An explainable recommendation method based on sentiment orientation

Abstract

An explainable recommendation method based on sentiment analysis, comprising: 1) constructing an actual user-item rating matrix; 2) obtaining the sentiment score of the user to the item from the comment by using a sentiment analysis method; 3) constructing a user-feature attention matrix and an item-feature quality matrix according to the feature word set extracted in step 2); 4) constructing a matrix decomposition model; 5) training the model in step 4); and 6) personalized recommendation. The present application improves the efficiency, further solves the problems of data sparsity and the possible unexplainability of the recommendation results, and realizes explainable recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of recommendation systems, more particularly, to an explainable recommendation method based on sentiment orientation. BACKGROUND

[0002] With the advent of the era of information explosion, the network provides more and more information to the user, and the user may often cause information overload, which greatly affects the user experience. In this situation, the recommendation system emerges as the times require, and it is an important means to solve this situation, and has good effect in various application occasions. The recommendation system can provide suggestions and recommendations for users, help users choose goods and services, and provide support for user decision-making. However, the data missing problem is one of the common severe challenges faced by the recommendation system. In the training data of the recommendation system, the user only interacts with a very small part of the items, resulting in that the preference scores of most user-item pairs are missing, which is also called the sparsity problem. In recent years, the research on the recommendation system mainly focuses on the model optimization design of user behavior data, and further improves the recommendation quality, but it also ignores one point-the explanation of the recommendation result, so that the user sometimes cannot believe the reason why they are recommended items or goods, or the user does not understand why a series of goods list is recommended to them.

[0003] At present, many recommendation algorithms are becoming mature, which promotes the development of the field of recommendation systems. The collaborative filtering algorithm is a well-known algorithm in the history of the development of the whole recommendation system, and has a pivotal position, and is still widely used today. Collaborative filtering is divided into two algorithms: user-based collaborative filtering and item-based collaborative filtering. The core idea is actually very simple "birds of a feather flock together, and people of the same group are divided". The so-called "birds of a feather flock together" is to calculate the list of the most similar items of each item, and we can recommend the similar items of the items that the user likes to the user, which is item-based collaborative filtering. The so-called "people of the same group are divided" is that we can recommend the items that the similar users like to the user (which the user has not operated), which is user-based collaborative filtering. The recommendation result of this method can solve the explainable problem, which not only provides the recommendation result for the user or system designer, but also explains why these items or goods are recommended, which increases the explainability of the recommendation system, helps to improve the transparency, persuasiveness, effectiveness, credibility and user satisfaction of the recommendation system, and is convenient for system designers to diagnose, debug and optimize the recommendation algorithm. However, it is not good at generalizing the sparse matrix problem, that is, the data missing problem mentioned above cannot be solved in this method.

[0004] The deep learning recommendation model AutoRec proposed by Suvash Sedhain et al. is based on an autoencoder, which encodes users or items and uses the generalization ability of the autoencoder to make recommendations. It uses a single hidden layer neural network structure, which is simple in structure and can realize fast training and deployment, but its expression ability is poor; the deep learning recommendation model Deep Crossing proposed by Ying Shan uses the classic deep learning framework of "embedding layer plus multiple hidden layers plus output layer" to automatically complete feature deep crossing, but this model uses fully connected hidden layers for feature crossing, which is not very targeted; and other deep learning recommendation system models can effectively solve the data missing problem without using sentiment orientation, but may not be able to completely explain the recommended results to the user. SUMMARY

[0005] In order to overcome the data sparsity problem of the recommendation system and the problem that the given recommendation results may not be explained, the present application provides an interpretable recommendation method based on sentiment analysis, which combines user reviews and scores of items, uses the user sentiment orientation revealed in the reviews to solve these problems, effectively improves the interpretability of the recommendation system, and enhances the credibility of the recommendation system.

[0006] The interpretable recommendation method based on sentiment orientation proposed by the present application can effectively solve the two problems mentioned above, namely, solving the data missing problem by combining user reviews and score estimates of missing scores, and solving the transparency problem by using user reviews of items to obtain their sentiment orientation scores to establish a model for prediction, which is based and interpretable. In addition, in the final target function optimization algorithm, an optional minimization algorithm is used to obtain the optimal solution to better perform the final recommendation.

[0007] The technical scheme adopted by the present application to solve the technical problems is:

[0008] An interpretable recommendation method based on sentiment analysis, the method comprising the following steps:

[0009] 1) Constructing an actual user-item rating matrix ;

[0010] 1.1) Selecting a rating dataset;

[0011] 1.2) Let be a user set, be an item set, and let represent an actual user-item rating matrix, where represents the rating of a certain user u to a certain item i. The rating is generally 1-5 points, and 0 indicates that the user has not rated the item.

[0012] 2) Obtain the sentiment score of the user on the item from the review using sentiment analysis method;

[0013] 2.1) Extract the feature word set of a user on an item from the review of the user on the item using natural language tools , opinion word set O;

[0014] 2.2) Match the sentiment score S according to the opinion word set O;

[0015] 2.3) Generate the feature sentiment pair (F, S);

[0016] 3) Construct the user-feature attention matrix and item-feature quality matrix respectively according to the feature word set extracted in step 2);

[0017] 3.1) Assuming that the feature word is mentioned times by the user , then we can define the user-feature attention matrix X as:

[0018] (1)

[0019] The definition is actually the role of the activation function, and the purpose is to scale to the same range as the actual user-item rating matrix A rating score;

[0020] 3.2) Assuming that the feature word is mentioned k times in the review about the item , and the average sentiment score of the feature word in the k mentions is ; then we can define the item-feature quality matrix Y as:

[0021] (2)

[0022] 4) Construct a matrix decomposition model;

[0023] 4.1) First, construct a decomposition model on the user feature attention matrix X and the item feature quality matrix Y, that is, estimate the hidden representations U1 and U2 of the user, feature and item based on the observed user-feature and item-feature relationships, and the objective function is as follows:

[0024] (3)

[0025] ​4.2) For the observable features in step 4.1), the score can not be fully explained, the user can consider some other potential or hidden factors when making a decision (i.e. scoring) on the product, so introduce a matrix containing some implicit factors The matrix of features, i.e. , , and use , to build a decomposition model for the actual user-item rating matrix A, the objective function is as follows:

[0026] (4)

[0027] 4.3) Combined with equation (3) and equation (4), the hidden factors can be estimated by the following optimization task:

[0028] (5)

[0030] 5) Train the model in step 4);

[0031] In the case of no closed form solution for equation (5), an optional minimization algorithm is used to iteratively optimize to find the optimal solution of U1, HI, U2, H2, V2, i.e. optimize the objective function for one parameter while fixing the other four parameters, the algorithm constantly updates the parameters until convergence or reaches the maximum number of iterations.

[0032] 6) Personalized recommendation;

[0033] 6.1) After obtaining the optimal solution of the decomposition model in step 5), any missing values of the user-feature attention matrix X, the item-feature quality matrix Y, and the actual user-item rating matrix A can be estimated, so the predicted user-feature attention matrix , the predicted item-feature quality matrix , and the predicted user-item rating matrix ;

[0034] 6.2) For a user , only the first few important item features are considered in the present invention, not all features, i.e. let the matrix The first k values of the row vector with the maximum value are . In the user interface, the ranking score of the item is:

[0035] (6)

[0036] where N = 5, The item is the user-item similarity score based on the k item features, .

[0037] According to The order of the goods is arranged, and finally a recommended list is generated for the user.

[0038] The technical concept of the present application is that, in view of the data sparsity problem existing in the recommendation system and the problem that part of the recommended results may not be explained, a sentiment-based interpretable recommendation method is proposed by combining the emotional tendency revealed in the user comments with the matrix decomposition technology.

[0039] The beneficial effects of the present application mainly lie in that, in the case where the closed solution does not often appear in the common matrix decomposition algorithm minimization objective function, the selectable minimization algorithm is introduced to obtain the optimal solution of each parameter, the efficiency is improved, and the problem of data sparsity and the problem that part of the recommended results may not be explained are further solved, and the interpretable recommendation is realized. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0041] Figure 1 The flowchart of the present application is shown.

[0042] Figure 2 The flowchart of the specific embodiment of the present application is shown. DETAILED DESCRIPTION

[0043] The present application will be further described below in combination with the drawings.

[0044] In the application scenario of the personalized commodity recommendation system, it is usually necessary to crawl the ratings and comments of users on a certain commodity and build a matrix model for recommendation on the basis of the ratings and comments. For example, on the Dianping website, a buyer user rates a store he / she has ever eaten and posts his / her comment content, and an actual user-item rating matrix model is built on the basis of this. However, the buyer user cannot have experienced all the stores, and therefore cannot have rated all the stores. Therefore, it is very important to derive the emotional tendency of the user to the commodity on the basis of the existing ratings and comments of the user to the commodity, so that the recommended result is not only accurate but also interpretable, thereby improving the transparency, persuasiveness, effectiveness, credibility and user satisfaction of the recommendation system, and facilitating the diagnosis, debugging and optimization of the recommendation algorithm by the system designer.

[0045] The personalized commodity recommendation method using the sentiment-based interpretable recommendation method. The method comprises the following steps:

[0046] 1) Constructing the real user-item rating matrix

[0047] 1.1) Selecting the rating dataset;

[0048] With the Yelp dataset, which consists of user reviews, there are 45981 users, 11537 items, and 229907 user-item reviews including ratings and reviews.

[0049] 1.2) Let be the set of users, be the set of items, and denote the real user-item rating matrix, where denotes the rating of item i by user u. Ratings are generally on a scale of 1-5, and 0 indicates that the user has not rated the item.

[0050] 2) Using sentiment analysis methods to obtain the sentiment score of users on items from reviews

[0051] 2.1) Using natural language tools to extract the feature word set of a user on an item from the user's review of the item and the opinion word set O.

[0052] 2.2) According to the opinion word set O, match the sentiment score S, which ranges from (strong to weak sentiment) and (weak to strong sentiment).

[0053] 2.3) Generate feature sentiment pairs (F, S)

[0054] 3) Construct the user-feature attention matrix and the item-feature quality matrix based on the feature word set extracted in step 2).

[0055] 3.1) Assuming that the feature word is mentioned times by user , then we can define the user-feature attention matrix X as:

[0056] (1)

[0057] This definition is actually the role of the activation function, and its purpose is to scale to the same range as the actual user-item rating matrix A rating score .

[0058] k-means clustering maps users to n fuzzy clusters in a Euclidean distance space to represent the degree of interest in different items. In this application scenario, n is 5. For the Yelp dataset, the five most frequent features in the five classes are, in order: food, lunch, service, meal, experience (class 1); service, menu, food, order, staff (class 2); place, area, location, restaurant, bar (class 3); price, order, food, service, menu (class 4); beer, drink, bar, order, wine (class 5) (since this dataset uses English, the features are represented in English).

[0059] 3.2) Hypothetical Feature Words In the relevant items The keyword was mentioned k times in the comments. The average sentiment score of the feature word across these k mentions is expressed as: Therefore, we can define the item-feature quality matrix Y as:

[0060] (2)

[0061] 4) Construct a matrix factorization model

[0062] 4.1) First, a decomposition model is constructed on the user feature attention matrix X and the item feature quality matrix Y. That is, the hidden representations U1 and U2 of users, features and items are estimated based on the observed relationships between user features and item features. The objective function is as follows:

[0063] (3)

[0064] Where r refers to the features mentioned in X and Y, namely, the observable features. and It is the regularization coefficient.

[0065] 4.2) The observable features in step 4.1) may not fully explain the ratings. Users may consider other potential or hidden factors when making decisions about products (i.e., ratings). Therefore, we introduce some implicit factors. The feature matrix is , and use , We construct a decomposition model for the actual user-item rating matrix A, with the following objective function:

[0066] (4)

[0067] 4.3) Combining equation (3) and equation (4), the hidden factors can be estimated by the following optimization task:

[0068] (5)

[0070] 5) Training the model in step 4)

[0071] In the case of no closed-form solution for equation (5), an optional minimization algorithm is used to iteratively optimize to find the optimal solution of U1, HI, U2, H2, V2, i.e., optimize the objective function with respect to one parameter while fixing the other four parameters, the algorithm keeps updating the parameters until convergence or reaches the maximum number of iterations, which is set to 100 to ensure convergence.

[0072] 6) Personalized recommendation

[0073] 6.1) After obtaining the optimal solution of the factorized model in step 5), any missing values of the user-feature attention matrix X, the item-feature quality matrix Y, and the actual user-item rating matrix A can be estimated, so the predicted user-feature attention matrix , the predicted item-feature quality matrix , and the predicted user-item rating matrix .

[0074] 6.2) For a user , only the top several important item features are considered in the present invention, not all features, i.e., let the matrix be the row vector with the top k values . In the user interface, the ranking score of an item is set to be:

[0075] (6)

[0076] where N = 5, is the user-item similarity score based on the k item features, .

[0077] According to the ranking order of , the final recommendation list of items is generated for the user.

[0078] The above-described embodiments are only to describe the preferred modes of the present invention, and not to limit the scope of the present invention, and various modifications and improvements to the technical solutions of the present invention made by those skilled in the art without departing from the design spirit of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An interpretable recommendation method based on sentiment analysis, comprising the following steps: 1) Construct a real user-item rating matrix ; 2) Use sentiment analysis to obtain users' sentiment scores for items from reviews; 3) Construct user-feature attention matrices based on the feature word sets extracted in step 2). And item-feature quality matrix ; 4) Construct a matrix factorization model; 5) Train the model from step 4); 6) Personalized recommendations; Step 1) specifically includes: 1.1) Select the rating dataset; 1.2) Let For user collection, To collect items, use This represents the actual user-item rating matrix, where This represents a user u's rating of an item i; the rating ranges from 1 to 5, with 0 indicating that the user has not rated the item. Step 2) specifically includes: 2.1) Extract a user's feature word set for a specific item from user reviews of that item using natural language processing tools. Opinion collection O; 2.2) Match the sentiment score S to the opinion word set O; 2.3) Generate feature sentiment pairs (F, S); Step 3) specifically includes: 3.1) Hypothetical Feature Words By user Mentioned Next, define the user-feature attention matrix X as follows: （1） This definition essentially describes the role of the activation function, its purpose being to... Scale to the same range as the actual user-item rating matrix A-scores. inside; 3.2) Hypothetical Feature Words In the relevant items The keyword was mentioned k times in the comments. The average sentiment score of the feature word across these k mentions is expressed as: The item-feature quality matrix Y can then be defined as: （2） Step 4) specifically includes: 4.1) First, a decomposition model is constructed on the user feature attention matrix X and the item feature quality matrix Y. That is, the hidden representations U1 and U2 of users, features and items are estimated based on the observed relationships between user features and item features. The objective function is as follows: （3） Where r refers to the observable features in X and Y. and These are the regularization term coefficients; 4.2) Introducing implicit factors into the observable features in step 4.1). The feature matrix is , and use , We construct a decomposition model for the actual user-item rating matrix A, with the following objective function: （4） 4.3) Combining equations (3) and (4), the hidden factors can be estimated through the following optimization task: （5） Step 5) specifically includes: when there is no closed-form solution in equation (5), an optional minimization algorithm is used to perform iterative optimization to find the optimal solution of U1, H1, U2, H2, V2, that is, to optimize the objective function for one parameter while fixing the other four parameters, the algorithm continuously updates the parameters until convergence or the maximum number of iterations is reached; Step 6) specifically includes: 6.1) After obtaining the optimal solution of the decomposition model in step 5), any missing values ​​of the user-feature attention matrix X, the item-feature quality matrix Y, and the actual user-item rating matrix A can be estimated. Therefore, the user-feature attention matrix can be predicted. Predicting the item-feature quality matrix Predicting user-item rating matrix ; 6.2) For users Let the matrix The largest of the k values ​​in a row vector is Set items in the user interface The ranking score is: （6） Where N=5, The item is a user-item similarity score based on these k item features. ; according to The items are sorted in order, and a recommendation list is generated for the user.

Images