Attribute fusion interactive recommendation model construction method and system based on enhanced graph convolution
By adopting enhanced graph convolution and attribute fusion interaction model in the recommendation system, the problem of insufficient evaluation of comment importance is solved, and high-order collaboration signals between user items are captured, which improves the accuracy and robustness of recommendations, especially in Excellent performance in negative user recommendations.
Patent Information
- Application Number
- CN202210106409.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-28
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-01-28
Smart Images

Figure CN114676315B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of information processing technology, and in particular to a method and system for constructing an attribute fusion interactive recommendation model based on enhanced graph convolution. Background Art
[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
[0003] The purpose of the recommendation system is to recommend relevant items to users. So far, the recommendation system has achieved great success in many applications. These applications collect a large number of user-item interaction records, which provides unprecedented opportunities for the recommendation system to achieve accurate recommendations, thereby greatly reducing the complexity of user decision-making. Comments are the most common displayed feedback, which represents the most realistic interactive experience of users on items. Some studies focus on using the comment text posted by users, extracting feature information from it, predicting scores for users and items, and making recommendations.
[0004] In recent years, in the field of rating prediction, review text has become an effective auxiliary information for modeling users and items. People use various deep learning techniques to learn potential feature information from text. For example, ConvMF and DeepCoNN use convolutional neural networks to extract feature information from review text to model users and items; MPRS concatenates user-item reviews and forms an interaction matrix, and uses convolutional neural networks to capture the interaction features between user and item reviews; TARMF uses a recurrent neural network to obtain the feature representation of each review text; MPCN uses a gating mechanism similar to a recurrent neural network, and is expected to capture important information in the text before modeling the review features. NARRE believes that different reviews have different importance, so it proposes a review-level attention mechanism to characterize the weight of each review. Acapnet and CARP use capsule networks to extract important information from different aspects of review text at multiple levels. The overall performance of these methods is satisfactory.
[0005] The research topic of graph neural network has received increasing attention in recent years. In recent years, researchers have designed a neural network structure for processing graph data, namely graph neural network, by drawing on the ideas of convolutional networks, recurrent networks and deep autoencoders. Furthermore, researchers have combined graph convolutional networks with traditional collaborative filtering techniques and proposed a series of recommendation models using rich structural information. For example, Wang et al. designed a neural graph collaborative filtering (NGCF) model to extract rich side information from the user-item interaction graph through graph neural networks to better learn the embedded representation of users and items. Wang et al. used graph convolutional networks to automatically mine the high-order structural information and semantic information of corresponding items in the knowledge graph and capture the user's potential remote interests, thereby recommending items that the user is more interested in. He et al. found that the feature changes and nonlinear activation of GCN (graph convolutional neural network) are not conducive to improving the recommendation performance. Then they simplified the design of GCN components and proposed a new model called LightGCN, which only contains the neighborhood aggregation component in GCN for collaborative filtering. Chen et al. found that removing the nonlinear structure in GCN can improve the recommendation performance. Therefore, the above-mentioned prior art proposed a residual network structure that can alleviate the over-smoothing problem in the multi-layer graph convolution aggregation process.
[0006] The rating prediction models mentioned above all try to use neural networks to accurately capture the important feature information of the review text, and have achieved good results. Most models ignore the fact that not every review can reflect the characteristics of the user (item). Some models are aware of this, but they only assign smaller weights to useless reviews by paying attention to the review text. However, this method has limitations. It is not convincing to judge the importance of the review text only by text features, and attention can only evaluate the usefulness of reviews in historical data, but cannot be associated with reviews that are not involved in training. Although the above graph convolutional neural network can capture high-order connectivity signals between user items, it is static when generating the user-item bipartite graph. In other words, they do not consider the temporal relationship between node neighbors. Therefore, it cannot be directly used in the recommendation model. Summary of the invention
[0007] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a method and system for constructing an attribute fusion interactive recommendation model based on enhanced graph convolution, which is used to solve the rating prediction task.
[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:
[0009] In the first aspect, a method for constructing an attribute fusion interactive recommendation model based on enhanced graph convolution is disclosed, including:
[0010] Construct a relational encoder based on enhanced graph convolution to extract high-level features between users and items from user-item interaction behaviors, and obtain dynamic feature representations of users and items that contain high-level collaborative signals at each moment;
[0011] A comment encoder that integrates attribute information is constructed to process and fuse the comment text and its corresponding attributes to obtain an accurate feature representation of the comment.
[0012] A further technical solution is to construct a relational encoder based on enhanced graph convolution as follows:
[0013] Construct a user-item bipartite graph as nodes through the interaction between users and items;
[0014] The self-attention aggregator fuses the temporal information of the interaction nodes in the user-item bipartite graph, and the user-item attention aggregator aggregates the neighbor nodes of users or items to learn the embedding representation of users and items at each moment;
[0015] Use enhanced graph convolution to perform multi-layer graph convolution on the user-item bipartite graph to learn the high-order signals of each node in the user-item bipartite graph;
[0016] Based on the learned embedding representations of users and items at each moment and the high-order signals of each node in the user-item bipartite graph, the dynamic feature representation of users and items at each moment containing high-order collaborative signals is obtained.
[0017] Further technical solutions, the steps of constructing a comment encoder integrating attribute information are as follows:
[0018] Use dilated convolution to extract review text features;
[0019] Processing attribute information;
[0020] The review text features are combined with attribute information, and then a feature interaction aggregation module based on self-supervised learning is used to judge the usefulness of the review and obtain an accurate feature representation of the review.
[0021] A further technical solution is that before constructing the user-item bipartite graph, a set of comment data corresponding to each user and item is required, including: a set of comments with a number of p published by user u and a set of comments with a number of q received by item v.
[0022] In a further technical solution, the nodes in the user-item bipartite graph are composed of two disjoint subsets;
[0023] Each edge of the user-item bipartite graph represents an interaction behavior;
[0024] A user’s neighbors is an ordered subset that includes the latest n interacted items before time t;
[0025] The neighbors of an item include the most recent m interacting users before time t.
[0026] In the second aspect, a system for constructing an attribute fusion interactive recommendation model based on enhanced graph convolution is disclosed, including:
[0027] A relational encoder building module based on enhanced graph convolution is used to extract high-level features between users and items from user-item interaction behaviors, and obtain dynamic feature representations of users and items that contain high-level collaborative signals at each moment;
[0028] The comment encoder building module that integrates attribute information is used to process and integrate the comment text and its corresponding attributes accordingly to obtain an accurate feature representation of the comment.
[0029] Thirdly, an attribute fusion interactive recommendation model based on enhanced graph convolution is disclosed, including:
[0030] A relational encoder based on enhanced graph convolution, including an embedding layer and an enhanced graph convolution layer;
[0031] The embedding layer is used to map the IDs of users and items and their corresponding time series information to low-dimensional vector representations;
[0032] The enhanced graph convolution layer uses graph convolution operations to aggregate neighbor information and update user and item embeddings;
[0033] and a review encoder that incorporates attribute information;
[0034] The comment encoder fused with attribute information processes the text in the comment set of users and items to obtain a multi-granular feature representation of the comment; processes the attribute information corresponding to the comment text and combines it with the multi-granular feature representation of the comment to obtain multiple comment features fused with attribute information;
[0035] Then, the comment features of multiple fused attribute information are regarded as nodes, and the beneficial feature interaction pairs are input into the addition layer to obtain the final comment feature representation containing usefulness information, thereby obtaining the attribute features of the fused comments and including the usefulness information of the comments.
[0036] In a further technical solution, the embedding layer uses the user's neighbor information to update the user embedding, uses the object's neighbor information to update the object embedding, and adds corresponding temporal information to the neighbors to enhance the feature representation of the neighbors.
[0037] In a further preferred technical solution, the embedding layer randomly initializes the user and item IDs into a d-dimensional embedding vector u i 、v j , forming two initial embedding matrices for users and items;
[0038] During the training process, the corresponding initial embedding vector is found from the embedding matrix through the user and item ID for subsequent training.
[0039] A further preferred technical solution is to add timing information to each interaction to enhance its feature representation for the initial embedding matrix:
[0040] Convert the time series information into the corresponding embedding vector, where the time information is defined as the time interval between the historical interaction and the latest interaction, and convert it into a time embedding t to reflect the actual impact of the historical interaction on the current state;
[0041] The sequential information is defined as the position of the historical interactions in the interaction sequence, which is converted into a position embedding p to reflect the different roles of each interaction in the interaction sequence.
[0042] In the fourth aspect, an attribute fusion interactive recommendation system based on enhanced graph convolution is disclosed, including:
[0043] The high-order connected features of users and items are used as attention to participate in the learning of the item review set, and the review feature representation of the item is obtained;
[0044] The attribute fusion interactive recommendation model based on enhanced graph convolution connects the comment features of users and items and the high-order connection features of users and items for prediction, and obtains the user's comment interaction on the item at the next moment.
[0045] One or more of the above technical solutions have the following beneficial effects:
[0046] The model of the present invention incorporates attribute information into text modeling in different attribute activation modes, and matches beneficial attribute feature interaction pairs for the comment text in a self-supervised manner to obtain accurate feature representation of the comment text; in addition, the present invention analyzes the high-order collaborative relationship between users and items, and designs an enhanced graph convolution method to capture the high-order collaborative signals between users and items, on this basis to model the dynamic characteristics of users and items.
[0047] The PGIR model proposed in the present invention consistently achieves better rating prediction results than existing models (including document-based methods and review-based methods). Furthermore, the present invention shows through analysis that the attribute activation method proposed by PGIR can well solve the problem of inaccurate recommendations for passive users, which not only improves the model performance, but also improves the robustness and interpretability of recommendations.
[0048] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0050] Figure 1 (a)- Figure 1 (b) is a user-item interaction diagram according to an embodiment of the present invention;
[0051] Figure 2 This is a model diagram of an embodiment of the present invention;
[0052] Figure 3 Schematic diagram of dilated convolution according to an embodiment of the present invention;
[0053] Figure 4 This is a diagram of the interactive fusion of features of an embodiment of the present invention. DETAILED DESCRIPTION
[0054] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.
[0055] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.
[0056] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.
[0057] Embodiment 1
[0058] This embodiment discloses a method for constructing an attribute fusion interactive recommendation model based on enhanced graph convolution. The attribute fusion interactive recommendation model based on enhanced graph convolution is named PGIR.
[0059] The source is explained below: The high-order collaborative signals of user items are hidden in the interaction between users and items. Below we analyze the high-order collaborative relationship between users and items through specific examples. Figure 1 It is a bipartite graph derived from user-item interaction information (reviews), showing the high-order connectivity between users and items. Figure 1 The bipartite graph in (a) consists of three users (u1, u2, u3) and their three interacting items (v1, v2, v3). Each user-item pair has a time (t) at which the interaction occurred. Figure 1(b) shows two tree structures with user u1 as a node, which represent the high-order connectivity of the user at time t5 and t4 respectively. Obviously, user u1 has different high-order connectivity at these two times, so time dynamics need to be considered when making recommendations to users. When only considering interactions (not considering good or bad reviews), it can be seen that at time t5, u1 and u2 have two paths, which indicates that the similarity between u1 and u2 is very high at this moment. If only first-order features are considered, the model cannot capture this high-order similarity. Therefore, it is necessary to model high-order information to better describe the dynamic preferences of users and the dynamic characteristics of items.
[0060] Figure 1 The interactions in are based on user comment behaviors. These comments are issued by different users and have different characteristics. Table 1 shows the comment texts corresponding to the interactions. As can be seen from Table 1, different users have different comment styles. Some are objective evaluations of the items themselves, while others are extremely subjective. In addition, the comment texts vary in length, where "..." indicates that the text has more than 200 words. In addition, not all comments can be used to represent user preferences and item quality. By evaluating the usefulness of comments, the recommendation model can focus on those valuable comments in a large amount of available information, thereby improving the performance of the model. Many studies use attention mechanisms to simulate the usefulness of comments. However, the above methods have limitations. Judging the importance of comment texts only by text features is not convincing, and attention can only evaluate the usefulness of comments in historical data, but cannot be associated with comments that are not involved in training.
[0061] Table 1 User comments on items
[0062]
[0063] In order to solve the above limitations, the present invention considers the attributes of comments to model the usefulness of comments. As shown in Table 1, the comments posted by users have some corresponding attribute information. In addition to the length attribute and score attribute of the comments shown in Table 1, the time when the comments were posted and the number of likes can also reflect the usefulness of the comments. In summary, it is feasible to use the attribute information of the comments to assist in judging the usefulness of the comment text. The attributes of these comments are associated with historical comments and comments that are not involved in the training, which makes up for the shortcomings of using only the comment text features.
[0064] In order to comprehensively utilize the high-order collaborative signals of user items and the attribute information of comments, the present invention proposes an attribute fusion interactive recommendation model (PGIR) based on enhanced graph convolution, which can effectively identify the high-order connectivity between user items, and can judge the usefulness of comments in combination with different comment attributes, so as to learn user preferences and item characteristics. PGIR consists of two main modules: a relation encoder based on enhanced graph convolution and a comment encoder fused with attribute information. The relation encoder based on enhanced graph convolution is used to capture the high-order connectivity relationship between user items. Specifically, the present invention constructs a bipartite graph through the interactive information between users and items, and designs an enhanced graph convolution. First, the present invention designs a self-attention aggregator before convolution, which learns the embedded representation of users and items at each moment by fusing the temporal information of the interactive nodes. Secondly, multi-layer graph convolution is performed on the user-item bipartite graph to learn the high-order signals of each node. Based on this, the present invention obtains the dynamic feature representation of user items containing high-order collaborative signals at each moment.
[0065] The comment encoder that integrates attribute information is used to determine the usefulness of comments and integrate attribute information. First, the present invention processes and integrates the comment text and its corresponding attributes accordingly. Secondly, the present invention also designs a feature interaction aggregation module based on self-supervised learning to determine the usefulness of comments and obtain accurate feature representations of comments. Specifically, the present invention takes into account that beneficial attribute pairs can more effectively reflect the usefulness of comments. For example, popular products often have many comments with high freshness. At this time, the (time, score) attribute pair can simply and effectively reflect the usefulness of comments; however, for items with low popularity, the present invention needs to mine valuable information from comments with low freshness. Usually, longer comments can extract more useful information. At this time, the (time, length) attribute pair may play a key role in judging the usefulness of comments. Based on this encoder, the present invention can obtain accurate feature representations of comments.
[0066] The model structure of the present invention consists of two important parts, which are named as follows: a relation encoder based on enhanced graph convolution and a comment encoder fused with attribute information. The relation encoder based on enhanced graph convolution is used to extract high-order features between users and items from user-item interaction behaviors, which are used as dynamic feature representations of users (items); the comment encoder fused with attribute information judges the usefulness of comments and obtains accurate feature representations of comments by fusing attribute information through a self-supervised feature interaction aggregation module. Figure 2 shown.
[0067] The present invention is a rating prediction task in a recommendation system, and its purpose is to predict more accurate ratings for user-item pairs based on user preferences and item characteristics.
[0068] Each user and item has a corresponding set of review data. represents the set of comments with number p published by user u, represents the set of reviews received by item v with the number q. i ={w1,w2,...,w t} represents a comment text with t words. At the same time, each comment can be represented by P = {P1, P2, ..., P k} to represent the k attributes of a comment. For example, the length attribute (length), time attribute (how long it was published), rating attribute (1 to 5), usefulness (likes), etc. of a comment are mapped to specific interval values, and the scalar corresponding to the attribute is used to define the validity of the comment. Taking length as an example, the longer the text, the closer its length attribute score is to 1, which indicates that the longer the comment, the more useful it is. Similarly, the usefulness of the comment is judged by comprehensively considering the scores of other attributes of the comment.
[0069] In order to capture the high-order features between users and items, the present invention constructs a bipartite graph G with user-item review interactions as nodes. The nodes are composed of two disjoint subsets U and V, and each edge of the graph G represents an interaction behavior. The user's neighbors are an ordered subset, including the latest n interactive items before time t, denoted as N u,t = {v t-n ,v t-n+1 ,...,v t}; Similarly, the neighbors of an item include the latest m interacting users before time t, denoted as N v,t = {u t-m ,u t-m+1 ,...,u t}.
[0070] The relation encoder based on enhanced graph convolution includes: embedding layer and enhanced graph convolution layer.
[0071] The role of the embedding layer is to map the IDs of users and items and their corresponding time series information to low-dimensional vector representations. Specifically, the present invention uses the user's neighbor information to update the user embedding, uses the item's neighbor information to update the item embedding, and adds corresponding time series information to the neighbors to enhance the feature representation of the neighbors.
[0072] First, the present invention randomly initializes the user and item IDs into a d-dimensional embedding vector u i 、v j , based on which two initial embedding matrices are formed for users and items and Wherein |U| and |V| represent the number of users and items, respectively. It should be noted that during the training process, the present invention searches for the corresponding initial embedding vector from the embedding matrix through the user and item IDs for subsequent training.
[0073] Secondly, in addition to the two key entities of users and items, the timing information of interactions is also crucial for interactive behaviors. Therefore, the present invention adds timing information to each interaction to enhance its feature representation. Specifically, the present invention converts the timing information into a corresponding embedding vector. Among them, the time information is defined as the time interval between the historical interaction and the latest interaction, which is converted into a time embedding t, which can reflect the actual impact of the historical interaction on the current state; the sequence information is defined as the position of the historical interaction in the interaction sequence, which is converted into a position embedding p, which reflects the different roles of each interaction in the interaction sequence.
[0074] Enhanced graph convolution layer: This paper uses graph convolution operations to aggregate neighbor information and update user and item embeddings. Since the model is a parallel network for users and items, this paper introduces it from the user perspective. Given the neighbor set N of user node u at time t u,t = {v t-n ,v t-n+1 ,...,v t The present invention simplifies the polymerization process of the first layer into the following two steps:
[0075]
[0076]
[0077] in, represents the representation of items in the neighbor set at the lth layer iteration, and are two trainable parameter matrices that can project different types of nodes into the same feature space. It should be noted that compared with most current GCNs, the enhanced graph convolution of the present invention redefines the neighbors of nodes, that is, n interactions before interaction time t are selected as neighbor nodes, which alleviates the over-smoothing problem of GCN to a certain extent; at the same time, the present invention incorporates timing information into the nodes to enhance their feature representation. Next, the present invention introduces the encoder in detail.
[0078] Temporal self-attention aggregation: The main function of the self-attention aggregator is to enhance the feature representation of node neighbors using the corresponding temporal information. For the user's neighbor set, the present invention uses each v j The lth layer of iterative self-attention aggregation is performed to obtain the element use To define the item embedding and its corresponding time embedding and position embedding. The aggregation process can be expressed as follows:
[0079]
[0080] In the above formula, the weight coefficient Calculated using the softmax function, defined as:
[0081]
[0082]
[0083] in, It is a limiting factor that prevents a certain inner product value from being too high.
[0084] After self-attention is performed on each neighbor node, the present invention performs self-attention on each element. Perform nonlinear activation to obtain the feature representation of fused temporal information, namely:
[0085]
[0086] in, is a trainable parameter matrix.
[0087] User attention aggregation: After the above self-attention aggregation, we get a set of item representations that incorporates temporal information. Since users have different preferences for different items, the present invention aggregates user attention on the collection to simulate this process, namely:
[0088]
[0089]
[0090] in, are trainable parameters, Is user u i For items v j The attention weight, is the user embedding iterated l-1 times. Based on the above information, the user's l-th layer neighbor aggregation is expressed as follows:
[0091]
[0092] Stacked convolution: Capturing the high-order connectivity between users and items is very important for modeling users' preferences for items. Based on the first-order connectivity, the present invention can stack L layers of enhanced graph convolution to obtain the high-order connectivity features of users (items). After the above operations, the present invention aggregates the neighbor information of the lth iteration. The final user representation is obtained by fusing the user features and neighbor information using formula (2), which fuses the user's l-hop neighbor collaboration signals. Finally, the present invention uses the embedding vector output by the last layer of the graph convolution layer as the user's high-order connectivity feature representation, that is, Furthermore, the present invention obtains the high-order connected feature representation of the object in the same way, namely
[0093] Comment encoder about fusion attribute information:
[0094] Text modeling: Since the review set of users and items (D u ,D v ), so the present invention still takes the user as the perspective and only introduces the modeling process of a single text. i Simplified to D, the length of the text is t. Multi-layer dilated convolution is used to extract multi-granularity information from the text and enrich the feature representation of the review text.
[0095] For a given review text D = [w1,w2,...,w t ], the model maps each word into a corresponding d-dimensional vector through the bag-of-words model, that is Then, the word embedding matrix D is fed into multi-layer dilated convolutions to capture the multi-granularity feature information of the review text.
[0096] Unlike ordinary CNNs, dilated convolutions expand the receptive field by skipping δ word embeddings each time. t And the convolution kernel K of size 2w+1, the principle of dilated convolution is as follows Figure 3 As shown in the figure, for the same 3×3 convolution kernel, by changing the expansion rate, the receptive field of the convolution kernel becomes 7×7 and 15×15.
[0097] The formula for dilated convolution can be expressed as:
[0098]
[0099] in, is a vector concatenation operation, b is a bias vector, and ReLU is a nonlinear activation function.
[0100] For multi-layer dilated convolution, the output of the previous layer is the input of the next layer. Therefore, by properly adjusting the dilation rate to expand the receptive field of the convolution kernel layer by layer, the module can aggregate semantic information in different receptive fields with very few parameters. In other words, the module captures long-term dependencies from text with a small number of parameters, and its parallel capability is better than that of recurrent neural networks.
[0101] At the same time, the output of each layer undergoes mean-pooling operation to represent the semantic feature representation of the text at that granularity. Therefore, after L layers of dilated convolution, the multi-granular feature representation of the comment is:
[0102]
[0103] in, is the granular feature representation of the text after l layers of dilated convolution and pooling, f is the number of convolution kernels, and in the present invention, f is set to be the same as the dimension d of the hidden layer, is a vector concatenation operation.
[0104] Attribute processing: In this layer, the model integrates attribute information into text features through the dot product method. The purpose of this step is to judge the usefulness of comments from different attribute perspectives (text length, release time, rating, usefulness (likes), etc.). That is, the comment features with integrated attributes can be expressed as:
[0105]
[0106] in, represents the feature representation of the comment incorporating the k-th attribute, and MLP() represents Convert to
[0107] In this work, four persuasive review attributes were screened to judge the usefulness of reviews from different perspectives.
[0108] Text length: Comments vary in length, so the length of the text is normalized based on the number of words in each comment. Generally, longer text is considered more useful.
[0109] Release time: First, subtract the oldest comment time from each comment to indicate the freshness of each comment, and then normalize the freshness to get the time attribute score of the comment. Generally, the most recent comments are considered to be more valuable for reference.
[0110] Rating: Each comment has a rating of 1-5. Normalizing the ratings alone has certain drawbacks. Because, generally speaking, not only high-rated comments are valuable for reference, but low-rated comments also reflect the user's strong interest preferences. Therefore, an activation function is designed. To handle the scoring attributes. is the average score of user ratings, representing the user's neutral preference. Through this activation function, both the user's positive and negative preferences get higher attribute scores. The usefulness of this activation function will be analyzed and studied in the experimental section.
[0111] Usefulness: Most e-commerce platforms provide a comment-like function. We normalize the number of likes for each comment to obtain the corresponding attribute score.
[0112] It is worth noting that the attributes mentioned in the above implementation examples are only several commonly used attributes. The above model is extensible, so other attributes can also be added to the model after being processed.
[0113] Feature interaction fusion: Considering that beneficial attributes can more effectively reflect the usefulness of comments, simply using attention to aggregate multiple comment features with fused attribute information may not achieve good results. Inspired by graph neural networks, this paper regards multiple comment features with fused attribute information as nodes, such as Figure 4 As shown in the figure. First, it is predicted whether there are edge connections between multiple nodes. If there are, it means that the two attribute features are beneficial feature interactions, which is helpful for judging the usefulness of the comments. Secondly, on the basis of the existence of edges, the comment features that integrate different attribute information are combined in pairs, and the beneficial feature interaction pairs are input into the addition layer to obtain the final comment feature representation containing usefulness information.
[0114] Specifically, taking two nodes that combine the attributes of P1 and P2 as an example, the following formula is used to predict the edge information between the two points:
[0115]
[0116] Among them, Pre() is a sigmoid activated MLP network, and its input is and The element product of α is limited to the range of [0,1]. A value of 0 indicates that the two nodes are not interacting beneficially. If it is non-zero, it indicates that the two nodes are interacting beneficially, and the value of α is the degree of benefit of the interaction between the two nodes. In the same way, the degree of benefit of the interaction between the two nodes can be obtained. Then, the benefit is multiplied by the corresponding attribute pair and input into the addition layer to obtain the final comment feature representation. This feature integrates the attribute features of the comment and contains the usefulness information of the comment. The specific formula is:
[0117]
[0118] In the same way, the comments in the user comment set are processed, and the final user comment feature matrix is expressed as Then, the user's high-order features are used as the comment feature matrix for attention aggregation, which can be formulated as:
[0119] s ui =(R i ) T tanh(W·E u +b) (15)
[0120]
[0121] Among them, α ui is the attention weight of user u to the i-th comment in the comment set. Based on the above information, the aggregation of the user comment feature matrix can be expressed as follows:
[0122]
[0123] Similarly, the above operation is performed on the item review set to obtain the item review feature representation R v .
[0124] Prediction and training:
[0125] Finally, PGIR combines the review features of users and items (R u ,R v ) and high-order connectivity features between users and items (E u ,E v ) are connected to make predictions. Predict the user's comment interaction on the item at the next moment, that is, the score. The user's score prediction for the item can be calculated as:
[0126]
[0127] in, is a trainable parameter matrix, is the corresponding bias vector, and σ is the ReLU activation function.
[0128] The PGIR model solves the rating prediction problem in the recommendation system. In short, for a user-item pair, it will predict a more accurate rating for the user-item pair. Therefore, the more commonly used mean square error loss is selected as the objective function, that is:
[0129]
[0130] Among them, τ is the training set, and y is the true rating given by user u to item v.
[0131] To prevent the edge weights in the feature interaction module from being too large, this module is limited in a self-supervisory manner. The purpose is to make the model detect as few beneficial attribute pairs as possible, so that the attribute pair information that best describes the usefulness of the review can be captured. Therefore, the final loss function can be formulated as:
[0132]
[0133] Among them, λ is the regularization weight factor, π ij Indicates the beneficial degree of the interaction between two attribute nodes in each comment.
[0134] For the rating prediction task, the present invention proposes an attribute fusion interactive recommendation model based on enhanced graph convolution. First, the user's interests and the characteristics of the items are constantly changing, and there are rich high-order collaborative signals hidden between the user and the items. A relational encoder based on enhanced graph convolution is designed to capture this information. It can aggregate neighbor information for users (items) at each time node and obtain its dynamic feature representation of fused high-order neighbor information. Secondly, we believe that the attribute information of the comment can play a role in identifying and correcting the comment text. Therefore, the present invention designs a comment encoder that integrates attribute information to judge the usefulness of the comment by integrating the attributes of the comment. The present invention also designs a feature interaction aggregation module based on self-supervised learning, which can match beneficial attribute pairs for the comment, and the module further improves the recommendation performance of the model. Finally, through a large number of experiments on five standard data sets, the results show that the PGIR model proposed by the present invention always achieves better rating prediction results than the existing models (including document-based methods and comment-based methods). Furthermore, the present invention shows through analysis and research that the attribute activation method proposed by PGIR can well solve the inaccurate recommendation of negative users, which not only improves the model performance, but also improves the robustness and interpretability of the recommendation.
[0135] The recommendation accuracy of the recommendation system for positive users is higher than that for negative users. Therefore, improving the recommendation accuracy of the model for negative users will greatly improve the recommendation performance of the model. Some works have also tried to make improvements in this regard, normalizing user ratings, integrating them into the review text features, and using them to model users and items, which has improved the recommendation performance of the model to a certain extent. However, simple normalization processing is not universal. The user's low rating behavior reflects the user's extremely strong preference, and normalization processing weakens the degree of this preference. Based on this, the present invention proposes an activation function for processing user attribute ratings, namely, the activation function
[0136] In order to verify the usefulness of the activation function proposed in the present invention, the present invention splits the Video data set. From the user's perspective, users with an average score greater than 4 points are defined as positive users, and the rest are negative users. It is found that the score span of negative users is very large, that is, negative users have both positive and negative evaluations, which is a great challenge for the model. The present invention designs different variants for the model to verify the conjecture: the No-rating model means that the rating attribute is not used, and the PGIR-nor model means that the normalized method is used to process the rating attribute. The experimental results are shown in Table 2.
[0137] Table 2 Impact of score activation method on model performance
[0138]
[0139]
[0140] It is noted that, with other experimental conditions unchanged, the scoring attribute has a great contribution to improving model performance. For the complete Video dataset, the activation method of the present invention has been greatly improved compared to the normalization processing method; judging from the results of the Video_pos dataset, the model is relatively simple to capture the interests of active users, and can easily make accurate recommendations, so the activation method of the present invention and the ordinary normalization processing method both show good prediction performance; by analyzing the results of the Video_neg dataset, the activation method of the present invention has been greatly improved compared to the method that does not use the scoring attribute, and truly captures the negative user preference information. At the same time, although the normalization method also improves the model performance, the activation method of the present invention still improves the model performance by more than 5% based on the normalization method. How to capture the preferences of negative users and accurately predict them has always been a difficulty in the recommendation system. Experiments have shown that the method of the present invention solves this problem to a certain extent. The improvement in the accuracy of prediction of negative users is one of the reasons for the improvement in our model performance.
[0141] Extensive experiments were conducted on five standard datasets on Amazon, and the results show that the proposed PGIR model achieves substantial improvement in rating prediction compared to recently proposed advanced models. In addition, the superiority of the proposed attribute activation method is experimentally demonstrated, which further improves the rating prediction performance of PGIR.
[0142] Embodiment 2
[0143] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the program.
[0144] Embodiment 3
[0145] The purpose of this embodiment is to provide a computer-readable storage medium.
[0146] A computer-readable storage medium stores a computer program, which executes the steps of the above method when executed by a processor.
[0147] Embodiment 4
[0148] The purpose of this embodiment is to provide a property fusion interactive recommendation model construction system based on enhanced graph convolution, including:
[0149] A relational encoder building module based on enhanced graph convolution is used to extract high-level features between users and items from user-item interaction behaviors, and obtain dynamic feature representations of users and items that contain high-level collaborative signals at each moment;
[0150] The comment encoder building module that integrates attribute information is used to process and integrate the comment text and its corresponding attributes accordingly to obtain an accurate feature representation of the comment.
[0151] In another implementation example, an attribute fusion interactive recommendation model based on enhanced graph convolution is also disclosed, including:
[0152] A relational encoder based on enhanced graph convolution, including an embedding layer and an enhanced graph convolution layer;
[0153] The embedding layer is used to map the IDs of users and items and their corresponding time series information to low-dimensional vector representations;
[0154] The enhanced graph convolution layer uses graph convolution operations to aggregate neighbor information and update user and item embeddings;
[0155] and a review encoder that incorporates attribute information;
[0156] The comment encoder that integrates attribute information processes the text in the comment set of users and items to obtain a multi-granular feature representation of the comment; processes the attribute information corresponding to the comment text, and combines it with the multi-granular feature representation of the comment to obtain multiple comment features of the integrated attribute information.
[0157] Then, the comment features that represent the fused attributes are regarded as nodes, and the beneficial feature interactions are input into the addition layer to obtain the final comment feature representation containing usefulness information, thereby obtaining the attribute features of the fused comments and including the usefulness information of the comments.
[0158] The steps involved in the apparatus of the above embodiment correspond to the method embodiment 1, and the specific implementation method can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.
[0159] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.
[0160] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.
Claims
1. Construction method of attribute fusion interactive recommendation model based on enhanced graph convolution, Its characteristics are: include: Construct a relational encoder based on enhanced graph convolution to extract high-level features between users and items from user-item interaction behaviors, and obtain dynamic feature representations of users and items that contain high-level collaborative signals at each moment; Construct a comment encoder that integrates attribute information, process and integrate the comment text and its corresponding attributes accordingly, and obtain an accurate feature representation of the comment; The steps to build a relational encoder based on enhanced graph convolution are: Construct a user-item bipartite graph as nodes through the interaction between users and items; The self-attention aggregator fuses the temporal information of the interaction nodes in the user-item bipartite graph, and the user-item attention aggregator aggregates the neighbor nodes of users or items to learn the embedding representation of users and items at each moment; Use enhanced graph convolution to perform multi-layer graph convolution on the user-item bipartite graph to learn the high-order signals of each node in the user-item bipartite graph; Based on the learned embedding representations of users and items at each moment and the high-order signals of each node in the user-item bipartite graph, the dynamic feature representation of users and items at each moment containing high-order collaborative signals is obtained; The steps to construct a comment encoder that incorporates attribute information are: Use dilated convolution to extract review text features; Processing attribute information; Combine the review text features with the attribute information, and then use the feature interaction aggregation module based on self-supervised learning to judge the usefulness of the review and obtain the accurate feature representation of the review; Before constructing the user-item bipartite graph, a set of comment data corresponding to each user and item is required, including: user u The number of releases is p The number of reviews and items v received is q A collection of reviews; The nodes in the user-item bipartite graph are composed of two disjoint subsets; Each edge of the user-item bipartite graph represents an interaction behavior; A user’s neighbors is an ordered subset that includes the latest n interacted items before time t; The neighbors of an item include the most recent m interacting users before time t.
2. Attribute fusion interactive recommendation model construction system based on enhanced graph convolution, Its characteristics are: include: A relational encoder building module based on enhanced graph convolution is used to extract high-level features between users and items from user-item interaction behaviors, and obtain dynamic feature representations of users and items that contain high-level collaborative signals at each moment; The comment encoder building module that integrates attribute information is used to process and integrate the comment text and its corresponding attributes to obtain accurate feature representation of the comment; The steps to build a relational encoder based on enhanced graph convolution are: Construct a user-item bipartite graph as nodes through the interaction between users and items; The self-attention aggregator fuses the temporal information of the interaction nodes in the user-item bipartite graph, and the user-item attention aggregator aggregates the neighbor nodes of users or items to learn the embedding representation of users and items at each moment; Use enhanced graph convolution to perform multi-layer graph convolution on the user-item bipartite graph to learn the high-order signals of each node in the user-item bipartite graph; Based on the learned embedding representations of users and items at each moment and the high-order signals of each node in the user-item bipartite graph, the dynamic feature representation of users and items at each moment containing high-order collaborative signals is obtained; The steps to construct a comment encoder that incorporates attribute information are: Use dilated convolution to extract review text features; Processing attribute information; Combine the review text features with the attribute information, and then use the feature interaction aggregation module based on self-supervised learning to judge the usefulness of the review and obtain the accurate feature representation of the review; Before constructing the user-item bipartite graph, a set of comment data corresponding to each user and item is required, including: user u The number of releases is p The number of reviews and items v received is q A collection of reviews; The nodes in the user-item bipartite graph are composed of two disjoint subsets; Each edge of the user-item bipartite graph represents an interaction behavior; A user’s neighbors is an ordered subset that includes the latest n interacted items before time t; The neighbors of an item include the most recent m interacting users before time t.
3. Attribute fusion interactive recommendation model based on enhanced graph convolution, Its characteristics are: include: A relational encoder based on enhanced graph convolution, including an embedding layer and an enhanced graph convolution layer; The embedding layer is used to map the IDs of users and items and their corresponding time series information to low-dimensional vector representations; The enhanced graph convolution layer uses graph convolution operations to aggregate neighbor information and update user and item embeddings; and a review encoder that incorporates attribute information; The comment encoder integrating attribute information processes the text in the comment set of users and items to obtain a multi-granular feature representation of the comment; Process the attribute information corresponding to the comment text and combine it with the multi-granularity feature representation of the comment to obtain multiple comment features that integrate the attribute information; Then, the review features of multiple fused attribute information are regarded as nodes, and the beneficial feature interaction pairs are input into the addition layer to obtain the final review feature representation containing usefulness information, thereby obtaining the attribute features of the fused review and including the usefulness information of the review; The embedding layer uses the user's neighbor information to update the user embedding, uses the item's neighbor information to update the item embedding, and adds corresponding temporal information to the neighbors to enhance the feature representation of the neighbors; The embedding layer randomly initializes the user and item ids as d dimensional embedding vector u i , v j , forming two initial embedding matrices for users and items; During the training process, the corresponding initial embedding vector is found from the embedding matrix through the user and item IDs for subsequent training; For the initial embedding matrix, add temporal information to each interaction to enhance its feature representation: Convert the timing information into the corresponding embedding vector, where the time information is defined as the time interval between the historical interaction and the latest interaction, and convert it into a time embedding t , reflects the actual impact of historical interactions on the current state; The sequential information is defined as the position of the historical interaction in the interaction sequence, which is converted into a position embedding p , reflecting the different roles of each interaction in the interaction sequence.
4. Attribute fusion interactive recommendation system based on enhanced graph convolution, Its characteristics are: include: The high-order connected features of users and items are used as attention to participate in the learning of the item review set, and the review feature representation of the item is obtained; The attribute fusion interactive recommendation model based on enhanced graph convolution connects the comment features of users and items and the high-order connection features of users and items to make predictions, and obtains the comment interaction of users on items at the next moment; The attribute fusion interactive recommendation model based on enhanced graph convolution is the model described in claim 3.
5. A computing device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, Its characteristics are: When the processor executes the program, the steps of the method according to claim 1 are implemented.
6. A computer-readable storage medium having a computer program stored thereon, Its characteristics are: When the program is executed by a processor, the steps of the method according to claim 1 are executed.
