An extractive explainable recommendation method based on topic-enhanced graph

Abstract

The application discloses an extractable interpretable recommendation method based on a topic-enhanced graph, comprising the following steps: 1. constructing a user-product graph by using the comments and scores of users on products, mining topics from the comments, and establishing a user-topic-product graph; 2. constructing a hybrid graph convolution method to fuse high-order collaborative signals and comment feature information, and encoding the implicit representation of users and products; 3. constructing a single-layer attention graph convolution to extract the explicit topic representation of users and products from the user-topic-product graph; and 4. fusing the implicit representation and the explicit topic representation of users and products, predicting the scores of users on products, and retrieving a plurality of related sentences from the historical comment texts of products as explanations. The application aims at the comment behavior of users, fuses and models the user preferences from two angles of implicit user-product collaborative signals and explicit topic information, predicts the scores of users on products, extracts sentences as explanations, and improves the trust degree of users on the recommendation results.

Description

Technical Field

[0001] This invention relates to the field of personalized and explainable recommendation, specifically an extractive and explainable recommendation method based on topic-enhanced graphs. Background Technology

[0002] In today's information age, with the continuous development of internet and big data technologies, the problem of information overload is becoming increasingly serious. Personalized recommendation systems recommend products that users may be interested in based on their historical behavior and preferences. However, traditional recommendation algorithms often lack interpretability, making it difficult to provide users with the reasons and basis for the recommendations, thus failing to meet user needs and build trust.

[0003] In recent years, explainable recommendation has become a hot topic in the field of recommender system research. Explainable recommendation not only provides users with the reasons and basis for recommendation results, but also helps to improve user trust in the recommender system and enhance the user experience. Existing extractive explainable methods aim to extract sentences from product-related explanatory text libraries as explanatory text. However, these methods neglect the topical information reflected by sentence semantics, making it difficult for user and product representations to reflect user preferences at different topical levels, thus limiting the effectiveness of explainability. Summary of the Invention

[0004] This invention addresses the shortcomings of existing extractive interpretability methods by proposing an extractive interpretability recommendation method based on topic enhancement graphs. The aim is to more fully mine comment information, achieve more accurate modeling of user-product interactions, thereby improving interpretability recommendation performance and user trust in the recommendation results.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The present invention provides an extractive, interpretable recommendation method based on topic enhancement graphs, characterized by the following steps:

[0007] Step 1: Obtain the user set Product Collection rating set And the comment collection ε; where, let N represents the i-th user. u Represents a set of users The total number of users, making v j ∈v represents the j-th product, N v Represents a product set The total number of products; Represents the i-th user u i For the j-th product v j The rating made e i,j∈ε represents the i-th user u i For the j-th product v j The given comments, and e i,j It contains k sentences; let Represents a set of sentences;

[0008] Step 2: Build a user-product graph with enhanced comments This represents users' rating and review behavior towards products; among which, Let R represent the rating matrix, where the element in the i-th row and j-th column is the i-th user u. i For the j-th product v j Rating r i,j ; Let E represent the comment matrix, where the element in the i-th row and j-th column is the i-th user u. i For the j-th product v j Comments i,j ;

[0009] Step 3: Cluster the features of all sentences in the sentence set s using a clustering method. Each cluster is considered as a topic, thus obtaining the topic corresponding to each sentence.

[0010] make Indicates from the set of sentences Retrieve the topic set, N t Represents a set of topics The total number of topics in the database; thus constructing a user-topic-product graph. This represents the interaction relationship between the user, the topic, and the product; where, This represents the topic relationship matrix C, which represents the interaction between users or products and topics. In the topic relationship matrix C, element c a,t Let c ∈{0,1} represent the relationship between the a-th user or product and topic t. If the comments of the a-th user or product involve topic t, then let c a,t =1 Otherwise, let c a,t =0; let express Let S be a sentence relation tensor, where K represents the number of sentences related to topic t in the comments of the a-th user or product. Let s be an element of tensor S. a,t,k This represents the k-th sentence related to topic t for user a or product a;

[0011] Step 4: Enhance the user-product graph of the review using L-layer graph convolution. Process the data to obtain the i-th user u. i implicit representation and the implicit representation of the j-th product

[0012] Step 5: For users, themes, and product images. Perform a graph convolution layer to obtain the i-th user u i explicit subject representation explicit subject representation of the j-th product

[0013] Step 6: Calculate the i-th user u i For the j-th product v j Predicted score And calculate the relevance score of the explainable sentences.

[0014] Step 7: Calculate the loss function for rating prediction and interpretable sentence extraction, and use it to train the model;

[0015] Step 7.1: Calculate the loss function for score prediction using equation (13).

[0016]

[0017] In equation (13), Represents a set of users and product collection The set of user-product pairs that constitutes this; Represents a set The total number of user-product pairs;

[0018] Step 7.2: Calculate the retrieval loss function of the interpretable sentence according to equation (14).

[0019]

[0020] In equation (11), This indicates that from comment e i,j The representation vector of a randomly sampled positive sentence. Represents the i-th user u i With the j-th product v j The implicit interaction representation vector, Represents the i-th user u i With the j-th product v j Explicit topic interaction representation, x - Indicates from the set of sentences The vector representing an unrelated negative sample sentence randomly sampled from the middle, where σ is the sigmoid activation function;

[0021] Step 7.3: Construct the total loss function according to equation (15).

[0022]

[0023] Step 7.4: Train the recommendation model consisting of L layers of graph convolutions and one layer of graph convolutions using gradient descent, and calculate the loss function. To update the model parameters until the loss function is reached. The process continues until convergence or the maximum number of iterations is reached, thus obtaining the optimal recommendation model. This model is used to predict a user's rating of the product under test and pushes the sentences with the highest relevance scores from the product's historical reviews to the corresponding user.

[0024] The extractive interpretable recommendation method based on topic enhancement graphs described in this invention is characterized in that step 4 is performed as follows:

[0025] Step 4.1, let This represents the initial representation vector of the i-th user. Let d represent the initial representation vector of the j-th product, where d represents the vector dimension;

[0026] Step 4.2: Convolution of the l-th layer graph using equation (1) to calculate the value of v from the j-th product. j Up to the i-th user u i Feature information propagation vector

[0027]

[0028] In equation (1), V represents the j-th product obtained from the (l-1)-th convolutional layer. j The representation vector; when l = 1, let and Let v represent the j-th product respectively. j and the i-th user u i The set of neighboring nodes; Comment e i,j The corresponding eigenvector; |·| represents calculating the number of elements in the set; The score r in the l-th convolutional layer represents the score r. i,j The specified parameter matrix; and Representing the rating r i,j Two linear mapping functions are used to map comment features e i,j Mapped to scalars between 0 and 1; MLP ri,j (·) represents the score r i,j A multilayer perceptron is used to process comment features e i,j Transform into the product feature vector space;

[0029] Step 4.3: Use equation (2) to aggregate the information propagation vector of the l-th layer graph convolution, thereby obtaining the i-th user u. i The representation vector of graph convolution at layer l

[0030]

[0031] In equation (2), Let be the parameter matrix of the graph convolution at layer l;

[0032] Step 4.4: Calculate the graph convolution pair of the l-th layer for the j-th product v according to the process in steps 4.1-4.3. j The representation

[0033] Step 4.5: After assigning l+1 to l, return to step 4.2 and continue until l = L, thus obtaining the implicit representation of the end user. and product implicit representation in, Represents the i-th user u i The representation vector of the graph convolution at the Lth layer, v represents the j-th product j The representation vector of the graph convolution at the Lth layer, and These represent two parameter matrices.

[0034] Step 5 is performed as follows:

[0035] Step 5.1: Calculate the propagation from topic t to the i-th user u using equation (3). i eigenvectors ω t→i :

[0036]

[0037] In equation (3), Let be the parameter matrix to be learned. Represents the i-th user u i The set of sentences in the comments that relate to topic t, s i,t,k express The kth sentence in the text, Sentence s i,t,k eigenvectors; Suppose the kth sentence s i,t,k The weights are calculated using equation (4):

[0038]

[0039] In equation (4), exp(·) represents the exponentiation operation, and T represents the vector transpose operation. Represents the i-th user u i A preliminary characterization of topic t is given, and the following is also given:

[0040]

[0041] In equation (5), Let be the representation vector of topic t, and ⊙ denote the Hadamard product of two vectors;

[0042] Step 5.2: Aggregate the topic feature vector using equation (5) to obtain the i-th user u. i explicit topic representation vector

[0043]

[0044] In equation (5), MLP(·) represents a multilayer perceptron. Represents the i-th user u i The collection of topics involved Represents the eigenvector ω t→i The weights are calculated using equation (7):

[0045]

[0046] In equation (7), Represents the i-th user u i The initial characteristics are as follows:

[0047]

[0048] Step 5.3: Perform steps 5.1 and 5.2 on the j-th product v. j The process yields explicit topic representations.

[0049] Step 6 is performed as follows:

[0050] Step 6.1: Calculate the i-th user u using equation (8). i For the j-th product v j Predicted score

[0051]

[0052] In equation (8), w is the parameter vector to be learned. Represents the i-th user u i With the j-th product v j The implicit interaction representation vector is obtained and calculated by equation (10):

[0053]

[0054] In equation (10), [;] represents the vector concatenation operation;

[0055] Step 6.2, use equation (11) to calculate the k-th sentence s k The i-th user u i and the j-th product v j Correlation score between

[0056]

[0057] In equation (11), x k Sentence s k The representation vector; Represents the i-th user u i With the j-th product v j The explicit topic interaction representation is obtained and calculated by equation (12):

[0058]

[0059] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the extractable interpretable recommendation method, and the processor is configured to execute the program stored in the memory.

[0060] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the extractable interpretable recommendation method.

[0061] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0062] 1. This invention addresses the issue of existing extractive interpretable models ignoring comment topic information by proposing an extractive interpretable recommendation method based on topic-enhanced graphs. Through steps 1 and 2, user-product graphs and user-topic-product graphs are constructed, improving the accuracy of user behavior modeling from both implicit user preference modeling and explicit topic user preference modeling perspectives.

[0063] 2. This invention targets comment information and mines topic information from sentence semantics in step 3. It connects users and products through topics as intermediate nodes and uses corresponding sentences as edge features. This modeling more accurately describes the relationship between users, products, topics and sentences. In step 5, the self-attention mechanism adaptively measures the contribution of different sentences and topics to the representation of users and products, thereby accurately capturing users' topic-level preferences and achieving a more accurate and interpretable sentence extraction effect.

[0064] 3. This invention targets high-order collaborative signals and comment information of users and products. By modeling collaborative signals and comment information separately in step 4, it avoids the disadvantage that comment information is not suitable for high-order graph convolution, thereby more accurately representing users and products. Attached Figure Description

[0065] Figure 1 This is a flowchart of an extractive interpretable recommendation method based on topic enhancement graphs according to the present invention. Detailed Implementation

[0066] In this embodiment, an extractive interpretable recommendation method based on topic enhancement graphs considers the fact that existing extractive interpretable methods ignore the actual situation where sentence semantics in comments reflect topic information. It accurately captures user preferences by implicitly modeling user-product collaborative signals and explicitly modeling user-topic-product information. While predicting user ratings for products, it retrieves sentences from candidate interpretive texts as interpretive results. Specifically, as follows... Figure 1 As shown, the method is performed according to the following steps:

[0067] Step 1: Obtain the user set Product Collection rating set And the comment collection ε; where, let N represents the i-th user. u Represents a set of users The total number of users, making Let N represent the j-th product. v Represents a product set The total number of products; Represents the i-th user u i For the j-th product v j The rating made e i,j ∈ε represents the i-th user u i For the j-th product v j The given comments, and e i,j ={s i,j,1 ,s i,j,2 ,…,s i,j,k}, s i,j,k Represents the i-th user u i For the j-th product v j The kth sentence in the comments; Represents a set of sentences;

[0068] Step 2: Build a user-product graph with enhanced comments This represents users' rating and review behavior towards products; among which, Let R represent the rating matrix, where the element in the i-th row and j-th column is the i-th user u. i For the j-th product vj Rating r i,j ; Let E represent the comment matrix, where the element in the i-th row and j-th column is the i-th user u. i For the j-th product v j Comments i,j ;

[0069] Step 3: Cluster all sentence features in the sentence set using clustering methods, treating each cluster as a topic, and obtaining the topic corresponding to each sentence. Let... Indicates from the set of sentences Retrieve the topic set, N t Represents a set of topics The total number of topics in the database; thus constructing a user-topic-product graph. This represents the interaction relationship between the user, the topic, and the product; where, This represents the topic relationship matrix C, which represents the interaction between users or products and topics. In the topic relationship matrix C, element c a,t Let c ∈{0,1} represent the relationship between the a-th user or product and topic t. If the comments of the a-th user or product involve topic t, then let c a,t =1 Otherwise, let c a,t =0; let express Let S be a sentence relation tensor, where K represents the number of sentences related to topic t in the comments of the a-th user or product. Let the elements s in the tensor S be... a,t,k This represents the k-th sentence related to topic t for the a-th user or product;

[0070] Existing methods use product attribute words labeled with expert knowledge as topic information. However, these attribute words are limited by domain adaptation issues and require interpretation of whether the text explicitly contains attribute words. However, text in reviews does not always contain explicit attribute words. This invention takes a semantic approach, treating sentences with similar semantics as having the same topic, thus avoiding the rigid requirement for attribute words. On the other hand, existing methods, when introducing topics and sentences into the user-product graph, simply connect sentences to corresponding topics. This inevitably leads to the introduction of irrelevant sentences when topic information propagates to user and product representations, reducing the quality of node representations. This invention proposes a user-topic-product graph and uses sentences as features on edges, more accurately modeling the relationships between users, products, topics, and sentences, and precisely mining user preferences for different topics.

[0071] Step 4: Enhance the user-product graph of the review using L-layer graph convolution. Processing is performed to obtain implicit representations of users and products:

[0072] Step 4.1, let This represents the initial representation vector of the i-th user. Let d represent the initial representation vector of the j-th product, where d represents the vector dimension;

[0073] Step 4.2: Convolution of the l-th layer graph using equation (1) to calculate the value of v from the j-th product. j Up to the i-th user u i Feature information propagation vector

[0074]

[0075] In equation (1), V represents the j-th product obtained from the (l-1)-th convolutional layer. j The representation vector; when l = 1, let and Let v represent the j-th product respectively. j and the i-th user u i The set of neighboring nodes; Comment e i,j The corresponding eigenvector; |·| represents calculating the number of elements in the set; The score r in the l-th convolutional layer represents the score r. i,j The specified parameter matrix; and Representing the rating r i,j Two linear mapping functions are used to map comment features e i,j Mapped to scalars between 0 and 1; For rating r i,j A multilayer perceptron is used to process comment features e i,j Transform into the product feature vector space;

[0076] Step 4.3: Use equation (2) to aggregate the information propagation vector of the l-th layer graph convolution, thereby obtaining the i-th user u. i The representation vector of graph convolution at layer l

[0077]

[0078] In equation (2), Let be the parameter matrix of the graph convolution at layer l;

[0079] Step 4.4: Calculate the graph convolution pair of the l-th layer for the j-th product v according to the process in steps 4.1-4.3. j The representation

[0080] Step 4.5: After assigning l+1 to l, return to step 4.2 and continue until l = L, thus obtaining the implicit representation of the end user. and product implicit representation in, Represents the i-th user u i The representation vector of the graph convolution at the Lth layer, v represents the j-th product j The representation vector of the graph convolution at the Lth layer, and These represent two parameter matrices respectively;

[0081] Existing methods, when incorporating comment information into the user-product graph, propagate comment information and neighbor nodes simultaneously during graph convolution operations. This approach, which is unsuitable for propagating comment information at higher orders, limits the number of graph convolution layers to one, restricting the model's utilization of higher-order collaborative signals. This invention calculates higher-order collaborative signals and comment information separately, fusing them in the final layer of the graph convolution, thereby more effectively capturing implicit user preferences.

[0082] Step 5: For users, themes, and product images. Perform a graph convolution layer to obtain explicit topic representations of users and products:

[0083] Step 5.1: Calculate the propagation from topic t to the i-th user u using equation (3). i eigenvectors ω t→i :

[0084]

[0085] In equation (3), Let be the parameter matrix to be learned. Represents the i-th user u i The set of sentences in the comments that relate to topic t, s i,t,k express The kth sentence in the text, Sentence s i,t,k eigenvectors; Suppose the kth sentence s i,t,k The weights are calculated using equation (4):

[0086]

[0087] In equation (4), exp(·) represents the exponentiation operation. This represents the vector transpose operation. Represents the i-th user u i A preliminary characterization of topic t is given, and the following is also given:

[0088]

[0089] In equation (5), Let be the representation vector of topic t, and ⊙ denote the Hadamard product of two vectors;

[0090] Step 5.2: Aggregate the topic feature vector using equation (5) to obtain the i-th user u. i explicit topic representation vector

[0091]

[0092] In equation (5), MLP(·) represents a multilayer perceptron. Represents the i-th user u i The collection of topics involved Represents the eigenvector ω t→i The weights are calculated using equation (7):

[0093]

[0094] In equation (7), Represents the i-th user u i The initial characteristics are as follows:

[0095]

[0096] This invention starts with sentence semantics, progressively aggregating rich sentence semantics from sentences to topics and then to users or products, learning user or product preferences or attributes across different topics. Taking a user as an example, graph convolution first weights and sums the user's historical comments as a specific topic representation for that user. For instance, even within the topic of clothing colors, one user might prefer light colors, while another prefers dark colors. Then, the weighted aggregation of the user's specific topic representations automatically mines the degree of the user's preference for different topics, thereby achieving more accurate and interpretable sentence extraction.

[0097] Step 5.3: Perform steps 5.1 and 5.2 on the j-th product v. j The process yields explicit topic representations.

[0098] Step 6: Predict the score and calculate the relevance score of the explainable sentence;

[0099] Step 6.1: Calculate the i-th user u using equation (8). i For the j-th product v j Predicted score

[0100]

[0101] In equation (8), w is the parameter vector to be learned. Represents the i-th user u i With the j-th product v j The implicit interaction representation vector is obtained and calculated by equation (10):

[0102]

[0103] In equation (10), [;] represents the vector concatenation operation;

[0104] Step 6.2, use equation (11) to calculate the k-th sentence s k The i-th user u i and the j-th product v j Correlation score between

[0105]

[0106] In equation (11), x k Sentence s k The representation vector; Represents the i-th user u i With the j-th product v j The explicit topic interaction representation is obtained and calculated by equation (12):

[0107]

[0108] Step 7: Calculate the loss function for rating prediction and interpretable sentence extraction, and use it to train the model;

[0109] Step 7.1: Calculate the loss function for score prediction using equation (13).

[0110]

[0111] In equation (13), Represents a set of users The set of user-product pairs consisting of the product set ν; Represents a set The total number of user-product pairs;

[0112] Step 7.2: Calculate the retrieval loss function of the interpretable sentence according to equation (14).

[0113]

[0114] In equation (11), This indicates that from comment e i,j The representation vector of a randomly sampled positive sentence, x- Indicates from the set of sentences The vector representing an unrelated negative sample sentence randomly sampled from the middle, where σ is the sigmoid activation function;

[0115] Step 7.3: Construct the total loss function according to equation (15).

[0116]

[0117] This invention effectively combines two tasks in extractive interpretable tasks: rating prediction and sentence extraction. On one hand, it uses textual information as a supervisory signal to train the recommendation model, enabling the model to learn not only user ratings of a product but also the reasons behind those ratings from the review text. On the other hand, it enhances the accuracy of the interpretable sentence extraction task by combining user and product rating behavior with the topic information in the reviews.

[0118] Step 7.4: Train the recommendation model consisting of L layers of graph convolutions and one layer of graph convolutions using gradient descent, and calculate the loss function. To update the model parameters until the loss function is updated. The optimal recommendation model is obtained by converging or reaching the maximum number of iterations.

[0119] Step 8: After the model training is complete, given a user u i And a product v j The corresponding score is predicted using equation (9). And using Equation (11) to calculate the relevance scores of all sentences in the product's historical reviews, and selecting the N most relevant sentences as explanatory text according to Equation (16):

[0120]

[0121] In equation (16), v represents the j-th product j A collection of historical commentary sentences, s j,k express The kth sentence in the text.

[0122] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0123] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. An extractive, interpretable recommendation method based on topic enhancement graphs, characterized in that, The procedure is as follows: Step 1: Obtain the user set Product Collection Rating Set and comment collection Among them, let Indicates the first One user, Represents a set of users The total number of users, making Indicates the first One product Represents a product set The total number of products; Indicates the first individual users For the first Product The rating made Indicates the first individual users For the first Product The comments given, and Include k A sentence; Represents a set of sentences; Step 2: Build a user-product graph with enhanced comments This represents users' rating and review behavior towards products; among which, Represents the rating matrix, The Middle Line number The column element is the first individual users For the first Product rating ; Represents the comment matrix, The Middle Line number The column element is the first individual users For the first Product Comments ; Step 3: Cluster the sentence set. Clustering is performed on the features of all sentences in the text, and each cluster is regarded as a topic, thus obtaining the topic corresponding to each sentence; make Indicates from the set of sentences Get the topic collection from the middle. Represents a set of topics The total number of topics in the database; thus constructing a user-topic-product graph. This represents the interaction relationship between the user, the topic, and the product; where, A topic relationship matrix representing the interaction between users or products and topics. In, elements Indicates the first Individual user or product and theme The relationship, if the first The topic mentioned in user or product reviews Then let Otherwise, let ;make express The sentence relation tensor in the text, where, Indicates the first The topic mentioned in user or product reviews The number of sentences, let the tensor one of the elements Indicates the first The topic involves individual users or products. The 1 sentence; Step 4, through Layered graph convolution for user-product graph comment enhancement Process it to obtain the first individual users implicit representation and the Implicit representation of a product ; Step 5: For users, themes, and product images. Perform a graph convolution layer to obtain the first layer. individual users explicit subject representation With the Explicit thematic representation of each product ; Step 6, calculate the first... individual users For the first Product Predicted score And calculate the relevance score of the explainable sentences. ; Step 7: Calculate the loss function for rating prediction and interpretable sentence extraction, and use it to train the model; Step 7.1: Calculate the loss function for score prediction using equation (13). : (13) In equation (13), Represents a set of users and product collection The set of user-product pairs that constitutes this; Represents a set The total number of user-product pairs; Step 7.2: Calculate the retrieval loss function of the interpretable sentence according to equation (14). : (14) In equation (14), Indicates from comments The representation vector of a randomly sampled positive sentence. Indicates the first individual users With the Product The implicit interaction representation vector, Indicates the first individual users With the Product Explicit topic interaction representation, Indicates from the set of sentences The representation vector of a randomly sampled, unrelated negative sample sentence. It is the sigmoid activation function; Step 7.3: Construct the total loss function according to equation (15). : (15) Step 7.4: Use gradient descent to process the data... The recommendation model, consisting of one layer of graph convolution and one layer of graph convolution, is trained, and the loss function is calculated. To update the model parameters until the loss function is reached. The process continues until convergence or the maximum number of iterations is reached, thus obtaining the optimal recommendation model. This model is used to predict a user's rating of the product under test and pushes the sentences with the highest relevance scores from the product's historical reviews to the corresponding user.

2. The extractive interpretable recommendation method based on topic enhancement graphs according to claim 1, characterized in that, Step 4 is performed as follows: Step 4.1, let Indicates the first The initial representation vector of each user, Indicates the first The initial representation vector of each product, where... Indicates the dimension of a vector; Step 4.2, the Layer graph convolution is calculated using equation (1) from the first layer. Product To the individual users Feature information propagation vector : (1) In equation (1), Indicates the first The first convolutional layer obtained Product The representation vector; when season ; and They represent the first Product and the individual users The set of neighboring nodes; Comment The corresponding feature vector; This indicates the number of elements in the set; Indicates the first Scoring in convolutional layers The specified parameter matrix; and Each represents a score Two linear mapping functions are used to map comment features. Mapped to scalars between 0 and 1; Rate A multilayer perceptron is used to process comment features. Transform into the product feature vector space; Step 4.3, use equation (2) to aggregate the first The information propagation vector of the layer graph convolution is obtained to obtain the first layer. individual users In the Representation vector of layer graph convolution : (2) In equation (2), For the first The parameter matrix of layer graph convolution; Step 4.4: Calculate the following steps according to steps 4.1-4.

3. Layer graph convolution pairs Product The representation ; Step 4.5, Assign to Then, return to step 4.2 and continue until... This allows us to obtain the implicit representation of the end user. and product implicit representation ;in, Indicates the first individual users In the The representation vector of layer graph convolution, Indicates the first Product In the The representation vector of layer graph convolution, and These represent two parameter matrices.

3. The extractive interpretable recommendation method based on topic enhancement graphs according to claim 2, characterized in that, Step 5 is performed as follows: Step 5.1: Calculate the value from the topic using equation (3). Spread to the individual users eigenvectors : (3) In equation (3), Let be the parameter matrix to be learned. Indicates the first individual users Topics mentioned in the comments A collection of sentences, express The Middle 1 sentence Sentence eigenvectors; Indicates the first 1 sentence The weights are calculated using equation (4): (4) In equation (4), This indicates exponentiation. This represents the vector transpose operation. Indicates the first individual users About the topic The initial characteristics are as follows: (5) In equation (5), Theme The representation vector, Denotes the Hadamard product of two vectors; Step 5.2: Aggregate the topic feature vectors using equation (5) to obtain the first... individual users explicit topic representation vector : (5) In equation (5), This represents a multilayer perceptron. Indicates the first users The collection of topics involved Representing the eigenvector The weights are calculated using equation (7): (7) In equation (7), Indicates the first user The initial characteristics are as follows: (8) Step 5.3: Follow steps 5.1 and 5.2 to process the... Product The process yields explicit topic representations. .

4. The extractive interpretable recommendation method based on topic enhancement graphs according to claim 3, characterized in that, Step 6 is performed as follows: Step 6.1: Calculate the first step using equation (9). individual users For the first Product Predicted score : (9) In equation (9), Let be the parameter vector to be learned. Indicates the first individual users With the Product The implicit interaction representation vector is obtained and calculated by equation (10): (10) In equation (10), This indicates a vector concatenation operation; Step 6.2, calculate the first step using equation (11). 1 sentence , No. individual users and the Product Correlation score between : (11) In equation (11), Sentence The representation vector; Indicates the first individual users With the Product The explicit topic interaction representation is obtained and calculated by equation (12): (12)。 5. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support a processor in executing the extractable interpretable recommendation method of any one of claims 1-4, the processor being configured to execute the programs stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the extractive interpretable recommendation method according to any one of claims 1-4.

Images