User interest recommendation method and system fusing collaborative filtering and temporal awareness

Abstract

The application provides a user interest recommendation method and system fusing collaborative transformation and time awareness, constructs a user item bipartite graph, aggregates neighbor information, and enriches the user item representation in the sequence by using the idea of collaborative filtering before the sequential pattern; then, additional time information is used to capture the time conversion mode between the items in the sequence, the representation learning in the sequential pattern is strengthened by considering the change of the user interest over time; finally, the learned item representation is specifically activated to a certain interest of the user through a target interaction network, a specific representation is formed by using a target item, and recommendation prediction is performed after fusing different representations, so that the specific interest of the user can be strengthened, and the recommendation result is more accurate and diversified.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of recommendation, and particularly relates to a user interest recommendation method and system fusing collaborative transformation and time awareness. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] Traditional sequential recommendation systems use user behavior data sequences to model user interests to recommend items that may interest him. Existing technologies mainly fall into two categories: one class uses the sequential pattern under the historical behavior, uses all the interactive items of the user to capture the user's intention; the other class uses the graph structure, from the perspective of the user-item interaction graph, obtains the user embedding by aggregating the node information for recommendation. In terms of structure, the overall framework of the two methods is different. The first one only considers the relationship between the current item and other items in the same sequence, extracts the overall features of the sequence to obtain the user interest representation. The other class considers mixed heterogeneous information, models the user interest by aggregating different information, and combines the collaboration signal by using the high-order connectivity in the user-item interaction graph.

[0004] The method in the sequential mode only uses the interaction project of the user itself, ignores the collaborative relationship between users, and when the behavior data of the user is small, the cold start problem occurs. Some methods consider user embedding in sequence mode modeling, but they only obtain user (item) embedding by simply mapping user (item) ID, and cannot encode the collaborative signal in user-item interaction into embedding, cannot capture higher-order information, and are difficult to produce better recommendation effect; the method based on the graph encodes the high-order information, but ignores the sequential mode of the interaction behavior, and the model defaults that the items in all interaction sequences are equally important, and cannot distinguish the importance of the early interaction item and the recent item. In addition, the existing sequence recommendation method only uses the sequential mode to simulate the conversion between items, and does not consider the temporal conversion mode hidden in the time signal. These signals are hidden in the user-item interaction and coexist with the sequential mode. Therefore, the time signal needs to be explicitly modeled within the sequential modeling framework to explore the influence of the temporal conversion mode on predicting the next item. Finally, the candidate item is usually rich, and the user may have multiple interests, however, not all the interacted items are helpful to infer the interest degree of the user to the candidate item, and the target item of the candidate may only be related to the specific interest represented by part of the items in the historical sequence. The existing method fuses the sequence representation into a fixed vector representation without considering the target item to be predicted, and since the user may have multiple interests, the target item may only be related to the specific interest represented by part of the items in the historical interaction sequence, and the fixed vector representation will limit the representation ability of the model. SUMMARY

[0005] To overcome the above-mentioned deficiencies of the prior art, the present application provides a user interest recommendation method fusing collaborative transformation and temporal awareness, which explicitly captures the item conversion between different sequences in the form of high-order connectivity by using global graph context embedding propagation; the global static representation extracted after collaborative transformation is used as the initial embedding of the sequential mode, the user-item interaction information is integrated into the embedding representation, and the sequential mode is enhanced; then, different forms of temporal embedding are fused to capture the dynamic interest change of the user over time; finally, the correlation between the historical items and the candidate items is measured to select the corresponding historical items, thereby realizing accurate user interest modeling and greatly improving the expressiveness of the model.

[0006] To achieve the above object, one or more embodiments of the present application provide the following technical scheme: a user interest recommendation method fusing collaborative transformation and temporal awareness, comprising the following steps:

[0007] constructing a user-item interaction graph using historical user-item data interaction sequences;

[0008] inputting the user node vector and the item node vector into a multi-layer graph convolution model to perform graph convolution operation, and outputting a final node embedding vector, the final node embedding vector including a user final embedding vector and an item final embedding vector;

[0009] obtaining a user behavior sequence based on the user final embedding vector and the item final embedding vector;

[0010] obtaining an interaction time sequence of the user-item based on the historical user-item data interaction sequence, obtaining different interaction time sequence representations of the interaction time sequence through multiple different encoding modes, inputting the different interaction time sequence representations and the user behavior sequence into a multi-head attention model to capture the behavior pattern of the user over time, and outputting a predicted recommended item.

[0011] Further, the user node vector is input into the multi-layer graph convolution model to output an intermediate user embedding vector of each layer, the intermediate user embedding vectors output by each layer are added after being assigned the same weight as the initial user node vector, and a final user embedding vector is output.

[0012] The item node vector is input into the multi-layer graph convolution model to output an intermediate item embedding vector of each layer, the intermediate item embedding vectors output by each layer are added after being assigned the same weight as the initial item node vector, and a final item embedding vector is output.

[0013] Further, the calculation formula of the multi-layer graph convolution model is:

[0014]

[0015]

[0016] wherein, l represents the number of layers of graph convolution, N u represents a set of items interacted by the user u, N i represents a set of users interacting with the item i, is a symmetric regularization term, is a vector representation of the item i after l-layer graph convolution, represents a vector representation of the user u after l-layer graph convolution.

[0017] Further, the interaction time sequence of the user-item is obtained based on the historical user-item data interaction sequence, and different interaction time sequence representations of the interaction time sequence are obtained through multiple different encoding modes, wherein the multiple different encoding modes are specifically:

[0018] a position embedding matrix is defined to indicate the relative position of the item in the interaction time sequence;

[0019] Convert the date in each interaction time sequence into an embedding vector to express the user's behavior pattern in a day, and count the number of days appearing;

[0020] Define a time interval matrix to represent the interaction time difference of each item in the interaction time sequence, and convert the time interval matrix into an embedding vector

[0021] Capture the user's periodic behavior habits using the sine and cosine functions, denoted as

[0022] Further, the calculation formula of the multi-head attention model is:

[0023]

[0024]

[0025]

[0026]

[0027] Wherein, E S is a user behavior sequence, All are parameter matrices, h is the number of heads, is a scale factor, is a position bias, and f is a calculation function.

[0028] Further, after the different interaction time sequence representations and the user behavior sequence are fused through the multi-head attention model, the output result is subjected to nonlinear transformation through the feedforward neural network.

[0029] Further, the correlation score of the target item and each item in the user behavior sequence is calculated, and the user sequence corresponding to the target item is obtained based on the correlation score.

[0030] Based on the user sequence and the predicted recommended item output, the final item embedding representation of the to-be-predicted position in the user behavior sequence is represented;

[0031] The similarity score of the final item embedding representation and the target item representation is calculated for item recommendation.

[0032] The second aspect of the application provides a user interest recommendation system that fuses collaborative transformation and time awareness, comprising:

[0033] The first data acquisition module: using the historical user-item data interaction sequence to construct a user-item interaction graph;

[0034] The graph convolution module: based on the user node vector representation and the item node vector representation in the user-item interaction graph, input into the multi-layer graph convolution model for graph convolution operation, output the final node embedding vector, the final node embedding vector includes a user final embedding vector and an item final embedding vector;

[0035] The second data acquisition module: based on the user final embedding vector and the item final embedding vector, obtain a user behavior sequence;

[0036] The fusion module: based on the user-item interaction sequence, obtain an interaction time sequence of the user-item, pass the interaction time sequence through a plurality of different encoding modes to obtain different interaction time sequence representations, input the different interaction time sequence representations and the user behavior sequence into a multi-head attention model to capture the behavior pattern of the user over time, and output an item representation;

[0037] The recommendation module: recommend by calculating the relevance of the output item representation and each item in the interaction sequence.

[0038] The third aspect of the application provides a computer readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above method.

[0039] The fourth aspect of the application provides an electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, which, when executed by the processor, complete the steps of the above method.

[0040] The above one or more technical solutions have the following beneficial effects:

[0041] The application fully utilizes the sequential pattern and high-order heterogeneous collaborative signal of user-item interaction, aggregates neighbor information by constructing a user-item interaction graph, and enriches the user-item representation within the sequence by using the idea of collaborative filtering before the sequential pattern. Then, the temporal transition pattern between items within the sequence is captured by using additional time information, the representation learning under the sequential pattern is strengthened by considering the change of user interest over time, and finally the learned item representation is specifically activated by a target interaction network. The specific representation is formed by using the target item, and the recommendation prediction is performed after the fusion of different representations, which can strengthen the specific interest of the user and make the recommendation result more accurate and diversified.

[0042] The advantages of the additional aspects of the application will be partially given in the following description, partially will become obvious from the following description, or will be understood by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification. The embodiments of the application, and their

[0044] Figure 1 Figure 1 is a schematic diagram of the overall framework of the first embodiment of the present application. DETAILED DESCRIPTION

[0045] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0046] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the application.

[0047] In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0048] Embodiment one

[0049] As shown in the figure, the embodiment discloses a user interest recommendation method combining collaborative transformation and time perception, characterized by comprising the following steps: Figure 1 Step 1: Constructing a user-item interaction graph using historical user-item data interaction sequences;

[0050] Step 2: Inputting user node vector representation and item node vector representation in the user-item interaction graph into a multi-layer graph convolution model for graph convolution operation, and outputting final node embedding vectors, wherein the final node embedding vectors include user final embedding vectors and item final embedding vectors;

[0051] Step 3: Obtaining user behavior sequences based on user final embedding vectors and item final embedding vectors;

[0052] Step 4: Obtaining user-item interaction time sequences based on historical user-item data interaction sequences, obtaining different interaction time sequence representations by using multiple different encoding methods on the interaction time sequences, inputting different interaction time sequence representations and user behavior sequences into a multi-head attention model to capture user behavior patterns over time, and outputting predicted items.

[0053]

[0054] ​In the embodiment, the item is a representation in the recommended dataset, different datasets have different items, the item can be a movie, a song, a commodity, etc., in short, the item is a category that the user has interacted with. If the recommended dataset used is a movie dataset, the recommendation task is to recommend movies that the user is interested in according to the movies the user has watched, and the item here is movie data.

[0055] In step 1 of the embodiment, a user-item interaction graph is constructed according to the interaction of the user and the item, the user-item interaction graph includes a user node (i.e., a user node vector e u,k ), an item node (i.e., an item node vector e i,k ), and an edge connecting the user node and the item node if the user and the item have an interaction record.

[0056] Wherein, the set of user node vectors is E U = [e u,1 , e u,2 , e u,3 , …, e u,|U| ], the set of item node vectors is E I = [e i,1 , e i,2 , e i,3 , …, e i,|I| ], for each user u, there is an interaction sequence and its corresponding interaction time sequence Each represents the kth interaction item of the user, and each represents the time when the user interacts at the kth time.

[0057] In step 2 of the embodiment, a message passing architecture of graph convolution is constructed to capture the collaborative signal along the graph structure, and based on the fully mined collaborative signal, the next interaction behavior of the user is predicted, after multiple layers of propagation, the collaborative signal is explicitly injected into the user item embedding to obtain a better effective embedding representation information. The collaborative signal here refers to the correlation between different users captured through the graph convolution structure, and other users similar to the user's interest can be found, and the similar users are used for recommendation.

[0058] The embodiment uses the idea of collaborative filtering, which is to find some similarity through the behavior of the group, such as the similarity between user interests or the similarity between items, and make decisions and recommendations for users through the similarity. If a user likes an item, we can recommend the item to another user who is similar to his interest.

[0059] In specific embodiments, the user node vector set and the item node vector set are input into a multi-layer graph convolution model, and after L-layer propagation, a final user embedding vector group and a final item embedding vector group are obtained.

[0060] First, the design of single-layer propagation is described, and the propagation rule is defined as:

[0061]

[0062]

[0063] Wherein, N u represents the set of items interacted by user u, N i represents the set of users interacted with item i, is a symmetric regularization term, which can avoid the size of embedding being too large with the increase of graph convolution operation, and l represents the number of layers of graph convolution.

[0064] The single-layer propagation is extended to multi-layer propagation, and the multi-layer graph convolution model in the embodiment includes L layers of convolution operation layers. The multi-layer graph convolution model adds the intermediate user embedding vector generated by each layer of propagation and the initial user node vector after assigning the same weight to them, aggregates all the layers of embedding, and obtains the final user embedding vector corresponding to the user node. The multi-layer graph convolution model adds the intermediate item embedding vector generated by each layer of propagation and the initial item node vector after assigning the same weight to them, aggregates all the layers of embedding, and obtains the final item embedding vector corresponding to the item node.

[0065] Wherein, the calculation formula for aggregating the embedding of each layer to obtain the final user embedding vector and the final item embedding vector is:

[0066]

[0067]

[0068] Wherein, e u is the final user embedding vector, e i is the final item embedding vector, and a l is the weight, represents the vector representation of user u after L-layer graph convolution, is the vector representation of item i after L-layer graph convolution.

[0069] The above formula obtains the update process of a single user embedding or item embedding, and the overall update is represented in the form of a matrix: the user-item interaction matrix is R∈R |U|×(|I|+1), |U| represents the number of users, |I| represents the number of items, the user sequence is normalized, so when the user sequence length is less than N, we fill it with [PAD] markers, so the column number of the matrix is |I| + 1. If the user u has an interaction with the item i, the corresponding position R ui of the matrix is marked as 1, otherwise 0. For the above process, we describe it as:

[0070]

[0071]

[0072]

[0073] where E l represents the user-item matrix representation after graph convolution of the l-th layer, E represents the user-item matrix representation after the weighted sum of each layer E l , A is the adjacency matrix of the user-item interaction graph, R is the user-item interaction matrix, and D is a diagonal matrix of (|u| + (|I| + 1)) x (|u| + (|I| + 1)) dimensions, where each diagonal value D ii represents the number of non-zero entries in the i-th row of the interaction matrix.

[0074] In steps 3 and 4 of the embodiment, since the interaction sequence length of each user is not the same, we set a fixed length N to normalize the user interaction sequence so that their lengths are the same. For sequences longer than N, we truncate the sequence and only take the user's last N interaction items; for sequences shorter than N, we fill them with [PAD] so that all user interaction sequences have a length of N. In this way, the user interaction sequence becomes The interaction time sequence is

[0075] In this embodiment, all user and item embedding vector representations are obtained from step 1, and then the item embedding representation in the sequence S u is found to obtain the matrix representation of the interaction sequence S u , the user behavior sequence and the interaction time sequence are For the interaction time sequence, four different encoding methods are proposed to process it: where R is used to indicate the dimension.

[0076] For the position embedding we define a learnable position embedding matrix to indicate the relative position of the item in the sequence. To express the user's behavior pattern in a day, we convert each time series of dates into an embedding vector. We count the number of days in the dataset and build a learnable embedding matrix M D ∈R |D|×d Here we count all the dates in the dataset and build a d-dimensional embedding matrix. For time intervals, we define a time interval matrix C ∈ R N×N where each element is defined as c ab = (t a -t b ) to represent the interaction time difference of each item in the interaction sequence. We convert the time interval matrix into an embedding vector to explicitly express the impact of time intervals. Inspired by the position encoding in the transformer, we use the sine and cosine functions to model the time signal to capture the user's periodic behavior habits, denoted as which can capture some periodic behaviors of the user.

[0077] The encoder converts the time interval c ab into a hidden vector

[0078]

[0079] where is the i-th value of the vector and ξ is an adjustable parameter.

[0080] Through the above four different forms of temporal modeling, the model can capture different temporal transformation relationships within the sequence.

[0081] For the user behavior sequence we perform embedding processing to obtain S u which represents a scalar. We find S u from all user-item embeddings obtained in step 1. Then the scalar becomes a vector

[0082] Concatenate the final embedding vector of the user obtained in step 2 with each item in the sequence to obtain the personalized representation of the sequence, denoted as:

[0083]

[0084] where represents the j-th item embedding in the interaction items of user x and the user embedding d = d u +d i , d u is the dimension of user embedding representation, d i is the dimension of item embedding representation, we make our model more personalized by introducing user embedding to the interaction sequence.

[0085] In this embodiment, the four different forms of time series and interaction sequences obtained by the above are fused to capture the user's behavior pattern over time. We calculate by the following formula:

[0086]

[0087]

[0088]

[0089]

[0090] wherein the number of heads of the multi-head attention model is 4, are parameter matrices, h is the number of heads, we set it to 4, d is a scaling factor to prevent the product from being too large, and the parameters is the position bias, d w = d k = d v = d t = d / h, f is a calculation function, and the specific is as follows:

[0091]

[0092] wherein A = (a ij ) ∈ R M×N , B = (b ilj ) ∈ R M×L×N , C = (c ij ) ∈ R M×N

[0093] In this embodiment, the first head of the attention layer performs formula (11), the second head performs formula (12), the third head performs formula (13), and the fourth head performs formula (14).

[0094] The output results of the multi-head attention model are fused as:

[0095] X = concat(X Pos , X Day , X Per , X Time ) W oX E R N×d (16)

[0096] where is the parameter matrix, through four different forms of time information coding, we get the final output result.

[0097] After multi-head attention, the adaptive combines the sequence of item embedding and four different forms of time embedding, and finally the output result is transformed through a feedforward neural network to the model:

[0098] FFN(X) = GELU(XW 1 +b 1 )W 2 +b 2 (17)

[0099] where, W 1 E R d×4d , b 1 E R 4d , W 2 E R 4d×d , b 2 E R d are all learnable parameters. Then we stack L attention layers, for ease of training, apply a residual connection to each sublayer:

[0100] y = x + Dropout(FFN(LayerNorm(x))) (18)

[0101] where, X represents the update representation of each sequence item embedding after the attention layer in formula (16), Y represents the output representation after the feedforward neural network.

[0102] Output [y1, y2,..., y N ] E R N×d After that, the output result is sent to the target interaction network. The existing work reflects the user interest by using the fused sequence representation, and in general, the recommendation result given by the model only meets the main interest of the user and ignores some marginal interest. By finding the historical items close to the candidate target item to strengthen the interest of the user, the model adaptively finds the items related to the target item in the sequence, and activates the corresponding interest of the user for the specific target item. In order to realize this process, a target interaction network is designed in this embodiment to calculate the attention score of each item v i for each target item v t .

[0103] Specifically, the attention score of the target item v t to be predicted and each item v i in the user interaction sequence is calculated.The relevance is normalized by a softmax function:

[0104]

[0105] where, is the embedding representation of the target item v t . is the output representation of the item at position j in the interaction sequence after passing through the attention layer and the feed-forward neural network, W ∈ R d×d is a parameter matrix, and the representation of the user sequence for the target item v t is:

[0106]

[0107] The resulting user representation varies with different target items v t .

[0108] By first randomly masking the interaction sequence, the sequence after masking is fed into the overall model, and then the representation of the item at the masked position is predicted.

[0109] Then, we fuse the user representation with the item representation at position k to be predicted through a gating mechanism:

[0110] m k = γ(GELU(y k W m +b m )) + (1 - γ)z t (21)

[0111] The above formula obtains the item representation m k at the masked position k in the interaction sequence, W m ∈ R d×d , b m ∈ R d is a learnable parameter matrix, GELU is an activation function, and γ is a learnable gating parameter.

[0112] The item representation at position k is predicted according to the obtained m k , and the similarity score between m k and the target item is calculated:

[0113]

[0114] where, is the probability value that the item at the masked position k in the user interaction sequence is item v t , mk The embedding representation of the final output is obtained from formula (21). For candidate target projects, e u This indicates user embedding.

[0115] The model is trained by analyzing user interaction sequences. A portion of the sequence is randomly masked using a specific probability ρ, transforming the sequence s into sv. The transformed sequence s′ and the time series are then fed into the model of this embodiment to calculate the predicted probability P(v|s,t). Finally, the loss is calculated as follows:

[0116]

[0117] s′ is the interaction sequence after the mask, and t is the time series. It is a shielding project v m The true value of , the probability value P(·) is determined by formula (22).

[0118] During the validation testing phase, the goal of predicting the next recommended item is achieved by masking the last position in the user interaction sequence.

[0119] Example 2

[0120] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0121] Example 3

[0122] The purpose of this embodiment is to provide a computer-readable storage medium.

[0123] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0124] Example 4

[0125] The purpose of this embodiment is to provide a user interest recommendation system that integrates cooperative transformation and temporal awareness, including:

[0126] First data acquisition module: Construct a user-project interaction graph using historical user-project data interaction sequences;

[0127] Graph Convolution Module: Based on the user node vector representation and the item node vector representation in the user-item interaction graph, the graph convolution operation is performed on the multi-layer graph convolution model to output the final node embedding vector, which includes the user final embedding vector and the item final embedding vector.

[0128] The second data acquisition module obtains a user behavior sequence based on the final user embedding vector and the final item embedding vector.

[0129] The fusion recommendation module obtains a user-item interaction time sequence based on the user-item interaction sequence, obtains different interaction time sequence representations by using multiple different encoding methods on the interaction time sequence, inputs the different interaction time sequence representations and the user behavior sequence into the multi-head attention model to capture the user's behavior pattern over time, and outputs a predicted recommended item.

[0130] The steps and methods involved in the devices of embodiments two, three and four correspond to the first method embodiment, and the specific implementation can be seen from the relevant description of embodiment one. The term "computer readable storage medium" should be understood to include a single medium or multiple media of one or more instruction sets; it should also be understood to include any medium capable of storing, encoding or carrying instruction sets for execution by a processor and causing the processor to perform any of the methods in the present application.

[0131] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, and alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be respectively manufactured into individual integrated circuit modules, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0132] Although the specific embodiments of the present application are described above in combination with the accompanying drawings, it is not a limitation on the scope of protection of the present application, and those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A method for recommending user interests by fusing collaborative filtering and temporal awareness, characterized in that, The method comprises the following steps: constructing a user-item interaction graph using historical user-item data interaction sequences; inputting user node vector representation and item node vector representation in the user-item interaction graph into a multi-layer graph convolution model for graph convolution operation, outputting final node embedding vectors, and the final node embedding vectors comprising user final embedding vectors and item final embedding vectors; obtaining a user behavior sequence based on the user final embedding vectors and the item final embedding vectors; obtaining an interaction time sequence of the user-item based on the historical user-item data interaction sequences, obtaining different interaction time sequence representations by using multiple different encoding methods on the interaction time sequence, and inputting the different interaction time sequence representations and the user behavior sequence into a multi-head attention model to capture the user behavior pattern over time, and outputting a pre-recommended item.

2. The method of claim 1, wherein the fusion of the co-transformed and time-aware user interest recommendation is based on a user's interest in a specific time period. inputting the user node vector into the multi-layer graph convolution model to output an intermediate user embedding vector of each layer, adding the intermediate user embedding vectors output by each layer after assigning the same weight to the initial user node vector, and outputting a final user embedding vector; inputting the item node vector into the multi-layer graph convolution model to output an intermediate item embedding vector of each layer, adding the intermediate item embedding vectors output by each layer after assigning the same weight to the initial item node vector, and outputting a final item embedding vector.

3. The method of claim 2, wherein the fusion of the collaborative transformation and the time-awareness of the user interest recommendation is based on a user's interest in a time period. The calculation formula of the multi-layer graph convolution model is: wherein, denotes the number of layers of graph convolution, denotes the set of items interacted by user u, denotes the set of users interacting with item i, denotes the set of users interacting with item i, is a symmetric regularization term, is the vector representation of item i after the layer graph convolution, denotes the vector representation of user u after the layer graph convolution.

4. The method of claim 1, wherein the fusion of the co-transformed and time-aware user interest recommendation is based on a user's interest in a specific time period. obtaining an interaction time sequence of the user-item based on the historical user-item data interaction sequences, obtaining different interaction time sequence representations by using multiple different encoding methods on the interaction time sequence, and inputting the different interaction time sequence representations and the user behavior sequence into a multi-head attention model to capture the user behavior pattern over time, and outputting a pre-recommended item. By defining a position embedding matrix to indicate the relative position of the item in the interaction time series; The date in each interaction time series is converted into an embedding vector to express the user's behavior pattern within a day by counting the number of days that appear; defining a time interval matrix to represent interaction time differences for each item in the interaction time series, converting the time interval matrix to an embedding vector ; The periodic behavior habits of the user are captured using the cosine and sine functions, denoted as .

5. The method of claim 4, wherein the fusion of the co-transformed and time-aware user interest recommendation is performed by a user interest recommendation engine. The calculation formula of the multi-head attention model is: wherein, is a user behavior sequence, , , , , are parameter matrices, h is a head number, is a scale factor, is a position deviation, is a calculation function; ; , is a dimension of a user embedding representation, is a dimension of an item embedding representation; a superscript T represents a device.

6. The method of claim 5, wherein the fusion of the co-transformed and time-aware user interest recommendation is performed by a user interest recommendation engine. After fusing the different interaction time sequence representations and the user behavior sequence through the multi-head attention model, the output result is subjected to nonlinear transformation through a feedforward neural network.

7. The user interest recommendation method integrating collaborative transformation and temporal awareness as described in claim 1, characterized in that, calculating the relevance score of the target item and each item in the user behavior sequence, obtaining a user sequence corresponding to the target item based on the relevance score; obtaining a final item embedding representation of the to-be-predicted position in the user behavior sequence based on the user sequence and the predicted recommended item output; calculating the similarity score of the final item embedding representation and the target item representation for item recommendation.

8. A user interest recommendation system that fuses collaborative filtering and temporal awareness, characterized in that, The method comprises the following steps: a first data acquisition module: constructing a user-item interaction graph using historical user-item data interaction sequences; a graph convolution module: inputting user node vector representation and item node vector representation in the user-item interaction graph into a multi-layer graph convolution model for graph convolution operation, outputting final node embedding vectors, and the final node embedding vectors comprising user final embedding vectors and item final embedding vectors; a second data acquisition module: obtaining a user behavior sequence based on the user final embedding vectors and the item final embedding vectors; The fusion recommendation module: based on the user-item data interaction sequence, the interaction time sequence of the user-item is obtained, the interaction time sequence is processed through multiple different encoding methods to obtain different interaction time sequence representations, and the different interaction time sequence representations and the user behavior sequence are input into a multi-head attention model to capture the user behavior pattern changing over time, and the output is a predicted recommended item.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the fusion collaborative transformation and time-aware user interest recommendation method of any one of claims 1-7.

10. A processing device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the fusion collaborative transformation and time-aware user interest recommendation method of any one of claims 1-7.

Images