A collaborative filtering recommendation method and system integrating product features

Abstract

The present invention discloses a collaborative filtering recommendation method and system integrating commodity features, which belongs to the field of intelligent commodity recommendation for new retail enterprises. It makes improvements on the basis of using collaborative filtering algorithm to calculate the similarity of items, integrates commodity features, calculates the similarity matrix based on commodity features, and combines it with the commodity similarity matrix of collaborative filtering algorithm. Finally, the predicted score of the commodity is obtained according to the similarity between items and the historical interest information of the user, so as to generate a recommendation list for the user; it solves the problems of cold start and low recommendation accuracy of traditional collaborative filtering algorithms. Through experimental verification, the algorithm proposed by the present invention has certain improvements in accuracy, recall rate, coverage and popularity compared with traditional collaborative filtering. Not only that, the accuracy of the present invention is also better than that of content-based recommendation algorithms, deep neural network-based recommendation algorithms and matrix decomposition-based recommendation algorithms.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent recommendation of commodities for new retail enterprises, and specifically relates to a collaborative filtering recommendation method and system integrating commodity features. Background Art

[0002] With the development of information technology and the Internet, the information disseminated by various industries and platforms has become more and more abundant and diverse. People have gradually moved from an era of information scarcity to an era of information overload, that is, the era of big data. In the era of big data, massive information perception is everywhere, which makes the value density of data obtained by users lower and lower, that is, it is increasingly difficult to find valuable information in a large amount of data. This poses a huge challenge to both information consumers and information producers: as an information consumer, it is very difficult to find the information of interest from a large amount of information; as an information producer, how to make the information produced stand out and attract the attention of the majority of users has become a world problem. Therefore, how to mine the value of data through algorithms and filter out the information needed by users from a large amount of data is a technical problem that needs to be solved urgently.

[0003] At the end of the 20th century, the introduction and development of search engines successfully solved the needs of some users, and the quality of life of users has been greatly improved. However, with the rapid development of e-commerce, there are more passive users who do not have clear search intentions. If they are asked to search for the information they want, they will inevitably waste a lot of time in a wide variety of information with huge volumes. This is unacceptable to users and is also a problem that troubles e-commerce platform merchants. In order to solve this problem, the recommendation system came into being. The emergence of the recommendation system has far-reaching significance for the development of the times and is a necessary technical tool to help people process effective information. As a filtering technology to solve "data overload", it plays an important role. The recommendation system uses the user's previous information data for analysis. When the user puts forward a demand, the recommendation algorithm and the analyzed information are used for correlation matching, and the information with higher matching degree is sorted and fed back to the user for recommendation. The recommendation system is mainly used to help users find information that they may be interested in, such as goods, books, music, movies, and even friends, thereby reducing the time for users to obtain effective information; at the same time, it can also help merchants reduce risks, open up new commodity markets, and maximize profits. Recommendation systems provide users with personalized and exclusive content and service recommendations to deal with the information overload problem that users often encounter. They play an increasingly important role in today's news dissemination, e-commerce, social networks, scientific research and other related fields, such as Google, Twitter, Taobao, Tmall, Netflix, etc. At present, personalized recommendation systems have become a hot topic of concern and research in the industry and academia, and their application areas are very wide, including e-commerce, social networks, video and music recommendations, etc.

[0004] As the maturity of recommendation systems increases, their application scope and fields have become more and more extensive, and their value is gradually increasing with the application. As a product of digitalization, it plays a key role in helping people identify, transform, and popularize information, helping people discover resources of potential interest to them and saving search time through recommendations. At present, the mainstream recommendation algorithms can be divided into content-based recommendations (CB), collaborative filtering recommendations (CF) and hybrid recommendations. Content-based recommendations directly use the feature information of items for recommendations, but the feature description of items is generally unstructured text data, which increases the difficulty of feature extraction; on the other hand, if the feature information of items is less, it will seriously affect the accuracy of recommendations. Collaborative filtering recommendations include item-based collaborative filtering (Item Collaborative Filtering, Item CF) and user-based collaborative filtering (User Collaborative Filtering, User CF). Item-based collaborative filtering and content-based recommendations are actually based on item similarity prediction recommendations, but the similarity calculation methods are different. The former is inferred from the user's historical preferences, while the latter is based on the attribute feature information of the item itself. Although collaborative filtering can solve the problem of feature extraction, it relies heavily on the user rating matrix and performs poorly on data sparsity and cold start problems. When a new item is added to the recommendation system, the similarity between the item and other items cannot be calculated because there is no purchase record, which affects the recommendation quality of the recommendation algorithm. In addition, this method uses a shallow model and cannot learn the deep features of users and items. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a collaborative filtering recommendation method and system that integrates product features, so as to solve the problems of existing traditional collaborative filtering recommendation algorithms, such as sparse data, low recommendation accuracy, cold start, and inability to learn the deep characteristics of users and projects.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention discloses a collaborative filtering recommendation algorithm integrating product features, including:

[0008] Step 1: Calculate the product similarity based on collaborative filtering and obtain the product similarity matrix W1;

[0009] Step 2: Extract product features using variance analysis, mutual information method, classification and regression tree embedding method, or packaging method based on recursive feature elimination cross-validation;

[0010] Step 3: Construct a product feature vector based on the Word2Vec model according to the extracted product features;

[0011] Step 4: Use the product feature vector to calculate the feature-based product similarity and obtain the product similarity matrix W2;

[0012] Step 5: Linearly combine the product similarity matrix W1 and the product similarity matrix W2 to obtain the final similarity expression of the products;

[0013] Step 6: Initialize parameters, calculate interest based on the parameters and generate a recommendation list;

[0014] Step 7: Evaluate the generated recommendation list based on accuracy, recall, coverage, and popularity. If it is better than the indicators of the traditional algorithm, output the generated recommendation list. If it is worse than the indicators of the traditional algorithm, execute step 1.

[0015] Preferably, in step 1, obtaining the product similarity matrix W1 includes:

[0016] 1) Create a reverse ranking list of user products based on the sales record table for the past year;

[0017] 2) Construct a product similarity matrix, i.e., a co-occurrence matrix C;

[0018] 3) Calculate the cosine similarity between items based on the co-occurrence matrix C and obtain the similarity matrix W1.

[0019] Preferably, in step 3, constructing a product feature vector based on the Word2Vec model includes:

[0020] 1) Merge the extracted features into a text, and use the entire text as the feature description information of the product;

[0021] 2) Input the processed product feature data into the Word2Vec model training, and convert the result into a vector after the training is completed;

[0022] 3) Obtain the feature vector representation of the product.

[0023] Preferably, in step 4, after obtaining the feature vector representation of the product, the cosine similarity of the products is calculated pairwise to generate a feature-based product similarity matrix W2.

[0024] Preferably, in step 4, the cosine similarity of the products is calculated using the following method:

[0025]

[0026] Among them, μ and ⅴ represent the feature vectors of the product, μ·ν is the dot product of the vectors, and ‖μ‖ is the modulus of the vector.

[0027] Preferably, in step 5, the product similarity matrix W1 and the product similarity matrix W2 are linearly combined to obtain the final similarity expression of the products as follows:

[0028] Sim=αSim p +βSim cf (2)

[0029] Where Sim is the final similarity of the product. p Sim is the product similarity based on features. cf is the product similarity based on collaborative filtering, and α and β are the parameters of linear combination.

[0030] Preferably, in step six, the main parameters of parameter initialization include: the number of single data iterations M, the number of similar items selected K, the number of recommended items N and the type of recommendation algorithm selected.

[0031] Preferably, in step 6, calculating the interest level and generating a recommendation list includes:

[0032] 1) Calculate the user u's interest in an item j through formula (3),

[0033]

[0034] Among them, S(j,k) is the set of k products that are most similar to product j, N(i) is the set of products that user u likes, and w ji is the similarity between product j and product i, r ui Indicates user u’s interest in product i;

[0035] 2) For p ui Sort in descending order, take the TopN products as the improved collaborative filtering recall result set, and recommend them to users.

[0036] Preferably, in step 7, the recommendation accuracy for a single user U is:

[0037]

[0038] The accuracy of the whole system is:

[0039]

[0040] Where M is the total number of test users;

[0041] The recommendation recall rate for a single user U is:

[0042]

[0043] The recall rate of the whole system is:

[0044]

[0045] The coverage calculation formula is as follows:

[0046]

[0047] Where P d (N) represents the number of different products that appear in all user recommendation lists, and P is the total number of products in the system;

[0048] The popularity calculation formula is as follows, where d(i,j) represents the distance between items i and j.

[0049]

[0050] The present invention discloses a collaborative filtering recommendation system integrating product features, including:

[0051] The product feature acquisition module is used to calculate the product similarity based on collaborative filtering to obtain the product similarity matrix W1; the variance analysis method, mutual information method, classification and regression tree embedding method or recursive feature elimination cross-validation packaging method are used to extract product features; the product feature vector W2 based on the Word2Vec model is constructed according to the extracted product features; the product feature vector is used to calculate the product similarity based on the product features to obtain the product similarity matrix;

[0052] The linear combination module is used to linearly combine the product similarity matrix W1 and the product similarity matrix W2 to obtain the final similarity expression of the products; the product similarity of the filter recommendation model is fitted with the product similarity based on the product features to obtain the final similarity expression of the products;

[0053] The initialization module is used to initialize the final similarity expression of the product parameters, calculate the interest degree based on the parameters and generate a recommendation list;

[0054] The evaluation module is used to evaluate the generated recommendation list based on accuracy, recall, coverage and popularity. If it is better than the indicators of the traditional algorithm, the generated recommendation list is output.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] The present invention discloses a collaborative filtering recommendation algorithm integrating commodity features. On the basis of using the collaborative filtering algorithm to calculate the similarity of items, the algorithm makes improvements, integrates commodity features, calculates a similarity matrix based on commodity features, and combines it with the commodity similarity matrix of the collaborative filtering algorithm. Finally, the predicted score of the commodity is obtained according to the similarity between items and the historical interest information of the user, thereby generating a recommendation list for the user. The present invention integrates commodity features, calculates similarity according to the commodity's own attributes, and improves the accuracy of similarity calculation to a certain extent. The present invention improves the traditional collaborative filtering method for calculating commodity similarity, combines commodity feature attributes and user historical purchase records to jointly calculate commodity similarity, solves the problem that some commodities are difficult to correctly find similar commodities due to too few sales records, and thus improves the accuracy and recall rate of recommendations. The present invention constructs a deep-level feature vector of commodities based on the Word2vec model, and the advantage is that the obtained word vector can well measure the similarity between words. In summary, the present invention solves the problems of cold start and low recommendation accuracy of the traditional collaborative filtering algorithm. Through experimental verification, the algorithm proposed in the present invention has improved accuracy, recall rate, coverage, and popularity compared with traditional collaborative filtering. Moreover, the accuracy of the present invention is also better than that of content-based recommendation algorithms, deep neural network-based recommendation algorithms, and matrix decomposition-based recommendation algorithms.

[0057] Furthermore, the present invention uses variance analysis, mutual information method, classification and regression tree embedding method, etc. to extract product features, thereby alleviating the impact of improper feature selection on algorithm accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart for realizing the present invention;

[0059] Figure 2 Flowchart for calculating product similarity based on item collaborative filtering;

[0060] Figure 3 Extract the product features.

[0061] Figure 4 This is the change graph of precision when α=1;

[0062] Figure 5 This is the recall change diagram when α=1;

[0063] Figure 6 This is the coverage change diagram when α=1;

[0064] Figure 7 This is the popularity change diagram when α=1;

[0065] Figure 8This is the change graph of precision when β=10;

[0066] Fig. 9 This is the recall change diagram when β = 10;

[0067] Fig.10 This is the coverage change diagram when β=10;

[0068] Fig.11 This is the popularity change diagram when β=10;

[0069] Fig.12 This is a system use case diagram. DETAILED DESCRIPTION

[0070] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0071] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0072] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0073] See also Figure 1 The overall implementation process of the collaborative filtering recommendation algorithm integrating product features described in the present invention is as follows:

[0074] Step 1: Calculate product similarity based on collaborative filtering

[0075] Collaborative filtering recommendation algorithms are divided into user-based collaborative filtering (UserCF) and item-based collaborative filtering (Item CF). Item CF uses the similarity between items to recommend, so if the number of users far exceeds the number of items, then you can consider using Item CF, such as shopping websites, because the data of their items is relatively stable, so when calculating the similarity of items, not only the amount of calculation is small, but also it does not need to be updated frequently; User CF is more suitable for news, blogs or micro-content recommendation systems, because its content update frequency is very high, especially in social networks, User CF is a better choice, which can increase the user's degree of conviction in the recommendation explanation. The present invention is based on the construction of a recommendation system for the sale of new retail enterprises. The number of users of the experimental data used is as high as one million, and the number of goods is only one thousand, so the recommendation model is constructed by collaborative filtering based on items.

[0076] The ItemCF algorithm does not use the content attributes of items to calculate the similarity between items. It mainly calculates the similarity between items by analyzing the user's behavior records. The algorithm believes that item A and item B have great similarity because most users who like item A also like item B. Its core is to generate the final recommendation result through the user's rating of the nearest neighbor of the target item. The user's rating of the target item is approximated by the weighted average of the user's rating of the nearest neighbor of the target item.

[0077] After understanding the principle of the ItemCF recommendation algorithm, combined with the characteristics of the existing data set, we will begin to build a product similarity matrix based on collaborative filtering, such as Figure 2 As shown, the implementation process is as follows:

[0078] 1) Create a user-item inverse list based on the sales record table data for the past year, that is, create a list of items purchased by each user.

[0079] 2) Construct a product similarity matrix, i.e., the co-occurrence matrix C. Traverse the user-product matrix in the first step, and each row represents a set of items that a user is interested in. For each set of items, add the items in it two by one to get a matrix. Finally, add these matrices to get the co-occurrence matrix C. Among them, C[i][j] records the number of users who like both item i and item j.

[0080] 3) Calculate the cosine similarity between items based on the co-occurrence matrix to obtain the similarity matrix W1. The traditional formula for calculating the similarity of commodities is shown in (10).

[0081]

[0082] Where N(i) represents the number of users who have purchased item i, N(j) represents the number of users who have purchased item j, and the numerator is the number of users who have purchased both item i and item j, represented by |N(i)∩N(j)|. Obviously, there is a problem with the above formula. If item j has been purchased by many people, the numerator will be very large, resulting in a large calculation result. Therefore, this formula will affect the similarity result, causing any item to have a high similarity with the popular item. To avoid this situation, the weight of the popular item must be reduced. The improved formula is shown in (11). The similarity W between the items is calculated according to the improved formula ij , and stored in the W[i,j] element of the matrix W.

[0083]

[0084] Step 2: Feature Selection

[0085] Feature selection is a very critical part of feature engineering, which directly affects the modeling accuracy and system simulation efficiency. The rationality of feature selection determines that the system model accurately reflects the actual situation. Therefore, how to effectively select the most suitable feature vector under limited data conditions has become a very urgent and complex topic. At present, people have adopted some methods and techniques to solve this problem, the purpose of which is to take the optimal subset from a large number of feature parameter sets to obtain the best performance parameter combination scheme, thereby providing more useful information for further research. The present invention adopts variance analysis method, mutual information method, Classification and Regression Trees, CART, and Recursive Feature Elimination Cross Validation, RFECV packaging method to select commodity feature data. The feature selection summary results of these four methods are shown in Table 1.

[0086] Table 1 Feature selection results

[0087]

[0088]

[0089] After the above feature extraction, we finally selected features such as product material name, secondary classification, tertiary classification, skin care series, category, applicable age, applicable part, applicable skin type, skin care efficacy, member price and retail price to build a collaborative filtering recommendation model that integrates product features.

[0090] Step 3: Construct product feature vector based on Word2Vec model

[0091] The feature extraction method selected by the present invention is based on the Word2Vec model, using the feature text of the product as a training data set, and obtaining the feature vector of the product after model training. Word2Vec is a well-known word embedding method, which can calculate the distributed word vector of each word in its given corpus environment. The word vector representation can characterize the semantics of each word to a certain extent. If the semantics of the words are similar, their word vectors are also close to each other in the vector space, which makes the vector modeling of the words more accurate, can improve the existing methods and improve the robustness. Word2Vec includes two training models, namely the continuous bag of words model CBOW and the Skipgram model. The CBOW model predicts the current word W(t) based on the 2n words in the context of the known word W(t); while the Skipgram model predicts the 2n words in the context based on the word W(t). The Skipgram model and the continuous bag of words model CBOW both include an input layer, a projection layer, and an output layer. The present invention uses the Skipgram model. The training goal of the Skipgram is to learn the distribution of word representation vectors, and its optimization goal is to maximize the likelihood function under the condition of a given word vector of the central word.

[0092] After understanding the principles and specific operations of the Word2Vec word embedding model, we will start to build a product feature word vector model based on the characteristics of the existing data set. Figure 3 As shown in the product feature extraction model:

[0093] 1) The extracted features are combined into a text, which is used as the feature description information of the product.

[0094] 2) Input the processed product feature data into the Word2Vec model training, and convert the results into vectors after the training is completed.

[0095] 3) Obtain the feature vector representation of the product.

[0096] Step 4: Calculate product similarity based on features

[0097] The extracted product feature vector can maintain the linear relationship between the product feature texts. In addition, the semantic similarity between two feature vectors can be measured by the cosine similarity between the two vectors. Therefore, after obtaining the feature vector representation of the product, the cosine similarity of the products is calculated pairwise to generate a feature-based product similarity matrix.

[0098] Formula (1) is a method for calculating cosine similarity, where μ and v represent the feature vectors of the product, μ·ν is the dot product of the vectors, and ‖μ‖ is the modulus of the vector.

[0099]

[0100] Step 5: Linearly combine similarity matrices

[0101] The linear combination of the feature-based product similarity matrix and the item-based collaborative filtering product similarity matrix is ​​as follows:

[0102] Sim=αSim p +βSim cf (2)

[0103] Where Sim is the final similarity of the product. p Sim is the product similarity based on features. cf is the product similarity based on collaborative filtering, and α and β are the parameters of linear combination.

[0104] Step 6: Parameter initialization

[0105] The main parameters are as follows: the number of single data iterations M, the number of similar items selected K, the number of recommended items N, and the type of recommendation algorithm selected.

[0106] Step 7: Calculate interest and generate recommendations

[0107] After obtaining the final product similarity, the user u's interest in an item j is calculated using the product similarity and the user's historical purchase behavior. The specific implementation steps are as follows:

[0108] 1) Item-CF calculates the interest of user u in an item j through formula (3). Where S(j,k) is the set of k items most similar to item j, N(u) is the set of items liked by user u, and w ji is the similarity between product j and product i, r ui represents the interest of user u in product i. The meaning of this formula is that the more similar an item is to an item that the user has been interested in historically, the more likely it is to be ranked higher in the user's recommendation list.

[0109]

[0110] 2) For p ui Sort in descending order, take the TopN products as the improved collaborative filtering recall result set, and recommend them to users.

[0111] Step 8: Evaluate the model

[0112] When a website provides recommendation services, if it cannot obtain explicit user evaluations of items, it can use the Top-N recommendation model to use the user's implicit feedback information to recommend the top N items for the item. The prediction accuracy of Top-N is generally measured by accuracy, recall, coverage, and popularity. To calculate the accuracy, first set the length of the recommendation list N, then calculate the list of items that the user may be interested in based on the recommendation model, rank the items based on the predicted interest, and select the top N to recommend them to the user. The user's behavior on a product that he has never used can be roughly divided into four situations: the system recommends to the user and the user likes it; the system recommends to the user, but the user does not like it; the user likes it, but the system does not recommend it; the user does not like it and the system does not recommend it. Table 2 summarizes these four possible situations, where N tp N is the number of items recommended by the system to the user and that the user likes very much; fp N is the number of items recommended by the system to the user but not liked by the user; fn The number of items that users like but the system does not recommend; N tn The number of items that the user does not like and the system does not recommend. u represents the number of items that the user likes, L represents the total number of items recommended by the system, and obviously: B u =N tp +N fn , L=N tp +N fp .

[0113] Table 2 Recommendations

[0114]

[0115] The accuracy is defined as the percentage of items recommended by the system and liked by users to the total number of recommendations. The recommendation accuracy for a single user U is:

[0116]

[0117] The accuracy of the whole system is:

[0118]

[0119] Where M is the total number of test users.

[0120] The recall rate is defined as the percentage of items recommended by the system and liked by users to all items liked by users in the system. The recall rate for a single user U is:

[0121]

[0122] The recall rate of the whole system is:

[0123]

[0124] Coverage rate refers to the percentage of all products recommended by the system to users. This indicator reflects the ability of the recommendation system to discover long-tail products. The larger the coverage rate, the more long-tail products the recommendation system can mine; conversely, it means that the recommendation system cannot mine more long-tail products. Therefore, coverage rate is an important indicator to measure the ability of the recommendation system to mine long-tail products. The calculation formula is as follows:

[0125]

[0126] Where P d (N) represents the number of different items that appear in all user recommendation lists, and P is the total number of items in the system.

[0127] Popularity refers to the popularity of the items recommended by the system to users. The less popular the items, the more likely they are to be considered novel by users. The calculation formula is as follows, where d(i,j) represents the distance between items i and j.

[0128]

[0129] The collaborative filtering recommendation that integrates product features proposed in the present invention is to generate a personalized recommendation list for the user, which contains a set of N products that are most similar to the products purchased by the user. Therefore, accuracy, recall, coverage and popularity are used as experimental evaluation indicators. Accuracy represents the degree of match between the actual purchased items of the user and the items recommended by the system. The higher the degree of match, the more accurate the recommendation result. Recall represents the proportion of correctly recommended items to the items liked by the user. The larger the value, the more items the system recommends correctly, and the higher the user satisfaction. Coverage corresponds to the types of products recommended by the system to the user. The higher the recommended coverage, the more types of products there are, and the stronger the system's ability to mine long-tail products. The lower the popularity, the fewer popular products there are in the product list, and the greater the possibility of recommending diverse and novel products.

[0130] Step 9: Model application

[0131] Through the above research, we obtained a collaborative filtering algorithm that integrates product features with a recommendation model that is superior to traditional algorithms. Next, we will build a new retail enterprise product intelligent recommendation system and apply the above model to the recommendation system. The system is mainly divided into background management module, data processing module, product display module, statistical recommendation module and personalized recommendation module. Among them, the personalized recommendation module is the core module, and its main task is to help users find valuable information for themselves, improve user satisfaction, increase corporate profits, and thus achieve a win-win situation for consumers and producers.

[0132] Confirmatory test

[0133] In order to illustrate the performance of the present invention, a comparative experiment was conducted using the method of the present invention and the traditional collaborative filtering-based method and content-based recommendation algorithm. The experiment mainly tested the algorithm on a selected data set. The data set comes from a retail enterprise, and the data consists of the enterprise's sales records, member data, and product data for one year. Among them, the sales record is a record of users purchasing goods in the past year, and the fields include user number, product number, quantity, amount, and transaction date. The member data set records the basic attributes of members, including gender, year of birth, date of birth, member level, and skin quality. Product data includes product name, classification, series, category, applicable age, retail price, etc.

[0134] Next, we will conduct an experimental verification on the recommendation effect of the collaborative filtering recommendation model that integrates product features. The selected data set is the sales records and product data of "a certain retail enterprise" in Beijing for one year. There are 346,432 sales records, which are divided into training set and test set in a ratio of 8:2, and the precision, recall, coverage and popularity indicators are selected for verification. The main innovation of the collaborative filtering recommendation model algorithm that integrates product features is to linearly combine the feature-based product similarity matrix and the project-based collaborative filtering product similarity matrix. Therefore, the values ​​of α and β are the key. The following sets the number of single data iterations M to 2, the number of similar items K to 20, and the number of recommended items N to 10. The result graphs of the precision, recall, coverage, and popularity values ​​of the collaborative filtering recommendation model that integrates product features when different α and β values ​​are taken are obtained. When α is 1, the changes in precision, recall, coverage, and popularity when β takes different values ​​are shown as follows: Figure 4 , 5 , 6, and 7. When β is 10, the changes in precision, recall, coverage, and popularity when α takes different values ​​are shown in Figure 8 , 9, 10, and 11. After continuous training and learning, when is 0.5 and is 10, the experimental accuracy of the collaborative filtering recommendation model recommendation algorithm integrating product features is the highest. In addition to using precision, recall, coverage, and popularity to measure the recommendation performance of the algorithm, it is also necessary to compare the collaborative filtering recommendation model recommendation algorithm integrating product features proposed in the present invention with the traditional recommendation algorithm to verify the accuracy of the method in this article. Table 3 is an experimental comparison of the recommendation algorithms of the present invention based on content, i.e., Content based, CB, based on deep neural networks, i.e., Deep Neural Networks, DNN, based on item collaborative filtering recommendation algorithm, i.e., Item Collaborative Filtering, Item CF, and based on matrix decomposition, i.e., Matrix Factorization, MF, under the "certain retail enterprise" data set.

[0135] Table 3 Experimental results

[0136]

[0137]

[0138] From the analysis of Table 3, we can see that although the popularity of the algorithm proposed in the present invention is slightly lower than that of CB, it has improved in accuracy, recall and coverage compared with Item CF, CB and MF. Although the collaborative filtering algorithm integrating product features proposed in the present invention is comparable to the accuracy of the DNN algorithm, it is better than DNN in recall, coverage and popularity. Based on the above analysis, the collaborative filtering recommendation algorithm integrating product features proposed in the present invention can provide personalized services to members, which is generally better than the existing recommendation algorithms.

[0139] The collaborative filtering recommendation algorithm integrating product features proposed in the present invention is mainly aimed at the problem that the traditional collaborative filtering recommendation user-item matrix is ​​too sparse and the cold start problem of new items. Combining the advantages of Word2Vec in feature extraction, a collaborative filtering recommendation model integrating product features is proposed and constructed. Based on the data set of "a certain retail enterprise", the model is described and experimentally simulated in detail, the performance of transaction data in Beijing under the model is analyzed, and the experimental results are compared with those of traditional recommendation algorithms, so that it is analyzed that the collaborative filtering recommendation model integrating product features generally has better recommendation effect.

[0140] Development and Application

[0141] This recommendation system analyzes the data generated in the e-commerce platform of "a certain retail enterprise", including purchase records, member data, and product data, to find out the associations between members, products, and members, so as to recommend products that the member may be interested in. First, the system cleans the member data, product data, and transaction data of "a certain retail enterprise" received, and imports them into the system to obtain the data required for the experiment; then, the data is processed by features, and the collaborative filtering recommendation model based on product features and the feature fusion recommendation model based on neural networks proposed in this paper are used to complete the recommendation and generate a personalized recommendation list. Finally, the recommendation results are displayed on the front-end page for members to browse. Because this system is an extension of the e-commerce platform of "a certain retail enterprise", it does not provide member registration and purchase functions.

[0142] like Fig.12 As shown in the system use case diagram, this system has three user roles: system administrator, e-commerce platform administrator, and ordinary user.

[0143] For system administrators, use cases include user information management, product information management, transaction data management, rating data management, and recommendation result management.

[0144] For e-commerce platform administrators, use cases include maintenance of recommendation models, data cleaning, and importing.

[0145] For ordinary users, the use cases include personal information management, product browsing, and statistical recommendation list viewing, including sales rankings, products with high browsing volume, etc. The most important is the viewing of personalized recommendation results. This use case is based on the user's historical purchasing behavior to provide each user with products that are more in line with the user's characteristics and interests.

[0146] The main purpose of this personalized recommendation system is to provide accurate personalized recommendations and monitor the recommendation effect. Therefore, this system needs to meet the needs of ordinary netizens for highly accurate recommendation of products and the needs of e-commerce platform users to monitor the recommendation effect of the system in real time. Based on the above analysis of the system and user needs, the recommendation system has the following functions.

[0147] (1) Identity recognition

[0148] This recommendation system is for members of the "certain retail enterprise" platform and does not provide member registration. Therefore, when importing member data, the system will use the member's mobile phone number as the login account and password. After entering the system, members can modify the initial password.

[0149] (2) Data import

[0150] The e-commerce platform administrator cleans the acquired data and imports it into the system database. Because the data is updated regularly, the e-commerce platform administrator needs to update and maintain the data regularly.

[0151] The e-commerce platform administrator imports the cleaned data into the system database, and can query and delete the system database in the background. Secondly, because this data is updated regularly, the e-commerce platform administrator needs to update and maintain this data regularly.

[0152] (3) Backstage management

[0153] The backend management function means that when the system administrator logs into the backend management system after being authenticated, he can manage member information, product information, transaction data and models.

[0154] (4) Statistical recommendations

[0155] Statistical recommendations include historical popular product statistics and product average score statistics. Historical popular product statistics are statistics of historical sales records, and the ten products with the highest sales volume are selected for display; product average score statistics are to calculate the average score of all product ratings based on the rating data, and select the ten products with the highest ratings for display.

[0156] (5) Personalized recommendations

[0157] The system will regularly train the recommendation model and store the recommendation results in the database. When members enter the system, personalized recommendation results will be displayed to them. The system intends to adopt three recommendation schemes, including recommending products with high similarity to historical purchases, recommending products with high predicted scores, and recommending products that members of similar types like.

[0158] In traditional collaborative filtering recommendations, the most important and most concerned area is the calculation of similarity. In most cases, the Person correlation coefficient and cosine similarity are effectively used. Therefore, the present invention makes improvements on the basis of using collaborative filtering algorithms to calculate item similarity, integrates product features, calculates a similarity matrix based on product features, and combines it with the product similarity matrix of the collaborative filtering algorithm. Finally, the predicted score for the product is obtained based on the similarity between items and the user's historical interest information, thereby generating a recommendation list for the user. The present invention calculates the user's interest in the product based on the user's purchase record of the product and the product's own characteristics, thereby judging the possibility of the user purchasing a certain product and generating a Top-N recommendation list. The prediction accuracy of Top-N recommendations is generally measured by accuracy, recall, and coverage. The accuracy is the percentage of correctly recommended items to the total number of recommendations, the recall is defined as the ratio of the items that users like in the recommendation list to all the items that users like in the system, and the coverage index refers to the proportion of all the items that the algorithm recommends to users. If the coverage of a recommendation system is low, then the system is likely to reduce user satisfaction due to the limitations of its recommendation range, because low coverage means that users have few items to choose from. Obviously, this indicator is related to the length N of the recommendation list.

[0159] Taking the data set of "a certain retail enterprise" as an example, the present invention is experimentally simulated. There are 346,432 rating data of the enterprise in Beijing for one year, which are divided into training set and test set in a ratio of 8:2, and the accuracy, recall, coverage and popularity indicators are selected for verification. The experimental results show that although the popularity of the collaborative filtering recommendation algorithm integrating product features proposed in the present invention is slightly worse than that of CB, it has a certain improvement in accuracy, recall and coverage compared with Item CF, CB and MF; although the collaborative filtering integrating product features proposed in the present invention is comparable to the accuracy of the DNN algorithm, it is better than DNN in recall, coverage and popularity. Based on the above analysis, the collaborative filtering recommendation algorithm integrating product features proposed in the present invention can provide personalized services to members, and its recommendation performance is generally better than the existing recommendation algorithms.

[0160] The present invention discloses a collaborative filtering recommendation system integrating product features, including:

[0161] The product feature acquisition module is used to calculate the product similarity based on collaborative filtering to obtain the product similarity matrix W1; the variance analysis method, mutual information method, classification and regression tree embedding method or recursive feature elimination cross-validation packaging method are used to extract product features; the product feature vector W2 based on the Word2Vec model is constructed according to the extracted product features; the product feature vector is used to calculate the product similarity based on the product features to obtain the product similarity matrix;

[0162] The linear combination module is used to linearly combine the product similarity matrix W1 and the product similarity matrix W2 to obtain the final similarity expression of the products; the product similarity of the filter recommendation model is fitted with the product similarity based on the product features to obtain the final similarity expression of the products;

[0163] The initialization module is used to initialize the final similarity expression of the product parameters, calculate the interest degree based on the parameters and generate a recommendation list;

[0164] The evaluation module is used to evaluate the generated recommendation list based on accuracy, recall, coverage and popularity. If it is better than the indicators of the traditional algorithm, the generated recommendation list is output.

[0165] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0167] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0168] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A collaborative filtering recommendation method integrating product features, characterized in that: include: Step 1: Calculate the product similarity based on collaborative filtering and obtain the product similarity matrix W1; Step 2: Extract product features using variance analysis, mutual information method, classification and regression tree embedding method, or packaging method based on recursive feature elimination cross-validation; Step 3: Construct a product feature vector based on the Word2Vec model according to the extracted product features; Step 4: Use the product feature vector to calculate the product similarity based on the product features to obtain the product similarity matrix W2; Step 5: Linearly combine the product similarity matrix W1 and the product similarity matrix W2 to obtain the final similarity expression of the products; Step 6: Initialize the final similarity expression of the product with parameters, calculate the interest degree based on the parameters and generate a recommendation list; Calculating interest and generating a recommendation list includes: 1) Calculate the user u’s interest in an item j using formula (3), Among them, S(j,k) is the set of k products that are most similar to product j, N(u) is the set of products that user u likes, and w ji is the similarity between product j and product i, r ui Indicates user u’s interest in product i; 2) For p ui Sort in descending order, take the top N products as the improved collaborative filtering recall result set, and recommend them to users; Step 7: Evaluate the generated recommendation list according to accuracy, recall, coverage and popularity. If it is better than the accuracy, recall, coverage and popularity of the traditional algorithm, output the generated recommendation list. If it is worse than the accuracy, recall, coverage and popularity of the traditional algorithm, execute step 1. The traditional algorithms include: Item CF, CB, MF and DNN. The recommendation accuracy for a single user U is: The accuracy of the whole system is: Where M is the total number of test users; The recommendation recall rate for a single user U is: The recall rate of the whole system is: The coverage calculation formula is as follows: Where P d (N) represents the number of different products that appear in all user recommendation lists, and P is the total number of products in the system; The popularity calculation formula is as follows, where d(i,j) represents the distance between items i and j; Among them, N tp N is the number of items recommended by the system to the user and that the user likes very much; fp N is the number of items recommended by the system to the user but not liked by the user; fn The number of items that users like but the system does not recommend; N tn The number of items that the user does not like and the system does not recommend; use B u represents the number of items that the user likes, L represents the total number of items recommended by the system, and obviously: B u =N tp +N fn , L=N tp +N fp .

2. The collaborative filtering recommendation method integrating product features according to claim 1, characterized in that: In step 1, obtaining the product similarity matrix W1 includes: 1) Create a reverse ranking list of user products based on the sales record table for the past year; 2) Construct a product similarity matrix, i.e., a co-occurrence matrix C; 3) Calculate the cosine similarity between items based on the co-occurrence matrix C and obtain the similarity matrix W1.

3. The collaborative filtering recommendation method integrating product features according to claim 1, characterized in that: In step 3, constructing a product feature vector based on the Word2Vec model includes: 1) Merge the extracted features into a text, and use the entire text as the feature description information of the product; 2) Input the processed product feature data into the Word2Vec model training, and convert the result into a vector after the training is completed; 3) Obtain product feature vector based on Word2Vec model.

4. The collaborative filtering recommendation method integrating product features according to claim 1, characterized in that: In step 4, after obtaining the feature vector representation of the product, the cosine similarity of the products is calculated pairwise to generate a feature-based product similarity matrix W2.

5. The collaborative filtering recommendation method integrating product features according to claim 4, characterized in that: In step 4, the cosine similarity of the products is calculated using the following method: Where μ and v represent the feature vectors of the product, μ·v is the dot product of the vectors, and ||μ|| is the modulus of the vector.

6. The collaborative filtering recommendation method integrating product features according to claim 1, characterized in that: In step 5, the product similarity matrix W1 and the product similarity matrix W2 are linearly combined to obtain the final similarity expression of the product as follows: Sim=αSim p +βSim cf (2) Where Sim is the final similarity of the product. p Sim is the product similarity based on features. cf is the product similarity based on collaborative filtering, and α and β are the parameters of linear combination.

7. The collaborative filtering recommendation method integrating product features according to claim 1, characterized in that: In step six, the main parameters for parameter initialization include: the number of single data iterations M, the number of similar items selected K, the number of recommended items N, and the type of recommendation algorithm selected.

8. A collaborative filtering recommendation system integrating product features, characterized in that: include: The product feature acquisition module is used to calculate the product similarity based on collaborative filtering and obtain the product similarity matrix W1; Variance analysis, mutual information method, classification and regression tree embedding method or recursive feature elimination cross-validation based packaging method are used to extract product features; a product feature vector W2 based on the Word2Vec model is constructed based on the extracted product features; the product feature vector is used to calculate the product similarity based on the product features to obtain a product similarity matrix; A linear combination module is used to linearly combine the product similarity matrix W1 and the product similarity matrix W2 to obtain a final similarity expression of the products; Fit the product similarity of the filtering recommendation model and the product similarity based on product features to obtain the final similarity expression of the products; The initialization module is used to initialize the final similarity expression of the product parameters, calculate the interest degree based on the parameters and generate a recommendation list; Calculating interest and generating a recommendation list includes: 1) Calculate the user u’s interest in an item j using formula (3), Among them, S(j,k) is the set of k products that are most similar to product j, N(u) is the set of products that user u likes, and w ji is the similarity between product j and product i, r ui Indicates user u’s interest in product i; 2) For p ui Sort in descending order, take the top N products as the improved collaborative filtering recall result set, and recommend them to users; An evaluation module is used to evaluate the generated recommendation list according to accuracy, recall, coverage and popularity, and output the generated recommendation list if it is better than the accuracy, recall, coverage and popularity of traditional algorithms; the traditional algorithms include: Item CF, CB, MF and DNN; The recommendation accuracy for a single user U is: The accuracy of the whole system is: Where M is the total number of test users; The recommendation recall rate for a single user U is: The recall rate of the whole system is: The coverage calculation formula is as follows: Where P d (N) represents the number of different products that appear in all user recommendation lists, and P is the total number of products in the system; The popularity calculation formula is as follows, where d(i,j) represents the distance between items i and j; Among them, N tp N is the number of items recommended by the system to the user and that the user likes very much; fp N is the number of items recommended by the system to the user but not liked by the user; fn The number of items that users like but the system does not recommend; N tn The number of items that the user does not like and the system does not recommend; use B u represents the number of items that the user likes, L represents the total number of items recommended by the system, and obviously: B u =N tp +N fn , L=N tp +N fp .

Images