A graph convolution recommendation method and device considering user-item demand relationship

Abstract

The present invention provides a graph convolutional recommendation method and device that considers user-item demand relationships. The method comprises: encoding a prediction timestamp to obtain a time code vector; inputting the time code vector into a user interest model to obtain a confidence score of the user's interest level in each item category at the prediction timestamp; generating graph structure data based on user-item interaction records; inputting the graph structure data into a graph convolutional network model to obtain the user's basic prediction score for each item; adjusting the basic prediction score using the interest confidence score to obtain the user's adjusted prediction score for each item; sorting the user's adjusted prediction scores for the items from high to low, and recommending the top K1 items in the sorted sequence to the user. The present invention also provides a graph convolutional recommendation device that considers user-item demand relationships. The present invention improves the recommendation accuracy of the recommendation system.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular to a graph convolution recommendation method and device that considers the relationship between user-item requirements. Background Art

[0002] Recommender systems are a method for mining data features. Based on historical user interactions, recommendation systems can predict items that users may be interested in, including but not limited to products, movies, music, and short videos. Recommender systems can analyze underlying behavioral patterns from large amounts of user data, significantly impacting both user internet experiences and businesses' precise delivery of targeted recommendations. Because recommendation systems' interaction record modeling closely matches graph models, recommendation methods based on graph neural networks have become a major research area.

[0003] Current mainstream graph neural network recommendation methods build a corresponding graph structure based on user-item interaction records. They then perform convolution operations to obtain latent feature encodings of the user and item. The predicted rating of the user for the item is determined based on the distance between the latent feature encodings of the user and item. Few current graph neural network recommendation methods consider how user demand for an item changes over time, and fewer solutions specifically quantify this demand. In real-world operations, both group and individual behavior are generally constrained by time cycles. For example, individuals' commutes, transportation, and entertainment preferences all have clear cyclical characteristics. Therefore, for users, their demand for an item is likely to have a cyclical relationship. From the user's perspective, this cyclical relationship can be called the user-item demand cycle, and from the item's perspective, the item-user lifecycle. Therefore, considering the relationship between user-item demand is essential for the performance of recommendation systems.

[0004] One existing approach to consider how user demand for items changes over time is to treat user interaction sequences as a user interaction function curve, with time as the horizontal axis and the items interacted with as the vertical axis. This user interaction function curve is then used to adjust the predicted user ratings for items output by graph neural network recommendation methods. However, this approach has the following difficulties:

[0005] (1) The items need to be quantified into numerical values ​​before they can be used as the vertical axis values ​​of the function. The conventional method is to use the item ID for quantification. However, the manual setting of the item ID has a significant impact on the final result. For example, the obtained user interaction function curve fluctuates too much, the information is confusing, and it is difficult to extract effective function features, thereby destroying the potential time characteristics.

[0006] (2) The timeline is too sparse, and the amount of user interaction data is difficult to ensure that it can support dense and uniform function needs, and the user's interaction with a single item may also be sparse.

[0007] It can be seen that the existing methods cannot effectively and accurately capture and quantify the changing relationship between users' demand for projects and time.

[0008] In addition, a reasonable interactive graph structure needs to be designed in the graph neural network recommendation method, because the design of the graph structure directly affects the final model performance. An overly simple graph structure may make it impossible to extract deep features, while an overly complex graph structure may lead to overfitting and a significant increase in computational overhead. Summary of the Invention

[0009] The present invention aims to at least solve the technical problem in the prior art that it is impossible to effectively and accurately capture and quantify the changing relationship between the degree of user demand for projects over time, which affects the accuracy of recommendations, and provide a graph convolution recommendation method and device that takes into account the relationship between user-project demand.

[0010] In order to achieve the above-mentioned purpose of the present invention, according to the first aspect of the present invention, the present invention provides a graph convolution recommendation method that considers the user-project demand relationship, including: encoding the user's prediction timestamp to obtain the user's time coding vector, wherein the time coding vector includes an hourly time feature vector, a weekly time feature vector, and a quarterly time feature vector; inputting the user's time coding vector into a pre-trained user interest model corresponding to the user to obtain a confidence score of the user's interest level in each project category at the prediction timestamp; generating graph structure data according to the user-project interaction record; inputting the graph structure data into a pre-trained graph convolutional network model to obtain the user's basic prediction score for each project; adjusting the user's basic prediction score for the project using the user's interest level confidence score for the project category to which each project belongs at the prediction timestamp to obtain the user's adjusted prediction score for each project; sorting the user's adjusted prediction scores for the projects from high to low, and recommending the first K1 projects in the sorted sequence to the user, where K1 is a positive integer.

[0011] To achieve the above-mentioned object of the present invention, according to a second aspect of the present invention, a graph convolution recommendation device that considers user-project demand relationships is provided, which is used to implement the graph convolution recommendation method that considers user-project demand relationships described in the first aspect of the present invention, comprising: a time encoding module that encodes a user's predicted timestamp to obtain a user's time encoding vector, wherein the time encoding vector includes an hourly time feature vector, a weekly time feature vector, and a quarterly time feature vector; a user interest model module that inputs the user's time encoding vector into a pre-trained user interest model corresponding to the user to obtain a confidence score of the user's interest level in each project category at the predicted timestamp; a graph structure data generation module that generates graph structure data based on user-project interaction records; a basic prediction score acquisition module that inputs the graph structure data into a pre-trained graph convolutional network model to obtain the user's basic prediction score for each project; a score adjustment module that adjusts the user's basic prediction score for the project using the user's interest level confidence score for the project category to which each project belongs at the predicted timestamp to obtain the user's adjusted prediction score for each project; and a recommendation module that sorts the user's adjusted prediction scores for the projects from high to low and recommends the top K1 projects in the sorted sequence to the user, where K1 is a positive integer.

[0012] The present invention encodes the timestamp that the user needs to predict to obtain a time coding vector, and the time coding vector includes an hourly time feature vector, a weekly time feature vector, and a quarterly time feature vector, which reflects the periodic relationship. The time coding vector is input into a trained user interest model, and the user interest model is used to adaptively capture the user's demand cycle for the project category, and obtain the user's interest confidence score for each project category at the prediction timestamp, and quantify the change relationship between the user's demand for the project category over time; the trained graph convolutional network model is used to analyze the graph structure data to obtain the user's basic prediction score for each project; the user's interest confidence score for each project category at the prediction timestamp is used to adjust the user's basic prediction score for each project under the project category, so that the obtained adjusted prediction score adds the time change of the user's demand relationship for the project, thereby improving the recommendation accuracy of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 1 is a flow chart of a graph convolutional recommendation method that considers the relationship between user-item requirements in a preferred embodiment of the present invention;

[0014] Figure 2 is a framework diagram of a user interest model in a preferred embodiment of the present invention;

[0015] Figure 3 It is a framework diagram of a graph convolutional network model in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0016] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0017] In the description of the present invention, it should be understood that the terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention.

[0018] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal communication between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0019] The execution subject of the graph convolution recommendation method that takes into account the user-project demand relationship provided by the present invention includes but is not limited to at least one of the electronic devices such as the server and the terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the graph convolution recommendation method that takes into account the user-project demand relationship can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] The present invention discloses a graph convolution recommendation method that considers the relationship between user-item requirements. In a preferred embodiment, referring to Figure 1 As shown, the following steps are included:

[0021] Step A includes:

[0022] Step A1: Encode the predicted timestamp of the user to obtain the time encoding vector of the user. The time encoding vector includes an hourly time feature vector, a weekly time feature vector, and a quarterly time feature vector.

[0023] In step A2, the user's time encoding vector is input into a pre-trained user interest model corresponding to the user to obtain a confidence score of the user's interest level in each item category at the predicted timestamp. Each user has a pre-trained user interest model.

[0024] Step B comprises:

[0025] Step B1: Generate graph-structured data based on user-item interaction records. User-item interaction records include interaction records between multiple users and multiple items. Specifically, they include information such as the timestamp, item category, and item number of each user's interaction with each item. The interaction record between a user u∈U and an item i∈I at timestamp t is a triple (u, i, t). For each user u, sort its interaction records with the items in ascending chronological order, obtaining the following data results:

[0026]

[0027] where i1,i2,…,i q are the items that user u has interacted with, t1, t2,…, t q is the interaction timestamp for .

[0028] In step B2, the graph structure data is input into the pre-trained graph convolutional network model to obtain the user's basic predicted score for each item.

[0029] Step C adjusts the user's basic predicted score for each item using the confidence score of the user's interest in the item category at the prediction timestamp to obtain an adjusted predicted score for each item. Specifically, the adjustment method is preferably, but not limited to, multiplying or adding the user's basic predicted score for the item by the confidence score of the item category to obtain the adjusted predicted score.

[0030] In step D, the user's adjusted predicted ratings for the items are sorted from high to low, and the top K1 items in the sorted order are recommended to the user, where K1 is a positive integer. Specifically, for user u, after executing steps AD above, the user's adjusted predicted ratings for all items are obtained. The top K1 items with the highest predicted ratings are extracted as a recommendation set and provided to user u.

[0031] In this embodiment, Figure 1An execution order of the steps is given, namely, steps A, B, C, and D. It should be noted that the execution order of the steps can also be steps B, A, C, and D, or steps A and B can be executed in parallel before steps C and D.

[0032] In this embodiment, in order to simply and effectively encode the predicted timestamp or timestamp into a time code vector with a periodic relationship, a timestamp encoding method is provided. Preferably, the timestamp encoding method in step A1 includes:

[0033] Step A01: round down the quotient of the user's predicted timestamp t divided by the first conversion coefficient T0 to obtain an integer time value f trans (t):

[0034]

[0035] Among them, f trans (t) is an integer in hours; T0>0.

[0036] Step A02: take the remainder of the integer time value divided by 24 as the hour time feature hour; round down the quotient of the integer time value divided by 24 to obtain the first intermediate value Take the remainder of the first middle value divided by 7 as the day of week feature dayOfWeek; round down the quotient of the integer time value divided by 2184 to obtain the second middle value Take the remainder of the second median value divided by 4 as the quarterly time feature quarter. Specifically:

[0037] hour=f trans (t)mod24

[0038]

[0039] dayOfWeek and quarter do not actually refer to the actual day of the week or quarter, but are a numbering system. Since these numbers can be aligned with real-world values ​​through sequential shifting, they also have potential characteristics.

[0040] Step A03, respectively perform vector conversion on the hour time feature hour, the week time feature dayOfWeek, and the quarter time feature quarter to obtain the hour time feature vector H hour , Weekly time feature vector H dayOfWeek , quarterly time feature vector H quarter It is preferred but not limited to use one-hot encoding to convert hourly time features, weekly time features, and quarterly time features into vector representations:

[0041] Hhour =one-hot(hour,24)

[0042] H dayOfWeek =one-hot(dayOfWeek,7)

[0043] H quarter =one-hot(quarter,4)

[0044] Step A04: Concatenate the hourly time feature vector, the weekly time feature vector, and the quarterly time feature vector to obtain the user's time coding vector H t :

[0045] H t =[H hour ;H dayOfWeek ;H quarter ].

[0046] In this embodiment, the user interest model preferably, but not limited to, utilizes an existing multi-layer perceptron structure, comprising a cascaded input layer, one or more fully connected layers, and an output unit. The input layer is used to convert the user's time-coded vector into an input representation; one or more fully connected layers learn the input representation to obtain hidden layer representations, with each fully connected layer acting as a hidden layer; and the output unit converts the hidden layer representation output by the last fully connected layer into a confidence score of the user's interest level in all item categories at timestamp t. The training process for each user's user interest model is as follows: extracting the user's item category-timestamp sequence in ascending chronological order from the user's item interaction records; encoding the timestamps in the item category-timestamp sequence to obtain a time-coded vector sequence; using the time-coded vector sequence as training data and the item category codes in ascending chronological order in the item category-timestamp sequence as the true labels; constructing a binary cross-entropy loss function; and using the training data to train the user interest model for that user until the binary cross-entropy loss function converges (the change is less than a preset change threshold). Upon completion of training, a trained user interest model for that user is obtained.

[0047] In this embodiment, the graph structure data may include a user-project interaction graph, where a node set in the user-project interaction graph includes user nodes and project nodes, and an edge exists between the interacting user nodes and project nodes. The graph convolutional network model is preferably, but not limited to, an existing GCN network model, i.e., a graph convolutional neural network model. Reference may also be made to the technical solution for obtaining project scores using a graph convolutional network model disclosed in Chinese Patent Publication No. CN116070034A, which will not be further described here.

[0048] In a preferred embodiment, in order to more accurately predict the item categories that the user may be interested in at a specific time stamp, and to have a simple and efficient learning effect, refer to Figure 2 As shown, the user interest model combines the time identity auxiliary matrix to extract the user's demand cycle characteristics. In this embodiment, the user interest model adopts a multi-layer perceptron structure, including an input layer, one or more fully connected layers and an output unit;

[0049] The input layer combines the temporal identity auxiliary matrix and the temporal encoding vector of user u Get input representation

[0050]

[0051] Where t represents the timestamp; the time identity auxiliary matrix includes the hour identity auxiliary matrix and the week identity auxiliary matrix ∥ indicates matrix concatenation operation; input indicates is the input feature of the first fully connected layer. and It can be initialized by Gaussian distribution, and updated by back propagation of the objective function by minimizing the user interest loss function during the training process of the user interest model. More than one fully connected layer learns the input representation to obtain the hidden layer representation. Assuming that the user interest model includes L layers (1 input layer and L-1 fully connected layers), the hidden layer representation of the output of the fully connected layer at the l+1 layer is for:

[0052]

[0053] in, represents the hidden layer representation or input representation of the l-th layer output; σ l+1 represents the activation function of the fully connected layer at the l+1th layer; FC l+1 represents the fully connected layer at the l+1th layer.

[0054] The output unit converts the hidden layer representation output by the last fully connected layer into a confidence score of the user's interest in each item category at timestamp t. The output unit is an activation function layer, preferably but not limited to a Sigmoid function, and outputs the confidence vector of the user's interest in the item category at timestamp t, such as the confidence vector of the user u's interest in the item category at timestamp t

[0055]

[0056] The hidden layer representation representing the L-th layer output of the user interest model. includes the confidence scores of user u’s interest in all item categories at timestamp t.

[0057] In a preferred embodiment, in order to encourage the user interest model to learn the similarities between project categories, it is further possible to accurately predict the project categories that the user may be interested in at a specific time stamp, and to have a simple and efficient learning effect. Figure 2 , the training method of the user interest model corresponding to user u includes:

[0058] Step 01: Based on the user-project interaction records, obtain the project category-timestamp sequence of user u in ascending order of time.

[0059]

[0060] Among them, t1, t2, t q They represent the timestamps of the interaction between user u and the project in the user-project interaction record; c1, c2, c q Respectively represent user u at timestamps t1, t2, t q The items interacted during the interaction belong to the item category, and each item category has more than one item; U represents the user set obtained through user-item interaction records.

[0061] Step 02: Encode the timestamps in the item category-timestamp sequence to obtain a time coding vector sequence:

[0062] Step 03: Build and initialize the user interest model. Build a network for the user interest model based on the multi-layer perceptron network structure and initialize the parameters of the fully connected layer in the user interest model.

[0063] Step 4: Train the user interest model using the temporal encoding vector sequence with the goal of minimizing the user interest loss function to obtain a user interest model corresponding to user u. During the training process, gradient descent is used to update the fully connected layer parameters in the user interest model and optimize the temporal identity auxiliary matrix.

[0064] In this embodiment, the user interest loss function is:

[0065]

[0066] Where T represents the timestamp set of the time-coded vector sequence; y u,tRepresents the real interest level vector of user u in the item category at timestamp t. The real interest level vector includes the real interest level of user u in all item categories at timestamp t. Specifically, based on the item category-timestamp sequence of user u, the real interaction item category of user u at each timestamp is obtained. The real interest level of the item category that user u actually interacted with at timestamp t is set to 1, and the real interest level of the item category that user u did not actually interact with at timestamp t is set to 0. At this time, the real interest level vector of user u in the item category at timestamp t is a sequence of the real interest levels (0 or 1) of all item categories. L represents the confidence vector of the user u's interest in the project category at timestamp t output by the user interest model. The confidence vector includes the confidence scores of the user u's interest in all project categories at timestamp t. BCE (,) represents the binary cross entropy loss function; λ represents the regularization coefficient.

[0067] And the regularization term R of user u at timestamp t u,t for:

[0068]

[0069] i and j both represent the project category index; Respectively The confidence score of user u’s interest in the i-th item category and the j-th item category at timestamp t; d H (c i ,c j ) means to obtain the item category c i and project category c j The encoded Hamming distance, |C| represents the number of item categories contained in the item category set C.

[0070] In a preferred embodiment, in order to enable the coding of project categories to ensure that similar project categories have a closer distance, the present invention provides a project category coding method, the method comprising:

[0071] Step a, constructing a project category similarity matrix based on user-project interaction records. Further preferably, it includes:

[0072] Step a01: Obtain the category similarity between any two item categories based on the user-item interaction records. i and project category c j The category similarity between is:

[0073]

[0074] in, Indicates the user-item interaction record with item category c i The set of users who have interacted; Indicates the user-item interaction record with item category c j The set of users who have interacted; Indicates that the interaction record is also related to the project category c i and project category c j The number of users who have interacted; Indicates the interaction record with item category c i The number of users who have interacted.

[0075] Step a02: The category similarities between all item categories form the item category similarity matrix S∈R |C|×|C| , the size of the item category similarity matrix is ​​|C|×|C|. |C| represents the size of the item category set.

[0076] Step b: perform eigenvalue decomposition on the item category similarity matrix to obtain the eigenvector matrix U and the eigenvalue diagonal matrix Σ: S = UΣU T . Among them, U∈R |C|×|C| ;Σ∈R |C|×|C| .

[0077] Step c: Select the first K2 largest eigenvalues ​​from the eigenvalue diagonal matrix Σ and the corresponding eigenvectors in the eigenvector matrix U to form a pseudo matrix U k ∈R |C|×K2 , each item category in the pseudo matrix includes K2-dimensional elements, where K2 is a positive integer. That is, K2 eigenvalues ​​are selected from the eigenvalue diagonal matrix Σ in descending order, and the K2 eigenvectors corresponding to these K2 eigenvalues ​​are extracted from the eigenvector matrix U to form the pseudo matrix U k .

[0078] Step d, for the pseudo matrix U k Normalization is performed, and each element of the normalized pseudo-matrix is ​​mapped to the range of {0, 1} using a sign function. After the mapping is completed, each item category is encoded into a K2-dimensional binary code, with each dimension being 0 or 1.

[0079] Normalized pseudo matrix The element at position (i, j) in the row and column for:

[0080]

[0081] The sign function sign(.) maps each element of the normalized pseudo-matrix to the range of {0, 1} in the following way: Mapping to H c,ij :

[0082]

[0083] In this embodiment, through steps b to d, the item categories are mapped to the vertices of a hypercube with K2 dimensions.

[0084] In a preferred embodiment, referring to Figure 3 As shown in the user-item interaction graph of the recommendation system, the information carried by the nodes is their IDs, and there is no semantic information other than the identification function. In order to enable the graph convolution process to capture features from distant nodes, preferably, in step B1, generating graph structure data based on user-item interaction records includes:

[0085] Step B11: extracting user set U, item set I and user-item interaction graph based on user-item interaction records. The nodes of the user-item interaction graph include user nodes and item nodes. There are edges connecting the interacting user nodes and item nodes.

[0086] Step B12, obtain the adjacency matrix A∈R of the user-item interaction graph (|U|+|I|)×(|U|+|I|) , and standardize the adjacency matrix to obtain the standard adjacency matrix |U| represents the number of users included in the user set U. |I| represents the number of items included in the item set I.

[0087]

[0088] Where D represents the degree matrix of the adjacency matrix A, Represents the inverse matrix of the arithmetic square root of the degree matrix D. The above normalization formula can avoid the numerical scaling problem in the graph convolution process.

[0089] Step B13: Calculate the user similarity between any two users and the project similarity between any two projects based on the standard adjacency matrix.

[0090] The user similarity between user u and user v is calculated as follows:

[0091]

[0092] The item similarity between items m and n is calculated as follows:

[0093]

[0094] in, Represents the standard adjacency matrix The row vector of the row where user u is located; Represents the standard adjacency matrix The row vector of the row where user v is located; Represents the standard adjacency matrix The column vector of the column where the item m is located; Represents the standard adjacency matrix The column vector of the column where item n is located; ∥·∥ means finding the L2 norm of the orientation vector.

[0095] Step B14, select the top-K3 most similar pairs of nodes among all user similarities and all project similarities, add similarity edges between each pair of nodes in the top-K3 pairs of nodes, and generate a user similarity graph and a project similarity graph. Specifically, K3 is a positive integer, and the user similarity of all user node pairs in the user set is calculated, and the project similarity of all project node pairs in the project set is calculated. K3 pairs of nodes are selected from the largest to the smallest according to the similarity values ​​among all user similarities and all project similarities. K3 pairs of nodes may include both user node pairs and project node pairs. After adding similarity edges to the user node pairs in the K3 pairs of nodes, a user similarity graph is formed, and after adding similarity edges to the project node pairs in the K3 pairs of nodes, a project similarity graph is formed. Figure 3 shown.

[0096] Step B15, as Figure 3 As shown in FIG, the user-item interaction graph, the user similarity graph, and the item similarity graph constitute graph structure data.

[0097] In this implementation, in order to enable the graph convolution process to capture features from distant nodes, similar edges between similar users and similar items are designed and added. Figure 3 , we obtain user similarity graphs and item similarity graphs, which can help graph convolutional networks extract feature information from similar nodes faster and more accurately.

[0098] In a preferred embodiment, in step B2, the graph structure data is input into a pre-trained graph convolutional network model to obtain the user's basic predicted score for each item, including:

[0099] Step B21: Initialize the embedding vectors of each user and item in the graph structure data using the standard normal distribution:

[0100] The embedding vector of user u is initialized as:

[0101] The embedding vector of item m is initialized as:

[0102] Step B22: Input the graph structure data of the initialized user embedding vector and the item embedding vector into the graph convolutional network model, and the graph convolutional network model obtains the feature representation of the user and the item respectively; wherein, the feature representation of user u is e u for:

[0103] Feature representation e of item m m for:

[0104] Step B23, based on the distance relationship between the feature representation vectors of the user and the item, calculate the basic prediction score of the user for each item, where the basic prediction score of user u for item m is for:

[0105] in, Represents the embedded representation of user u output by the kth layer of the graph convolutional network model, when k is greater than 0:

[0106]

[0107] The embedded representation of item m output by the kth layer of the graph convolutional network model:

[0108]

[0109] Among them, K represents the maximum index of the graph convolutional network model layer; U represents the user set; I represents the item set; α k Represents the weight coefficient of the kth layer of the graph convolutional network model; N u Represents the set of neighbor nodes of user u in the graph structure data; N m Represents the set of neighbor nodes of item m in the graph structure data; S u The set of similar user nodes (i.e., connected neighbor user nodes) of user u in the user similarity graph representing graph structure data; S v The set of similar user nodes (i.e., connected neighbor user nodes) of user v in the user similarity graph representing graph structure data; S m The set of similar project nodes (i.e., connected neighboring project nodes) of project m in the project similarity graph representing graph structure data; S n A set of similar item nodes (i.e., connected neighbor item nodes) of item n in an item similarity graph representing graph structured data.

[0110] In this embodiment, the training process of the graph convolutional network model may include: obtaining user-project interaction records corresponding to different historical timestamps, generating multiple graph structure data according to the above-mentioned graph structure data generation method and serving as training sets; using user-project interaction records corresponding to different historical timestamps as real label data; constructing the network structure of the above-mentioned graph convolutional network model and initializing the network parameters; and using the training set and real label data to perform supervised training on the initialized graph convolutional network model.

[0111] In this implementation, similar users and items are analyzed, similar edges between nodes are added based on the degree of similarity, and the graph convolution process is explicitly guided to more quickly extract the potential feature representations of users and items. This structural design avoids situations where a small number of convolutions fail to obtain high-order information about the graph structure, and where multiple convolutions extract overly cluttered feature information. Furthermore, it can, to a certain extent, reduce the long-tail distribution problem in recommendation systems where unpopular items are not noticed. The simplified network structure and computational approach can maintain high computational performance even with large amounts of data.

[0112] In a preferred embodiment, in step C, user u's adjusted predicted score for item m is for:

[0113]

[0114] in, represents the basic prediction score of user u on item m; γ represents the enhancement adjustment factor, γ≥1; β represents the suppression adjustment factor, β∈[0,1]; represents the confidence score of user u’s interest in the i-th item category to which item m belongs at timestamp t.

[0115] In this embodiment, due to the correlation between periodic functions and trigonometric functions and the common sense laws of social operation, the adjustment method provided in this embodiment assumes that user interest follows a triangular variation pattern, where the peak value of the function corresponds to an interest confidence score of 1 for the project category, and the lowest value corresponds to a score of 0. Specifically, when the confidence score of user u's interest in the project category to which the project belongs is close to 1, it indicates that user u currently has a high interest in projects in that category. Therefore, the basic prediction score is enhanced by multiplying it by an enhancement adjustment factor γ greater than 1. Conversely, when the confidence score of user u's interest in the project category to which the project belongs is close to 0, it indicates that user u currently has a low interest in projects in that category. In this case, the basic prediction score is suppressed by multiplying it by a suppression adjustment factor β less than 1. By introducing the minimum and maximum adjustment factors β and γ, the magnitude of the score adjustment can be flexibly controlled to avoid excessive penalties or rewards.

[0116] The present invention also discloses a graph convolution recommendation device that considers the user-item requirement relationship, which is used to implement the above-mentioned graph convolution recommendation method that considers the user-item requirement relationship. In a preferred embodiment, the device includes:

[0117] A time encoding module encodes the user's predicted timestamp to obtain the user's time encoding vector, wherein the time encoding vector includes an hourly time feature vector, a weekly time feature vector, and a quarterly time feature vector;

[0118] The user interest model module inputs the user's time encoding vector into the pre-trained user interest model corresponding to the user to obtain the confidence score of the user's interest level in each item category at the predicted timestamp;

[0119] Graph structure data generation module, which generates graph structure data based on user-project interaction records;

[0120] The basic prediction score acquisition module inputs graph structure data into a pre-trained graph convolutional network model to obtain the user's basic prediction score for each project;

[0121] The score adjustment module adjusts the user's basic prediction score for each item using the user's confidence score of the item category to which the item belongs at the prediction timestamp, and obtains the user's adjusted prediction score for each item;

[0122] The recommendation module sorts the user's adjusted prediction scores for the items from high to low, and recommends the first K1 items in the sorted sequence to the user, where K1 is a positive integer.

[0123] In this embodiment, the time coding module, user interest model module, graph structure data generation module, basic prediction score acquisition module, score adjustment module, and recommendation module correspond to step A1, step A2, step B1, step B2, step C, and step D respectively, and will not be repeated here.

[0124] The present invention takes into account the changes in the user's demand relationship for projects over time and designs a user interest model. A method for encoding and converting timestamp data is designed and trained to reduce the messy multi-dimensional timestamp data to feature encoding with a periodic relationship. A K2-dimensional cube mapping method for encoding and converting project categories is designed and trained to ensure that similar project categories are closer, thereby ensuring that similar project categories have a closer distance. Based on the user's historical interaction records, the interaction time and project category are feature encoded. Combined with the time identity matrix and the MLP model in the user interest model, it is possible to adaptively capture the user's demand cycle for projects, thereby realizing the prediction of project categories that users may be interested in at specific times.

[0125] This paper designs a graph-structured convolutional network. By analyzing similar users and similar projects, similar edges between nodes are added according to similar weights, and the graph convolution process is explicitly guided to more quickly extract the potential feature representations of users and projects. This structural design can avoid the situation where a small number of convolutions cannot obtain high-order information of the graph structure and multiple convolutions will extract too much cluttered feature information. It can also, to a certain extent, reduce the long-tail distribution problem in the recommendation system where unpopular projects are not paid attention to. The simplified network structure and calculation method can maintain high computing performance even with large-scale data volumes.

[0126] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "example," "specific example," "one implementation," "a preferred implementation," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0127] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A graph convolution recommendation method considering the relationship between user-item requirements, characterized by: include: Encoding the predicted timestamp of the user to obtain a time coding vector of the user, wherein the time coding vector includes an hourly time feature vector, a weekly time feature vector, and a quarterly time feature vector; Input the user's time encoding vector into the pre-trained user interest model corresponding to the user to obtain the confidence score of the user's interest level in each item category at the predicted timestamp; generate graph structure data based on the user-item interaction records; Input the graph structure data into the pre-trained graph convolutional network model to obtain the user's basic prediction score for each item; The user's confidence score of the project category to which each project belongs at the prediction timestamp is used to adjust the user's basic prediction score for the project, thereby obtaining the user's adjusted prediction score for each project; Sort the user's adjusted predicted scores of the items from high to low, and recommend the first K1 items in the sorted sequence to the user, where K1 is a positive integer; The user interest model adopts a multi-layer perceptron structure. The training method of the user interest model corresponding to user u includes: Based on the user-project interaction records, obtain the project category-timestamp sequence of user u in ascending time order; Encode the timestamps in the item category-timestamp sequence to obtain a time coding vector sequence; Build and initialize user interest model; The user interest model corresponding to user u is trained using the temporal coding vector sequence with the goal of minimizing the user interest loss function. Among them, the user interest loss function is: Where T represents the timestamp set of the time-coded vector sequence; y u,t Represents the real interest level vector of user u in the item category at timestamp t. The real interest level vector includes the real interest level of user u in all item categories at timestamp t. Specifically, based on the item category-timestamp sequence of user u, the real interaction item category of user u at each timestamp is obtained. The real interest level of the item category that user u actually interacted with at timestamp t is set to 1, and the real interest level of the item category that user u did not actually interact with at timestamp t is set to 0. The obtained real interest level vector of user u in the item category at timestamp t is a sequence of the real interest levels of all item categories. L represents the confidence vector of the user u's interest in the project category at timestamp t output by the user interest model. The confidence vector includes the confidence scores of the user u's interest in all project categories at timestamp t. BCE (,) represents the binary cross entropy loss function; λ represents the regularization coefficient; The regularization term R of user u at timestamp t u,t for: i and j both represent the project category index; Respectively The confidence score of user u’s interest in the i-th item category and the j-th item category at timestamp t; d H (c i ,c j ) means to obtain the item category c i and project category c j The encoded Hamming distance, |C| represents the number of item categories contained in the item category set C.

2. The graph convolution recommendation method considering the user-item requirement relationship according to claim 1, characterized in that: The step of encoding the predicted timestamp of the user to obtain the time coding vector of the user includes: Rounding down the quotient of the user's predicted timestamp divided by the first conversion coefficient to obtain an integer time value; Take the remainder of the integer time value divided by 24 as the hour time feature; Round down the quotient of the integer time value divided by 24 to obtain the first middle value, and take the remainder of the first middle value divided by 7 as the day of the week feature; Round down the quotient of the integer time value divided by 2184 to obtain the second middle value, and take the remainder of the second middle value divided by 4 as the quarterly time feature; Perform vector conversion on hourly time features, weekly time features, and quarterly time features respectively to obtain hourly time feature vectors, weekly time feature vectors, and quarterly time feature vectors; The user's time coding vector is obtained by concatenating the hourly time feature vector, the weekly time feature vector, and the quarterly time feature vector.

3. The graph convolution recommendation method considering the user-item requirement relationship according to claim 1, characterized in that: The user interest model includes an input layer, one or more fully connected layers and an output unit; The input layer combines the temporal identity auxiliary matrix and the temporal encoding vector of user u. Get input representation Where t represents the time stamp; the time identity auxiliary matrix includes the hour identity auxiliary matrix and the week identity auxiliary matrix || represents matrix concatenation operation; The one or more fully connected layers learn the input representation to obtain hidden layer representations respectively; The output unit converts the hidden layer representation output by the last fully connected layer into a confidence score of the user's interest in each item category at timestamp t.

4. The graph convolution recommendation method considering the user-item requirement relationship according to claim 1, characterized in that Methods for obtaining project category codes include: Constructing an item category similarity matrix based on user-item interaction records; Perform eigenvalue decomposition on the item category similarity matrix to obtain the eigenvector matrix and eigenvalue diagonal matrix; Select the first K2 largest eigenvalues ​​from the eigenvalue diagonal matrix and the eigenvectors in the eigenvector matrix corresponding to them to form a pseudo-matrix, wherein each item category in the pseudo-matrix includes K2-dimensional elements, where K2 is a positive integer; The pseudo matrix is ​​normalized, and each element of the normalized pseudo matrix is ​​mapped to the range of {0, 1} using a sign function. After the mapping is completed, each item category is encoded into a K2-dimensional binary code.

5. The graph convolution recommendation method considering the user-item requirement relationship according to claim 4, characterized in that: The constructing of the project category similarity matrix based on the user-project interaction records includes: Based on the user-item interaction records, the category similarity between any two item categories is obtained, and the item category c i and project category c j The category similarity between is: in, Indicates the user-item interaction record with item category c i The set of users who have interacted; Indicates the user-item interaction record with item category c j The set of users who have interacted; The category similarities between all item categories constitute the item category similarity matrix, and the size of the item category similarity matrix is ​​|C|×|C|.

6. The graph convolution recommendation method considering the user-item requirement relationship according to claim 1, 2, 3, 4, or 5, characterized in that: Generating graph structure data according to user-project interaction records includes: Extracting user sets, item sets, and a user-item interaction graph based on user-item interaction records, wherein the nodes of the user-item interaction graph include user nodes and item nodes; Obtain the adjacency matrix of the user-item interaction graph and standardize the adjacency matrix to obtain the standard adjacency matrix; Calculate the user similarity between any two users and the item similarity between any two items based on the standard adjacency matrix; Select the top-K3 most similar pairs of nodes from all user similarities and all item similarities, add similarity edges between each pair of nodes in the top-K3 pairs of nodes, and generate user similarity graph and item similarity graph; The user-item interaction graph, user similarity graph, and item similarity graph constitute graph structure data.

7. The graph convolution recommendation method considering the user-item requirement relationship according to claim 6, characterized in that: The graph structure data is input into the pre-trained graph convolutional network model to obtain the user's basic prediction score for each item, including: The graph convolutional network model obtains feature representations of users and items respectively; Among them, the feature representation of user u is e u for: Feature representation e of item m m for: Calculate the user's basic prediction score for each item, where user u's basic prediction score for item m is for: in, Represents the embedded representation of user u output by the kth layer of the graph convolutional network model, when k is greater than 0: The embedding representation of item m output by the kth layer of the graph convolutional network model: Among them, K represents the maximum index of the number of layers of the graph convolutional network model; U represents the user set; I represents the item set; α k Represents the weight coefficient of the kth layer of the graph convolutional network model; N u Represents the set of neighbor nodes of user u in the graph structure data; N m Represents the set of neighbor nodes of item m in the graph structure data; S u The set of similar user nodes of user u in the user similarity graph representing graph structure data; S v The set of similar user nodes of user v in the user similarity graph representing graph structure data; S m S represents the set of similar project nodes of project m in the project similarity graph of graph structure data; n A set of similar item nodes for item n in an item similarity graph representing graph-structured data.

8. The graph convolution recommendation method considering the user-item requirement relationship according to any one of claims 1, 2, 3, 4, 5, or 7, characterized in that: User u's adjusted predicted rating for item m for: in, represents the basic prediction score of user u on item m; γ represents the enhancement adjustment factor, γ≥1; β represents the suppression adjustment factor, β∈[0,1]; represents the confidence score of user u’s interest in the i-th item category to which item m belongs at timestamp t.

9. A graph convolution recommendation device that considers user-item requirement relationships, used to implement the graph convolution recommendation method that considers user-item requirement relationships according to any one of claims 1-8, characterized in that: include: A time encoding module encodes the user's predicted timestamp to obtain the user's time encoding vector, wherein the time encoding vector includes an hourly time feature vector, a weekly time feature vector, and a quarterly time feature vector; The user interest model module inputs the user's time encoding vector into the pre-trained user interest model corresponding to the user to obtain the confidence score of the user's interest level in each item category at the predicted timestamp; Graph structure data generation module, which generates graph structure data based on user-project interaction records; The basic prediction score acquisition module inputs graph structure data into a pre-trained graph convolutional network model to obtain the user's basic prediction score for each project; The score adjustment module adjusts the user's basic prediction score for each item using the user's confidence score of the item category to which the item belongs at the prediction timestamp, and obtains the user's adjusted prediction score for each item; The recommendation module sorts the user's adjusted prediction scores for the items from high to low, and recommends the first K1 items in the sorted sequence to the user, where K1 is a positive integer.

Images