A review usefulness prediction method considering seed information and causality

Abstract

The application discloses a review usefulness prediction method considering seed information and causality, comprising the following steps: obtaining review texts and corresponding non-text data, thereby constructing a review data set D; obtaining user review preferences, thereby constructing a seed topic word distribution φ s ; constructing a Bayesian seed topic regression model based on the review data set D and the seed topic word distribution φ s ; initializing all parameters in steps S2 and S3 based on the review data set D, and performing parameter inference on a document topic distribution, a topic word distribution and a review usefulness prediction distribution by using an EM algorithm and a Gibbs sampling method. The application allows users to guide the theme discovery process by adding seed information, thereby quickly and accurately mining beneficial themes that are concerned by users, and meanwhile, the prediction accuracy is improved by jointly modeling the review texts and the review related data, so that the application can be widely applied to the fields of causality inference and linguistics.

Description

Technical Field

[0001] This invention relates to the field of comment usefulness prediction technology, and more specifically to a comment usefulness prediction method that considers seed information and causal relationships. Background Technology

[0002] The rapid development of the internet has made online interaction a global, everyday phenomenon, bringing immense convenience to people's lives. At the same time, the resulting massive amounts of textual comment data have become readily available. This data contains behavioral characteristics of society, markets, businesses, and consumers. Therefore, in order to realize the value of this data, researchers must be able to extract underlying data features to measure, understand, and interpret the causes and consequences of the market behaviors it represents.

[0003] To extract latent data features from unstructured comment text data, researchers have proposed topic modeling techniques. Topic modeling can help us extract implicit semantic information from text. However, for marketers, simply obtaining the latent semantic features of the text is not enough; they hope to gain valuable user feedback from these features and use it to improve products or services to enhance their core competitiveness.

[0004] However, existing methods for predicting comment usefulness using fusion topic models have three problems. First, extracting latent data features from text data using traditional topic models is an unsupervised process, and its biggest limitation is that the extracted data features are not always meaningful or of interest to users. Second, most current research focuses on the correlation between data to achieve better prediction results, neglecting the crucial role of causal relationships in intervening in and predicting comment usefulness. Third, some comment usefulness prediction studies that consider causal relationships often use a two-stage approach for causal inference, but since the text itself is considered as estimated data, if the text and the predicted data are estimated separately, the resulting causal effects will be biased. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting the usefulness of reviews that considers seed information and causal relationships. This method aims to solve the problems of existing text-based causal inference methods, which cannot jointly estimate data related to review text, review length, review rating, and overall rating, and cannot obtain coherent and meaningful thematic features from text data. This will improve prediction accuracy, and the review usefulness prediction results based on causal inference can effectively and directly guide marketers in making decisions, thereby improving the accuracy of those decisions.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The present invention provides a method for predicting the usefulness of comments that considers seed information and causal relationships, characterized by including:

[0008] S1. Obtain the comment text and its corresponding non-text data to construct the comment dataset D;

[0009] S11. Obtain the comment dataset D containing a set of comment texts W = {W1, W2, ..., W...} containing K common topics. d ,…,W D}, where |D| represents the number of comment texts in the comment text set W; W d Let d be the text of the comment. w d,n W represents the text of the dth comment. d The nth word in N d W represents the text of the dth comment. d The total number of words in the text;

[0010] Construct a dictionary V = {w1,...,w1,...} for the set of comment texts W. v ,...,w |V|}, where w v Let |V| represent the v-th word in dictionary V, and |V| represent the total number of unique words in the comment text set W.

[0011] S12. Obtain the non-text data set U = {Y, L, C1, C2} from the comment dataset D, where Y represents the comment usefulness rating in the comment dataset D, and Y = {y1, y2, ..., y d ,...,y |D|}, y d W represents the text of the dth comment. d The corresponding usefulness score of the comments; L represents the length of the comments in the comment dataset D, and L={l1,l2,...,l d ,...,l |D|}, l d W represents the text of the dth comment. d The corresponding comment length; C1 represents the comment rating in comment dataset D, and C1 = {c 1,1 ,c 1,2 ,...,c 1,d ,...,c 1,|D|}, c 1,d W represents the text of the dth comment. d The corresponding review rating; C2 represents the overall rating in the review dataset D, and C2 = {c 2,1 ,c 2,2 ,...,c 2,d ,...,c 2,|D|}, c2,d W represents the text of the dth comment. d The corresponding overall rating;

[0012] S2. Obtain user comment preferences to construct seed topic keyword distribution φ s ;

[0013] S21. Based on user interests and preferences, generate the word distribution under the k-th seed topic according to equation (1). This yields the word distribution across all seed topics.

[0014]

[0015] In equation (1), Let be the prior parameters of the Dirichlet distribution corresponding to the k-th seed topic, where ~ indicates that it follows the distribution, Dir(·) represents the Dirichlet distribution, and K represents the number of topics;

[0016] S3, Based on the comment dataset D and seed topic word distribution φ s Construct a Bayesian seed topic regression model;

[0017] S31. Define the word distribution under all regular topics as follows: in, Let represent the word distribution under the k-th regular topic, and Satisfying equation (2):

[0018]

[0019] In equation (2), Let be the prior parameters of the Dirichlet distribution corresponding to the k-th regular topic;

[0020] S32. Define the topic distribution of all comment texts as θ={θ1,...,θ d ,...,θ D}, where θ d W represents the text of the dth comment. d The topic distribution, and θ d Satisfying equation (3):

[0021] θ d ~Dir(α) d (3)

[0022] In equation (3), α d For the dth comment text W d The prior parameters of the corresponding Dirichlet distribution;

[0023] S33. Generate the d-th comment text W according to formula (4). dThe nth word w d,n Corresponding topic z d,n Thus, the d-th comment text W is obtained. d Theme This leads to the topic set Z = [z1,…,z] of the comment dataset D. d ,…,z |D| ];

[0024] z d,n ~Mult(θ) d (4)

[0025] In equation (6), Mul(·) represents a multinomial distribution;

[0026] S34. Calculate the text W of the dth comment. d The average topic is denoted as in, W represents the text of the dth comment. d The average result of the k-th topic in the dataset is obtained; thus, the average topic of the comment dataset D is obtained.

[0027]

[0028] S35. Generate the text of the dth comment according to formula (5). d The nth word w d,n Corresponding topic z d,n Indicator variable x d,n :

[0029]

[0030] In equation (5), Beta(·) represents the Beta distribution, and Bern(·) represents the Bernoulli distribution. Indicates the indicator variable x d,n The prior parameters of the corresponding Bernoulli distribution, x d,n W represents the text of the dth comment. d The nth word w d,n It is the word distribution from the k-th seed topic. The word distribution in the middle is generated from the k-th regular topic. Generate in; when x d,n When W = 0, it indicates that the d-th comment text is W. d The nth word w d,n Distribution from the k-th regular keyword Generated in, and w d,n Obtain the parameter as The multinomial distribution, i.e. When x d,n When W = 1, it indicates that the d-th comment text is W. dThe nth word is distributed from the kth seed topic word. Generated in, and w d,n Obtain the parameter as The multinomial distribution, i.e.

[0031] S36, Definition This indicates that all comment topics are included. The overall regression design matrix consists of comment length L, comment rating C1, and overall rating C2; where, Let R represent a matrix containing comment length L, comment rating C1, and overall rating C2, where R represents a real matrix and dim represents the dimension of the matrix.

[0032] S37. Assume there exists an error ε, and that it satisfies equation (6), and construct a regression equation y = Aω + ε, which satisfies equation (7):

[0033] ε~N(0,σ 2 I)(6)

[0034]

[0035] In equations (6) and (7), σ 2 The regression variance is denoted by ω, the regression weights are ω, I represents the identity matrix, N(·) represents the Gaussian distribution, IG(·) represents the inverse gamma distribution, and the mean m0 and covariance matrix S are also given. o These are the hyperparameters of a Gaussian distribution, with shape a0 and scale b. o It is a hyperparameter of the inverse gamma distribution;

[0036] S38. Considering the relationship between topic Z and matrix B, split matrix B into the first matrix B. 1,d With the second matrix B 2,d Among them, B 1,d B represents a matrix indicating relevance to topic Z. 2,d If the matrix represents a matrix that is not relevant to topic Z, then the usefulness score y of the review is... d Satisfying equation (8):

[0037]

[0038] In equation (8), ω z ω represents the weight vector of the average topic. zb Represents the first matrix B 1,d The weight vector, ω b Represents the second matrix B 2,d The weight vector, b represents the Kronecker product. 1,d Represents the first matrix B 1,d In the d-th line, b 2,dRepresents the second matrix B 2,d In the d-th row of the diagram, T represents transpose;

[0039] S4. Based on the comment dataset D, initialize all parameters in steps S2 and S3, and use the EM algorithm and Gibbs sampling method to infer the parameters of document topic distribution, topic word distribution, and comment usefulness prediction distribution.

[0040] S41. Calculate the text W of the dth comment using formula (9). d Topic distribution θ d :

[0041]

[0042] In equation (9), This indicates that the text of comment W is removed in the dth comment section. d After the nth word, the dth comment text W d The number of words in the kth regular topic;

[0043] S42. Calculate the word distribution under the k-th regular topic using equation (10).

[0044]

[0045] In equation (10), This indicates that the text of comment W is removed in the dth comment section. d After the nth word, the vth word w in dictionary V v The number of words assigned to the k-th regular topic in the comment text set W;

[0046] S43. Calculate the usefulness of the comment using formula (11). d Predicted distribution:

[0047]

[0048] In equation (11), the weight variable ω′ z And satisfy This indicates the removal of the dth comment text W. d The nth word w d,n Corresponding topic z d,n The subsequent topic results, b d This represents the d-th row in matrix B, where ∝ indicates proportionality to, n d,n W represents the text of the dth comment. d The nth word w d,n Word count vectors assigned to K topics. W represents the text of the dth comment. d Remove the nth word wd,n The remaining words are assigned to word count vectors for K topics.

[0049] 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 comment usefulness prediction method, and the processor is configured to execute the program stored in the memory.

[0050] 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 comment usefulness prediction method.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. This invention defines two types of topics: seed topics related to user preferences and regular topics. These two types of topics are distinguished by introducing a binary indicator variable. Seed topics are mainly provided by marketers based on their own preferences and are used to filter information in the comment text that is not related to the marketer's attention, thereby more effectively meeting the marketer's differentiated needs for attention preferences.

[0053] 2. This invention organically combines seed topic modeling and causal inference methods to identify potential signals in comment text that may affect the usefulness of the comment. Compared with other text analysis methods, topic modeling technology can perform text analysis without human intervention. Integrating it with causal inference methods can achieve the goal of obtaining coherent and meaningful topic features while exploring the key role of causal relationships in intervening in and predicting the usefulness of comments.

[0054] 3. This invention provides a method for predicting comment usefulness that considers seed information and causal relationships, jointly estimating comment text, comment length, comment rating, and overall rating to more accurately predict comment usefulness. Traditional two-stage methods often extract latent semantic information from the comment text in the first stage and construct a regression model in the second stage to model semantics and other relevant data. The comment usefulness prediction method proposed in this invention better reflects the reality that, since the text itself is used as estimation data, it cannot be estimated separately from the prediction data, thus avoiding potential regression weight bias and ensuring the accuracy of the results. Attached Figure Description

[0055] Figure 1 This is a model diagram of the usefulness prediction method for comments in this invention;

[0056] Figure 2 This is a flowchart illustrating the usefulness prediction method of the present invention.

[0057] Figure 3 This is a structural diagram of the topic sampling method for the usefulness prediction method of this invention.

[0058] Figure 4 This is a structural block diagram of the usefulness prediction method of the present invention. Detailed Implementation

[0059] In this embodiment, as Figure 1 As shown, a method for predicting the usefulness of comments considering seed information and causal relationships is presented. This method can jointly model comment text, comment length, comment rating, and overall rating. Simultaneously, the corpus text in the method can be represented by a topic probability, i.e., the text content is vectorized, and the latent features of the text content are obtained. This embodiment of the invention integrates a topic modeling method, using a seed topic model to vectorize each document and output the topic probability of that document to ensure the usefulness of the output topic. This comment usefulness prediction method includes: 1. Obtaining the required comment data through web crawling, including comment text, comment usefulness score, comment length, comment rating, and overall rating; 2. Performing data cleaning and preprocessing on the crawled text dataset; 3. When inputting the preprocessed text data into the Bayesian seed topic regression model, incorporating seed topic word distribution knowledge generated based on user interest preferences to constrain topic generation, thereby learning meaningful topics; 4. After obtaining the document's topic word distribution, inputting it along with comment length, comment rating, and overall rating into the regression model, iterating using the EM algorithm, and evaluating using mean squared error until convergence.

[0060] like Figure 2 As shown, specifically, a method for predicting the usefulness of comments that considers seed information and causal relationships includes:

[0061] S1. Obtain the comment text and its corresponding non-text data to construct the comment dataset D;

[0062] In this embodiment, the comment text data can come from movie websites, game platforms, e-commerce platforms, etc., without strict limitations. For example, taking e-commerce platform comment data as an example, firstly, the text and non-text data of user comments on products are obtained from the e-commerce platform. After preprocessing the data, a response data set is constructed.

[0063] S11. Obtain the comment dataset D containing a set of comment texts W = {W1, W2, ..., W...} containing K common topics. d ,...,W |D|}, where |D| represents the number of comment texts in the comment text set W; W d Let d be the text of the comment. w d,n W represents the text of the dth comment.d The nth word in N d W represents the text of the dth comment. d The total number of words in the text;

[0064] Construct a dictionary V = {w1,...,w1,...} for the set of comment texts W. v ,...,w |V|}, where w v Let |V| represent the v-th word in dictionary V, and |V| represent the total number of unique words in the comment text set W.

[0065] S12. Obtain the non-text data set U = {Y, L, C1, C2} from the comment dataset D, where Y represents the comment usefulness rating in the comment dataset D, and Y = {y1, y2, ..., y d ,...,y |D|}, y d W represents the text of the dth comment. d The corresponding usefulness score of the comments; L represents the length of the comments in the comment dataset D, and L={l1,l2,...,l d ,...,l |D|}, l d W represents the text of the dth comment. d The corresponding comment length; C1 represents the comment rating in comment dataset D, and C1 = {c 1,1 ,c 1,2 ,...,c 1,d ,...,c 1,|D|}, c 1,d W represents the text of the dth comment. d The corresponding review rating; C2 represents the overall rating in the review dataset D, and C2 = {c 2,1 ,c 2,2 ,...,c 2,d ,...,c 2,|D|}, c 2,d W represents the text of the dth comment. d The corresponding overall rating;

[0066] S2. Obtain user comment preferences to construct seed topic keyword distribution φ s ;

[0067] S21. Based on user interests and preferences, generate the word distribution under the k-th seed topic according to equation (1). This yields the word distribution across all seed topics.

[0068]

[0069] In equation (1), Let be the prior parameters of the Dirichlet distribution corresponding to the k-th seed topic, where ~ indicates that it follows the distribution, Dir(·) represents the Dirichlet distribution, and K represents the number of topics;

[0070] S3, Based on the comment dataset D and seed topic word distribution φ s Construct a Bayesian seed topic regression model;

[0071] S31. Define the word distribution under all regular topics as follows: in Let represent the word distribution under the k-th regular topic, and Satisfying equation (2):

[0072]

[0073] In equation (2), Let be the prior parameters of the Dirichlet distribution corresponding to the k-th regular topic;

[0074] S32. Define the topic distribution of all comment texts as θ={θ1,...,θ d ,...,θ |D|}, where θ d W represents the text of the dth comment. d The topic distribution, and θ d Satisfying equation (3):

[0075] θ d ~Dir(α) d (3)

[0076] In equation (3), α d For the dth comment text W d The prior parameters of the corresponding Dirichlet distribution;

[0077] S33. Generate the d-th comment text W according to formula (4). d The nth word w d,n Corresponding topic z d,n Thus, the d-th comment text W is obtained. d Theme This leads to the topic set Z = [z1,…,z] of the comment dataset D. d ,…,z |D| ];

[0078] z d,n ~Mult(θ) d (4)

[0079] In equation (6), Mul(·) represents a multinomial distribution;

[0080] S34. Calculate the text W of the dth comment. d The average topic is denoted as in, W represents the text of the dth comment. d The average result of the k-th topic in the dataset is obtained; thus, the average topic of the comment dataset D is obtained.

[0081] S35. Generate the text of the dth comment according to formula (5). d The nth word w d,n Corresponding topic z d,n Indicator variable x d,n :

[0082]

[0083] In equation (5), Beta(·) represents the Beta distribution, and Bern(·) represents the Bernoulli distribution. Indicates the indicator variable x d,n The prior parameters of the corresponding Bernoulli distribution, x d,n W represents the text of the dth comment. d The nth word w d,n It is the word distribution from the k-th seed topic. The word distribution in the middle is generated from the k-th regular topic. Generate in; when x d,n When W = 0, it indicates that the d-th comment text is W. d The nth word w d,n Distribution from the k-th regular keyword Generated in, and w d,n Obtain the parameter as The multinomial distribution, i.e. When x d,n When W = 1, it indicates that the d-th comment text is W. d The nth word is distributed from the kth seed topic word. Generated in, and w d,n Obtain the parameter as The multinomial distribution, i.e.

[0084] S36, Definition This indicates that all comment topics are included. The overall regression design matrix consists of comment length L, comment rating C1, and overall rating C2; where, Let R represent a matrix containing comment length L, comment rating C1, and overall rating C2, where R represents a real matrix and dim represents the dimension of the matrix.

[0085] S37. Assume there exists an error ε, and that it satisfies equation (6), and construct a regression equation y = Aω + ε, which satisfies equation (7):

[0086] ε~N(0,σ 2 I) (6)

[0087]

[0088] In equations (6) and (7), σ 2 The regression variance is denoted by ω, the regression weights are ω, I represents the identity matrix, N(·) represents the Gaussian distribution, IG(·) represents the inverse gamma distribution, and the mean m0 and covariance matrix S are also given. o These are the hyperparameters of a Gaussian distribution, with shape a0 and scale b. o It is a hyperparameter of the inverse gamma distribution;

[0089] S38. Considering the relationship between topic Z and matrix B, split matrix B into the first matrix B. 1,d With the second matrix B 2,d Among them, B 1,d B represents a matrix indicating relevance to topic Z. 2,d If the matrix represents a matrix that is not relevant to topic Z, then the usefulness score y of the review is... d Satisfying equation (8):

[0090]

[0091] In equation (8), ω z ω represents the weight vector of the average topic. zb Represents the first matrix B 1,d The weight vector, ω b Represents the second matrix B 2,d The weight vector, b represents the Kronecker product. 1,d Represents the first matrix B 1,d In the d-th line, b 2,d Represents the second matrix B 2,d In the d-th row of the diagram, T represents transpose;

[0092] In this embodiment, the topic model construction process in step S3 of the Bayesian seed topic regression model includes:

[0093] To extract coherent and meaningful data features from text data, this method chooses to feed some additional information to the model. This allows users to provide topics they are interested in or want to learn about, and to provide relevant seed words. For example... Figure 3 As shown, the model proposed in this embodiment is based on the traditional LDA model, and the topic word distribution φ is adjusted. k Expanded to a mixture of two multinomial distributions: seed keyword distribution Distribution of conventional keywords Controlling words using indicator variables and beta distribution is based on π. k The probability from the seed keyword distribution The generation in the middle is still based on 1-π k The probability from the regular keyword distribution Generated in [the context of] [the process]. Among them, the distribution of seed keywords [is shown]. Constrained to generate words only from the corresponding seed set, while the regular topic word distribution... It can generate any word, including the seed word.

[0094] It should also be noted that this invention, based on the topic model, incorporates seed topic word distribution knowledge generated based on user interest preferences; according to the seed topic word distribution and the preprocessed document corpus dataset, a Bayesian seed topic regression model is used to perform supervised representation learning on the comment text, see [link to related documentation]. Figure 4 Furthermore, the correlation information generated during the estimation process is used not only for regression prediction but also for exploring next-generation topics.

[0095] S4. Based on the comment dataset D, initialize all parameters in steps S2 and S3, and use the EM algorithm and Gibbs sampling method to infer the parameters of document topic distribution, topic word distribution, and comment usefulness prediction distribution.

[0096] S41. Calculate the text W of the dth comment using formula (9). d Topic distribution θ d :

[0097]

[0098] In equation (9), This indicates that the text of comment W is removed in the dth comment section. d After the nth word, the dth comment text W d The number of words in the kth regular topic;

[0099] S42. Calculate the word distribution under the k-th regular topic using equation (10).

[0100]

[0101] In equation (10), This indicates that the text of comment W is removed in the dth comment section. d After the nth word, the vth word w in dictionary V v The number of words assigned to the k-th regular topic in the comment text set W;

[0102] S43. Calculate the usefulness of the comment using formula (11).d Predicted distribution:

[0103]

[0104] In equation (11), the weight variable ω z 'satisfy This indicates the removal of the dth comment text W. d The nth word w d,n Corresponding topic z d,n The subsequent topic results, b d This represents the d-th row in matrix B, where ∝ indicates proportionality to, n d,n W represents the text of the dth comment. d The nth word w d,n Word count vectors assigned to K topics. W represents the text of the dth comment. d Remove the nth word w d,n The remaining words are assigned to word count vectors for K topics.

[0105] 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 comment usefulness prediction method. The processor is configured to execute the program stored in the memory.

[0106] In this embodiment, a computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the above-described comment usefulness prediction method.

[0107] Example inference:

[0108] The comment usefulness prediction method in this embodiment identifies potential topic distributions and regression parameters most likely to generate observation data, and obtains the joint distribution of the model by multiplying all nodes in the model, as shown in Equation (12).

[0109]

[0110] In equation (12), if p(x|z)=1, then p(w|z,x,φ)=p(w|z,x,φ) s If p(x|z) = 0, then p(w|z,x,φ) = p(w|z,x,φ). r ).

[0111] The inference task in this embodiment is to compute the posterior distribution of the latent variable (z,x), i.e., p(Z,X|W,B,y,ω,σ), given the observed data (y,B,W) and other priors. 2 ).

[0112] p(Z,X|W,B,y,ω,σ 2 )∝p(Z,X|W)·p(y|Z,B,ω,σ 2 (13)

[0113] The topic assignment in one document is independent of the topic assignments in all other documents, therefore:

[0114]

[0115] In equation (12), This indicates the removal of the dth comment text W. d The nth word w d,n The corresponding indicator variable x d,n The result of the indicator variable after, and the first part of the expression It is a sampling distribution of a seed topic model, Part Two This refers to the usefulness of the comment. d The predicted distribution.

[0116] 1. Regarding the formula in the first part In other words:

[0117]

[0118] Equation (15) can be further expanded from p(Z,X,W)=p(W|Z,X)p(X|Z)p(Z), therefore, we need to obtain We only need to calculate p(W|Z,X), p(X|Z), and p(Z) separately.

[0119]

[0120]

[0121]

[0122] In equation (16), in w represents the v-th word in dictionary V. v The number of words assigned to the k-th regular topic in the comment text set W; W represents the text of the dth comment. d The nth word w d,n The corresponding seed topic is z d,n The probability of.

[0123] In equation (18), in W represents the text of the dth comment. d The number of words in the k-th topic.

[0124] Substitute formulas (16)-(18) into the equations... Meanwhile, since it is the dth comment text W d The nth word w d,n Sampling is performed, therefore the multiplication sign is canceled out, i.e.:

[0125]

[0126]

[0127] 2. For the formula derivation in Part Two:

[0128] Given y ~ N(Aω,σ) 2 I), Since B is observable, it can be viewed as a linear model with document-specific regression parameters. Definition Therefore, the usefulness rating of the review is y. d Satisfying equation (20):

[0129]

[0130] At the same time, due to therefore:

[0131]

[0132] In summary, the method of this invention organically combines seed topic modeling and causal inference methods. (See also...) Figure 4 This approach achieves control over user interests and preferences while avoiding potential regression weight biases. It ensures the accuracy of the results and meets marketers' differentiated needs for attention preferences. This solves the problems of existing text-based causal inference methods, which cannot jointly estimate comment text, comment length, comment rating, and overall rating data, and cannot obtain coherent and meaningful thematic features from text data. At the same time, the comment usefulness prediction results based on causal inference can effectively and directly guide marketers' decision-making, thus solving the problem that traditional correlation-based inference ignores the key role of causality in intervening in and predicting comment usefulness.

Claims

1. A method for predicting the usefulness of reviews that considers seed information and causal relationships, characterized in that, include: S1. Obtain the comment text and its corresponding non-text data to construct the comment dataset D; S11. Obtain the comment dataset D containing a set of comment texts W = {W1, W2, ..., W...} containing K common topics. d ,…,W |D| }, where |D| represents the number of comment texts in the comment text set W; W d Let d be the text of the comment. w d,n W represents the text of the dth comment. d The nth word in N d W represents the text of the dth comment. d The total number of words in the text; Construct a dictionary V = {w1,...,w1,...} for the set of comment texts W. v ,...,w |V| }, where w v Let |V| represent the v-th word in dictionary V, and |V| represent the total number of unique words in the comment text set W. S12. Obtain the non-text data set U = {Y, L, C1, C2} from the comment dataset D, where Y represents the comment usefulness rating in the comment dataset D, and Y = {y1, y2, ..., y d ,...,y |D| }, y d W represents the text of the dth comment. d The corresponding usefulness score of the comments; L represents the length of the comments in the comment dataset D, and L={l1,l2,...,l d ,...,l |D| }, l d W represents the text of the dth comment. d The corresponding comment length; C1 represents the comment rating in comment dataset D, and C1 = {c 1,1 ,c 1,2 ,...,c 1,d ,...,c 1,|D| }, c 1,d W represents the text of the dth comment. d The corresponding review rating; C2 represents the overall rating in the review dataset D, and C2 = {c 2,1 ,c 2,2 ,...,c 2,d ,...,c 2,|D| }, c 2,d W represents the text of the dth comment. d The corresponding overall rating; S2. Obtain user comment preferences to construct seed topic keyword distribution φ s ; S21. Based on user interests and preferences, generate the word distribution under the k-th seed topic according to equation (1). This yields the word distribution across all seed topics. In equation (1), Let be the prior parameters of the Dirichlet distribution corresponding to the k-th seed topic, where ~ indicates that it follows the distribution, Dir(·) represents the Dirichlet distribution, and K represents the number of topics; S3, Based on the comment dataset D and seed topic word distribution φ s Construct a Bayesian seed topic regression model; S31. Define the word distribution under all regular topics as follows: in, Let represent the word distribution under the k-th regular topic, and Satisfying equation (2): In equation (2), Let be the prior parameters of the Dirichlet distribution corresponding to the k-th regular topic; S32. Define the topic distribution of all comment texts as θ={θ1,...,θ d ,…,θ |D| }, where θ d W represents the text of the dth comment. d The topic distribution, and θ d Satisfying equation (3): i d ~Dir(a d ) (3) In equation (3), α d For the dth comment text W d The prior parameters of the corresponding Dirichlet distribution; S33. Generate the d-th comment text W according to formula (4). d The nth word w d,n Corresponding topic z d,n Thus, the d-th comment text W is obtained. d Theme This leads to the topic set Z = [z1,…,z] of the comment dataset D. d ,…,z |D| ]; z d,n ~Mult(θ d ) (4) In equation (6), Mul(·) represents a multinomial distribution; S34. Calculate the text W of the dth comment. d The average topic is denoted as in, W represents the text of the dth comment. d The average result of the k-th topic in the dataset is obtained; thus, the average topic of the comment dataset D is obtained. S35. Generate the text of the dth comment according to formula (5). d The nth word w d,n Corresponding topic z d,n Indicator variable x d,n : In equation (5), Beta(·) represents the Beta distribution, and Bern(·) represents the Bernoulli distribution. Indicates the indicator variable x d,n The prior parameters of the corresponding Bernoulli distribution, x d,n W represents the text of the dth comment. d The nth word w d,n It is the word distribution from the k-th seed topic. The word distribution in the middle is generated from the k-th regular topic. Generate in; when x d,n When W = 0, it indicates that the d-th comment text is W. d The nth word w d,n Distribution from the k-th regular keyword Generated in, and w d,n Obtain the parameter as The multinomial distribution, i.e. When x d,n When W = 1, it indicates that the d-th comment text is W. d The nth word is distributed from the kth seed topic word. Generated in, and w d,n Obtain the parameter as The multinomial distribution, i.e. S36, Definition This indicates that all comment topics are included. The overall regression design matrix consists of comment length L, comment rating C1, and overall rating C2; where, Let R represent a matrix containing comment length L, comment rating C1, and overall rating C2, where R represents a real matrix and dim represents the dimension of the matrix. S37. Assume there exists an error ε, and that it satisfies equation (6), and construct a regression equation y = Aω + ε, which satisfies equation (7): ε~N(0,σ 2 I)(6) In equations (6) and (7), σ 2 The regression variance is denoted by ω, the regression weights are ω, I represents the identity matrix, N(·) represents the Gaussian distribution, IG(·) represents the inverse gamma distribution, and the mean m0 and covariance matrix S are also given. o These are the hyperparameters of a Gaussian distribution, with shape a0 and scale b. o It is a hyperparameter of the inverse gamma distribution; S38. Considering the relationship between topic Z and matrix B, split matrix B into the first matrix B. 1,d With the second matrix B 2,d Among them, B 1,d B represents a matrix indicating relevance to topic Z. 2,d If the matrix represents a matrix that is not relevant to topic Z, then the usefulness score y of the review is... d Satisfying equation (8): In equation (8), ω z ω represents the weight vector of the average topic. zb Represents the first matrix B 1,d The weight vector, ω b Represents the second matrix B 2,d The weight vector, b represents the Kronecker product. 1,d Represents the first matrix B 1,d In the d-th line, b 2,d Represents the second matrix B 2,d In the d-th row of the diagram, T represents transpose; S4. Based on the comment dataset D, initialize all parameters in steps S2 and S3, and use the EM algorithm and Gibbs sampling method to infer the parameters of document topic distribution, topic word distribution, and comment usefulness prediction distribution. S41. Calculate the text W of the dth comment using formula (9). d Topic distribution θ d : In equation (9), This indicates that the text of comment W is removed in the dth comment section. d After the nth word, the dth comment text W d The number of words in the kth regular topic; S42. Calculate the word distribution under the k-th regular topic using formula (10). In equation (10), This indicates that the text of comment W is removed in the dth comment section. d After the nth word, the vth word w in dictionary V v The number of words assigned to the k-th regular topic in the comment text set W; S43. Calculate the usefulness of the comment using formula (11). d Predicted distribution: In equation (11), the weight variable ω z ′, and satisfy This indicates the removal of the dth comment text W. d The nth word w d,n Corresponding topic z d,n The subsequent topic results, b d This represents the d-th row in matrix B, where ∝ indicates proportionality to, n d,n W represents the text of the dth comment. d The nth word w d,n Word count vectors assigned to K topics. W represents the text of the dth comment. d Remove the nth word w d,n The remaining words are assigned to word count vectors for K topics.

2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing the comment usefulness prediction method of claim 1, the processor being configured to execute the programs stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the comment usefulness prediction method of claim 1.

Images