Multi-granularity interest-based multi-behavior sequence recommendation method and application thereof

Abstract

The invention relates to a multi-granularity interest-based multi-behavior sequence recommendation method and application thereof, and the method comprises the steps: collecting multi-behavior historical interaction data of a user, forming a historical article interaction sequence, constructing a full connection graph, and extracting high-order multi-behavior dependency representation; user global behavior representation is extracted based on a linear multi-head self-attention mechanism, a sequence of multi-behavior dependency representation is divided into sub-sessions, multi-granularity user interest representation of all the sub-sessions is extracted based on a multi-granularity multi-head self-attention mechanism, a result is input into a nonlinear activation layer for representation fusion, and a final embedded representation is obtained; and based on the final embedding representation and the candidate article embedding vector, obtaining a predicted score value of the user on the article at the position, and recommending the predicted score value. The method is applied to user recommendation systems. According to the method, the problem of data scarcity of a recommendation system is solved, deep user interest expressions and behavior modes are mined, the limitation that existing multi-behavior dependency representation is insufficient is solved, the calculation complexity is reduced, the user interest expressions are enriched, and accurate personalized recommendation services are provided.

Description

Technical Field

[0001] The present invention relates to the technical field of calculation, inference or counting, and in particular to a multi-behavior sequence recommendation method based on multi-granularity interests in the field of deep learning and its application. Background Art

[0002] The recommendation system uses e-commerce websites to provide customers with product information and suggestions, help users decide what products to buy, and simulate sales staff to help customers complete the purchase process. It can automatically provide users with personalized recommendations, help users alleviate the problem of information overload, and also bring greater profit margins to companies. With the iterative update of algorithms and the development of e-commerce, the recommendation system has become an indispensable part of many personalized service fields such as e-commerce, fund recommendations, and short video recommendations.

[0003] As the core of the recommendation system, the recommendation algorithm plays a vital role in generating recommendation lists. In recent years, a number of sequential recommendation methods based on user historical behavior have emerged. These methods predict the items that users are interested in at the next moment by mining the personalized interest expressions of users hidden in historical behaviors, giving the recommendation system the advantages of real-time and personalization.

[0004] However, the sequence recommendation methods in the prior art often have the following defects:

[0005] (1) Only a single user behavior is considered, such as payment. From the perspective of model training, this results in a lack of data resources for training sequential recommendation methods, which leads to problems such as underfitting in deep learning-based sequential recommendation methods. Furthermore, the user's interest expression is actually reflected in a deeper level in multiple historical interactions, including but not limited to clicks, favorites, and adding to shopping carts, which cannot be expressed by recommendation methods that only consider a single payment behavior.

[0006] (2) The expression of user interests is incomplete and insufficient. This is mainly reflected in the fact that these methods only capture user interests at a single granularity. For example, based on the user's payment behavior of purchasing a pair of basketball shoes, the recommendation method can only infer the user's interest in basketball shoes, which is a single-granular expression of user interest. However, by combining multiple historical interactions, if the user's purchase of basketball shoes is combined with the user's purchase of related products such as basketball shirts, the recommendation method can infer the user's interest in the sport of basketball. Based on this, related products such as basketball shorts can also be recommended. Existing methods cannot reflect the multi-granular characteristics of user interests.

[0007] In summary, existing sequential recommendation algorithms only consider a single type of user historical interactions and capture user interest expressions at a single granularity, which makes it difficult for sequential recommendation methods to provide accurate personalized recommendation services. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention provides a multi-behavior sequence recommendation method based on multi-granularity interests and its application.

[0009] The technical solution adopted by the present invention is a multi-behavior sequence recommendation method based on multi-granularity interests, the method comprising the following steps:

[0010] S1 collects multi-behavior historical interaction data of a preset number of users to form a corresponding historical item interaction sequence;

[0011] S2 creates initial embedding vectors of items and behaviors based on the historical item interaction sequence, which are item representation vectors and behavior representation vectors;

[0012] S3 constructs a fully connected graph G based on the interaction relationship of the historical item interaction sequence;

[0013] S4 performs convolution processing on the fully connected graph G using a graph convolutional neural network to extract high-order multi-behavior dependency representations;

[0014] S5 extracts the user global behavior representation hidden in the multi-behavior dependency representation based on a linear multi-head self-attention mechanism;

[0015] S6 divides the sequence of the multi-behavior dependency representations into a plurality of sub-sessions, and extracts multi-granular user interest representations in all sub-sessions based on a multi-granular multi-head self-attention mechanism;

[0016] S7 inputs the results obtained in S5 and S6 into a nonlinear activation layer for representation fusion to obtain the final embedded representation of the historical item interaction sequence;

[0017] S8 multiplies the final embedding representation corresponding to any position after the fusion of the representation in S7 by the transpose of the embedding vector of the candidate item to obtain the predicted score of the user for the item at that position, and selects several corresponding items from high to low according to the predicted score for recommendation.

[0018] Preferably, in S1, the multi-behavior historical interaction data includes a user ID, an item ID, and interaction behavior data between the user and the item.

[0019] Preferably, in S2, creating initial embedding vectors of items and behaviors includes the following steps:

[0020] S2.1 obtain the historical item interaction sequence, and select a maximum sequence length N so that the length of all sequences n≤N;

[0021] S2.2 Based on the learnable item embedding matrix and the behavior embedding matrix Embed the items and behaviors involved in the sequence separately to obtain the item representation vector and behavior representation vector Among them, k i , k b are the number of items and behavior types respectively, d is the dimension of the vector, and a [pad] vector representation is created to fill the missing interaction representation; here, the first dimension k in the item embedding matrix i The “+1” of +1 is the vector representation of [pad]. When the number of historical interactions of a user is less than N, the gap is represented by item [pad] and behavior [pad];

[0022] S2.3 The historical item interaction sequence is obtained by adding the item sequence V and the behavior sequence B accordingly.

[0023] Preferably, in S3, a fully connected graph G based on the historical item interaction sequence H is constructed, and the product of the item semantic similarity and the behavior semantic similarity of the connected interactions is calculated to obtain an adjacency matrix A.

[0024] Preferably, in S4, the adjacency matrix A of the fully connected graph G obtained in S3 is added to the identity matrix to obtain the adjacency matrix with self-loops added right After row and column normalization, graph convolution operations are performed to obtain high-order multi-behavior dependency representations.

[0025] Preferably, S5 comprises the following steps:

[0026] S5.1 injecting sequence information into the high-order multi-behavior dependency representation to obtain an updated sequence;

[0027] S5.2 inputs the updated sequence into an improved linear multi-head self-attention mechanism, wherein the improved linear multi-head self-attention mechanism obtains the product of the key matrix and the value matrix and performs regularization operations on the query matrix and the key matrix;

[0028] The improved linear multi-head self-attention mechanism here ensures equivalence with the original attention mechanism compared to the traditional multi-head self-attention mechanism;

[0029] S5.3 obtains the expression of the user's global sequence pattern.

[0030] Preferably, S6 comprises the following steps:

[0031] S6.1 Representing higher-order multi-behavior dependencies Divided into o sessions {S1, ..., S o}, o is a hyperparameter;

[0032] S6.2 For each session Si , based on the idea of ​​Deep Set network and permutation invariance, g query matrices are generated to represent the user's multi-granular interests;

[0033] S6.3 Steps of using multi-head self-attention mechanism to output session S i The representation of each historical interaction in the multi-granular query matrix is ​​added to the average value of the session sequence information representation to obtain the multi-granular interest representation of the session.

[0034] S6.4 Concatenate the multi-granularity interest representations of all sessions to obtain the multi-granularity interest representation at time granularity o

[0035] Preferably, in S7, the user global behavior pattern obtained in S5 and the multi-granularity user interest features in all sub-sessions at different time granularities obtained in S6 are input into a nonlinear activation module for feature fusion to obtain the final embedded representation of the historical item interaction sequence; the different time granularities here include coarse time granularity and fine time granularity.

[0036] Preferably, in S8, the user's interest expression at position t in the sequence The score of each item at that position is taken as the product of the transposed embedding vectors of all candidate items, and the K items with the highest scores are recommended.

[0037] An application of the multi-behavior sequence recommendation method based on multi-granularity interests is applied to a user recommendation system.

[0038] The present invention relates to a multi-behavior sequence recommendation method based on multi-granularity interests and its application. The method collects multi-behavior historical interaction data of a preset number of users to form a corresponding historical item interaction sequence. Based on the historical item interaction sequence, initial embedding vectors of items and behaviors are created to be item representation vectors and behavior representation vectors. A fully connected graph G is constructed based on the interaction relationship of the historical item interaction sequence. A graph convolutional neural network is used to perform convolution processing on the fully connected graph G to extract high-order multi-behavior dependency representations. Based on a linear multi-head self-attention mechanism, a user global behavior representation hidden in the multi-behavior dependency representation is extracted and a sequence of the multi-behavior dependency representation is divided into a plurality of sub-sessions. Based on a multi-granularity multi-head self-attention mechanism, a multi-granularity user interest representation in all sub-sessions is extracted. The result is input into a nonlinear activation layer for representation fusion to obtain a final embedding representation of the historical item interaction sequence. The final embedding representation corresponding to any position after the representation fusion is multiplied by the transpose of the candidate item embedding vector to obtain the predicted score value of the user at the position for the item, and a plurality of corresponding items are selected from high to low according to the predicted score value for recommendation. The method is applied to a user recommendation system.

[0039] The beneficial effects of the present invention are:

[0040] (1) It uses multi-type interaction data to effectively solve the data scarcity problem of the recommendation system, and can also mine deep user interest expressions and behavior patterns from multi-type interaction data;

[0041] (2) We construct a fully connected graph at the interaction level, integrate item semantics and behavior semantics into the adjacency matrix of the fully connected graph, and use a graph convolutional neural network to extract multi-behavior dependency representations at the interaction level, which effectively solves the limitation of the existing methods that the multi-behavior dependency representations are insufficient.

[0042] (3) Based on the user's historical item interaction sequence, comprehensive sequence behavior patterns are extracted, and the global sequence behavior pattern is extracted using a linear self-attention mechanism, which greatly reduces the computational complexity. The multi-granularity multi-head self-attention mechanism is used to extract the multi-granularity interest expression of users in different time granularity sessions, enriching the expression of user interests. The nonlinear activation layer is used to fuse multi-dimensional user interest expressions and behavior patterns; ultimately, the recommendation method can provide accurate personalized recommendation services. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the method of the present invention;

[0044] Figure 2 is a schematic diagram of S2 in the method of the present invention;

[0045] Figure 3 is a schematic diagram of S3 in the method of the present invention;

[0046] Figure 4 is a schematic diagram of S4 in the method of the present invention;

[0047] Figure 5 is a schematic diagram of S5 in the method of the present invention;

[0048] Figure 6 is a schematic diagram of S6 in the method of the present invention;

[0049] Figure 7 Schematic diagram of S7 in the method of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific implementation described herein is only used to explain the present invention and is not intended to limit the present invention.

[0051] like Figure 1 As shown, the multi-behavior sequence recommendation method based on multi-granularity interest Transformer provided by the embodiment of the present invention includes:

[0052] S1 Data collection and dataset creation;

[0053] S2 creates initial embedding vectors for items and actions;

[0054] S3 builds a fully connected graph at the interaction level;

[0055] S4 extracts high-order multi-behavior dependencies at the interaction level;

[0056] S5 extracts the user's global sequence pattern representation;

[0057] S6 extracts multi-granular interest representation based on local sessions;

[0058] S7 represents fusion;

[0059] S8 rating prediction.

[0060] The purpose of the present invention is to solve the technical problem that the existing sequence recommendation methods only consider a single type of user historical interaction and capture a single granularity of user interests, resulting in inaccurate personalized recommendations, and to provide a multi-behavior sequence recommendation method based on a multi-granularity interest Transformer.

[0061] The following will explain in detail each part of the model architecture.

[0062] Create initial embedding vectors for items and actions:

[0063] Get the user's multi-type (multi-behavior) historical item interaction sequence output in step S1. The interaction behavior data here includes but is not limited to click, favorite, add to cart, purchase, etc. Select the maximum length of the sequence N so that the length of all sequences n≤N; Based on the learnable item embedding matrix and the behavior embedding matrix where k i , k b are the number of items and the number of behaviors, respectively. The embedding matrix also creates a [pad] vector representation to fill in the missing interaction representations. The items and behaviors involved in the sequence are embedded separately to obtain the item representation sequence and behavior representation sequence Historical Item Interaction Sequence It is obtained by adding the item sequence V and the behavior sequence B, and its formula is expressed as:

[0064] V={v1,……,v N} (1)

[0065] B={b1,……,b N} (2)

[0066]

[0067] H={I1,……,I N} (4)

[0068] Construct a fully connected graph at the interaction level:

[0069] Based on the historical interaction sequence H generated in step S2, a fully connected graph G at the historical interaction level is constructed. In order to reflect the multi-behavior dependency at the interaction level, the adjacency matrix of the fully connected graph G is Calculated by the following formula:

[0070] V i,j =v i ⊙v j (5)

[0071] B i,j =b i ⊙b j (6)

[0072]

[0073] The adjacency matrix calculated by this step integrates the item semantic dependency and behavior semantic dependency of the interaction pairs, so that the multi-behavior dependency at the interaction level is reflected in the adjacency matrix.

[0074] Extract high-order multi-behavior dependencies at the user interaction level:

[0075] First, add the adjacency matrix A obtained in step S3 to the identity matrix to obtain the adjacency matrix with self-loops added Ensure that its own information is not lost during the transmission process, then use formula (9) to Perform row and column standardization respectively. The left and right sides of are multiplied by the inverse matrix of the degree matrix, so as to determine that it aggregates the surrounding neighbor information in a weighted average manner during the information transmission process. Finally, the graph convolution operation is performed, and the formula is as follows:

[0076]

[0077] Among them, H (l) is the input of the graph convolutional neural network, where l represents the lth layer, W represents the weight matrix of each layer, and D is the degree matrix of the adjacency matrix A.

[0078] Extract user global sequence pattern representation:

[0079] First, construct the sequence information embedding matrix The sequence information of the corresponding position is injected into the multi-behavior dependency sequence representation obtained in step S4. Here, the sequence information refers to the position code p. The initialization method is consistent with the item embedding matrix and the behavior embedding matrix. The sequence information of the i-th position in the sequence is p i It is represented by the following definition formula:

[0080]

[0081] Then H (l) A linear multi-head self-attention mechanism is input to obtain the expression of the user's global sequence pattern. First, in order to make the linear attention mechanism equivalent to the dot product attention mechanism, L2 regularization operations are performed on the query matrix and the key matrix on the rows and columns respectively. At the same time, in order to ensure the stability of training, the activation function ELU is used to prevent extremely small negative numbers from appearing. The formula is as follows:

[0082]

[0083] Among them, ρ1 and ρ2 are L2 regularization operators, W Q , W K , W V are the linear transformation matrices in the attention mechanism, which are used to map the input to different spaces; i and j represent the regularization of each row and column of the matrix, respectively, and d is the dimension of sequence representation.

[0084] Extracting multi-granular interest representations based on local sessions:

[0085] Express the multi-behavior dependencies in step S4 Divide into o sessions For each session S i , capturing the multi-granularity interest representation hidden in the conversation through a multi-granularity multi-head self-attention mechanism;

[0086] Different from the original query matrix, which only has one spatial transformation W q , it is proposed here to use multiple W qg , which reflects the characteristics of multi-granularity; first, based on the idea of ​​Deep Set and permutation invariance, g query matrices are generated to represent the user's multi-granular interests, and the g query matrices are concatenated to obtain Then the steps of the multi-head self-attention mechanism are adopted: first, the spatial transformation matrices of the query matrix, key matrix and value matrix h heads are constructed respectively Then calculate the multi-head weight α of the key matrix on the multi-granularity query matrix hTo characterize the interest expression of each historical interaction in the session in the multi-granularity query matrix, and perform a 1d pooling operation (m represents the norm of 1d pooling) to make the weight smoother, and then multi-head weight α h Multiply it with the value matrix, concatenate the results of multiple attention heads, and add the average value of the session sequence information representation to obtain the multi-granular interest representation of the session. Finally, the interest representations of each session are spliced ​​to obtain multi-granularity interest expressions at time granularity o

[0087] The formula is defined as follows:

[0088]

[0089] Representation Fusion:

[0090] The obtained global behavior pattern is expressed as H Lin and multi-granularity interest expression at different time granularities Input nonlinear activation module to perform feature fusion The formula is as follows:

[0091]

[0092] Among them, W d To represent the fusion weight matrix;

[0093] Here, different time granularities refer to debugging at coarse granularity and fine granularity for each dataset.

[0094] Rating prediction:

[0095] The user at position t expresses his interest The product of the embeddings of all items is used as the score of each item at that position, and the K items with the highest scores are recommended. The score for position t in the sequence is as follows:

[0096]

[0097] Model training:

[0098] The model proposed in the embodiment of the present invention adopts the masked language model (MLM) as its training strategy, and its core idea is:

[0099] 1. Randomly mask some interactions in the sequence and replace them with a special token ([mask]);

[0100] 2. Use the context of the masked interaction in the sequence to predict the interaction;

[0101] 3. Finally, the model parameters are optimized by continuously narrowing the gap between the predicted results and the original results.

[0102] The embodiment of the present invention uses a cross entropy loss function to calculate the loss in the model training process. The specific formula is as follows:

[0103]

[0104] where p i,t is the true probability that item i interacts with the user at position t. If an interaction p occurs i,t =1, otherwise p i,t =0.

[0105] The present invention also relates to an application of a multi-behavior sequence recommendation method based on multi-granularity interests, which is applied to a user recommendation system.

[0106] To prove the effectiveness of the embodiments of the present invention, experiments were conducted on two international public datasets, Taobao and IJCAI CONTEST. The experiments used HR@K, NDGC@K, and MRR to evaluate the experimental results, and compared the method proposed by the present invention with the existing high-level methods. The experimental results are shown in Table 1.

[0107] Table 1 Comparison of experimental results between the method of the present invention and the existing method

[0108]

[0109] It can be concluded from Table 1 that the method proposed in the present invention has significant improvements over the existing optimal method on the Taobao dataset and the IJCAI dataset, with improvements of 15.3%, 16.5%, 14.4%, 14.6%, 13.5% and 9.0%, 8.8%, 10.4%, 9.4%, 4.8% on HR@5, NDGC@5, HR@10, NDGC@10, and MRR, respectively.

[0110] The application process of the present invention also involves a computer-readable storage medium, on which a multi-behavior sequence recommendation program based on multi-granularity interests is stored. When the program is executed by a processor, the multi-behavior sequence recommendation method based on multi-granularity interests is implemented; it also involves a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the multi-behavior sequence recommendation method based on multi-granularity interests is implemented.

[0111] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0112] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0113] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0114] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0115] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0116] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A multi-behavior sequence recommendation method based on multi-granularity interests, characterized by: The method comprises the following steps: S1 collects multi-behavior historical interaction data of a preset number of users to form a corresponding historical item interaction sequence; S2 creates initial embedding vectors of items and behaviors based on the historical item interaction sequence, which are item representation vectors and behavior representation vectors; S3 constructs a fully connected graph G based on the interaction relationship of the historical item interaction sequence; S4 performs convolution processing on the fully connected graph G using a graph convolutional neural network to extract high-order multi-behavior dependency representations; S5 extracts the user global behavior representation hidden in the multi-behavior dependency representation based on a linear multi-head self-attention mechanism; S6 divides the sequence of the multi-behavior dependency representations into a plurality of sub-sessions, and extracts multi-granular user interest representations in all sub-sessions based on a multi-granular multi-head self-attention mechanism; S7 inputs the results obtained in S5 and S6 into a nonlinear activation layer for representation fusion to obtain the final embedded representation of the historical item interaction sequence; S8 multiplies the final embedding representation corresponding to any position after the fusion of the representation in S7 by the transpose of the embedding vector of the candidate item to obtain the predicted score of the user for the item at that position, and selects several corresponding items from high to low according to the predicted score for recommendation.

2. The method for recommending multiple behavior sequences based on multi-granularity interests according to claim 1, characterized in that: In S1, the multi-behavior historical interaction data includes a user ID, an item ID, and interaction behavior data between the user and the item.

3. The method for recommending multiple behavior sequences based on multi-granularity interests according to claim 1, characterized in that: In S2, creating the initial embedding vectors for items and actions includes the following steps: S2.1 obtain the historical item interaction sequence, and select a maximum sequence length N so that the length of all sequences n≤N; S2.2 Based on the learnable item embedding matrix and the behavior embedding matrix Embed the items and behaviors involved in the sequence separately to obtain the item representation vector and behavior representation vector Among them, k i , k b are the number of items and behavior types respectively, d is the dimension of the vector, and a [pad] vector representation is created to fill the missing interaction representation; S2.3 The historical item interaction sequence is obtained by adding the item sequence V and the behavior sequence B accordingly.

4. The method for recommending multiple behavior sequences based on multi-granularity interests according to claim 3, characterized in that: In S3, a fully connected graph G based on the historical item interaction sequence H is constructed, and the product of the item semantic similarity and the behavior semantic similarity of the connected interactions is calculated to obtain the adjacency matrix A.

5. The method for recommending multiple behavior sequences based on multi-granularity interests according to claim 1, characterized in that: In S4, the adjacency matrix A of the fully connected graph G obtained in S3 is added to the identity matrix to obtain the adjacency matrix with self-loops added. right After row and column normalization, graph convolution operations are performed to obtain high-order multi-behavior dependency representations.

6. The method for recommending multiple behavior sequences based on multi-granularity interests according to claim 1, characterized in that: S5 includes the following steps: S5.1 injecting sequence information into the high-order multi-behavior dependency representation to obtain an updated sequence; S5.2 inputs the updated sequence into an improved linear multi-head self-attention mechanism, wherein the improved linear multi-head self-attention mechanism obtains the product of the key matrix and the value matrix and performs regularization operations on the query matrix and the key matrix; S5.3 obtains the expression of the user's global sequence pattern.

7. The method for recommending multiple behavior sequences based on multi-granularity interests according to claim 1, characterized in that: S6 includes the following steps: S6.1 Representing higher-order multi-behavior dependencies Divided into o sessions {S1, ..., S " }, o is a hyperparameter; S6.2 For each session S i , based on the idea of ​​Deep Set network and permutation invariance, g query matrices are generated to represent the user's multi-granular interests; S6.3 Steps of using multi-head self-attention mechanism to output session S i The representation of each historical interaction in the multi-granular query matrix is ​​added to the average value of the session sequence information representation to obtain the multi-granular interest representation of the session. S6.4 Concatenate the multi-granularity interest representations of all sessions to obtain the multi-granularity interest representation at time granularity o 8. The method for recommending multiple behavior sequences based on multi-granularity interests according to claim 1, characterized in that: In S7, the user global behavior pattern obtained in S5 and the multi-granularity user interest features in all sub-sessions at different time granularities obtained in S6 are input into the nonlinear activation module for feature fusion to obtain the final embedded representation of the historical item interaction sequence.

9. The method for recommending multiple behavior sequences based on multi-granularity interests according to claim 1, characterized in that: In S8, the user at position t in the sequence expresses his interest The score of each item at that position is taken as the product of the transposed embedding vectors of all candidate items, and the K items with the highest scores are recommended.

10. An application of the multi-behavior sequence recommendation method based on multi-granularity interests according to any one of claims 1 to 9, characterized in that: Applied to user recommendation system.

Images