A Sequential Recommendation Method for Shared Account Scenarios

Abstract

The present invention discloses a sequential recommendation method for shared account scenarios, belonging to the technical field of sequential recommendation for shared accounts, and comprising the following steps: constructing a recommendation model based on a capsule graph convolutional network and deep subspace clustering for personalized recommendation, the recommendation model including a main capsule graph construction module, a capsule graph convolutional network, an account-level dynamic routing module, a sequential decoder based on deep subspace clustering, and a prediction module; the method includes the following steps: establishing a sequential interaction graph; further processing the sequential interaction graph into a main capsule graph; learning the vector representations of shared accounts and mixed sequences, and generating personalized recommendation results. The present invention can mine the unique preferences of different potential users in the shared account sequence and generate more accurate item recommendation results for shared accounts.

Description

Technical Field

[0001] The present invention belongs to the technical field of shared account sequence recommendation, and particularly relates to a sequence recommendation method for a shared account scenario. Background Art

[0002] In video platforms and e-commerce networks, users' historical interaction behaviors are usually recorded by the platform in chronological order. As one of the core functions of the platform, the recommendation system captures users' long-term behavior patterns by analyzing users' historical behavior sequences, and then provides personalized recommendation results to improve users' experience and the platform's activity. However, with the continuous in-depth research, researchers have found that users tend to share an account among family members or close friends. In this scenario, the historical interaction behaviors of multiple potential users are mixed and recorded in the same shared account sequence, bringing challenges to the current recommendation system. By using methods such as capsule networks, graph convolutional networks, and subspace clustering, the diverse interests of potential users in the shared account can be effectively modeled, thereby improving the accuracy of the recommendation system.

[0003] In the prior art, the π-net (a shared account sequence recommendation method based on recurrent neural networks) and the PSJNet method (a shared account sequence recommendation method based on hierarchical recurrent networks) learn the representation of the shared account through recurrent neural networks and clustering networks, but this method cannot analyze the relationship between potential users in the account and account behaviors in a fine-grained manner. The DA-GCN (a shared account sequence recommendation method based on graph neural networks) and the TiDA-GCN (a shared account sequence recommendation method based on graph neural networks and self-attention mechanism) capture the interests of potential users in the shared account through graph neural networks and multi-head self-attention mechanisms respectively, ignoring the excessive parameter overhead and training cost of the model. Generally speaking, the existing technical methods cannot distinguish the ownership of each interaction in the shared account mixed sequence and are limited by high computational complexity and parameter overhead. Summary of the Invention

[0004] To solve the above problems, the present invention proposes a sequence recommendation method for shared account scenarios, and constructs a recommendation model based on a capsule graph convolutional network and deep subspace clustering. Specifically, a lightweight capsule graph convolutional network is first proposed to identify the ownership of each interaction of potential users. In this part, a primary capsule graph is constructed to identify the interaction ownership of different potential users. A graph convolutional network with an attention mechanism is used to propagate messages on the graph, realizing a fine-grained distinction of the preferences of different potential users. Subsequently, an account-level dynamic routing mechanism is designed to merge the preferences of potential users and generate an account-level vector representation. In addition, a sequence decoder based on deep subspace clustering is proposed. This sequence decoder uses a low-rank subspace basis to simulate the interests of potential users and adopts a contrastive learning strategy to align the preferences of potential users in the sequence representation.

[0005] The technical solution of the present invention is as follows:

[0006] A sequence recommendation method for shared account scenarios constructs a recommendation model based on a capsule graph convolutional network and deep subspace clustering for personalized recommendation. The recommendation model includes a primary capsule graph construction module, a capsule graph convolutional network, an account-level dynamic routing module, a sequence decoder based on deep subspace clustering, and a prediction module; the method includes the following steps:

[0007] Step 1: Obtain the interaction relationship between the shared account and the items and the sequential dependence relationship between the items in the shared account scenario, and establish a sequential interaction graph based on these two relationships;

[0008] Step 2: Further process the sequential interaction graph into a primary capsule graph through the primary capsule graph construction module;

[0009] Step 3: Learn the vector representations of potential users in the shared account and the vector representation of the mixed sequence from the primary capsule graph through the capsule graph convolutional network;

[0010] Step 4: Aggregate the vector representations of potential users in the shared account into the vector representation of the shared account through the account-level dynamic routing module;

[0011] Step 5: Align the user interests with the vector representation of the mixed sequence through the sequence decoder based on deep subspace clustering;

[0012] Step 6: Generate personalized recommendation results based on the vector representation of the shared account and the vector representation of the mixed sequence.

[0013] Further, the specific process of step 1 is as follows:

[0014] Obtain the item set in the shared account scenario as , where represents the The number of projects is ; the set of shared accounts is , where is the th shared account, and the number of shared accounts is ; the set of hybrid sequences is which represents the hybrid sequence of the th shared account;

[0015] The established sequence interaction graph is , where represents the set of nodes, represents the set of edges, and each edge in the set represents a relationship between two different nodes in the sequence interaction graph; there are two types of relationships in the sequence interaction graph, namely, the interaction relationship between the shared account and the project and the sequence dependency relationship between the projects; the above two relationships are represented by the adjacency matrix as follows:

[0016] (1);

[0017] where represents the adjacency matrix storing the sequence dependency relationship between the projects; if the th project is the predecessor node of the th project in the hybrid sequence, then the element in the th row and th column of , otherwise, ; the th project corresponds to the th row, and the th project corresponds to the th column; represents the adjacency matrix containing the interaction relationship between the shared account and the project. If there is an interaction between the th shared account and the th project , then the element corresponding to the th row and th column in , otherwise, ; the th shared account corresponds to the th row, and the th project corresponds to the th column; is the transpose symbol.

[0018] Furthermore, in step 2, the main capsule graph construction module includes a linear attention module and a one-dimensional convolution module; the specific working process of the main capsule graph construction module is as follows:

[0019] Step 2.1: Convert the feature representation of items in the sequence interaction graph into the capsule representation of items through the linear attention module. The specific process is as follows:

[0020] (2);

[0021] Among them, is the capsule representation of the item in the 0th layer of the main capsule graph obtained after conversion. The number of layers of the main capsule graph corresponds to the number of layers of the sequence interaction graph; , are the embedding dimensions of the feature representation and the capsule representation respectively; , , are respectively The embedding representations of the query, key, and value initialized by the value of, is the feature representation of the item in the 0th layer of the sequence interaction graph obtained by random initialization; is the softmax function; is the dimension transformation matrix of the linear attention module; is the bias term of the linear attention module;

[0022] Step 2.2: Convert the feature representation of the shared account in the sequence interaction graph into the capsule representation of the shared account through the pointwise one-dimensional convolution module. The specific calculation process is as follows:

[0023] (3);

[0024] Among them, is the capsule representation of the shared account in the 0th layer of the main capsule graph obtained after conversion; is the feature representation of the shared account in the 0th layer of the sequence interaction graph obtained by random initialization; represents the one-dimensional convolution operation; represents the one-dimensional convolution kernel of the one-dimensional convolution module; represents the bias of the one-dimensional convolution module; is a hyperparameter that controls the number of potential users within the shared account;

[0025] Step 2.3: Assume that each shared account contains potential users, and divide the capsule representation of each shared account into the capsule representations of potential users. Specifically: the th shared account in the 0th layer of the main capsule graph obtained after conversion The capsule representation of is divided into the capsule representations of potential users , which is the th potential user in the th shared account in the 0th layer of the main capsule graph obtained after conversion; the construction of the main capsule graph is completed on the premise of retaining the relationships in the sequence interaction graph.

[0026] Furthermore, in step 3, the capsule graph convolutional network includes several graph convolutional layers, and each layer needs to perform a graph convolution operation through a graph convolutional network with an attention mechanism. The specific working process is as follows:

[0027] Step 3.1: Calculate the information obtained from the capsule representation of potential users during the th layer of graph convolution and update the capsule representation of potential users; the specific process is as follows:

[0028] Step 3.1.1: Through the graph convolutional network with attention, calculate the correlation between the capsule representation of potential users in the shared account and the capsule representation of items that have interacted with the shared account on the main capsule graph. The specific calculation process is as follows:

[0029] (4);

[0030] where is the correlation between and is the th potential user in the th shared account; is the th item that has interacted with the th shared account; is the exponential function with base e; represents the set of items that have interacted with the th shared account ;

[0031] Step 3.1.2: Calculate the information passed from the capsule representation of items to potential users during the th layer of graph convolution, specifically:

[0032] (5);

[0033] where is during the The capsule represents the information transmitted to ; represents the first learnable weight during the -th layer graph convolution; is for the -th layer graph convolution and is the -th item that has interacted with the -th shared account 's capsule representation; is the second learnable weight during the -th layer graph convolution; is for the -th layer graph convolution and 's correlation;

[0034] Meanwhile, calculate the self-connection information of as follows:

[0035] (6);

[0036] where is the self-connection information of during the -th layer graph convolution; is the third learnable weight during the -th layer graph convolution; is for the -th layer graph convolution and is the -th potential user in the -th shared account 's capsule representation;

[0037] Step 3.1.3. Update the capsule representation of as follows:

[0038] (7);

[0039] where is the updated capsule representation of after the -th layer graph convolution; represents the set of all items that have interacted with ;

[0040] Step 3.2. Calculate the information obtained from the capsule representation of the item during the -th layer graph convolution and update the capsule representation of the item; the specific calculation process is as follows:

[0041] (8);

[0042] (9);

[0043] Among them, represents the th potential user in the th shared account who has an interaction with the th project; th potential user; represents the message passed from the capsule representation at the th layer of graph convolution to the capsule representation at the th layer; th layer; represents the fourth learnable weight that controls the amount of information at the th layer of graph convolution; is the capsule representation at the th layer of graph convolution; th layer; represents the project adjacent to in the mixed sequence; represents the message passed from the capsule representation at the th layer of graph convolution from the th capsule representation to the th capsule representation; represents the fifth learnable weight that controls the amount of information at the th layer of graph convolution; is the capsule representation at the th layer of graph convolution; th layer;

[0044] Step 3.3. The update process of the capsule representation of the project is as follows:

[0045] (10);

[0046] Among them, is the updated th capsule representation after the th layer of graph convolution; represents the set of all potential users who have interacted with ; ; represents the set of projects adjacent to in all mixed sequences;

[0047] Step 3.4. Adopt a multi-layer aggregation protocol to aggregate the representations obtained by graph convolution on each layer to obtain the complete project vector representation and potential user vector representation. The specific calculation process is as follows:

[0048] (11);

[0049] (12);

[0050] Among them, is the complete vector representation; is the complete vector representation; is the total number of layers of graph convolution;

[0051] Step 3.5: Calculate the vector representation of the final required mixed sequence. The specific calculation process is as follows:

[0052] (13);

[0053] Among them, is the vector representation of the mixed sequence.

[0054] Furthermore, the specific process of step 4 is as follows:

[0055] Step 4.1: The account-level dynamic routing module iterates a total of times. The dynamic routing process of the th iteration is as follows:

[0056] (14);

[0057] Among them, represents the account-level capsule representation at the th iteration; represents the function for compressing routing information; represents the function for compressing routing information; is the coupling coefficient between the capsule representation of the th potential user and the account capsule representation at the th iteration;

[0058] Step 4.2: After completing one routing process, the coupling coefficient needs to be updated. The specific process is as follows:

[0059] (15);

[0060] Among them, is the coupling coefficient between the capsule representation of the th potential user and the account capsule representation at the th iteration; is the learnable weight matrix in the account-level dynamic routing module; is the element-wise inner product operation;

[0061] Step 4.3: After the iteration ends, the complete account capsule representation is obtained, that is, the account-level capsule representation at the th iteration ; ;

[0062] Step 4.4. Calculate the vector representation of the shared account. The specific calculation process is as follows:

[0063] (16);

[0064] Among them, is the vector representation of the shared account.

[0065] Furthermore, the specific process of Step 5 is as follows:

[0066] Step 5.1. Initialize the subspace basis to calculate the affinity between the sequence representation and the subspace basis; different subspaces in the subspace basis represent the user interests of different potential users under the same shared account; represents the th subspace; is the total number of subspaces; the specific calculation process of the affinity is as follows:

[0067] (17);

[0068] Among them, represents the vector representation of the th item in; represents the subspace affinity between and; represents the hyperparameter that controls the calculation smoothness;

[0069] Step 5.2. Use the contrastive learning method to align the vector representation of the mixed sequence enhanced by the affinity with the vector representation of the original mixed sequence; the specific process is as follows:

[0070] First, assign the subspace affinity to the item vector representation in the mixed sequence:

[0071] (18);

[0072] Among them, represents the vector representation of the th item after enhancement, represents the vector representation of the mixed sequence after enhancement;

[0073] Subsequently, calculate the contrastive learning loss between the vector representation of the enhanced mixed sequenceand the vector representation of the regularized original mixed sequence , and the specific calculation process of the contrastive learning loss is as follows:

[0074] (19);

[0075] Among them, and are respectively the th item and the regularized embedding vector representation of the th item in; represents the vector representation of the th item after enhancement; represents the temperature coefficient that controls the influence from negative example pairs to positive example pairs;

[0076] Finally, for all mixed sequences, the complete contrastive learning loss is expressed as :

[0077] (20);

[0078] Step 5.3, use and to refine by means of a residual link; the vector representation of the refined complete mixed sequence is obtained through the following calculation process:

[0079] (21);

[0080] Among them, represents the vector representation of the refined complete mixed sequence; is the standard L2 regularization function; represents the learnable weight matrix in the residual link method.

[0081] Furthermore, the specific process of step 6 is as follows:

[0082] Step 6.1, input and into the prediction module for prediction, and the specific process is as follows:

[0083] (22);

[0084] Among them, represents the predicted probability distribution, is the th item; is the transformation matrix in the prediction module that maps the predicted value to the dimension of the candidate item set; is the bias term that adjusts the activation function threshold in the prediction module;

[0085] Sort all items in descending order according to the size of the predicted probability distribution, and the first This item is the personalized recommendation result for the shared account, where is a pre-set threshold;

[0086] Step 6.2: Optimize the learnable parameters of the prediction model using cross-entropy loss. The calculation process of the cross-entropy loss is as follows:

[0087] (23);

[0088] Step 6.3: Add the contrastive learning loss in Step 5.2 , and the complete loss function of the prediction model is :

[0089] (24);

[0090] where, is a hyperparameter that controls the participation degree of the self-supervised signal in the contrastive learning loss.

[0091] The beneficial technical effects brought by the present invention: By constructing the main capsule graph and designing a graph convolutional network with an attention mechanism, the interaction ownership of different potential users in the shared account is identified, realizing the fine-grained distinction of the preferences of different potential users; by designing a sequence decoder for contrastive learning enhanced deep subspace clustering, the diverse interests of potential users are simulated, and the sequence representation is aligned with the preferences of potential users. The capsule graph convolutional network enhanced by the present invention using deep subspace clustering can simultaneously model the potential users in the shared account in a fine-grained manner and effectively consider the diverse interests of potential users in the process of sequence representation learning, improving the accuracy of sequence modeling, and thus recommending more accurate items for the shared account. The items recommended in the present invention include but are not limited to electronic goods, video programs, etc., so the application scenario of the present invention is wide. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 is a flowchart of the sequence recommendation method for the shared account scenario of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0093] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments:

[0094] The present invention takes the shared account hybrid sequence as the research object and the core goal is to mine the behavior preferences in the shared account hybrid sequence. The key technical problems to be solved include: First, it is difficult to capture the preferences of different potential users in the shared account in a fine-grained manner. Second, the diverse interests of potential users are not fully considered in the process of sequence representation learning. Solving these two aspects of problems can more accurately model the shared account hybrid sequence.

[0095] Therefore, the specific key problems to be solved by the present invention are as follows:

[0096] Key technical problem 1: It is difficult to capture the preferences of different potential users in a shared account in a fine-grained manner;

[0097] The present invention proposes a graph capsule convolutional network to identify the ownership of each interaction in the mixed sequence of a shared account by potential users. In this part, a capsule graph is constructed to represent different potential users and their interaction relationships with items (items include but are not limited to electronic goods, video programs, etc.) in a fine-grained manner. By propagating messages on the capsule graph through a graph convolutional method with an attention mechanism, this component realizes the fine-grained discrimination of the preferences of different potential users.

[0098] Key technical problem 2: The diverse interests of potential users are not fully considered during the sequence representation learning process;

[0099] The present invention designs a sequence representation alignment method based on deep subspace clustering. This part uses a low-rank subspace basis to simulate the interests of potential users and uses subspace affinity to refine the sequence representation. By applying a contrastive learning strategy, this part realizes the alignment of potential users with the diverse interests of potential users in the subspace.

[0100] As Figure 1 shown, the present invention proposes a sequence recommendation method for the shared account scenario, constructs a recommendation model based on a capsule graph convolutional network and deep subspace clustering, and performs personalized recommendations according to the interaction behavior information and serialized preference information from the mixed behavior sequence of the shared account; the recommendation model includes a main capsule graph construction module, a capsule graph convolutional network, an account-level dynamic routing module, a sequence decoder based on deep subspace clustering, and a prediction module; the method includes the following steps:

[0101] Step 1: Obtain the interaction relationship between the shared account and items and the sequence dependence relationship between items in the shared account scenario, and establish a sequence interaction graph according to the above two relationships; the specific process is as follows:

[0102] Obtain the item set in the shared account scenario as , where represents the th item, is the total number of items; the shared account set is , where is the th shared account, is the total number of shared accounts; the mixed sequence set is , where represents the mixed sequence of the th shared account;

[0103] The established sequence interaction graph is , where represents the set of nodes, represents the set of edges, and each edge in the set represents a relationship between two different nodes in the sequence interaction graph. There are two types of relationships in the sequence interaction graph, namely the interaction relationship between the shared account and the project and the sequence dependency relationship between projects. The above two relationships are represented by the adjacency matrix as follows:

[0104] (1);

[0105] Among them, represents the adjacency matrix storing the sequence dependency relationship between projects. If the th project is the predecessor node of the th project in the mixed sequence, then the element in the th row and th column of , otherwise, ; the th project corresponds to the th row, and the th project corresponds to the th column; represents the adjacency matrix containing the interaction relationship between the shared account and the project. If there is an interaction between the th shared account and the th project , then the element corresponding to the th row and th column of , otherwise, . The th shared account corresponds to the th row, and the th project corresponds to the th column. is the transpose symbol.

[0106] Step 2. On the basis of the sequence interaction graph, the sequence interaction graph is further processed into a primary capsule graph through the primary capsule graph construction module. The primary capsule graph construction module includes a linear attention module and a one-dimensional convolutional module; the specific working process of the primary capsule graph construction module is as follows:

[0107] Step 2.1. Convert the feature representation of the projects in the sequence interaction graph into the capsule representation of the projects through the linear attention module. The specific process is as follows:

[0108] (2);

[0109] Among them, is the capsule representation of the item at the 0th layer in the main capsule diagram obtained after conversion, and the number of layers of the main capsule diagram corresponds to the number of layers of the sequence interaction diagram; and are the embedding dimensions of the feature representation and the capsule representation respectively; and and are respectively the embedding representations of the query, key, and value initialized by the value of, is the feature representation of the item at the 0th layer in the sequence interaction diagram obtained by random initialization; is the softmax function; is the dimension transformation matrix of the linear attention module; is the bias term of the linear attention module.

[0110] Step 2.2. Convert the feature representation of the shared account in the sequence interaction diagram into the capsule representation of the shared account through the pointwise one-dimensional convolution module. The specific calculation process is as follows:

[0111] (3);

[0112] Among them, is the capsule representation of the shared account at the 0th layer in the main capsule diagram obtained after conversion; is the feature representation of the shared account at the 0th layer in the sequence interaction diagram obtained by random initialization; represents the one-dimensional convolution operation; represents the one-dimensional convolution kernel of the one-dimensional convolution module; represents the bias of the one-dimensional convolution module; is a hyperparameter that controls the number of potential users within the shared account.

[0113] Step 2.3. Assume that each shared account contains potential users. For example, where represents the th potential user in the th shared account. Divide the capsule representation of each shared account into capsule representations of potential users. For example, divide the capsule representation of the th shared account at the 0th layer in the main capsule diagram obtained after conversion into capsule representations of potential users , , For the th potential user in the th shared account in the 0th layer of the main capsule diagram obtained after conversion; the construction of the main capsule diagram is completed while preserving the relationships in the sequence interaction diagram. The capsule representation of;

[0114] Step 3. Learn the vector representations of potential users in the shared account and the vector representation of the mixed sequence from the main capsule diagram through the capsule graph convolutional network; the capsule graph convolutional network contains several graph convolutional layers, and each layer performs a graph convolution operation through the graph convolutional network with an attention mechanism. The specific working process is as follows:

[0115] Step 3.1. Calculate the information obtained from the capsule representation of potential users during the th layer of graph convolution and update the capsule representation of potential users. Taking as an example, the specific process is as follows:

[0116] Step 3.1.1. Through the graph convolutional network with attention, calculate the correlation between the capsule representation of potential users in the shared account and the capsule representation of items that have interacted with the shared account on the main capsule diagram. The specific calculation process is as follows:

[0117] (4);

[0118] Among them, is the and correlation between; is the th potential user in the th shared account; is the capsule representation of; is the th item that has interacted with the th shared account; is the capsule representation of; is the exponential function with base e; represents the set of items that have interacted with the th shared account.

[0119] Step 3.1.2. After obtaining the correlation information, calculate the information passed from the capsule representation of items to potential users during the th layer of graph convolution. Specifically:

[0120] (5);

[0121] Among them, is during the th layer of graph convolution; The capsule represents the information transmitted to ; represents the first learnable weight during the -layer graph convolution, which controls the amount of information transmitted from adjacent items; is the -th item that has interacted with the -th shared account during the -layer graph convolution The capsule representation of; is the second learnable weight during the -layer graph convolution, which controls the influence of the attention mechanism; is the -th layer graph convolution and The correlation between;

[0122] Meanwhile, calculate the self-connection information of The specific process is as follows:

[0123] (6);

[0124] Among them, is the self-connection information of during the -layer graph convolution; is the third learnable weight during the -layer graph convolution, which controls the amount of information retained from ; is the -th item in the -th shared account during the -th potential user The capsule representation of;

[0125] Step 3.1.3, update the capsule representation of The update process is as follows:

[0126] (7);

[0127] Among them, is the updated capsule representation of after the -layer graph convolution; represents the set of all items that have interacted with ;

[0128] Step 3.2, calculate the information obtained from the capsule representation of the item during the -layer graph convolution, and update the capsule representation of the item. Take the -th item For example, the specific calculation process is as follows:

[0129] (8);

[0130] (9);

[0131] Among them, represents the th potential user in the th shared account who has an interaction with the th project; represents the th message passed from the capsule representation at the th layer of graph convolution to the capsule representation at ; represents the fourth learnable weight that controls the amount of information at the th layer of graph convolution; represents the capsule representation at the th layer of graph convolution; represents the th capsule representation at the th layer of graph convolution; represents the item adjacent to in the mixed sequence; represents the th message passed from the capsule representation at the th layer of graph convolution to the capsule representation at ; represents the fifth learnable weight that controls the amount of information at the th layer of graph convolution; represents the capsule representation at the th layer of graph convolution.

[0132] Step 3.3. After obtaining the messages from adjacent potential users and items, the update process of the capsule representation of the item is as follows:

[0133] (10);

[0134] Among them, is the updated th capsule representation after the th layer of graph convolution; represents the set of all potential users who have interacted with ; represents the set of items adjacent to in all mixed sequences. represents the set of items adjacent to

[0135] Step 3.4: Adopt a multi-layer aggregation protocol to aggregate the representations obtained through graph convolution on each layer, thereby obtaining the complete project vector representation and potential user vector representation. The specific calculation process is as follows:

[0136] (11);

[0137] (12);

[0138] Among them, is the vector representation of the complete ; is the vector representation of the complete ; is the total number of layers of graph convolution.

[0139] Step 3.5: Obtain the vector representation of the final required mixed sequence by aggregating all the learned project vector representations. The specific calculation process is as follows:

[0140] (13);

[0141] Among them, is the vector representation of the mixed sequence.

[0142] Step 4: Aggregate the vector representations of potential users in the shared account into the vector representation of the shared account through the account-level dynamic routing module; the specific process is as follows:

[0143] Step 4.1: During the account-level dynamic routing process, it is necessary to calculate the connection strength between different potential user capsules and the shared account capsule, which is represented by a randomly initialized coupling coefficient. The account-level dynamic routing module performs a total of iterations, being a hyperparameter. The dynamic routing process of the th iteration is as follows:

[0144] (14);

[0145] Among them, represents the account-level capsule representation of at the th iteration; represents the function for compressing routing information; is the coupling coefficient between the capsule representation of the th potential user and the account capsule representation at the th iteration.

[0146] Step 4.2: After completing one routing process, it is necessary to update the coupling coefficient. The specific process is as follows:

[0147] (15);

[0148] Wherein, is the coupling coefficient between the capsule representation of the th potential user and the account capsule representation at the th iteration; is the learnable weight matrix in the account-level dynamic routing module; is the element-wise inner product operation.

[0149] Step 4.3. After the account-level dynamic routing, the model obtains the complete account capsule representation, that is, the account-level capsule representation at the th iteration .

[0150] Step 4.4. By aggregating all the learned account representations, calculate the vector representation of the shared account. The specific calculation process is as follows:

[0151] (16);

[0152] Wherein, is the vector representation of the shared account.

[0153] Step 5. Align the user interest with the vector representation of the mixed sequence through a sequence decoder based on deep subspace clustering; the specific process is as follows:

[0154] Step 5.1. Initialize the subspace basis to calculate the affinity between the sequence representation and the subspace basis; different subspaces in the subspace basis represent the user interests of different potential users under the same shared account; represents the th subspace; is the total number of subspaces. Taking the mixed sequence of the th shared account as an example, the vector representation of is

[0155] (17);

[0156] Wherein, represents the vector representation of the th item in ; represents the subspace affinity between and ;

[0157] Step 5.2: Align the vector representation of the mixed sequence enhanced by affinity with the vector representation of the original mixed sequence using the contrastive learning method. The specific process is as follows:

[0158] First, assign the subspace affinity to the item vector representation in the mixed sequence:

[0159] (18);

[0160] where, represents the vector representation of the th item after enhancement, represents the vector representation of the enhanced mixed sequence.

[0161] Subsequently, calculate the contrastive learning loss between the vector representation of the enhanced mixed sequence and the vector representation of the regularized original mixed sequence. The specific process of calculating the contrastive learning loss is as follows:

[0162] (19);

[0163] where, , are the regularized embedding vector representations of the th item and the th item in respectively; represents the vector representation of the th item after enhancement; represents a pair of positive examples in the process of calculating the contrastive learning loss; represents a pair of negative examples in the process of calculating the contrastive learning loss; represents the temperature coefficient that controls the influence from the negative example pair to the positive example pair.

[0164] Finally, for all mixed sequences, the complete contrastive learning loss is expressed as :

[0165] (20);

[0166] Step 5.3: Refine and using the residual connection method. The vector representation of the complete refined mixed sequence can be obtained through the following calculation process:

[0167] (21);

[0168] where, The vector representation of the refined complete hybrid sequence; is the standard L2 regularization function; represents the learnable weight matrix in the residual link method.

[0169] Step 6. Based on the learned vector representation of the shared account and the vector representation of the hybrid sequence, generate personalized recommendation results; the specific process is as follows:

[0170] Step 6.1. Input the learned vector representation of the shared account and the vector representation of the refined complete hybrid sequence into the prediction module for prediction, and the specific process is as follows:

[0171] (22);

[0172] where, represents the predicted probability distribution, is the th item; is the transformation matrix in the prediction module that maps the predicted value to the dimension of the candidate item set; is the bias term that adjusts the activation function threshold in the prediction module.

[0173] Sort all items in descending order according to the size of the predicted probability distribution, and the first items in the sorted item sequence are the personalized recommendation results for the shared account, where is a pre-set threshold.

[0174] Step 6.2. Optimize the learnable parameters of the prediction model using cross-entropy loss, and the calculation process of the cross-entropy loss is as follows:

[0175] (23);

[0176] Step 6.3. Add the contrastive learning loss used in Step 5.2, and the complete loss function of the prediction model is :

[0177] (24);

[0178] where, is a hyperparameter that controls the participation degree of the self-supervised signal from the contrastive learning loss.

[0179] The recommendation model based on capsule graph convolutional network and deep subspace clustering constructed by the present invention is abbreviated as the LightGC²N model. To prove the feasibility and superiority of the present invention, a series of comparative experiments were conducted. In terms of experimental design, the experimental datasets adopted by the present invention include four datasets: Hvideo-E, Hvideo-V, Hamazon-M, and Hamazon-B. All these datasets come from real shared account scenarios and record the mixed behavior sequences of shared accounts. For the four datasets, the ratio of the training set to the test set is set to 2:8. The specific experiments carried out are as follows:

[0180] Experiment 1: Evaluated the recommendation performance of the model LightGC²N of the present invention in terms of the MRR (Mean Reciprocal Rank) evaluation metric compared with nine competitive baseline methods. The specific baseline methods adopted are as follows: NCF (a recommendation method based on neural collaborative filtering network), LightGCN (a recommendation method based on lightweight graph convolutional network), HRNN (a sequential recommendation method based on recurrent neural network), NAIS (an attention-based sequential recommendation method), TGSRec (a sequential recommendation method based on time-aware collaborative filtering), π-Net (a shared account sequence recommendation method based on recurrent neural network), PSJ-Net (a shared account sequence recommendation method based on hierarchical recurrent network), DA-GCN (a shared account sequence recommendation method based on graph neural network), TiDA-GCN (a shared account sequence recommendation method based on graph neural network and self-attention mechanism).

[0181] The experimental results of different methods on four real shared account sequence recommendation datasets are shown in Table 1. Among them, the higher the value of MRR, the higher the model accuracy. It can be observed from Table 1 that the model LightGC²N of the present invention has achieved the best performance on all four datasets. The performance of LightGC²N is better than other state-of-the-art shared account sequence recommendation methods (π-Net, PSJNet, DA-GCN, DA-GCN). This observation indicates that it is of great significance to capture the fine-grained differences between potential users in shared accounts. At the same time, LightGC²N is better than other graph-based shared account sequence recommendation methods (i.e., DA-GCN and TiDA-GCN), which proves the advantage of the graph capsule attention network proposed by the present invention in modeling complex relationships in the shared account sequence recommendation scenario. In addition, LightGC²N has achieved the best performance on all datasets, which proves the superiority of the graph capsule convolutional network and subspace alignment method proposed by the present invention in the shared account sequence recommendation scenario.

[0182] Table 1 Experimental results of different methods on four real shared account sequence recommendation datasets (%)

[0183] 。

[0184] Experiment 2: The present invention also conducted ablation experiments to verify the performance of the model LightGC²N of the present invention. Specifically, by deleting the test module, an ablation experiment analysis was carried out on the model LightGC²N. The present invention designed two variant models to respectively verify the degree to which the capsule graph convolutional network module and the sequence decoder module based on deep subspace clustering improve the performance of LightGC²N. The symbol usage is shown in Table 2.

[0185] Table 2 Explanation of Variant Symbols in Ablation Experiments

[0186] 。

[0187] The same dataset and experimental metrics (i.e., MRR) as in Experiment 1 were used to compare their performance on four datasets. The experimental results are shown in Table 3:

[0188] Table 3 Experimental Results of Ablation Experiment Variants (%)

[0189] 。

[0190] As can be seen from Table 3, the LightGC²N model achieved the best results on four different datasets, verifying the effectiveness of the graph capsule attention network and the deep subspace sequence decoder. Specifically, LightGC²N performed better than , proving that the fine-grained distinction of different potential user preferences indeed improved the prediction performance of the model in the shared account sequence recommendation task. In addition, LightGC²N performed better than , proving the effectiveness of the deep subspace clustering module in aligning sequence representations with potential user preferences.

[0191] The above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions, or substitutions made by those skilled in the art within the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A sequence recommendation method for shared account scenarios, characterized in that: A recommendation model based on capsule graph convolutional network and deep subspace clustering is constructed for personalized recommendation. The recommendation model includes a main capsule graph construction module, a capsule graph convolutional network, an account-level dynamic routing module, a sequence decoder based on deep subspace clustering, and a prediction module. The method includes the following steps: Step 1: Obtain the interaction relationship between the shared account and the project and the sequence dependency relationship between the projects in the shared account scenario, and establish a sequence interaction diagram based on these two relationships; Step 2: The sequence interaction graph is further processed into a main capsule graph through the main capsule graph construction module; Step 3: Learn the vector representation of potential users in the shared account and the vector representation of the mixed sequence from the main capsule graph through the capsule graph convolutional network; Step 4: Aggregate the vector representations of potential users in the shared account into the vector representation of the shared account through the account-level dynamic routing module; Step 5: Align user interests with the vector representation of the mixed sequence through a sequence decoder based on deep subspace clustering; Step 6: Generate personalized recommendation results based on the vector representation of the shared account and the vector representation of the mixed sequence; The specific process of step 1 is as follows: Get the project collection in the shared account scenario as ,in, Indicates projects, is the total number of projects; the shared account set is ,in, For the Shared accounts, is the total number of shared accounts; the mixed sequence set is ,in, Indicates A mixed sequence of shared accounts; The sequence interaction diagram established is ,in, Represents a collection of nodes, Represents a set of edges, each edge in the set represents a relationship between two different nodes in the sequence interaction graph; the sequence interaction graph contains two types of relationships, namely the interaction relationship between shared accounts and projects and the sequence dependency relationship between projects; the above two relationships are represented by the adjacency matrix express: (1)； in, represents the adjacency matrix that stores the sequential dependencies between items; if Projects In the mixed sequence, Projects The predecessor node of Middle Row, No. Elements of a column ,otherwise, ;No. Items corresponding to Row, No. Items corresponding to List; represents the adjacency matrix containing the interaction relationship between shared accounts and projects. If Shared accounts With Projects There is interaction between them, then Middle Row, No. The element corresponding to the column ,otherwise, ;No. The shared account corresponds to Row, No. Items corresponding to List; is the transpose symbol; In step 2, the main capsule graph construction module includes a linear attention module and a one-dimensional convolution module; the specific working process of the main capsule graph construction module is: Step 2.1: Convert the feature representation of the items in the sequence interaction graph into the capsule representation of the items through the linear attention module. The specific process is as follows: (2)； in, The capsule representation of the item at layer 0 in the main capsule graph obtained after conversion. The number of layers in the main capsule graph corresponds to the number of layers in the sequence interaction graph. , They are the embedding dimensions of feature representation and capsule representation respectively; , , They are The embedded representation of the query, key, and value initialized by the value of It is the feature representation of the items in the 0th layer of the sequence interaction graph obtained by random initialization; is the softmax function; is the dimension transformation matrix of the linear attention module; is the bias term of the linear attention module; Step 2.2: Convert the feature representation of the shared account in the sequence interaction graph into the capsule representation of the shared account through a point-by-point one-dimensional convolution module. The specific calculation process is as follows: (3)； in, The capsule representation of the shared account at layer 0 in the main capsule graph obtained after conversion; is the feature representation of the shared account at layer 0 in the randomly initialized sequence interaction graph; Represents a one-dimensional convolution operation; Represents the one-dimensional convolution kernel of the one-dimensional convolution module; Represents the bias of the one-dimensional convolution module; is a hyperparameter that controls the number of potential users in a shared account; Step 2.3: Assume that each shared account contains potential users, and divide the capsule representation of each shared account into The capsule representation of a potential user is as follows: Shared accounts Capsules indicate Divide Capsule representation of potential users , is the first layer of the main capsule image after conversion. of shared accounts Potential users The capsule representation of the sequence interaction graph is completed under the premise of retaining the relationship in the sequence interaction graph. In step 3, the capsule graph convolution network contains several graph convolution layers, each of which needs to be subjected to a graph convolution operation through a graph convolution network with an attention mechanism. The specific working process is as follows: Step 3.1: Calculate the The information obtained by the capsule representation of potential users during layer graph convolution is used to update the capsule representation of potential users. The specific process is as follows: Step 3.1.

1. Calculate the correlation between the capsule representations of potential users in the shared account and the capsule representations of items that have interacted with the shared account on the main capsule graph through the graph convolutional network with attention. The specific calculation process is as follows: (4)； in, for and The correlation between For the of shared accounts Potential users The capsule indicates; For the The shared account has interacted Projects The capsule indicates; is an exponential function with base e; Indicates Shared accounts The collection of items that have interacted; Step 3.1.2, calculate the The capsule of the project during the layer graph convolution represents the information conveyed to the potential user, specifically: (5)； in, For the When layer graph convolution The capsule indicates the information conveyed; Indicated in The first learnable weight during layer graph convolution; For the When the layer graph is convolved with The shared account has interacted Projects The capsule indicates; For the The second learnable weight during layer graph convolution; For the When layer graph convolution and The correlation between At the same time, calculation The self-connection information is as follows: (6)； in, For the When layer graph convolution Self-connection information; For the The third learnable weight during layer graph convolution; For the When layer graph convolution of shared accounts Potential users The capsule indicates; Step 3.1.3: The capsule indicates that an update is being performed. The update process is as follows: (7)； in, To conduct the Updated after layer graph convolution The capsule indicates; Indicates all The collection of items that have been interacted with; Step 3.2: Calculate the The capsule representation of the project obtains information during the layer graph convolution and updates the capsule representation of the project; the specific calculation process is as follows: (8)； (9)； in, Indicates shared account with Projects There is interaction potential users; Indicated in When layer graph convolution The capsule indicates The capsules represent the messages delivered; Indicated in The fourth learnable weight that controls the amount of information during layer graph convolution; For the When layer graph convolution The capsule indicates; Indicates that in a mixed sequence Adjacent projects; Indicated in When layer graph convolution is performed from The capsule indicates The capsules represent the messages delivered; Indicated in The fifth learnable weight that controls the amount of information during layer graph convolution; For the When layer graph convolution The capsule indicates; Step 3.3: The capsule representation update process of the project is as follows: (10)； in, To conduct the Updated after layer graph convolution The capsule indicates; Indicates all Potential users who have interacted A collection of; Indicates that in all mixed sequences A collection of adjacent items; Step 3.4: Use a multi-layer aggregation protocol to aggregate the representations obtained through graph convolution on each layer to obtain a complete item vector representation and potential user vector representation; the specific calculation process is as follows: (11)； (12)； in, For complete The vector representation of ; For complete The vector representation of ; is the total number of layers of graph convolution; Step 3.5: Calculate the vector representation of the final required mixed sequence. The specific calculation process is as follows: (13)； in, is the vector representation of the mixed sequence; The specific process of step 4 is as follows: Step 4.1: Account-level dynamic routing module iterative calculation sequence The dynamic routing process of this iteration is as follows: (14)； in, Indicates At iteration Account-level capsule representation of ; Indicates the function of compressing routing information; For the The iteration The coupling coefficient between the capsule representation of each potential user and the capsule representation of the account; Step 4.2: After completing a routing process, the coupling coefficient needs to be updated. The specific process is as follows: (15)； in, For the The first iteration The coupling coefficient between the capsule representation of each potential user and the capsule representation of the account; It is the learnable weight matrix in the account-level dynamic routing module; It is the element inner product operation; Step 4.3: After the iteration, the complete account capsule representation is obtained, i.e. At iteration Account-level capsule representation ; Step 4.4: Calculate the vector representation of the shared account. The specific calculation process is as follows: (16)； in, is the vector representation of the shared account; The specific process of step 5 is as follows: Step 5.1: Initialize the subspace basis The affinity between the calculated sequence representation and the subspace basis is represented; different subspaces in the subspace basis represent the user interests of different potential users under the same shared account; Indicates subspaces; is the total number of subspaces; the specific calculation process of affinity is as follows: (17)； in, express Middle A vector representation of items; express and The subspace affinity between them; represents the hyperparameter that controls the smoothness of the calculation; Step 5.2: Use contrastive learning method to align the vector representation of the affinity-enhanced mixed sequence with the vector representation of the original mixed sequence; the specific process is: First, subspace affinities are assigned to the vector representations of items in the mixed sequence: (18)； in, Indicates the enhanced A vector representation of items, The vector representation of the enhanced mixed sequence; Then, the vector representation of the enhanced mixed sequence The vector representation of the original mixed sequence after regularization Compute the loss of contrastive learning between , the specific contrastive learning loss calculation process is as follows: (19)； in, , They are Middle Projects, The regularized embedding vector representation of each item; Indicates the enhanced A vector representation of each item; represents the temperature coefficient that controls the influence from negative pairs to positive pairs; Finally, for all mixed sequences, the complete contrastive learning loss is expressed as : (20)； Step 5.3: and The residual link method is used for refinement; the vector representation of the refined complete mixed sequence is obtained through the following calculation process: (21)； in, A vector representation of the complete mixed sequence after refinement; is the standard L2 regularization function; Represents the learnable weight matrix in the residual link method; The specific process of step 6 is as follows: Step 6.1: and Input to the prediction module for prediction. The specific process is as follows: (22)； in, represents the predicted probability distribution, For the projects; It is the transformation matrix in the prediction module that maps the predicted value to the dimension of the candidate item set; is the bias term for adjusting the threshold of the activation function in the prediction module; Sort all items in descending order according to the size of the predicted probability distribution. The items are the personalized recommendation results for the shared account, is a preset threshold; Step 6.2: Use cross entropy loss to optimize the learnable parameters of the prediction model. The calculation process is: (23)； Step 6.3, add the contrastive learning loss in step 5.2 , the complete loss function of the prediction model is : (24)； in, is a hyperparameter that controls the degree of involvement of the self-supervision signal from the contrastive learning loss.

Images