function, (i, t, u n1 ,u n2 ) is represented as the item i and user node u in the dataset C n1 An interaction occurs within time t, and item i interacts with user node u n2 No interaction occurs; user node u n1 and user node u n2 are any two users in the set.

[0115] Example

[0116] like Figure 1-2 As shown in the figure, the initial static embedding of the user is generated using different models according to the specific downstream task and does not change after generation. All user-item interactions are divided into Q time windows by time. Each time window includes all interaction records within that time period. All interactions within a time window are constructed into a hypergraph, where all interactions of a certain item are constructed as a hyperedge, thus obtaining Q hypergraphs. These Q hypergraphs are respectively input into Q variant hypergraph convolutional network models. All hyperedges connected to the node are encoded through convolution operations, and then the information of all hyperedges connected to the point is aggregated to obtain the embedding of each node. The initial embedding of the node consists of two parts, one from the previous gating unit and the other from the initial static embedding. If the user node u appears for the first time, the embedding is the corresponding initial static embedding.

[0117] After several convolutional layers, Q user short-term preference embeddings are obtained, and these Q are fused to obtain the user's long-term preference embedding; all node embeddings on the hyperedge are aggregated to obtain the hyperedge embedding, which is the dynamic embedding of the project's popularity; in the fusion layer, the user's long-term and short-term preference embeddings and the project's dynamic embedding of popularity are combined to obtain the initial interaction embedding, which is input into the self-attention layer; after introducing the position encoding of the attention mechanism in the interaction embedding, the attention score is calculated, attention aggregation is performed, and the BPR loss function is used for model training and optimization. Finally, the trained model is used for prediction in the prediction layer.

[0118] Experiments were conducted on the Amazon dataset according to the above process, with NDCG@5 and MRR selected as evaluation indicators. The model of the present invention achieved approximately 3% and 7% improvement in recommendation performance compared to the HyperRec model, respectively.

[0119] The above is only a description of the preferred embodiment of the present invention. Ordinary technicians in this technical field can make several modifications and optimizations based on the above disclosure without departing from the above basic principles. These improvements and optimizations should be regarded as the scope of protection understood by the present invention.

Claims

1. A recommendation method based on a variant hypergraph convolutional network, characterized by: The following steps are involved: Step 1: Collect and process data; Divide all user-project interaction data into several time windows according to time, construct a hypergraph of all project-user interaction data in each time window, and construct a hyperedge for the interaction data between the same project and user in each time window; Step 2: All hypergraphs are taken as input and fed into the variant hypergraph convolutional network model in chronological order. The static embedding layer obtains the initial static embedding of the user node based on the input hypergraph. The dynamic embedding layer of the user's long-term and short-term preferences obtains the dynamic embedding of the user node's long-term and short-term preferences through multi-layer convolution operations based on the initial static embedding. Step 3: In the project popularity dynamic embedding layer, all user nodes on each hyperedge in several hypergraphs are aggregated to obtain the project popularity model, and then the project popularity dynamic embedding is obtained; Step 4: The fusion layer combines the dynamic embedding of user node long-term and short-term preferences and the dynamic embedding of project popularity to generate user-project interaction embedding; Step 5: Use the attention mechanism to aggregate the interaction sequences in the user-item interaction embedding; The prediction layer performs interactive prediction based on the output of the attention mechanism aggregation and introduces dynamic embedding and static embedding to obtain the recommendation results.

2. The recommendation method based on a variant hypergraph convolutional network according to claim 1, characterized in that: In the step 1, the user set is recorded as U = {u1, u2, ..., u N }, the item set is recorded as I = {i1, i2, ..., i P }; Divide all user project interaction data into several time windows T = {t1, t2, ..., t Q }, each time window builds a hypergraph Each hypergraph is expressed as: Where, is the set of user nodes in the hypergraph; is the set of hyperedges in the hypergraph; For size The diagonal matrix of represents the weights between hyperedges; For size The association matrix represents the connection relationship between user nodes and hyperedges.

3. The recommendation method based on a variant hypergraph convolutional network according to claim 2, characterized in that: When a user node In the time window t q Inside and a hyperedge When connected, On the contrary, when a user node In the time window t q Inside and a hyperedge When there is no connection, Set the number of hyperedges connected to user node u by matrix Indicates that the number of user nodes connected by hyperedges is represented by the matrix express; Where, is the number of hyperedges, is the number of user nodes.

4. The recommendation method based on a variant hypergraph convolutional network according to claim 1, characterized in that: All hypergraphs are taken as input and fed into the variant hypergraph convolutional network model in chronological order. First, the hyperedges connecting all items and user nodes are encoded through convolution operations. Then, the hyperedge information connecting all items and user nodes is aggregated to obtain the dynamic embedding of each user node. The convolution operation of the first layer of convolution is represented by a matrix: Where, is the initial embedding of the first convolution user node u, is a matrix The transpose of P 0 is the trainable weight matrix between the initial static embedding and the first convolutional layer, τ(·) is the activation function; Through the multi-layer hypergraph convolutional network, the information of high-order neighbors can be recursively aggregated. The output of the L-th layer convolution is expressed as: Where, is the initial embedding of the L-th convolution user node u, P (L-1) is the trainable weight matrix between the initial static embedding and the Lth convolutional layer; At the same time, in order to prevent the numerical instability caused by the stacking of multiple convolutional layers, the L-th layer convolution is symmetrically normalized; Where, is a matrix to the power of -1 / 2, is a matrix to the power of -1.

5. The recommendation method based on a variant hypergraph convolutional network according to claim 4, characterized in that: Since some preferences of the previous time period will be retained in the next time period, a gating unit is set to represent the initial embedding of the user node in each time window: Where W R and z R are respectively represented as the transformation matrix and vector of the gate control unit, g is the percentage value used to control the historical preference of the gate control unit, It is represented as the dynamic embedding of the user node in the previous hypergraph containing the user node u, e u is the static embedding of the user node, is the vector z R is the transpose of , and σ is the activation function.

6. The recommendation method based on a variant hypergraph convolutional network according to claim 5, characterized in that: This is the first time that user node u appears in the hypergraph, that is, there is no user node u in any previous hypergraph, then g = 0, e u is the correspondence of user node u in static embedding; the u ∈E,E=[e1,e2,...,e N ] Where, e1, e2, ..., e N Represent user nodes u1, u2, ..., u N The corresponding initial static embedding.

7. The recommendation method based on a variant hypergraph convolutional network according to claim 3, characterized in that: Aggregate all user nodes on each hyperedge in several hypergraphs to obtain the popularity model of the project, and then obtain the dynamic popularity embedding of the project. The calculation process of the popularity model of the project is as follows: Where, is a matrix to the power of -1 / 2, is a matrix to the power of -1 / 2, is the dynamic embedding of the output, P L is the trainable matrix of the Lth layer; The resulting matrix In the time window t q The popularity of all items within is dynamically embedded.

8. The recommendation method based on a variant hypergraph convolutional network according to claim 7, characterized in that: The fusion layer combines the embedding of the user node's long-term and short-term preferences with the item's popularity embedding to generate the user-item interaction embedding; the calculation process is as follows: Where, e u and They are the static embedding of user nodes and the dynamic embedding of people, t q The popularity embedding of item i in the time window, W F and z represent the transformation matrix and vector of the fusion layer respectively.

9. The recommendation method based on a variant hypergraph convolutional network according to claim 5, characterized in that: The attention mechanism is used to aggregate the interaction sequences in the user-item interaction embedding. All the interaction data between a certain item i and the user node in each time window are represented by the interaction sequence L i express; Where, For project i and user node In the time window interactions that occur within is the number of elements in the sequence; For each interaction of item i, a positional encoding p is introduced k , where the position code p is introduced in the kth interaction k as follows: After introducing position encoding, the interaction sequence of item i is 10. The recommendation method based on a variant hypergraph convolutional network according to claim 5, characterized in that: Based on the obtained interaction sequence, calculate the attention score of the last interaction and each previous interaction; The calculation process is as follows: Where W Q and W K are all exchange matrices, d is the embedding size, For the last interaction, is the jth interaction in the interaction sequence, and att is the attention mechanism operation; The result of attention aggregation represents the popularity embedding of item i after the last interaction. The calculation process of attention aggregation is as follows: Where W V is the interaction matrix of the attention mechanism.

11. The recommendation method based on a variant hypergraph convolutional network according to claim 5, characterized in that: Interactively predict the output of the attention mechanism aggregation through the prediction layer, while introducing dynamic embedding and static embedding; Where, is the dynamic embedding of the user node in the previous hypergraph containing the user node u, e u is the initial static embedding, is the transpose of the attention-aggregated popularity embedding in the previous hypergraph containing item i.

12. The recommendation method based on a variant hypergraph convolutional network according to claim 5, characterized in that: The variant-based hypergraph convolutional network model is optimized using the BPR loss function. The BPR loss function is calculated as follows: Where, ||θ|| 2 To use L2 regularization for the parameter θ, λ is used to control the degree of regularization, δ is the Sigmoid function, (i, t, u n1 ,u n2 ) is represented as the item i and user node u in the dataset C n1 An interaction occurs within time t, and item i interacts with user node u n2 No interaction occurs.

Images