A product recommendation method, system, device and storage medium

Abstract

The present invention discloses a product recommendation method, system, device and storage medium, S1, constructing a target user-product interaction matrix and a target user-product category interaction matrix, and then constructing directed weighted graph adjacency matrices of the two interaction matrices respectively; S2, processing the two directed weighted graph adjacency matrices respectively and in turn through a graph neural network, an attention network and a threshold recurrent neural network to obtain a product vector and a product category vector, and splicing the product vector and the product category vector through a vector fusion network to obtain a final vector; S3, calculating the final vector and all candidate products to obtain a score ranking, and obtaining a prediction list 1; S4, obtaining the recent behavior data of multiple users, constructing a multi-user-product interaction matrix, recommending products with which similar users have interactive behaviors to the target user, and obtaining a recommendation list 2; S5, linearly combining the two recommendation lists to obtain a final recommendation list, and pushing the products in the final recommendation list to the target user.

Description

Technical Field

[0001] The present invention belongs to the field of recommendation algorithms and relates to a commodity recommendation method, system, device and storage medium. Background Art

[0002] In recent years, with the rapid development of technologies such as cloud computing, big data, and the Internet of Things, the emergence of various applications in the Internet space has triggered an explosive growth in the scale of data. Big data contains rich value and huge potential, which will bring revolutionary development to human society, but it also brings serious "information overload" problems. How to quickly and effectively obtain valuable information from complex data has become a key problem in the current development of big data. As an effective way to solve the problem of "information overload", the recommendation system has become a hot topic in academia and industry and has been widely used, forming many related research results. The recommendation system mines the products (such as information, services, products, etc.) that users are interested in from massive data through recommendation algorithms based on user needs and interests, and recommends the results to users in the form of personalized lists. At present, recommendation systems have been successfully applied in many fields, including e-commerce (such as Amazon, eBay, Netflix, Alibaba, etc.), information retrieval (such as iGoogle, MyYahoo, Baidu, etc.), social networks (Face-book, Twitter, Tencent, etc.), location services (such as Foursquare, Yelp, Dianping, etc.), news push (such as Google News, GroupLens, Toutiao, etc.), etc. Traditional recommendation methods mainly include collaborative filtering, content-based recommendation methods and hybrid recommendation methods. Among them, the most classic algorithm is collaborative filtering, such as matrix factorization, which uses the interactive information between users and products to recommend to users. Collaborative filtering is currently the most widely used recommendation algorithm and has won many awards in the Netflix Grand Prix in recent years. However, it also encounters serious data sparsity (the products rated by a user only account for a very small part of the total number of products) and cold start (new users and new products often have no rating data) problems. In addition, the classic collaborative filtering method uses a shallow model and cannot learn the deep features of users and products. Content-based recommendation methods use the items that users have selected to find other items with similar attributes for recommendation. However, this method requires effective feature extraction. Traditional shallow models rely on artificially designed features, and their effectiveness and scalability are very limited, which restricts the performance of content-based recommendation methods. As more and more data on the Internet can be perceived and acquired, multi-source heterogeneous data including images, texts, and labels contain rich user behavior information and personalized demand information. Hybrid recommendation methods that integrate multi-source heterogeneous auxiliary information are increasingly valued because they can alleviate the data sparsity and cold start problems in traditional recommendation systems. However, since auxiliary information often has complex characteristics such as multimodality, data heterogeneity, large scale, data sparsity, and uneven distribution, research on hybrid recommendation methods that integrate multi-source heterogeneous data still faces severe challenges.

[0003] In recent years, deep learning has become a hot trend in Internet big data and artificial intelligence. Deep learning has made breakthrough progress in image processing, natural language understanding, and speech recognition, bringing new opportunities for the research of recommendation systems. On the one hand, deep learning can characterize massive data related to users and products by learning a deep nonlinear network structure. It has a strong ability to learn the essential characteristics of data sets from samples and can obtain deep feature representations of users and products. On the other hand, deep learning can automatically learn features from multi-source heterogeneous data, thereby mapping different data to the same latent space, and can obtain a unified representation of data. On this basis, it can integrate traditional recommendation methods for recommendation, effectively utilize multi-source heterogeneous data, and alleviate the data sparsity and cold start problems in traditional recommendation systems. In the past three years, the research on recommendation systems based on deep learning has begun to receive increasing attention from the international academic and industrial communities.

[0004] In summary, recommendation algorithms currently face the following problems: First, since the products that users have interacted with only account for a very small part of the total number of products, they encounter serious data sparsity problems; second, since new users and new products often have no rating data, there is a cold start problem. Third, traditional recommendation methods use shallow models that cannot learn the deep features of users and products. Their effectiveness and scalability are very limited, which restricts the performance of the recommendation system. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a product recommendation method, system, device and storage medium to improve the accuracy of personalized product recommendations and alleviate the cold start problem and matrix sparsity problem existing in the recommendation system.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A product recommendation method includes the following processes:

[0008] S1, construct the target user-product interaction matrix and the target user-product category interaction matrix, and then construct the product interaction directed weighted graph adjacency matrix and the product category interaction directed weighted graph adjacency matrix respectively;

[0009] S2, the adjacency matrix of the product interaction directed weighted graph and the adjacency matrix of the product category interaction directed weighted graph are processed by the graph neural network, the attention network and the threshold recurrent neural network respectively to obtain the product vector and the product category vector, and the product vector and the product category vector are concatenated by the vector fusion network to obtain the final vector;

[0010] S3, calculate the final vector and all candidate products to get the score ranking and obtain the prediction list Top-N1;

[0011] S4, obtain the recent behavior data of multiple users, build a multi-user-product interaction matrix, recommend products with which similar users have interacted to the target user, and obtain the recommendation list Top-N2;

[0012] S5: Linearly combine the recommendation lists Top-N1 and Top-N2 to obtain the final recommendation list Top-N, and push the products in the final recommendation list Top-N to the target user.

[0013] Preferably, in S1, the rows in the target user-product interaction matrix represent the interaction behaviors of user 1, the columns represent items 1 to 7, and the values ​​in the target user-product interaction matrix are the interaction behaviors of user 1 on item i, where 0 represents no interaction, 1 represents click-to-browse, and 2 represents purchase; each node in the product interaction directed weighted graph represents a project, and each directed edge represents the user's interaction behavior.

[0014] Preferably, in S1, the rows in the target user-product category interaction matrix represent the interaction behaviors of user 1, the columns represent the three categories of items, and the values ​​in the target user-product category interaction matrix are the interaction behaviors of user 1 on item category i, where 0 represents no interaction, 1 represents click-to-browse, and 2 represents purchase; each node in the product category interaction directed weighted graph represents an item category, and each edge represents the user's interaction with category 2 after interacting with category 1.

[0015] Preferably, in S2, the conversation sequence data in the adjacency matrix of the directed weighted graph of product interactions and the adjacency matrix of the directed weighted graph of product category interactions are processed and learned through a graph neural network to capture the spatial structure information and node relationship information between conversation sequences, and obtain product vectors and product category vectors with more potential information.

[0016] Preferably, in S2, the attention network assigns different attention weights to each product and product category according to the attention mechanism, assigns more weights to the most relevant products, extracts the most critical product and category information, captures the dependencies between interactive interests, and thus obtains the overall preference.

[0017] Preferably, in S2, the threshold recurrent neural network performs sequence processing on the session data to obtain dependency and timing information between session click items.

[0018] Preferably, in S4, the multi-user-product interaction matrix uses [0, 1, 2] as the interaction interval, representing no interaction, click-to-browse, and purchase, respectively; user-based collaborative filtering is used to obtain a user similarity matrix by calculating the cosine similarity between the user and other users, and products with which similar users have interacted are recommended to the target user, and the recommendation list Top-N2 is obtained based on the basic attribute information registered by the user.

[0019] A product recommendation system, comprising:

[0020] A matrix construction module is used to construct a target user-product interaction matrix and a target user-product category interaction matrix, and then respectively construct a product interaction directed weighted graph adjacency matrix and a product category interaction directed weighted graph adjacency matrix;

[0021] The vector acquisition module is used to process the adjacency matrix of the product interaction directed weighted graph and the adjacency matrix of the product category interaction directed weighted graph through the graph neural network, the attention network and the threshold recurrent neural network respectively to obtain the product vector and the product category vector, and then concatenate the product vector and the product category vector through the vector fusion network to obtain the final vector;

[0022] The first prediction list acquisition module is used to calculate the final vector and all candidate products to obtain the score ranking and obtain the prediction list Top-N1;

[0023] The second prediction list acquisition module is used to obtain the recent behavior data of multiple users, build a multi-user-product interaction matrix, recommend products with which similar users have interacted to the target user, and obtain the recommendation list Top-N2;

[0024] The product push module is used to linearly combine the recommendation lists Top-N1 and Top-N2 to obtain the final recommendation list Top-N, and push the products in the final recommendation list Top-N to the target users.

[0025] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the above-mentioned commodity recommendation methods when executing the computer program.

[0026] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any one of the above-mentioned product recommendation methods are implemented.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The present invention aims to solve the problem that the traditional recommendation method cannot learn the deep-level features of users and products by using shallow models. The present invention uses a deep learning model to learn the explicit feedback data, implicit feedback data of users, display feedback data and implicit feedback data of products, and category data of products. In view of the data sparsity problem, a session-based graph neural network that is good at extracting graph structures is used, and a graph neural network is used to process product interaction information. By processing and learning the session sequence data through a multi-layer graph neural network layer, the spatial structure information and node relationship information between session sequences can be effectively captured, alleviating the matrix sparsity problem. For the cold start problem, the cold start problem can be effectively solved by combining the advantages of GRU and Attention in processing session sequences and the powerful interpretability of collaborative filtering. Deep learning is combined with existing recommendation methods, and the two methods are integrated to construct a recommendation system module. The purpose is to solve the problem of low precision of personalized recommendations and alleviate the cold start problem and matrix sparsity problem in the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a structural diagram of the recommendation algorithm of the present invention;

[0030] Figure 2 This is the attention network model diagram;

[0031] Figure 3 It is a structural diagram of the threshold recurrent neural network;

[0032] Figure 4 It is a conversation interaction sequence diagram;

[0033] Figure 5 It is a conversation interaction category sequence diagram;

[0034] Figure 6 is the target user-product interaction matrix. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0036] It should be noted that the words "front", "rear", "left", "right", "up" and "down" used in the following description refer to directions in the drawings, and the words "inside" and "outside" refer to directions toward or away from the geometric center of a specific component, respectively.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0038] like Figure 1 As shown, the product recommendation method described in the present invention is a session-based recommendation system that uses the interaction sequence information of the user on a series of products within a certain period of time as the input of the model. The model models the target user's interaction sequence information, predicts the products that the target user may be interested in, and outputs a recommendation list.

[0039] Step 1: First, construct the interaction sequence information between users and products into a target user-product interaction matrix, and convert the target user-product interaction matrix into a directed weighted graph such as Figure 4 As shown in the figure, the rows in the target user-item interaction matrix represent the interaction behavior of user 1, and the columns represent items 1 to 7. The values ​​in the interaction matrix are the interaction behaviors of user 1 on item i, where 0 means no interaction, 1 means click to browse, and 2 means purchase. Each node in the directed weighted graph represents an item, and each directed edge represents the user's interaction behavior. For example: User 1 clicks on item 3, then clicks on item 5 and item 2, then buys item 1, then clicks on item 4, item 6, and item 7, then returns to buy the item 2 that he clicked on, and finally clicks on item 6 to end this session in a closed loop.

[0040] Step 2: Secondly, construct the interaction sequence category information between users and products into a target user-product category interaction matrix, and convert the target user-product category interaction matrix into a directed weighted graph such as Figure 5 As shown in the figure, the rows in the target user-product category interaction matrix represent the interaction behaviors of user 1, and the columns represent the three categories of products. The values ​​in the target user-product category interaction matrix are the interaction behaviors of user 1 on product category i, where 0 represents no interaction, 1 represents click browsing, and 2 represents purchase. Each node in the figure represents a product category, and each edge represents the user interacting with category 2 after interacting with category 1. The rows in the interaction matrix represent the interaction behaviors of user 1, and the columns represent products 1 to 7. For example: user 1 clicked on category 2 and then clicked on category 1 continuously, then purchased category 2, then clicked on category 3, category 2, and category 1, then went back to purchase the category 1 that he had clicked on, and finally clicked on category 2 to end this session in a closed loop.

[0041] Finally, the product interaction directed weighted graph adjacency matrix and the product category interaction directed weighted graph adjacency matrix are constructed for the target user-product interaction matrix and the target user-product category interaction matrix respectively to store the directed weighted graph data, and the adjacency matrix is ​​used as the input of the graph neural network.

[0042] Step 3: Graph neural networks are good at extracting feature information such as graph structure, so graph neural networks are used to process product interaction information. The adjacency matrix of the directed weighted graph of product interactions and the adjacency matrix of the directed weighted graph of product category interactions are input into two graph neural networks respectively. The two graph neural networks obtain the weight information between product nodes and the weight information between product category nodes respectively. The graph neural network layer processes and learns the session sequence data in the adjacency matrix of the directed weighted graph of product interactions and the adjacency matrix of the directed weighted graph of product category interactions, which can capture the spatial structure information and node relationship information between session sequences and obtain the product vector I with more potential information. i and product category vector T i , the product vector I i and product category vector T i As input to the attention network.

[0043] Step 4: Figure 2 As shown in the figure, the attention layer is used to automatically learn the characteristics of the session. Using the attention layer can help the model make more accurate judgments. The output vector of the graph neural network is used as the input of the Attention layer, and the product vector I i and product category vector T i The information is input into the attention layer, and different attention weights are assigned to each product and product category according to the attention mechanism. More weights are assigned to the most relevant products, and the most critical product and category information is extracted to capture the dependencies between interactive interests, thereby obtaining the overall preference. The product and product category vectors with added weight information are used as the input of the threshold recurrent neural network.

[0044] Step 5: Figure 3 As shown in the figure, the gated recurrent neural network, i.e., the GRU network, is used to extract the features of the product information before and after the session. The product and product category vectors with weighted information are input into the GRU network, and the gated neural unit is used to process the session data in sequence to obtain the dependency and timing information between the session click items. This takes into account the information of the previously interacted products while the information of the current product is interacted with. Update gate z t The calculation formula is shown in formula (4-1): Reset gate r t The calculation formula is shown in formula (4-2):

[0045] z t =σ(W z At +U z h t-1 ) (4-1)

[0046] r t =σ(W r A t +U r h t-1 ) (4-2)

[0047]

[0048]

[0049] where h t-1 represents the hidden state at the previous moment, r t Represents the value of the reset gate, z t represents the value of the update gate, A t Indicates the current input, h t-1 Represents the hidden state of the previous moment, and the weight matrix U is used to control the degree of memory. Indicates output. t The value is between 0 and 1. t The closer the value of is to 1, the more the input information at this moment is integrated with the memory of the previous moment, so the more information needs to be remembered. t The closer the value of is to 0, the less the input information at this moment is integrated with the memory of the previous moment, and the more information needs to be forgotten. Update gate z t The value is used to determine the degree of selective memory, deleting irrelevant memories and selecting important memories to pass on. t The value of is also between 0 and 1. t The closer the value of is to 1, the more state information is brought in from the previous moment, and the more is remembered. t The closer the value of is to 0, the less state information is brought in from the previous moment, and the more irrelevant memories are deleted. The product vector and product category vector obtained by the threshold recurrent neural network are used as the input of the vector fusion network.

[0050] Step 6: Use the vector fusion network to concatenate the product vector and the product category vector to obtain the final vector V i , taking product categories as multi-granular auxiliary information of products.

[0051] Step 7: Obtain the recent behavior data of multiple users and construct a multi-user-product interaction matrix such as Figure 6As shown: The matrix uses [0,1,2] as the interaction interval, representing no interaction, click browsing and purchase respectively. By using user-based collaborative filtering, the user similarity matrix is ​​obtained by calculating the cosine similarity between the user and other users, and the products with which similar users have interacted are recommended to the target user. The recommendation list Top-N2 is obtained based on the basic attribute information registered by the user.

[0052] For example, users U1 and U5 are similar users, and user U5 has no interactive behavior. The item I1 that U1 has browsed and the item I5 that U1 has purchased are recommended to U5.

[0053] Step 8: The final vector V i The score ranking is calculated with all candidate products, and the predicted output of the session is obtained through the softmax activation function. The cross entropy function is used as the loss function to evaluate the inconsistency between the predicted output and the true value, and finally the model's prediction list Top-N1 is obtained.

[0054] Step 9: Linearly combine Top-N1 and Top-N2 to get the final recommendation list Top-N, and push the products in the final recommendation list Top-N to the target user.

[0055] In order to evaluate the recommendation performance of the proposed session-based multi-granularity graph neural network combined with collaborative filtering recommendation model, it will be compared with the baseline model, canceling collaborative filtering, attention mechanism and GRU respectively, and verifying the effectiveness of the model through ablation experiments. The following model in RecBole is used as the baseline model:

[0056] (1) ItemKNN (2004): A top-N recommendation algorithm based on items. This model is used to calculate the similarities between items and to combine these similarities to calculate the similarities between items and candidate recommended items.

[0057] (2) Improved-GRU-Rec (2016): A session-based recommendation model based on an improved RNN, which uses data augmentation and a method to account for changes in the input data distribution to capture the user's current interests.

[0058] (3) NGCF (2019): Graph Neural Network Collaborative Filtering, which uses the target user-item interaction bipartite graph structure to model and uses the expression of high-order connectivity of the graph structure to explicitly inject collaborative signals into the embedding process.

[0059] (4) SRGNN (2019): Session-based graph neural network recommendation, using graph neural network combined with self-attention mechanism to complete session-based recommendation tasks. Use graph structure data to store drawing sequences. GNN captures the complex transformation of products, and then uses attention network to represent each session as a combination of global preference and current interest of the session.

[0060] (1) Comparison of SMGNN-CF and Baseline Methods

[0061] SMGNN-CF was compared with other recommendation models Item-KNN, NGCF, Improved-GRU-Rec), SR-GNN and GCSAN on the Yochoose dataset. The performance comparison of the experimental results is shown in Table 1:

[0062] Table 1 Comparative experimental results

[0063]

[0064] Item-KNN uses only the traditional collaborative filtering algorithm and can only make recommendations based on shallow features. Therefore, the recommendation effect is the worst. After applying NGCF to GNN and collaborative filtering, the P@20 and MRR@20 of NGCF are higher than those of Item-KNN. This verifies the importance of neural networks in learning better user and product representations. The recurrent neural network RNN ​​is used in the Improved-GRU-Rec model. This model makes recommendations based on the temporal information of the session. The Improved-GRU-Rec performs better than Item-KNN and NGCF. In SR-GNN, GNN is used to capture the complex node transformations of the product, and the attention network is used to represent each session as the global preference and the current interest of the session. Compared with the first three recommendation models, SR-GNN performs better. The SMGNN-CF proposed in this paper combines graph neural network, attention network, and threshold recurrent neural network with traditional collaborative filtering. The graph neural network is used to learn the complex changes of each node in the graph structure, the attention network is used to capture the dependencies between interactive interests, the threshold recurrent neural network forms the priority of the current session, and the recommendation is made based on the temporal information of the session and the dependencies between sessions. From the experiments, it can be seen that the SMGNN-CF proposed in this paper has better recommendation effect than the other four models.

[0065] (2) Impact of ablation experiments on the recommendation effect of the SMGNN-CF model

[0066] This experiment designed an ablation experiment to analyze the performance of each module in SMGNN-CF. The experiment was conducted on the Yochoose dataset to verify the influence of the traditional collaborative filtering layer, category graph neural network layer, GRU layer and attention layer on the experimental results. The experimental results are shown in Table 2.

[0067] Table 2 Impact of ablation experiments on the model

[0068]

[0069] From the results in Table 2, we can see that after SMGNN-CF removes the CF layer, the recommendation effect of the model decreases by about 8%, indicating that traditional collaborative filtering can improve the recommendation effect. After the model removes the product category graph neural network, the recommendation effect of the model decreases by 5%, indicating that product categories as coarse-grained auxiliary information for recommendation can improve the recommendation effect. After the model removes the Attention layer, the recommendation effect of the model decreases by 11%, indicating that the attention mechanism can capture the interest dependency between conversation sequences, thereby improving the recommendation effect. After the model removes the GRU layer, the recommendation effect of the model decreases most significantly, decreasing by 21%, indicating that the GRU network can have the greatest impact on the model, and the recommendation effect of the model depends to a large extent on the GRU network, which also shows that the GRU network is most effective in processing conversation sequence tasks. The results of the ablation experiment show that the model proposed in this paper combines traditional collaborative filtering, GRU network and Attention to effectively improve the effect of the recommendation algorithm and alleviate the problems of data sparsity and cold start.

[0070] This model overcomes the following problems: First, it uses graph neural networks to process user interaction data and learn more abstract and dense latent vectors of deep representations of users and products. This alleviates the problem that shallow models cannot learn the deep features of users and products and the data sparsity problem. Second, it uses deep learning combined with traditional recommendation algorithms to alleviate the cold start problem. Overall, it improves the recommendation effect.

[0071] This recommendation method can be applied to major e-commerce websites, music video websites, and social networks. First, the recommendation algorithm can better explore the user's deep interest needs and accurately reach the user's interest points, thereby bringing more accurate product recommendations. The recommendation system accurately recommends products of interest to users, increases the exposure of products to users who are interested in them, and also allows more unpopular but high-quality products to be exposed, thereby promoting the sales of the products. The recommendation system can make products more efficiently distributed, improve the operating efficiency of the entire platform, and effectively save company resources. Second, it provides software services to merchants, such as data analysis and personalized product recommendations. By efficiently distributing products, products are recommended to users who are willing to buy them, promoting the sale of products, and thus obtaining more commissions from merchants.

[0072] The following are device embodiments of the present invention, which can be used to perform method embodiments of the present invention. For details not disclosed in the device embodiments, please refer to the method embodiments of the present invention.

[0073] In yet another embodiment of the present invention, a product recommendation system is provided, which can be used to implement the above-mentioned product recommendation method. Specifically, the product recommendation system includes a matrix construction module, a vector acquisition module, a first prediction list acquisition module, a second prediction list acquisition module and a product push module.

[0074] Among them, the matrix construction module is used to construct the target user-product interaction matrix and the target user-product category interaction matrix, and then construct the product interaction directed weighted graph adjacency matrix and the product category interaction directed weighted graph adjacency matrix respectively.

[0075] The vector acquisition module is used to process the adjacency matrix of the product interaction directed weighted graph and the adjacency matrix of the product category interaction directed weighted graph through the graph neural network, attention network and threshold recurrent neural network respectively to obtain the product vector and product category vector, and then concatenate the product vector and product category vector through the vector fusion network to obtain the final vector.

[0076] The first prediction list acquisition module is used to calculate the final vector Vi and all candidate products to obtain a score ranking and obtain a prediction list Top-N1;

[0077] The second prediction list acquisition module is used to obtain the recent behavior data of multiple users, build a multi-user-product interaction matrix, recommend products with which similar users have interactive behaviors to the target user, and obtain the recommendation list Top-N2.

[0078] The product push module is used to linearly combine the recommendation lists Top-N1 and Top-N2 to obtain the final recommendation list Top-N, and push the products in the final recommendation list Top-N to the target users.

[0079] In another embodiment of the present invention, a terminal device is provided, the terminal device includes a processor and a memory, the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGAs), or other processors. GateArray, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, which are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding functions; the processor described in the embodiment of the present invention can be used for the operation of the product recommendation method, including: S1, constructing a target user-product interaction matrix and a target user-product category interaction matrix, and then constructing a product interaction directed weighted graph adjacency matrix and a product category interaction directed weighted graph adjacency matrix respectively; S2, converting the product interaction directed weighted graph adjacency matrix and the product category interaction directed weighted graph adjacency matrix After being processed by graph neural network, attention network and threshold recurrent neural network respectively, product vector and product category vector are obtained, and the product vector and product category vector are concatenated by vector fusion network to obtain the final vector; S3, the final vector is calculated with all candidate products to obtain the score ranking, and the prediction list Top-N1 is obtained; S4, the recent behavior data of multiple users are obtained, and a multi-user-product interaction matrix is ​​constructed, and the products with which similar users have interactive behaviors are recommended to the target user, and the recommendation list Top-N2 is obtained; S5, the recommendation list Top-N1 is linearly combined with Top-N2 to obtain the final recommendation list Top-N, and the products in the final recommendation list Top-N are pushed to the target user.

[0080] In another embodiment, the present invention further provides a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understandable that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory.

[0081] The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the product recommendation method in the above embodiment; one or more instructions in the computer-readable storage medium are loaded by the processor and execute the following steps: S1, construct a target user-product interaction matrix and a target user-product category interaction matrix, and then construct a product interaction directed weighted graph adjacency matrix and a product category interaction directed weighted graph adjacency matrix respectively; S2, the product interaction directed weighted graph adjacency matrix and the product category interaction directed weighted graph adjacency matrix are respectively subjected to a graph neural network, an attention network and a threshold recurrent neural network in sequence. Processing, obtain product vectors and product category vectors, and concatenate the product vectors and product category vectors through the vector fusion network to obtain the final vector; S3, calculate the final vector and all candidate products to obtain the score ranking, and obtain the prediction list Top-N1; S4, obtain the recent behavior data of multiple users, build a multi-user-product interaction matrix, recommend products with similar users to the target user, and obtain the recommendation list Top-N2; S5, linearly combine the recommendation lists Top-N1 and Top-N2 to obtain the final recommendation list Top-N, and push the products in the final recommendation list Top-N to the target user.

[0082] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0083] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes 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 generate 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.

[0084] 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.

[0085] 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.

[0086] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0087] It should be understood that the above description is for illustration and not for limitation. Many embodiments and many applications beyond the examples provided will be apparent to those skilled in the art upon reading the above description. Therefore, the scope of the present teachings should not be determined with reference to the above description, but rather with reference to the foregoing claims and the full scope of equivalents to which such claims are entitled. For the purpose of comprehensiveness, all articles and references, including disclosures of patent applications and publications, are incorporated herein by reference. The omission of any aspect of the subject matter disclosed herein in the foregoing claims is not intended to be a waiver of such subject matter, nor should it be considered that the applicant has not considered such subject matter to be part of the disclosed inventive subject matter.

Claims

1. A product recommendation method, characterized in that: The process includes: S1, construct the target user-product interaction matrix and the target user-product category interaction matrix, and then construct the product interaction directed weighted graph adjacency matrix and the product category interaction directed weighted graph adjacency matrix respectively; S2, the adjacency matrix of the product interaction directed weighted graph and the adjacency matrix of the product category interaction directed weighted graph are processed by the graph neural network, the attention network and the threshold recurrent neural network respectively to obtain the product vector and the product category vector, and the product vector and the product category vector are concatenated by the vector fusion network to obtain the final vector; S3, calculate the final vector and all candidate products to get the score ranking and obtain the prediction list Top-N1; S4, obtain the recent behavior data of multiple users, build a multi-user-product interaction matrix, recommend products with which similar users have interacted to the target user, and obtain the recommendation list Top-N2; S5: Linearly combine the recommendation lists Top-N1 and Top-N2 to obtain the final recommendation list Top-N, and push the products in the final recommendation list Top-N to the target user.

2. The commodity recommendation method according to claim 1, characterized in that: In S1, the rows in the target user-item interaction matrix represent the interaction behavior of user 1, and the columns represent items 1 to 7. The values ​​in the target user-item interaction matrix are the interaction behaviors of user 1 on item i, where 0 represents no interaction, 1 represents click browsing, and 2 represents purchase; Each node in the commodity interaction directed weighted graph represents an item, and each directed edge represents a user's interaction behavior.

3. The commodity recommendation method according to claim 1, characterized in that: In S1, the rows in the target user-item category interaction matrix represent the interaction behaviors of user 1, the columns represent the three categories of items, and the values ​​in the target user-item category interaction matrix are the interaction behaviors of user 1 on item category i, where 0 represents no interaction, 1 represents click browsing, and 2 represents purchase; In the directed weighted graph of product category interactions, each node represents an item category, and each edge represents a user interacting with category 2 after interacting with category 1.

4. The commodity recommendation method according to claim 1, characterized in that: In S2, the conversation sequence data in the adjacency matrix of the directed weighted graph of product interactions and the adjacency matrix of the directed weighted graph of product category interactions are processed and learned through a graph neural network to capture the spatial structure information and node relationship information between conversation sequences, and obtain product vectors and product category vectors with more potential information.

5. The commodity recommendation method according to claim 1, characterized in that: In S2, the attention network assigns different attention weights to each product and product category according to the attention mechanism, assigns more weights to the most relevant products, extracts the most critical product and category information, captures the dependencies between interactive interests, and thus obtains the overall preference.

6. The commodity recommendation method according to claim 1, characterized in that: In S2, the gated recurrent neural network performs sequence processing on the session data to obtain the dependency and timing information between session click items.

7. The commodity recommendation method according to claim 1, characterized in that: In S4, the multi-user-item interaction matrix uses [0, 1, 2] as the interaction interval, representing no interaction, click-to-browse, and purchase, respectively. User-based collaborative filtering is used to obtain the user similarity matrix by calculating the cosine similarity between the user and other users. Items with which similar users have interacted are recommended to the target user, and the recommendation list Top-N2 is obtained based on the basic attribute information registered by the user.

8. A product recommendation system, characterized in that: include: A matrix construction module is used to construct a target user-product interaction matrix and a target user-product category interaction matrix, and then respectively construct a product interaction directed weighted graph adjacency matrix and a product category interaction directed weighted graph adjacency matrix; The vector acquisition module is used to process the adjacency matrix of the product interaction directed weighted graph and the adjacency matrix of the product category interaction directed weighted graph through the graph neural network, the attention network and the threshold recurrent neural network respectively to obtain the product vector and the product category vector, and then concatenate the product vector and the product category vector through the vector fusion network to obtain the final vector; The first prediction list acquisition module is used to calculate the final vector and all candidate products to obtain the score ranking and obtain the prediction list Top-N1; The second prediction list acquisition module is used to obtain the recent behavior data of multiple users, build a multi-user-product interaction matrix, recommend products with which similar users have interacted to the target user, and obtain the recommendation list Top-N2; The product push module is used to linearly combine the recommendation lists Top-N1 and Top-N2 to obtain the final recommendation list Top-N, and push the products in the final recommendation list Top-N to the target users.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the product recommendation method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the product recommendation method according to any one of claims 1 to 7 are implemented.

Images