User comment quality double-view reasoning method and system based on state transition learning

Abstract

The invention relates to the technical field of user comment analysis, and particularly provides a user comment quality double-view reasoning method and system based on state transition learning. The method comprises the following steps: generating an edge set based on a calculated association relationship, and establishing a multi-attribute user comment heterogeneous graph; based on the multi-attribute user comment heterogeneous graph, formulating a user quality dependence rule, and designing a state transition learning algorithm to iteratively infer user quality; based on the user quality, establishing a comment usefulness dependency rule, utilizing a linear function to balance the content usefulness CBH and the content influence IBH, and predicting the comment quality of a general perspective under all products; based on a multi-attribute user comment heterogeneous graph, a designed deep network with a graph aggregation unit and a multi-head self-attention mechanism is utilized to predict the comment quality of a specific view angle under a given user, and the method comprehensively evaluates direct and indirect relationships among the user, a voter and related comments of the user and the voter. And a consumer can be helped to quickly distinguish effective and false information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of user comment analysis, and particularly relates to a user comment quality double-view reasoning method and system based on state transition learning. BACKGROUND

[0002] On e-commerce websites, user-generated reviews have great value for potential consumers. However, the number of reviews is very large, which makes it difficult for consumers to find useful reviews. In addition, consumers are also disturbed by a large amount of invalid and false information. Therefore, identifying and utilizing useful reviews is crucial to help consumers make decisions. In today's practice, many e-commerce websites sort product reviews according to the posting time and / or product ratings of the reviews. However, these review rating methods only consider some direct and simple factors, resulting in their limited help for consumers.

[0003] Existing research on the usefulness of reviews mainly focuses on text analysis, data mining, etc. Many studies ignore the quality of users, especially the behavior of reviewers and voters. These methods have not been able to effectively evaluate the actual usefulness of reviews, but only rely on the text content of the reviews, which cannot fully reflect the real usefulness of the reviews. For example, it is easy for malicious users or water armies to take advantage of it to fake reviews to mislead consumers. Therefore, many e-commerce websites (such as Amazon) use crowdsourcing mechanisms to invite users to vote to evaluate the usefulness of reviews. Such a voting question may be: "Does this review help you?", Users can choose: "Yes" or "No" to vote for or against the review. And the voting result may be: "20 people think the following review is helpful", such a voting result only reflects the number of positive votes (evaluations), and does not consider the evaluations of all users.

[0004] In summary, on the one hand, good users tend to write useful reviews, and potential consumers are more likely to believe reviews written by well-known users; on the other hand, good users tend to vote reasonably on reviews, such as voting for useful reviews, and it is difficult for consumers to distinguish between valid and false information without comprehensively evaluating the direct and indirect relationships between users, voters, and their related reviews. SUMMARY

[0005] Therefore, the present application provides a user comment quality double-view reasoning method and system based on state transition learning, which comprehensively evaluates the direct and indirect relationships between users, voters, and their related reviews, and helps consumers quickly distinguish between valid and false information.

[0006] In a first aspect, the present application provides a user comment quality double-view reasoning method based on state transition learning, the method comprising:

[0007] Step 1, obtain user's review data of product, calculate user-user, user-review and review-review three kinds of association relationship; Step 2, generate edge set based on the calculated association relationship, establish multi-attribute user review heterogeneous graph; Step 3, based on the multi-attribute user review heterogeneous graph, formulate user quality dependent rule, design state transition learning algorithm to iteratively infer user quality; Step 4, based on user quality, establish review usefulness dependent rule, use linear function to weigh content usefulness CBH and content influence IBH, predict the review quality of general perspective under all products; Step 5, based on the multi-attribute user review heterogeneous graph, use the designed deep network with graph aggregation unit and multi-head self-attention mechanism to predict the review quality of specific perspective under given user.

[0008] Optionally, the step 1 comprises: Step 1.1, obtain the review data of product of the user of the platform; Step 1.2, extract user set And review set , calculate Cartesian product , traverse all user-review pairs in the Cartesian product ; When Write from , define the association relationship between And As , wherein Is the attribute label of the current association relationship; When Voted on , define the association relationship between And As , wherein 1 is the attribute label of the current association relationship; Step 1.3, take out user set , calculate unordered product , traverse all user-user pairs in the unordered product , wherein ; When And Voted on the same review, define the association relationship between And As , wherein Is the attribute label of the current association relationship; When Voted on the review written by , define the association relationship between And The relationship between them is ,in The attribute labels for the current association; Step 1.4: Retrieve the comment collection Calculate the unordered product Iterate through and compute all comment-comment pairs in the unordered product. ,in ;when and All were voted on by the same user, defining the comments. and The relationship between them is ,in The attribute label for the current relationship; when and All were written by the same user and defined. and The relationship between them is ,in For the attribute labels of the current relationship; when writing users A vote was held to define and The relationship between them is ,in The attribute label for the current relationship.

[0009] Optionally, step 2 includes: Step 2.1: Calculate the user set and comment collection The union of the nodes forms a node set. Calculate the union of all relationships between users and comments to form a multi-attribute edge set. ;in, The union of user-comment pairs; The union of user-user pairs; The union of comments and comment pairs; Step 2.2: Construct the node adjacency matrix ,in The number of nodes (elements) in the data; using Represents a node and nodes The relationship between them, if Then the matrix The Line 1 Columns are assigned a value of 1, otherwise a value of 0; a heterogeneous graph of multi-attribute user reviews is built based on the node adjacency matrix.

[0010] Optionally, step 3 includes: Step 3.1: Formulate the user quality dependency rules as follows: A1: User Quality depends on voting support The user who commented; A2: User Quality depends on and Users who voted for the same comment; Step 3.2: Calculate and generate based on user quality dependency rules All user sets that it depends on , Quality and The quantity and quality of users in the data are both positively correlated; Step 3.3: Initialize each user mass fraction is ,in ; Design a state transition function to support iteration; its expression is: (1); in, For the outgoing frequency of relevant users, express The quality score at the next iteration. These are the mass fraction iteration coefficients, used to adjust the initial mass fraction term. With subsequent iterations of quality score terms Importance percentage; and They represent The quality-dependent and quantity-dependent characteristics; Step 3.4: Begin iterative learning; first, generate users using the state transition function. quality score Then calculate the first Next and first The absolute difference in quality scores between each iteration | |; If the absolute difference in quality scores for all users is less than the threshold If the iteration fails, the iteration terminates; otherwise, continue to the next iteration. iteration For hyperparameters; Step 3.5: Replace the state transition function used in the iterative process with a learnable transition function, the expression of which is: (2); Among them, the mass vector is used. To replace the first in the state transition function Quality score of the next iteration ; The transition probability matrix, used to capture the quality dependencies between users, is expressed as follows: (3).

[0011] Optionally, step 4 includes: Step 4.1: Establish the comment usefulness dependency rules as follows: B1: Comments Its usefulness depends on the quality of its content; B2: Comments Its usefulness depends on the quality of the writer; B3: Comments Its usefulness depends on other factors. Comments from co-authors or voters; B4: Comments Its usefulness depends on Comments written by users who voted; Step 4.2: Based on the comment usefulness dependency rule, integrate comment content and related influence through linear combination to predict comments from a general perspective across all products. quality Its expression is: (4); in, Content Influence (IBH) is related to rules B2-B4; For content usefulness CBH, it is associated with rule B1. These are parameters used to weigh the impact of content influence against content usefulness. Step 4.3: Calculate based on lightweight semantic feature method The CBH score is calculated by first identifying product keywords for five themes using entity extraction technology. These keywords include descriptive terms, functions, performance, appearance, and technical specifications. Then, the CBH score is generated based on the keyword frequency in the review content, expressed as: (5); in, For the first The keywords in the comments Normalized frequency in terms of content The weights that need to be learned; Step 4.4, Calculation IBH, based on B3 and B4, extracts points from the heterogeneous graph of multi-attribute user comments. Comment nodes, i.e., extraction and Comment nodes that have a relationship are then used. Indicates pointing to The collection of all comment nodes; based on Define the dependencies of IBH: C1: The more comment elements in a comment node, the more comment nodes there are. The greater the in-degree, The higher the IBH; C2: The higher the IBH of the comment element, the better. The higher the IBH; Step 4.5, Attribute Tags The edges do not involve third-party voting and only reflect the self-dependency of the same writer, therefore they need to be cleaned. Exclude and exist Comment nodes with type relationships; Step 4.6: Initialize each comment IBH is ,in ; Design the state transition function of IBH to support iteration, its expression is: (6); in, for The author's user quality score For comments The degree of exit, exist IBH at the next iteration These are state transition coefficients, used to adjust the initial state terms. With subsequent state items Importance percentage; and These represent the quantity and quality characteristics of IBH, respectively. Step 4.7: Begin iterative learning; first, use the IBH state transition function to generate... IBH score Then calculate with the first The absolute difference of IBH in each order | |; If the absolute difference of IBH for all comments is less than the threshold If the condition is met, the calculation stops; otherwise, continue to the next step. iteration For hyperparameters; Step 4.8: To achieve state transition learning, the iterative process is transformed into matrix operations. During the iteration, a learnable transition function is used to replace the IBH state transition function, with the following expression: (7); The above equation converges according to the Markov chain convergence theorem, and the IBH score of the final iteration step is obtained. for A diagonal matrix, its first... Each diagonal value corresponds to a comment. The author's user quality score; for The transition probability matrix is ​​expressed as follows: (8).

[0012] Optionally, step 5 includes: Step 5.1: Generate initial feature vectors for users and comments; given by... Written by the author 1 comment and by Voters vote to construct a Author Comments Matrix If the user Wrote a comment ,but Otherwise Similarly, construct a Voter Comment Matrix If a voting value exists, then Equal to the vote value, otherwise equal to ;use The OK represent The initial writer characteristics, using The OK represent Initial voter characteristics, using The List represent Initial characteristics; Step 5.2: Establish graph aggregation cells and calculate hidden states; for any target node Given a set of neighbor nodes Calculate the hidden state of the target node. Its expression is: (9); in, For activation function, and These are the learnable weights and biases, respectively; the aggregation function. Defined as , for eigenvectors, As weight, Defined as ,in For the target node eigenvectors; aggregation functions and the internal weights of the function Substituting the calculation expression into formula (9) and combining them, we obtain the graph aggregation unit: (10); Step 5.3, User Defined as three potential roles: comment writer, comment voter, and interactor, where interaction refers to indirect interaction formed through voting with other users; As the target node of the graph aggregation unit, the hidden state vectors of the three roles are calculated based on the graph aggregation unit. , and ; Step 5.4: Calculation based on graph aggregation unit Its expression is: (11); Among them, neighbor node group for All comments written; The feature vector is ,in for The initial feature vector, To use convolutional neural networks (CNNs) The content consists of complete sentences and word sequence pre-trained word feature vectors; target nodes. The feature vector is its initial writer feature. ; Step 5.5: Calculation based on graph aggregation unit Its expression is: (12); Among them, neighbor node group for All comments that have been voted on; The feature vector is ,in The corresponding element values ​​of the voter comment matrix one-hot vector, To use CNN The content consists of complete sentences and word sequence pre-trained word feature vectors; target nodes. The feature vector is its initial voter feature. ; Step 5.6: Calculation based on graph aggregation unit Its expression is: (13); in, For calculation and generation based on user quality dependency rules All user sets it depends on; The feature vector is Target node The feature vector is its initial voter feature. ; Step 5.7: Comments The influencing factors are defined as three categories: writer influence, voter influence, and indirect interaction influence. Interaction refers to the indirect interaction influence formed through co-writers and other comments. As the target node of the graph aggregation unit, the hidden state vectors of the three influences are calculated based on the graph aggregation unit. , and ; Step 5.8: Calculation based on graph aggregation unit Its expression is: (11); Among them, neighbor node group for The author; The feature vector is ,in for The initial writer feature vector, To use CNN The content consists of complete sentences and word sequence pre-trained word feature vectors; target nodes. The feature vector is its initial feature ; Step 5.9: Calculation based on graph aggregation unit Its expression is: (12); Among them, neighbor node group for All voters; The feature vector is ,in The corresponding element values ​​of the voter comment matrix one-hot vector, for Initial voter characteristics; target node The feature vector is its initial feature ; Step 5.10: Calculation based on graph aggregation unit Its expression is: (13); in, For pointing The collection of all comment nodes; The feature vector is Target node The feature vector is its initial feature ; Step 5.11: Design the State Fusion Layer It is used to capture and fuse the correlations between hidden states; to fuse user hidden states. , and To achieve this, we first use a multi-head self-attention mechanism to learn the intermediate representations of the hidden states. , and ; to learn The middle representation To achieve the goal, the process is as follows: ; in, The weight matrix is ​​a learnable matrix; for and The correlation between them is the upper-level attention, where Its expression is: ; in, The lower-level attention function is defined as follows: ,in Indicates the inner product. and The weight matrix is ​​then learned; a standard residual connection network is then used to combine the original hidden state vector. and the updated intermediate representation , recorded as Its expression is: ; in, For activation function, The projection matrix is ​​learnable; Obtain in the same way and Finally, , and The concatenation is performed, and the output of the state fusion layer is expressed as follows: ; Step 5.12: Based on the output of the state fusion layer, model each user. The final expression And every comment The final expression Its expression is: ; ; Step 5.13: Using a Multilayer Perceptron (MLP) and Predicting comment quality from a specific user's perspective Its expression is: ; in,[ , ]for and splicing; Step 5.14: All parameters of the model are initialized randomly based on a uniform distribution. The loss function used for model training is expressed as follows: ; The objective of the above equation is to minimize the total loss between the predicted usefulness score and the corresponding true value, where for Predicted comment quality This corresponds to the actual value.

[0013] Secondly, this invention provides a dual-perspective inference system for user review quality based on state transition learning, the system comprising: The data acquisition module is used to acquire user review data for the product; The relationship calculation module is used to calculate three types of relationships: user-user, user-comment, and comment-comment. The multi-attribute heterogeneous graph construction module is used to generate edge sets based on computational relationships and build a multi-attribute user review heterogeneous graph. The user quality inference module is used to formulate user quality dependency rules based on heterogeneous graphs of multi-attribute user reviews and design state transition learning algorithms to iteratively infer user quality. The general perspective review quality inference module is used to establish review usefulness dependency rules based on user quality, and use a linear function to weigh content usefulness (CBH) and content influence (IBH) to predict the general perspective review quality for all products. The Specific Perspective Comment Quality Inference Module is designed to predict the comment quality of a given user from a specific perspective based on a heterogeneous graph of multi-attribute user comments and by utilizing a deep network with graph aggregation units and a multi-head self-attention mechanism.

[0014] The technical solution provided by this invention includes the following steps: acquiring user review data for products; calculating three types of relationships: user-user, user-review, and review-review; generating edge sets based on the calculated relationships; establishing a multi-attribute user review heterogeneous graph; formulating user quality dependency rules based on the multi-attribute user review heterogeneous graph; designing a state transition learning algorithm to iteratively infer user quality; establishing review usefulness dependency rules based on user quality; using a linear function to balance content usefulness (CBH) and content influence (IBH) to predict review quality from a general perspective for all products; and using a designed deep network with graph aggregation units and a multi-head self-attention mechanism based on the multi-attribute user review heterogeneous graph to predict review quality from a specific perspective for a given user. This method comprehensively evaluates the direct and indirect relationships between users, voters, and their related reviews, helping consumers quickly distinguish between valid and false information. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the dual-perspective inference method for user review quality based on state transition learning provided in this embodiment of the invention; Figure 2 A flowchart illustrating the iterative inference of user quality based on a state transition learning algorithm provided in this embodiment of the invention; Figure 3 A flowchart for predicting review quality from a general perspective across all products, provided as an embodiment of the present invention; Figure 4 This is a structural diagram for predicting the quality of comments from a specific perspective for a given user, as provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0020] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0021] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0022] Figure 1 A flowchart of the dual-perspective inference method for user review quality based on state transition learning provided in this embodiment of the invention is shown below. Figure 1 As shown, the method includes: Step 1: Obtain user review data for the product and calculate three types of relationships: user-user, user-review, and review-review.

[0023] In this embodiment of the invention, step 1 includes: Step 1.1: Obtain user review data for the product from the platform; In this embodiment of the invention, Amazon China platform is selected to obtain user review data for the product. It should be noted that the platform here is only an example, and other platforms can also be selected.

[0024] Step 1.2: Extract the user set and comment collection Calculate the Cartesian product Iterate through and calculate all user-comment pairs in the Cartesian product. ;when Written by Then define and The relationship between them is ,in The attribute label for the current relationship; when right A vote was held to define and The relationship between them is ,in 1 represents the attribute label of the current association; Step 1.3: Retrieve the user set Calculate the unordered product Iterate through and compute all user-user pairs in the unordered product. ,in ;when and A vote was cast on the same comment, defining and The relationship between them is ,in The attribute label for the current relationship; when For writing self The comments were voted on, defining and The relationship between them is ,in The attribute labels for the current association; Step 1.4: Retrieve the comment collection Calculate the unordered product Iterate through and compute all comment-comment pairs in the unordered product. ,in ;when and All were voted on by the same user, defining the comments. and The relationship between them is ,in The attribute label for the current relationship; when and All were written by the same user and defined. and The relationship between them is ,in For the attribute labels of the current relationship; when writing users A vote was held to define and The relationship between them is ,in The attribute label for the current relationship.

[0025] Step 2: Generate edge sets based on the calculated relationships and establish a heterogeneous graph of multi-attribute user comments.

[0026] In this embodiment of the invention, step 2 includes: Step 2.1: Calculate the user set and comment collection The union of the nodes forms a node set. Calculate the union of all relationships between users and comments to form a multi-attribute edge set. ;in, The union of user-comment pairs; The union of user-user pairs; The union of comments and comment pairs; Step 2.2: Construct the node adjacency matrix ,in The number of nodes (elements) in the data; using Represents a node and nodes The relationship between them, if Then the matrix The Line 1 Columns are assigned a value of 1, otherwise a value of 0; a heterogeneous graph of multi-attribute user reviews is built based on the node adjacency matrix.

[0027] Step 3: Based on the heterogeneous graph of multi-attribute user reviews, formulate user quality dependency rules and design a state transition learning algorithm to iteratively infer user quality.

[0028] In embodiments of the present invention, such as Figure 2 As shown, step 3 includes: Step 3.1: Formulate the user quality dependency rules as follows: A1: User Quality depends on voting support The user who commented; A2: User Quality depends on and Users who voted for the same comment; Step 3.2: Calculate and generate based on user quality dependency rules All user sets that it depends on , Quality and The quantity and quality of users in the data are both positively correlated; Step 3.3: Initialize each user mass fraction is ,in ; Design a state transition function to support iteration; its expression is: (1); in, For the outgoing frequency of relevant users, express The quality score at the next iteration. These are the mass fraction iteration coefficients, used to adjust the initial mass fraction term. With subsequent iterations of quality score terms Importance percentage; and They represent The quality-dependent and quantity-dependent characteristics; Step 3.4: Begin iterative learning; taking the t-th iteration as an example, the calculation process for each iteration is introduced. First, the user is generated using the state transition function. quality score Then calculate the first Next and first The absolute difference in quality scores between each iteration | |; If the absolute difference in quality scores for all users is less than the threshold If the iteration fails, the iteration terminates; otherwise, continue to the next iteration. iteration For hyperparameters; Step 3.5: Replace the state transition function used in the iterative process with a learnable transition function, the expression of which is: (2); Among them, the mass vector is used. To replace the first in the state transition function Quality score of the next iteration ; The transition probability matrix, used to capture the quality dependencies between users, can be quickly constructed from the user-user edges in the user review graph, and its expression is: (3).

[0029] In this embodiment of the invention, after substitution, the first term in formula (2) It is unchangeable, the second item. It satisfies the existence condition of the stationary distribution of the Markov chain, ensuring that it will inevitably converge to a unique steady-state solution after sufficient iterations. The uniqueness of the steady-state solution can guarantee that the final user quality score is consistent under different initial values.

[0030] Step 4: Based on user quality, establish a comment usefulness dependency rule, and use a linear function to weigh content usefulness (CBH) and content influence (IBH) to predict comment quality from a general perspective across all products.

[0031] In embodiments of the present invention, such as Figure 3 As shown, step 4 includes: Step 4.1: Establish the comment usefulness dependency rules as follows: B1: Comments Its usefulness depends on the quality of its content; B2: Comments Its usefulness depends on the quality of the writer (user); B3: Comments Its usefulness depends on other factors. Comments from co-authors or voters; B4: Comments Its usefulness depends on Comments written by users who voted; Step 4.2: Based on the comment usefulness dependency rule, where B1 is related to the comment content, and B2-B4 involve the influence from the writer, voters, and related comments; by linearly combining the comment content and related influences, predict comments from a general perspective across all products. quality Its expression is: (4); in, Content Influence (IBH) is related to rules B2-B4; For content usefulness CBH, it is associated with rule B1. These are parameters used to weigh the impact of content influence against content usefulness. Step 4.3: Calculate based on lightweight semantic feature method The CBH score is calculated by first identifying product keywords for five themes using entity extraction technology. These keywords include descriptive terms, functions, performance, appearance, and technical specifications. Then, the CBH score is generated based on the keyword frequency in the review content, expressed as: (5); in, For the first The keywords in the comments Normalized frequency in terms of content The weights to be learned; the usefulness of a comment is evaluated by considering the frequency of specific keywords appearing in the comment and the importance weight of those specific keywords. Step 4.4, Calculation IBH, based on B3 and B4, extracts points from the heterogeneous graph of multi-attribute user comments. Comment nodes, i.e., extraction and Comment nodes that have a relationship are then used. Indicates pointing to The collection of all comment nodes; based on Define the dependencies of IBH: C1: The more comment elements in a comment node, the more comment nodes there are. The greater the in-degree, The higher the IBH; C2: The higher the IBH of the comment element, the better. The higher the IBH; Step 4.5, Attribute Tags The edges do not involve third-party voting and only reflect the self-dependency of the same writer, therefore they need to be cleaned. Exclude and exist Comment nodes with type relationships; Step 4.6: Initialize each comment IBH is ,in ; Design the state transition function of IBH to support iteration, its expression is: (6); in, for The author's user quality score For comments The degree of exit, exist IBH at the next iteration These are state transition coefficients, used to adjust the initial state terms. With subsequent state items Importance percentage; and These represent the quantity and quality characteristics of IBH, respectively. Step 4.7: Begin iterative learning; taking the t-th iteration as an example, the calculation process for each iteration is introduced. First, the IBH state transition function is used to generate... IBH score Then calculate with the first The absolute difference of IBH in each order | |; If the absolute difference of IBH for all comments is less than the threshold If the condition is met, the calculation stops; otherwise, continue to the next step. iteration For hyperparameters; Step 4.8: To achieve state transition learning, the iterative process is transformed into matrix operations. During the iteration, a learnable transition function is used to replace the IBH state transition function, with the following expression: (7); The above equation converges according to the Markov chain convergence theorem, and the IBH score of the final iteration step is obtained. for A diagonal matrix, its first... Each diagonal value corresponds to a comment. The author's user quality score; for The transition probability matrix can be quickly constructed from the comment-comment edges in the user comment graph, and its expression is: (8).

[0032] Step 5: Based on the heterogeneous graph of multi-attribute user reviews, use the designed deep network with graph aggregation units and multi-head self-attention mechanism to predict the review quality from a specific perspective for a given user.

[0033] In embodiments of the present invention, such as Figure 4 As shown, step 5 includes: Step 5.1: Generate initial feature vectors for users and comments; given by... Written by the author 1 comment and by Voters vote to construct a Author Comments Matrix If the user Wrote a comment ,but Otherwise Similarly, construct a Voter Comment Matrix If a voting value exists, then Equal to the vote value, otherwise equal to ;use The OK represent The initial writer characteristics, using The OK represent Initial voter characteristics, using The List represent Initial characteristics; Step 5.2: Establish graph aggregation cells and calculate hidden states; for any target node Given a set of neighbor nodes Calculate the hidden state of the target node. Its expression is: (9); in, For activation function, and These are the learnable weights and biases, respectively; the aggregation function. Defined as , for eigenvectors, As weight, Defined as ,in For the target node eigenvectors; aggregation functions and the internal weights of the function Substituting the calculation expression into formula (9) and combining them, we obtain the graph aggregation unit: (10); Step 5.3, User Defined as three potential roles: comment writer, comment voter, and interactor, where interaction refers to indirect interaction formed through voting with other users; As the target node of the graph aggregation unit, the hidden state vectors of the three roles are calculated based on the graph aggregation unit. , and ; Step 5.4: Calculation based on graph aggregation unit Its expression is: (11); Among them, neighbor node group for All comments written; The feature vector is ,in for The initial feature vector, To use convolutional neural networks (CNNs) The content consists of complete sentences and word sequence pre-trained word feature vectors; target nodes. The feature vector is its initial writer feature. ; Step 5.5: Calculation based on graph aggregation unit Its expression is: (12); Among them, neighbor node group for All comments that have been voted on; The feature vector is ,in The corresponding element values ​​of the voter comment matrix one-hot vector, To use CNN The content consists of complete sentences and word sequence pre-trained word feature vectors; target nodes. The feature vector is its initial voter feature. ; Step 5.6: Calculation based on graph aggregation unit Its expression is: (13); in, For calculation and generation based on user quality dependency rules All user sets it depends on; The feature vector is Target node The feature vector is its initial voter feature. ; Step 5.7: Comments The influencing factors are defined as three categories: writer influence, voter influence, and indirect interaction influence. Interaction refers to the indirect interaction influence formed through co-writers and other comments. As the target node of the graph aggregation unit, the hidden state vectors of the three influences are calculated based on the graph aggregation unit. , and ; Step 5.8: Calculation based on graph aggregation unit Its expression is: (11); Among them, neighbor node group for The author; The feature vector is ,in for The initial writer feature vector, To use CNN The content consists of complete sentences and word sequence pre-trained word feature vectors; target nodes. The feature vector is its initial feature ; Step 5.9: Calculation based on graph aggregation unit Its expression is: (12); Among them, neighbor node group for All voters; The feature vector is ,in The corresponding element values ​​of the voter comment matrix one-hot vector, for Initial voter characteristics; target node The feature vector is its initial feature ; Step 5.10: Calculation based on graph aggregation unit Its expression is: (13); in, For pointing The collection of all comment nodes; The feature vector is Target node The feature vector is its initial feature ; Step 5.11: Design the State Fusion Layer It is used to capture and fuse the correlations between hidden states; to fuse user hidden states. , and To achieve this, we first use a multi-head self-attention mechanism to learn the intermediate representations of the hidden states. , and ; to learn The middle representation To achieve the goal, the process is as follows: ; in, The weight matrix is ​​a learnable matrix; for and The correlation between them is the upper-level attention, where Its expression is: ; in, The lower-level attention function is defined as follows: ,in Indicates the inner product. and The weight matrix is ​​then learned; a standard residual connection network is then used to combine the original hidden state vector. and the updated intermediate representation , recorded as Its expression is: ; in, For activation function, The projection matrix is ​​learnable; Obtain in the same way and Finally, , and The concatenation is performed, and the output of the state fusion layer is expressed as follows: ; Step 5.12: Based on the output of the state fusion layer, model each user. The final expression And every comment The final expression Its expression is: ; ; Step 5.13: Using a Multilayer Perceptron (MLP) and Predicting comment quality from a specific user's perspective Its expression is: ; in,[ , ]for and splicing; Step 5.14: All parameters of the model are initialized randomly based on a uniform distribution. The loss function used for model training is expressed as follows: ; The objective of the above equation is to minimize the total loss between the predicted usefulness score and the corresponding true value, where for Predicted comment quality This corresponds to the actual value.

[0034] This invention provides a dual-perspective inference system for user review quality based on state transition learning. The system includes: a data acquisition module, a relation calculation module, a multi-attribute heterogeneous graph construction module, a user quality inference module, a general perspective review quality inference module, and a specific perspective review quality inference module. The system comprises the following modules: a data acquisition module for acquiring user reviews of products; a relationship calculation module for calculating user-user, user-review, and review-review relationships; a multi-attribute heterogeneous graph construction module for generating edge sets based on calculated relationships and establishing a multi-attribute user review heterogeneous graph; a user quality inference module for formulating user quality dependency rules based on the multi-attribute user review heterogeneous graph and designing a state transition learning algorithm for iterative inference of user quality; a general perspective review quality inference module for establishing review usefulness dependency rules based on user quality, using a linear function to balance content usefulness (CBH) and content influence (IBH) to predict review quality from a general perspective for all products; and a specific perspective review quality inference module for predicting review quality from a specific perspective for a given user based on the multi-attribute user review heterogeneous graph using a designed deep network with graph aggregation units and a multi-head self-attention mechanism.

[0035] This invention comprehensively assesses the direct and indirect relationships between users, voters, and their related comments. It constructs a multi-attribute heterogeneous graph of user (consumer) comment data using a structured representation, capturing complex indirect or direct dependencies and reducing the difficulty of natural language text analysis. Based on the multi-attribute heterogeneous graph, it constructs a transition probability matrix, formulates user quality dependency rules, and learns and infers stable user quality scores through state transitions. It also formulates comment usefulness dependency rules, calculates comment content usefulness (CBH) scores using a lightweight semantic feature method, and calculates comment content influence (IBH) scores based on Markov convergence using multivariate relationships in the heterogeneous graph. A learnable linear function is designed to balance IBH and CBH, obtaining general-perspective comment quality inference for a given product. Finally, it integrates multiple potential influencing factors using graph aggregation units to generate hidden user state representations, and uses multi-head self-attention to fusion states to predict comment quality from a specific perspective for a given user, helping consumers quickly distinguish between valid and false information.

[0036] The technical solution provided by this invention includes the following steps: acquiring user review data for products; calculating three types of relationships: user-user, user-review, and review-review; generating edge sets based on the calculated relationships; establishing a multi-attribute user review heterogeneous graph; formulating user quality dependency rules based on the multi-attribute user review heterogeneous graph; designing a state transition learning algorithm to iteratively infer user quality; establishing review usefulness dependency rules based on user quality; using a linear function to balance content usefulness (CBH) and content influence (IBH) to predict review quality from a general perspective for all products; and using a designed deep network with graph aggregation units and a multi-head self-attention mechanism based on the multi-attribute user review heterogeneous graph to predict review quality from a specific perspective for a given user. This method comprehensively evaluates the direct and indirect relationships between users, voters, and their related reviews, helping consumers quickly distinguish between valid and false information.

[0037] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A state transition learning based user review quality dual-view reasoning method, characterized in that, The method comprises: Step 1, obtaining user review data of a product, and calculating user-user, user-review and review-review three types of association relationships; Step 2, generating an edge set based on the calculated association relationships, and establishing a multi-attribute user review heterogeneous graph; Step 3, formulating a user quality dependency rule based on the multi-attribute user review heterogeneous graph, designing a state transition learning algorithm to iteratively infer user quality; Step 4, establishing a review usefulness dependency rule based on user quality, and using a linear function to weigh content usefulness CBH and content influence IBH to predict the review quality of a general perspective under all products; Step 5, based on the multi-attribute user review heterogeneous graph, using the designed deep network with graph aggregation unit and multi-head self-attention mechanism to predict the review quality of a specific perspective under a given user.

2. The method of claim 1, wherein, The step 1 comprises: Step 1.1, obtaining user review data of a product on a platform; Step 1.2: Extract the user set and comment collection Calculate the Cartesian product Iterate through and calculate all user-comment pairs in the Cartesian product. ;when Written by Then define and The relationship between them is ,in The attribute label for the current relationship; when right A vote was held to define and The relationship between them is ,in 1 represents the attribute label of the current association; Step 1.3: Retrieve the user set Calculate the unordered product Iterate through and compute all user-user pairs in the unordered product. ,in ;when and A vote was cast on the same comment, defining and The relationship between them is ,in The attribute label for the current relationship; when For writing self The comments were voted on, defining and The relationship between them is ,in The attribute labels for the current association; Step 1.4, take out the review set , calculate the disordered product , traverse all the review-review pairs in the disordered product , wherein ; when and are both voted by the same user, define the association relationship between and as , wherein is the attribute label of the current association relationship; when and are both written by the same user, define the association relationship between and as , wherein is the attribute label of the current association relationship; when the user who writes has voted on , define the association relationship between and as , wherein is the attribute label of the current association relationship.

3. The method of claim 2, wherein, The step 2 comprises: Step 2.1, compute the union of the user set and the comment set , forming the node set ; compute the union of all the associations between users and comments, forming the multi-attribute edge set ; where, is the union of user-comment pairs; is the union of user-user pairs; is the union of comment-comment pairs; Step 2.2: Construct the node adjacency matrix ,in The number of nodes in; using Represents a node and nodes The relationship between them, if Then the matrix The Line number Columns are assigned a value of 1, otherwise a value of 0; a heterogeneous graph of multi-attribute user reviews is built based on the node adjacency matrix.

4. The method of claim 3, wherein, The step 3 comprises: Step 3.1, formulating a user quality dependency rule as follows: A1 : a user the quality of the vote depends on the users supporting the comment; A2: the user the quality depends on and users who have voted for the same review; Step 3.

2. Compute generation based on user quality dependency rules All user sets that are dependent , Quality is positively correlated with both the number and quality of users in Step 3.3, initialize each user with a mass fraction of where ; A state transition function is designed to support iteration, and its expression is: （1）； wherein, is the out-degree of the relevant user, denotes the quality score at the next iteration, is the quality score iteration coefficient, used to adjust the importance of the initial quality score term to the subsequent iteration quality score term ; and denote the quality-dependent and quantity-dependent characteristics of , respectively; Step 3.4, start iteration learning; first generate user quality score , then calculate the absolute difference between the quality score of the th iteration and the th iteration | | ; if the absolute difference of the quality score of all users is less than a threshold , terminate the iteration, otherwise continue the th iteration, is a hyperparameter; Step 3.5: replace the state transition function used in the iteration process with a learnable transition function, and its expression is: （2）； where the quality vector is used to replace the th iteration quality score, in the state transition function; is the transition probability matrix, which is used to capture the quality dependence between users, and its expression is: （3）。 5. The method of claim 4, wherein, The step 4 comprises: Step 4.1, establishing a review usefulness dependency rule as follows: B1: review The usefulness of a review depends on the quality of its content; B2: Reviews usefulness depends on the quality of the writers; B3: Reviews The usefulness of the reviews of other reviews that have common authors or voters; B4: Reviews The usefulness of the application depends on the availability of reviews written by users who have voted on the item. Step 4.2, based on the review usefulness dependency rule, integrate the review content and the related influence by linear combination to predict the review of general perspective under all products of the quality The expression is: （4）； wherein, is a content impact IBH, associated with rules B2-B4; is a content usefulness CBH, associated with rules B1, is a trade-off parameter for weighing the impact of content impact and content usefulness; Step 4.

3. Calculate based on lightweight semantic feature method CBH; first, the product keywords of five topics are identified by entity extraction technology to calculate the CBH score, and the product keywords of the five topics include description words, functions, performance, appearance and technical indicators; then, the CBH score is generated according to the keyword frequency in the review content, and the expression is: （5）； wherein, is the number of keywords in the content of the review, is the normalized frequency of the keyword in the content of the review, is the number of keywords in the content of the review, is the weight to be learned; Step 4.4, computing IBH based on B3 and B4, extract the multi-attribute user review heterogeneous graph directed to review nodes, i.e., extract the review nodes associated with , then use to represent the set of all review nodes directed to ; based on , define the dependent factors of IBH: C1: The more the comment elements in the comment node The larger the in-degree of the comment node The higher the IBH of the comment node C2: The higher the IBH of the middle comment element, The higher the IBH of the middle comment element; Step 4.5, attribute tag Edges do not involve third-party votes, only reflect the self-reliance of the same writer, so they need to be cleaned , excluding review nodes with Exist type relationship; Step 4.6, initialize each review IBH for wherein ; A state transition function of IBH is designed to support iteration, and its expression is: （6）； wherein, is a user quality score of the author, is a comment out-degree, is IBH at the n-th iteration, is a state transition coefficient for adjusting the importance ratio of the initial state item to the subsequent state items; and represent the IBH number and quality characteristics, respectively;​ Step 4.7, start iterative learning; first generate IBH scores using IBH state transition function for all reviews then compute absolute difference of IBH scores for the first and second iteration ; if all absolute difference of IBH scores are less than a threshold then compute stop, otherwise continue the next iteration, for hyperparameters; Step 4.8, to realize state transition learning, the iteration process is converted into matrix operation, and the state transition function of IBH is replaced with a learnable transition function in the iteration, and its expression is: （7）； The above equation converges according to the Markov chain convergence theorem, and the IBH score of the final iteration step is obtained. for A diagonal matrix, its first... Each diagonal value corresponds to a comment. The author's user quality score; for The transition probability matrix is ​​expressed as follows: （8）。 6. The method of claim 5, wherein, The step 5 comprises: Step 5.1: Generate initial feature vectors for users and comments; given by... Written by the author 1 comment and by Voters vote to construct a Author Comments Matrix If the user Wrote a comment ,but Otherwise Similarly, construct a Voter Comment Matrix If a voting value exists, then Equal to the vote value, otherwise equal to ;use The OK represent The initial writer characteristics, using The OK represent Initial voter characteristics, using The List represent Initial characteristics; Step 5.2, establish graph aggregation unit, calculate hidden state; for any target node , given its set of neighbor nodes , calculate the hidden state of the target node , whose expression is: （9）； wherein, is an activation function, and are learnable weights and biases, respectively; an aggregation function is defined as , , is a feature vector of is a weight, is defined as wherein is a feature vector of the target node ; substituting the computational expressions of the aggregation function and the intra-function weights into equation (9) and combining, a graph aggregation unit is obtained: （10）； Step 5.3, the user is defined as three potential roles of comment writer, comment voter and interactor, where the interaction refers to indirect interaction formed by voting with other users; the as the target node of the graph aggregation unit, based on the graph aggregation unit to calculate the hidden state vector of the three roles , and ; Step 5.4, graph aggregation unit based computation whose expression is: （11）； Among them, neighbor node group for All comments written; The feature vector is ,in for The initial feature vector, To use convolutional neural networks (CNNs) The content consists of complete sentences and word sequence pre-trained word feature vectors; target nodes. The feature vector is its initial writer feature. ; Step 5.5, graph aggregation unit computation whose expression is: （12）； wherein the group of neighbor nodes is all the comments voted by the target node ; the feature vector of the target node is , wherein is the corresponding element value of the voter comment matrix ; is the one-hot vector of ; the feature vector of the target node is the initial voter feature of the target node ; Step 5.6, graph aggregation unit based computation whose expression is: （13）； wherein, is computed based on the user quality dependent rules all the user sets on which the target node depends; the feature vector of the target node is its initial voter feature vector; Step 5.7, comment on the impact of the factors received The impact factors are defined as three categories of writer impact, voter impact and indirect interaction impact, where the interaction refers to the indirect interaction impact formed by the common writers and other comments; the As the target node of the graph aggregation unit, the hidden state vector of the three influences is calculated based on the graph aggregation unit 、 and ; Step 5.8, graph aggregation unit computation based on whose expression is: （11）； Among them, neighbor node group for The author; The feature vector is ,in for The initial writer feature vector, To use CNN The content consists of complete sentences and word sequence pre-trained word feature vectors; target nodes. The feature vector is its initial feature ; Step 5.

9. Calculate based on graph aggregation unit The expression is: （12）； wherein the group of neighbor nodes is all voters of the feature vector of is the corresponding element value of the voter review matrix the one-hot vector of is the initial voter feature of the feature vector of the target node is its initial feature Step 5.10, computing based on graph aggregation unit whose expression is: （13）； wherein, is a set of all review nodes pointing to ; is a feature vector of ; a feature vector of the target node is its initial feature ; Step 5.11, Designing the state fusion layer , to capture and fuse inter-state dependencies; to fuse user hidden states , and , first use multi-head self-attention mechanism to learn intermediate representations of hidden states , and ; to learn intermediate representations of , the process is as follows:​ ； wherein, is a learnable weight matrix; is and a correlation upper attention between whose expression is: ； where is the lower attention function, defined as where denotes the inner product, and are learnable weight matrices; then, a standard residual connection network is used to combine the original hidden state vector and the updated intermediate representation , denoted as with the expression: ； wherein, is an activation function, is a learnable projection matrix; In the same way, we obtain and ; finally, we concatenate , and as the output of the state fusion layer, whose expression is: ； Step 5.

12. Model the final representation of each user based on the output of the state fusion layer and the final representation of each review whose expression is:​ ； ； Step 5.13, predicting the review quality for a given user and a particular perspective using a multi-layer perceptron, MLP, by and predicting the review quality for a given user and a particular perspective whose expression is: ； wherein, , ] is a splice of and . Step 5.14, all parameters of the model are randomly initialized based on uniform distribution, and the loss function expression used in model training is: ； The objective of the above formula is to minimize the total loss between the predicted usefulness score and the corresponding true value, where is the predicted review quality, is the corresponding true value. 7.A state transition learning based user review quality dual-view reasoning system, characterized in that, The system comprises: A data acquisition module for acquiring user review data of a product; A relationship calculation module for calculating user-user, user-review and review-review three types of association relationships; A multi-attribute heterogeneous graph construction module for generating an edge set based on the calculated association relationships, and establishing a multi-attribute user review heterogeneous graph; A user quality inference module for formulating a user quality dependency rule based on the multi-attribute user review heterogeneous graph, and designing a state transition learning algorithm to iteratively infer user quality; A general perspective review quality inference module for establishing a review usefulness dependency rule based on user quality, and using a linear function to weigh content usefulness CBH and content influence IBH to predict the review quality of a general perspective under all products; A specific perspective review quality inference module for predicting the review quality of a specific perspective under a given user based on the multi-attribute user review heterogeneous graph and using the designed deep network with graph aggregation unit and multi-head self-attention mechanism.

Images