CN111401936A - A Recommendation Method Based on Comment Space and User Preferences - Google Patents

Abstract

The invention belongs to the technical field of commodity recommendation, and discloses a recommendation method based on comment space and user preference, comprising: step 1, constructing a commodity comment data corpus and a user comment data corpus; Construct a product review space and a user preference space respectively; step 3, according to the user's position in the user preference space and the constructed product review space, obtain a product recommendation for the user. The invention uses the commodity review data to quantify the commodity attributes and maps the commodities to the attribute space, and combines the method of the user's historical shopping preference for commodity recommendation. For the existing recommendation system, it avoids the low recommendation efficiency caused by the sparse matrix. With the continuous increase of the amount of comment information, the present invention can more accurately recommend products that meet the user's shopping preference to the user by mining the relationship between the user's shopping preference and commodity attributes.

Description

A Recommendation Method Based on Comment Space and User Preferences

technical field

The invention belongs to the technical field of commodity recommendation, and in particular relates to a recommendation method based on comment space and user preference.

Background technique

While the Internet brings convenience to people, it also brings people complicated and heavy screening tasks, so that people have to spend a lot of time to select the information they really need. In terms of online shopping, it also brings people a sea of information. When users are faced with a variety of commodity information, they are always unable to start, and they often need to spend several times more time than before to choose the one they like and suit themselves. commodity. Similarly, information overload will also bring great challenges to the merchants on the Internet. The information released by the merchants through the Internet will soon be overwhelmed by other information. The merchants need to constantly update their product information to achieve the purpose of sales. . In order to solve this problem, various online shopping platforms have launched various recommendation algorithms (Li Lin, Liu Jinxing, Meng Xiangfu, etc. Product recommendation model integrating rating matrix and review text [J]. Journal of Computers, 2018, 41(7): 1559-1573.), the purpose is to recommend suitable products for users, reduce users' shopping time, and improve users' decision-making efficiency.

With the development and progress of the times, the challenges of the e-commerce industry will also increase greatly, and the quality of the recommendation system of e-commerce websites will also become a way for merchants to compete (Li Yuqi, Chen Weizheng, Yan Hongfei, etc. Based on the network to represent the personality of learning Chemical Commodity Recommendation [J]. Chinese Journal of Computer, 2019(8):7.). As far as the current recommendation system is concerned, the accuracy and real-time performance of the internal recommendation algorithm are not high enough. Therefore, it is still necessary to propose a feasible recommendation algorithm to improve and improve the existing recommendation system.

SUMMARY OF THE INVENTION

Aiming at the problem that the accuracy and real-time performance of the internal recommendation algorithm of the existing recommendation system are not high enough, the present invention proposes a recommendation method based on comment space and user preference.

In order to achieve the above object, the present invention adopts the following technical solutions:

A recommendation method based on comment space and user preferences, including:

Step 1: Build a corpus of product review data and a corpus of user review data;

Step 2: Construct a product review space and a user preference space based on the product review data corpus and the user review data corpus respectively;

Step 3: According to the user's position in the user preference space and the constructed product review space, the product recommendation for the user is obtained.

Further, the building a corpus of product review data includes:

Collect the comment data of all users on the same product;

The Jieba word segmentation module is used to perform word segmentation on the comment data of all users on the same product, and remove stop words;

Using a dictionary-based classification algorithm, according to the four attributes of quality, price, appearance and express delivery, the review data of all users for the same product is classified to complete the construction of the product review data corpus.

Further, the construction of the user comment data corpus includes:

Collect the comment data of all products by the same user;

The Jieba word segmentation module is used to perform word segmentation on the comment data of all products by the same user, and remove stop words;

Using a dictionary-based classification algorithm, according to the four attributes of quality, price, appearance and express delivery, the same user's review data for all products is classified to complete the construction of the user review data corpus.

Further, building a product review space based on the product review data corpus includes:

Construct four-dimensional space coordinates according to the four attributes of quality, price, appearance and express delivery;

Use the KNN classification algorithm to classify the review data in the product review data corpus into two categories: positive and negative;

The positive and negative review data describing the same attribute of the same product are formed into a sequence according to the probability of positive sentiment in descending order, and the first comment data marked as positive in the sequence corresponding to the positive review data is set as this. The positive biased reference review on the attribute feature is the unit 1 on this attribute feature; the last sentence marked as negative in the corresponding sequence of negative review data is designated as the negative reference review on the attribute feature, which is the attribute feature. unit -1;

Use the cosine similarity algorithm to calculate the cosine value between the review data to be processed and the reference review on this attribute feature: if the classification label of the review data is positive, calculate the difference between the review data and the positive biased reference review on this attribute feature The obtained cosine value is the coordinate value of the review data in the attribute dimension; if the classification label of the review sentence is negative, calculate the cosine value between the review data and the negative bias reference review on this attribute feature. , the inverse of the obtained cosine value is the coordinate value of the comment data in this attribute dimension;

In the same review data, the calculated coordinate values of each attribute dimension are summarized into a coordinate point in a four-dimensional space. For the attribute features not mentioned in the review data, the value on the dimension corresponding to the unmentioned attribute feature is calculated. is 0, the coordinate value of the comment data in the four-dimensional space coordinate is obtained;

Average the values of all the comment data of the same product in the dimension of the same product, and obtain the coordinate value of each attribute of each product in the four-dimensional space coordinates;

The coordinate value of each attribute of each commodity in the spatial coordinates, that is, each commodity point, is presented in the four-dimensional spatial coordinates to complete the construction of the commodity comment space.

Further, after the completion of the product review space construction, the method further includes:

The K-means clustering algorithm is used to cluster the generated product points, and the products with similar attribute characteristics are placed in a cluster, and the cluster center is saved.

Further, constructing the user preference space based on the user comment data corpus includes:

Construct four-dimensional space coordinates according to the four attributes of quality, price, appearance and express delivery;

Quantify the user's emotional tendency according to the user's degree of attention to each attribute feature, where the degree of attention is the ratio of the number of comment data for each attribute of the same user to the total number of comment data, and obtain the spatial coordinates of each user's emotional tendency;

Display all the spatial coordinate points representing the user in the four-dimensional spatial coordinate to complete the construction of the user preference space.

Further, the step 3 includes:

The user's position in the user preference space is taken as the point of the user's shopping preference, and the Euclidean distance is used to calculate the distance between the point of the user's shopping preference and the center point of each cluster in the product review space. products are recommended to users.

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

The invention uses the commodity review data to quantify the commodity attributes and map the commodities to the attribute space, and carries out commodity recommendation in combination with the method of the user's historical shopping preference. For the existing recommendation system, the problem of low recommendation efficiency caused by sparse matrix is avoided, and with the continuous increase of the amount of review information, the judgment of the attribute characteristics of the product is more accurate. The user's comments are analyzed independently by each user to obtain their shopping preferences, and the accuracy of the recommendation for the user will not be affected by the analysis results of other users to achieve personalized recommendation. By mining the relationship between the user's shopping preference and commodity attributes, the present invention can more accurately recommend commodities that meet the user's shopping preference to the user, and the experimental results show that only using historical comment data can achieve good results.

Description of drawings

1 is a basic flowchart of a recommendation method based on comment space and user preference according to an embodiment of the present invention;

2 is a schematic diagram of cosine similarity calculation of a recommendation method based on comment space and user preference according to an embodiment of the present invention;

3 is a schematic diagram of a product review space mapping of a recommendation method based on review space and user preference according to an embodiment of the present invention;

Fig. 4 is the elbow method SSE change curve diagram of a recommendation method based on comment space and user preference according to an embodiment of the present invention;

5 is a schematic diagram of user preference space mapping of a recommendation method based on comment space and user preference according to an embodiment of the present invention;

FIG. 6 is an example diagram of a product recommendation of a recommendation method based on comment space and user preference according to an embodiment of the present invention;

FIG. 7 is an example diagram of commodities and comments of a recommendation method based on comment space and user preference according to an embodiment of the present invention;

8 is an example diagram of users and comments of a recommendation method based on comment space and user preferences according to an embodiment of the present invention;

FIG. 9 is a comparison diagram of the accuracy rates of different algorithms of a recommendation method based on comment space and user preference according to an embodiment of the present invention.

Detailed ways

The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments:

As shown in Figure 1, a recommendation method based on comment space and user preferences includes:

Step S101: constructing a product review data corpus and a user review data corpus;

Step S102: respectively constructing a product review space and a user preference space based on the product review data corpus and the user review data corpus;

Step S103: According to the user's position in the user preference space and the constructed product review space, obtain a product recommendation for the user.

Specifically, the building a corpus of product review data includes:

Collect the comment data of all users on the same product; as an implementable manner, the comment data used in this embodiment is mainly divided into the following two parts:

①Public data: This part of the data refers to the comment data published on the Internet. Use Amazon product review data, which can identify both products and reviews, and users and reviews;

②Semi-public data: This part of the data refers to the data provided by the official organizing committee of the competition. Another part of the experimental data used is provided by the official organizing committee of the Big Data and Computer Intelligence Competition.

The Jieba word segmentation module is used to perform word segmentation on the comment data of all users on the same product, and remove stop words; the three modes of the Jieba word segmentation module include: full mode, precise mode and search engine mode. As an implementable manner, this embodiment uses the full mode of Jieba word segmentation to segment the comment data. In the process of Chinese text processing, in order to save storage space and improve operating efficiency, some words need to be removed after word segmentation, and these words are called stop words.

Using a dictionary-based classification algorithm, according to the four attributes of quality, price, appearance and express delivery, the review data of all users for the same product is classified to complete the construction of the product review data corpus. A comment dictionary means that when users comment on a certain attribute feature of a product, the comments often contain a few words that describe this attribute feature, mainly adjectives, nouns, adverbs, etc. The classification method adopted in this embodiment is a classification algorithm based on a (comment) dictionary. The key of the dictionary-based review data classification algorithm lies in the construction of the dictionary. The four dictionaries used in this embodiment are based on the existing dictionaries and are improved by manual annotation. Before classifying, by analyzing the review data, it is determined that people generally care about the four attribute features of the product: quality, price, appearance and express delivery. Therefore, in this embodiment, these four attribute features are used as the four aspects to measure the product features. . Part of the contents of the four dictionaries are shown in Table 1.

Table 1 Parts of the four dictionaries

quality price Exterior express delivery solid price cut nice deliver goods adulterated Cheap pretty Ship high quality affordable face value speed poor quality discount atmosphere delivery Copycat cost-effective noble logistics defective selling price handsome quick Foot goods high nice good service

By traversing the processed comment data for matching, the category described by the processed text is further identified. In the classification results, there will be a phenomenon that one line of data corresponds to multiple category labels, such as: "The book is very good, the color is very good, the quality is excellent, and the delivery is super fast." In this comment, this comment data not only describes the quality attributes of this product , also describes the appearance attributes and express attributes of this product.

Specifically, the construction of the user comment data corpus includes:

Collect the review data of all products by the same user; specifically, the used review data is consistent with the original data for constructing the product review data corpus;

The Jieba word segmentation module is used to perform word segmentation on the comment data of all products by the same user, and remove stop words;

Using a dictionary-based classification algorithm, according to the four attributes of quality, price, appearance and express delivery, the same user's review data for all products is classified to complete the construction of the user review data corpus.

Specifically, building a product review space based on the product review data corpus includes:

Four-dimensional space coordinates are constructed according to the four attributes of quality, price, appearance and express delivery; specifically, the present invention extracts four commodity attributes that are generally concerned by users as classification objects when classifying the review data, and the construction of the multi-dimensional space coordinate system is also based on the commodity These four attribute characteristics of the coordinate axis are established. Each coordinate axis in this spatial coordinate system represents an attribute feature of the product. The positive direction of the coordinate axis indicates that the review has a positive attitude toward this attribute feature, and the negative direction indicates that the review has a negative attitude toward this attribute feature. Comments were neutral or did not mention this attribute.

Use the KNN classification algorithm to classify the review data in the product review data corpus into two categories: positive and negative;

All the review data describing the same attribute feature of the same product is formed into a sequence, and the first comment data marked as positive in the sequence is designated as the positive bias reference review on the attribute feature, which is the unit on this attribute feature. 1; Set the last sentence marked as negative as a negative reference comment on the attribute feature, which is the unit-1 on the attribute feature;

Use the cosine similarity algorithm to calculate the cosine value between the review data to be processed and the reference review on this attribute feature: if the classification label of the review data is positive, calculate the difference between the review data and the positive biased reference review on this attribute feature The obtained cosine value is the coordinate value of the review data in the attribute dimension; if the classification label of the review sentence is negative, calculate the cosine value between the review data and the negative bias reference review on this attribute feature. , the inverse of the obtained cosine value is the coordinate value of the comment data in this attribute dimension;

In the same review data, the calculated coordinate values of each attribute dimension are summarized into a coordinate point in a four-dimensional space. For the attribute features not mentioned in the review data, the value on the dimension corresponding to the unmentioned attribute feature is calculated. is 0, the coordinate value of the comment data in the four-dimensional space coordinate is obtained;

Average the values of all the comment data of the same product in the dimension of the same product, and obtain the coordinate value of each attribute of each product in the four-dimensional space coordinates;

The coordinate value of each attribute of each commodity in the spatial coordinates, that is, each commodity point, is presented in the four-dimensional spatial coordinates to complete the construction of the commodity comment space.

Specifically, the quantitative representation of the sentimental tendency of product reviews refers to evaluating the attribute feature described by each review of the product, and finally obtaining a value to represent the sentimental tendency of the review to this attribute feature. The spatial coordinates generated by the present invention It is based on sentiment analysis of each comment. When determining the emotional tendency, this embodiment uses three tools: SnowNLP class library based on Python, KNN classification algorithm, and cosine similarity;

The SnowNLP class library is used to calculate the probability of the positive sentiment of the sentence after each comment is cut by attributes, and obtain a value that represents the positive sentiment probability of this sentence on a certain attribute feature. For example: a sentence that describes the quality attribute characteristics "the quality of this product is good", the probability value of the positive emotion of this sentence is 0.749396374210698. For the same commodity, obtain a descending sequence S of sentence positivity describing the same attribute feature;

However, the SnowNLP class library can only obtain the probability of being biased towards the positive, which is not enough for the present invention to construct the comment space, and the present invention also proposes the following solutions:

The KNN classification algorithm is a theoretically mature classification algorithm and one of the simplest machine learning algorithms. The core idea of the algorithm is to analyze which category the most recent K objects of the object to be classified belong to, and which category the object to be classified belongs to. Since it is only related to the recent K samples, the KNN classification algorithm has higher requirements on the training set. If the error rate of the training set sample data is high, the accuracy of the final classification result will also be low. In this embodiment, the value of K is set to 9, so that the situation where the number of two classes is the same can be well prevented. Calculate the distance between all product reviews and the reviews that have been classified in the training set, and obtain the K closest reviews. Calculate the number of categories to which the K comments belong, and divide the comments that need to be classified into the categories that the K comments belong to;

To find the two reviews corresponding to 1 and -1 on the axis, we use SnowNLP and KNN classification algorithms on the contest data with sentiment polarity annotations. SnowNLP can only obtain the probability that the current comment is biased towards positive (positive sentiment), and we take the comment data with the highest probability in the S sequence as the comment corresponding to 1. When obtaining the comment corresponding to -1, due to the limitations of the algorithm, SnowNLP obtains the sequence at the beginning, and the probability difference is obvious; at the end, the probability value is not very different. In order to find the comment corresponding to -1, the present invention will use the KNN classification method to classify the contest data marked with sentiment polarity, and use SnowNLP to process all the negative comment data sets separately. The probability value is obviously different, whichever is the smallest The value as -1 corresponds to the comment.

Specifically, the present invention uses the KNN classification algorithm to process each sentence after the public data set is processed with the help of the contest data marked with emotional polarity, and divides these comment sentences into positive and negative according to the emotional polarity of each sentence Two categories. After the classification is completed, the sequence S corresponding to the positive and negative comment sentences of each attribute of each product is obtained through the SnowNLP class library. Among all the sentences describing the same attribute feature of the same product, the first sentence in the sequence S corresponding to the positive comment sentence is selected. The sentence marked as positive is designated as the positive biased reference sentence on the attribute feature, which is the unit 1 on this attribute feature, and the last sentence marked as negative in the sequence S corresponding to the negative comment sentence is designated as the attribute feature. The negative reference sentence is the unit-1 on the attribute characteristic;

Cosine similarity is an algorithm that estimates the similarity between two vectors by calculating the cosine value of the angle between two vectors in space. The cosine value has a minimum value of -1 and a maximum value of 1. When the cosine value between the two vectors is closer to 1, it means that the angle between the two vectors is closer to 0 degrees, and the two vectors are more similar; when the cosine value between the two vectors is closer to -1, indicating that the closer the angle between the two vectors is to 180 degrees, the greater the difference between the two vectors. For example, in two-dimensional space, define the coordinates of vector a as (x ₁ , y ₁ ) and the coordinates of vector b as (x ₂ , y ₂ ), and place the two vectors in the same two-dimensional space, as shown in the figure 2 shown;

Then the cosine value of the angle between the vector a and the vector b, the calculation formula is shown in (1):

If a and b are two vectors in n-dimensional space, this formula still holds. Suppose a and b are two n-dimensional vectors. Then the calculation formula of the cosine value of the vector a and the vector b is as follows. Therefore, the present invention regards two reviews as two n-dimensional vectors, and the cosine similarity can be used to calculate the similarity between the two reviews:

The present invention uses the cosine similarity algorithm to calculate the cosine value between the comment sentence to be processed and the standard comment sentence on this attribute feature. If the classification label of the comment statement is positive, the cosine value of the comment statement and the attribute feature is positively biased with reference to the comment to calculate the cosine value, and the obtained value is the coordinate value of the comment statement in the attribute dimension; If the classification label is negative, the cosine value of the comment statement and the negative bias on this attribute feature is calculated with reference to the comment, and the inverse of the obtained value is the coordinate value of the comment statement on the attribute dimension;

In the same comment, the calculated coordinate values of each attribute dimension are summarized into a coordinate point in a four-dimensional space. If the comment does not mention a certain attribute feature, the value of the comment on this attribute dimension is 0, and the comment is obtained. The coordinate value in space coordinates;

The invention adopts the method of taking the mean value to obtain the coordinates of the commodity in the comment space coordinate system. After coordinate representation of each comment, get the specific value of each comment's emotional inclination on each attribute, such as: comment "The quality of this toy is too bad, and it is sold so expensive, bad comment", this The result of a comment after coordinate representation is (-0.91, -0.85, 0, 0). This coordinate indicates that this review has a value of -0.91 in the dimension of the quality attribute feature, a value of -0.85 in the dimension of the price attribute feature, and a value of 0 in the two attribute feature dimensions of appearance and express delivery;

Each product contains multiple reviews, and each review corresponds to a coordinate value in the spatial coordinate system. The present invention adds up the corresponding attributes in the coordinate values of all comments of each product to obtain an average value, as shown in formula (3):

All commodities are displayed in a four-dimensional space, in which the first three dimensions are quality, price, and appearance, which are displayed through the coordinate axis, and the fourth-dimensional express attribute indicates the value of the fourth-dimensional value through the change of color depth. As shown in Figure 3.

Specifically, after the completion of the product review space construction, the method further includes:

K-means clustering algorithm is used to cluster the generated commodity points, commodities with similar attribute characteristics are placed in a cluster, and the cluster center is saved; Clustering, placing products with similar attribute characteristics in a cluster is convenient for recommending operations for users. When using the K-means clustering algorithm for clustering operation, the K value needs to be determined first, and the K value cannot be determined only by analyzing the commodity points formed above. Determine the size of the K value. As shown in Figure 4;

The core indicator of the elbow method is SSE, that is, the sum of squares of errors, and its core idea is:

a. With the continuous increase of the K value, the samples are divided more finely, the degree of aggregation between the data objects in each cluster will become higher and higher, and the SSE value will also decrease;

b. When the K value is less than the optimal number of clusters, with the increase of the K value, the degree of aggregation between the data in each cluster will greatly increase, and the SSE value will also greatly decrease. When the K value reaches the optimal number of clusters, with the increase of the K value, the increase in the degree of aggregation between the data in each cluster will decrease, the decrease in the SSE value will also decrease, and then it will become flat. The K value closest to the extreme point of SSE is the optimal number of clusters;

According to the principle of the elbow method, when the K value is 7, the change of the SSE value is relatively slow. Therefore, for the clustering results of the data set used in this embodiment, the optimal number of clusters is 7 ;

In the present invention, after using the K-means algorithm to obtain several clusters, the coordinates of the center points of several clusters are saved, and the coordinates of other points in the clusters can be temporarily ignored, but it is necessary to mark which cluster these points belong to and which cluster each The product name corresponding to the coordinate point.

Specifically, building a user preference space based on the user comment data corpus includes:

Four-dimensional space coordinates are constructed according to the four attributes of quality, price, appearance and express delivery; in order to achieve the purpose of synchronizing with commodity attributes and improve the accuracy of recommendation, the user's preference space coordinate system needs to be the same as the commodity comment space coordinate system. A four-dimensional space coordinate system is established, and the four axes represent the user's attention to quality, price, appearance, and express delivery. The same coordinate system can combine the user's shopping preference with the attribute characteristics of the product, and can achieve a high accuracy rate when recommending;

Quantify the user's emotional tendency according to the user's degree of attention to each attribute feature, where the degree of attention is the ratio of the number of comment data for each attribute of the same user to the total number of comment data, and obtain the spatial coordinates of each user's emotional tendency; For all the comments of a single user, some comments are related to quality attributes, and some comments are related to price attributes. To get the user's attention to quality attributes. For example, if you want to get a user's attention to the quality attribute, you need to count all the comment sentences describing the quality in the user, and calculate the ratio of these comment sentences to the total number of comment sentences. degree of attention. The calculation formula is shown in (4):

Among them, (n _quality , n _price , n _appearance , n _express ) represent the number of comment sentences describing quality, price, appearance, and express attributes respectively, and n represents the total number of comment sentences of the user;

All the spatial coordinate points representing the user are displayed in the four-dimensional spatial coordinates to complete the construction of the user preference space; the present invention obtains the spatial coordinates of the emotional inclination of a single user through the quantitative representation of the user's emotional inclination, and this coordinate is the user's preference space in the preference space. Location. All spatial coordinate points representing users are displayed in the user preference space, in which the first three dimensions represent the user's emotional tendencies towards the quality, price, and appearance attributes of commodities respectively, and the fourth dimension represents the user's emotional tendencies towards the properties of commodity express delivery. The magnitude of the value is represented by the intensity of the color. As shown in Figure 5.

Specifically, the step S103 includes:

The user's position in the user preference space is taken as the point of the user's shopping preference, and the Euclidean distance is used to calculate the distance between the point of the user's shopping preference and the center point of each cluster in the product review space, and the cluster closest to the point of the user's shopping preference is calculated. products are recommended to users. When the present invention performs the recommendation operation, according to the position of the user in the preference space, the product in the cluster closest to the user is recommended to the user. Because when the products are mapped to the multi-dimensional space coordinate system, these products have been clustered, and the center point coordinates of each cluster are calculated and retained; when users are mapped to the multi-dimensional space coordinate system, each Coordinate points representing the user's shopping preferences. The present invention uses Euclidean distance to calculate the distance between the points of the user's shopping preference and the center points of these clusters. Among them, Euclidean distance, also known as Euclidean distance, is a widely used distance definition, which refers to the distance between two points in n-dimensional space. The calculation formula is shown in (5).

Among them, n represents an n-dimensional space. The smaller the distance, the more the products in this cluster are in line with the user's shopping preferences; the larger the distance, the less the products in this cluster are in line with the user's shopping preferences;

The product in the cluster with the smallest distance is recommended to the user, and the recommendation results are shown in Figure 6, Figure 7, and Figure 8. Fig. 6 shows some products recommended by the user whose ID is 15905 according to the method of the present invention. Since the present invention processes commodity data and user data in parallel, there is no relationship between commodities and users during processing, and in the recommendation result, the commodities that the user once gave good reviews are re-recommended to the user, and the recommendation result is highly credible.

In order to verify the effect of the present invention, the recommendation result is measured by the correct rate p, where p represents the proportion of the products recommended by the recommendation system to the user, the products actually liked by the user, and the larger the value, the better the effect of the recommendation algorithm. The calculation formula of P is shown in (6):

Among them, R(u) represents the list of products recommended by the recommendation system, T(u) represents the user's behavior list, that is, the user's favorite products, and the absolute value represents the length of the list.

The method of the present invention is compared with the recommendation accuracy p of SVD, recommendation algorithm based on association rules, recommendation algorithm based on commodity attribute value, and traditional algorithm (collaborative filtering algorithm based on user attributes and commodity classification). Specifically, the length of the recommendation list is selected to be 100, that is, |R(u)|=100. The recommended numbers of users are 5, 10, 15, and 20, respectively, and the calculation results are shown in Figure 9. It can be concluded from FIG. 9 that the recommendation method of the present invention has the highest accuracy rate, but as the number of recommended users continues to increase, the accuracy rate is constantly changing. The reason is that, with the increasing number of recommended users, the number of liked items among the recommended users is different, and when calculating the correct rate, the magnitude of the increase of the numerator is also different, and the final calculated correct rate is constantly changing.

To sum up, the recommendation system of the online shopping platform recommends the most suitable product to the user by discovering some connections between the user and the product. By mining the relationship between the user's shopping preference and the commodity attributes, the present invention can more accurately recommend the commodity conforming to the user's shopping preference to the user. The invention uses the commodity review data to quantify the commodity attributes and map the commodities to the attribute space, and carries out commodity recommendation in combination with the method of the user's historical shopping preference. For the existing recommendation system, the problem of low recommendation efficiency caused by sparse matrix is avoided, and with the continuous increase of the amount of comment information, the judgment of the attribute characteristics of the product is more accurate. Each user's comments are analyzed in the way of independent analysis among users, and their shopping preferences are obtained by this method, so that the correctness of the recommendation for the user will not be affected by the analysis results of other users, so as to realize personalized recommendation. Experimental results show that only using historical review data can achieve good results.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made. It should be regarded as the protection scope of the present invention.

Claims (7)

1. a recommendation method based on comment space and user preference, is characterized in that, comprises: Step 1: Build a corpus of product review data and a corpus of user review data; Step 2: Build a product review space and a user preference space based on the product review data corpus and the user review data corpus respectively; Step 3: According to the user's position in the user preference space and the constructed product review space, the product recommendation for the user is obtained. 2. A recommendation method based on comment space and user preference according to claim 1, wherein the building a corpus of commodity comment data comprises: Collect the comment data of all users on the same product; The Jieba word segmentation module is used to perform word segmentation on the comment data of all users on the same product, and remove stop words; Using a dictionary-based classification algorithm, according to the four attributes of quality, price, appearance and express delivery, the review data of all users for the same product is classified to complete the construction of the product review data corpus. 3. A kind of recommendation method based on comment space and user preference according to claim 1, is characterized in that, described constructing user comment data corpus comprises: Collect the comment data of all products by the same user; The Jieba word segmentation module is used to perform word segmentation on the comment data of the same user on all products, and remove stop words; Using a dictionary-based classification algorithm, according to the four attributes of quality, price, appearance and express delivery, the same user's review data for all products is classified to complete the construction of the user review data corpus. 4. A kind of recommendation method based on comment space and user preference according to claim 1, is characterized in that, constructing commodity comment space based on commodity comment data corpus comprises: Construct four-dimensional space coordinates according to the four attributes of quality, price, appearance and express delivery; Use the KNN classification algorithm to classify the review data in the product review data corpus into two categories: positive and negative; The positive and negative review data describing the same attribute of the same product are formed into a sequence according to the probability of positive sentiment in descending order, and the first comment data marked as positive in the sequence corresponding to the positive review data is set as this. The positive biased reference review on the attribute feature is the unit 1 on this attribute feature; the last sentence marked as negative in the corresponding sequence of negative review data is designated as the negative reference review on the attribute feature, which is the attribute feature. unit -1; Use the cosine similarity algorithm to calculate the cosine value between the review data to be processed and the reference review on this attribute feature: if the classification label of the review data is positive, calculate the difference between the review data and the positive biased reference review on this attribute feature The obtained cosine value is the coordinate value of the review data in the attribute dimension; if the classification label of the review sentence is negative, calculate the cosine value between the review data and the negative bias reference review on this attribute feature. , the inverse of the obtained cosine value is the coordinate value of the comment data in this attribute dimension; In the same review data, the calculated coordinate values of each attribute dimension are summarized into a coordinate point in a four-dimensional space. For the attribute features not mentioned in the review data, the value on the dimension corresponding to the unmentioned attribute feature is calculated. is 0, the coordinate value of the comment data in the four-dimensional space coordinate is obtained; Average the values of all the comment data of the same product in the dimension of the same product, and obtain the coordinate value of each attribute of each product in the four-dimensional space coordinates; The coordinate value of each attribute of each commodity in the spatial coordinates, that is, each commodity point, is presented in the four-dimensional spatial coordinates to complete the construction of the commodity comment space. 5. A recommendation method based on comment space and user preference according to claim 4, characterized in that, after the completion of the product comment space construction, further comprising: The K-means clustering algorithm is used to cluster the generated product points, and the products with similar attribute characteristics are placed in a cluster, and the cluster center is saved. 6. A kind of recommendation method based on comment space and user preference according to claim 1, is characterized in that, constructing user preference space based on user comment data corpus comprises: Construct four-dimensional space coordinates according to the four attributes of quality, price, appearance and express delivery; Quantify the user's emotional tendency according to the user's degree of attention to each attribute feature, where the degree of attention is the ratio of the number of comment data for each attribute of the same user to the total number of comment data, and obtain the spatial coordinates of each user's emotional tendency; Display all spatial coordinate points representing users in four-dimensional spatial coordinates to complete the construction of user preference space. 7. A kind of recommendation method based on comment space and user preference according to claim 5, is characterized in that, described step 3 comprises: The user's position in the user preference space is taken as the point of the user's shopping preference, and the Euclidean distance is used to calculate the distance between the point of the user's shopping preference and the center point of each cluster in the product review space. products are recommended to users.

Images