CN115618127A - A Neural Network Recommender System Collaborative Filtering Algorithm - Google Patents

Abstract

The invention discloses a collaborative filtering algorithm for a neural network recommendation system. The invention first uses a convolutional neural network to establish a convolutional neural network model for some previous behaviors of users and the attributes of items themselves, and extracts the characteristics of users and their item attributes. Fit the score in the way of full connection, introduce the idea of association rules, generate the item association matrix according to the item association relationship, generate the candidate item set, propose the revised Pearson correlation coefficient, establish the user-rating matrix, optimize the matrix, and calculate The similarity of the scores between users is ranked according to the similarity of the scores according to the order from high scores to low scores, and the nearest neighbors in the front are selected to calculate the target user's rating prediction value for items without ratings, and then select the prediction The N items with the highest value are recommended to users, and the movie recommendation algorithm based on convolutional neural network and recurrent neural network has good results in the corresponding evaluation indicators.

Description

A Neural Network Recommender System Collaborative Filtering Algorithm

technical field

The invention relates to the technical field of medical devices, in particular to a collaborative filtering algorithm for a neural network recommendation system.

Background technique

The advancement of science and technology has made human life more and more convenient, and people can see the beauty of the world without leaving home. They will share the interesting things they have encountered on social platforms, and use the Internet to search a large number of content they are interested in. The amount of behavioral data generated by each person is huge, which makes the information on the network soar. In such an environment of massive data, it is difficult for people to accurately and quickly obtain data that is meaningful to them. To solve this problem, it seems that the more effective solution is the recommendation system. By analyzing a series of behaviors generated by users when they visit webpages, for example, the movie recommendation system integrates the user's movie viewing information, browsing information and movie ratings, and then the user's possible feelings Interested in the movie recommended to him. A mature movie recommendation will provide online users with real-time recommendations, which will increase the user's stickiness to the website or application.

In today's society, recommendation systems are ubiquitous. Fields such as e-commerce, books, music, social networking, and movies all have their own recommendation algorithms. The initial recommendation is to sort according to the user's browsing data, and then recommend the top-ranked products to the user. Take the movie field as an example. Now there is a comedy film, and many users have clicked on this comedy film, so the ranking of this comedy film is high. And now there is a user who is very interested in suspense films but does not like comedy films, it is very inappropriate to recommend comedy films to this user at this time. Therefore, when making online real-time recommendations to users, if this method is used, it may lead to the loss of a large number of users, and they will not be loved by users. Such products cannot survive in this environment. At present, many video and movie websites attach great importance to the quality and efficiency of recommendation algorithms. Some video and movie companies will spend some money to hold some competitions to obtain better recommendation algorithms. For example, Netflix spent a lot of money to hold a video recommendation competition, and applied the recommendation method that performed better in the competition to their company's movie recommendation project. For example, Association Rules (Association Rules), Singular Value Decomposition (Singular Value Decomposition, SVD), Collaborative Filtering (Collaborative Filtering) and other recommendation methods, the birth of these methods has brought new vitality to the field of movie recommendation systems. Of course, the application of these methods has also achieved very good recommendation results. However, traditional collaborative filtering algorithms cannot combine contextual connections to make recommendations, and cannot effectively capture changes in users' hobbies, resulting in a decrease in the accuracy of recommendations, and when faced with the cold start problem of new users, there will be items to be predicted Insufficient number of neighbors and other issues. Therefore, a collaborative filtering algorithm for neural network recommendation system is proposed.

Contents of the invention

The purpose of the present invention is to provide a collaborative filtering algorithm for a neural network recommendation system, which solves the problem that the traditional collaborative filtering algorithm cannot make recommendations in combination with contextual connections, and cannot effectively capture changes in users' hobbies, which leads to a decrease in the accuracy of recommendations. decline, and when faced with the cold start problem of new users, there will be problems such as insufficient neighbors of the project to be predicted.

Above-mentioned technical purpose of the present invention is achieved through the following technical solutions:

A neural network recommendation system collaborative filtering algorithm, comprising the following steps:

Include the following steps:

S1. First, use the convolutional neural network to establish a convolutional neural network model for some of the user's previous behaviors and the attributes of the project itself, so as to provide a more accurate recommendation to the user;

The data set mainly contains two static objects, users and projects, among which user-related attributes include user ID, gender, occupation and age, and project attributes include project ID, project type and project name. First, it needs to be digitally encoded , convert the discrete variable into a continuous value vector, and the category field in the data, such as the item category attribute, will be represented by characters. The usual processing method is to use one-hot encoding to convert it into a value;

S2. Extract the features of users and their item attributes, and fit the scores through full connection;

First, convert the attribute values into numbers for representation, and use the converted number as the index N of the embedding matrix, and then assign a fixed-size potential factor to each index, which is the size of the vector. Usually, it is selected. Then the size of the matrix dimension is (N}32), and the user feature vector and item feature vector can be obtained after the data-related attributes are operated, and the score value can also be fitted at this time;

S3. Introduce the idea of association rules, generate an item association matrix according to the item association relationship, and use the item association matrix to generate candidate item sets;

The degree of relevance indicates the strength of the relationship between items. This method defines the degree of relevance as the possibility that the user _browses another item after browsing one item. The number of users co-viewed by i and project j is calculated by the ratio of the number of users browsed by project i, as follows:

Among them, N _i , N _j represent the number of users who rated item i and item j respectively, and the item correlation matrix G=nxn can be established according to the item correlation degree, where r _ij =(i,j=1,2...n) is the correlation degree between project j and project i, the main diagonal elements of the matrix are all 0, and the correlation degree between projects is generally different, and the size of the correlation degree indicates the strength of the relationship between projects, and the project correlation matrix is generally is asymmetric, that is, r _ij ≠ r _ji ;

The correlation degree is introduced into the selection process of candidate item sets, and the item correlation matrix is used instead of the similarity matrix to generate candidate item sets;

S4. Aiming at the problem of poor similarity accuracy in the case of sparse data, a modified Pearson correlation coefficient is proposed, a user-rating matrix is established, and the matrix is optimized using the forgetting function to calculate the score between the target user and other users. Similarity, according to the order from high score to low score, ranked by the similarity of the score, select the top users as the nearest neighbors of the target user, and calculate the score prediction of the target user for the unrated item according to the score information of the neighbors value, and then select the N items with the highest predicted value to recommend to the user;

When selecting co-rating data, the number of users who co-rated should also be considered. If the proportion of co-rating users among the two item scoring users is higher, the two items will be more similar to a certain extent. For example, item i ₁ is compared with 10 item i ₂ is rated by 12 users, item i ₃ is rated by 50 users, if items i ₁ and i ₂ , and items i ₁ and i ₃ have 8 common rating users, then item i ₁ is considered It is more similar to i _2. Similarly, the project i ₁ and i ₂ have 8 co-rated users, which means that the two items are more similar than the case of 5 co-rated users. Therefore, a modified Pearson correlation coefficient is proposed, and the formula is as follows :

分别为项目i和项目j的平均评分；Among them, U _i , U _j , U represent the user sets who rate item i, item j and both item i and item j respectively; |U _i |, |U _j |, |U| represent sets U _i , U _j , the number of elements in U, r _ui , r _uj are user u’s ratings on item i and item j respectively,

are the average ratings of item i and item j respectively;

The algorithm is as follows:

(1) environment setting

Assuming that user u is a new user to the current system, the historical behavior information is empty, but he already has historical behavior information in other systems (such as social networking sites, business sites, video sites) (can be authorized or the current system has cross-site cookie);

(2) Construction of user relationship network

Build a relationship network between users based on their historical information in other systems, and divide them into communities, and use a similarity calculation formula when building a relationship network based on user purchases or rating information. The formula is:

和

分别表示r_u和r_v中非零元素的标准差[20]；Among them, r _u and r _v represent the rating vectors of users u and v respectively, covar(r _u , r _v ) represents the covariance of r _u and r _v ,

and

Respectively represent the standard deviation of the non-zero elements in r _u and r _v [20];

It is worth noting that a similarity of 0 means that no edge should be added between the corresponding two users, and the overlapping elements between the two vectors should be no less than three, otherwise the similarity between the corresponding two users will be considered to be 0;

If a user has friends in a social networking site, that is, has a circle of friends, then it can be directly divided into communities;

(3) Community division

The community division adopts the method of principal module maximization (PMM). The basic principle is to maximize the modularity index Q. Q measures the community division of the real network by measuring the difference between the random network with the same degree distribution and the interconnection within the community. The strength of , is defined as:

其中

in

S is the affiliation matrix of node u and community C, if node u belongs to community C, then S _uC is 1, otherwise it is 0, B is the module matrix, A is the adjacency matrix of node relationship, d is the degree sequence of all nodes, m is the total number of edges in the node relationship network, and Q less than 0 means that the division effect is very poor;

Considering the multi-dimensional relationship between users, including the user's rating of the project, the friendship between users, etc., the PMM divides the community according to the multi-dimensional relationship, and the elements in the division result S represent the user u. How high is the degree of belonging to community C;

(4) Changes in the strategy of selecting similar neighbors

Previously, among all users, the k neighbors with the highest similarity were selected, but this method only selects users in the same community as the target user u, and the target user u’s predicted score for item j:

瓦为用户u的平均评分,a为常数，C为用户v所在社区，r_V,j为用户v对项目j的评分，

氏为用户v的平均评分。in,

W is the average rating of user u, a is a constant, C is the community of user v, r _V,j is the rating of user v on item j,

is the average rating of user v.

In summary, the present invention has the following beneficial effects:

The experimental simulation proves that the algorithm performs model verification on the public movie dataset MovieLens. Through a series of comparative experimental results, it can be seen that the movie recommendation algorithm based on convolutional neural network and recurrent neural network studied by this method has good results in the corresponding evaluation indicators.

Description of drawings

Fig. 1 is the comparison between recommendation based on user community division and traditional collaborative filtering recommendation in the present invention;

Fig. 2 is UBCF-IBCF prediction accuracy comparison among the present invention;

Fig. 3 is UBCF-IBCF recommended accuracy comparison among the present invention;

Fig. 4 is the size of candidate item set in the present invention;

Fig. 5 is the proportion of items of interest to users in the present invention;

Fig. 6 is a partial user-rating table in the present invention;

Fig. 7 is a partial user information table in the present invention;

Fig. 8 is a table of items in the present invention;

Fig. 9 is the comparison of traditional and improved Pearson's correlation coefficient among the present invention;

Fig. 10 is the comparison of traditional and improved vector cosine similarity among the present invention;

Fig. 11 is the comparison of traditional and improved modified vector cosine similarity in the present invention;

Fig. 12 is the traditional and improved Pearson correlation comparison in the present invention;

Fig. 13 is the framework of the convolutional neural network model in the present invention.

detailed description

Embodiment, a neural network recommendation system collaborative filtering algorithm, comprises the following steps:

Include the following steps:

S1, as shown in Figure 13, first use the convolutional neural network to establish a convolutional neural network model for some of the user's previous behaviors and the attributes of the project itself, so as to provide a more accurate recommendation to the user;

The data set mainly contains two static objects of user and project, wherein user-related attributes include user ID, gender, occupation and age, and project attributes include project ID, project type and project name. In this embodiment, the project is For movies, movie attributes include movie ID, movie type and movie name. The movie year is a new feature attribute extracted from the movie name. Of course, the item can also be books and music. The first is to digitally encode it, convert discrete variables into continuous value vectors, and use characters to represent category fields in the data, such as item category attributes. The usual processing method is to use one-hot encoding, and convert it to converted to a value;

S2. Extract the features of users and their item attributes, and fit the scores through full connection;

First, convert the attribute values into numbers for representation, and use the converted number as the index N of the embedding matrix, and then assign a fixed-size potential factor to each index, which is the size of the vector. Usually, it is selected. Then the size of the matrix dimension is (N}32), and the user feature vector and item feature vector can be obtained after the data-related attributes are operated, and the score value can also be fitted at this time;

S3. Introduce the idea of association rules, generate an item association matrix according to the item association relationship, and use the item association matrix to generate candidate item sets;

The degree of relevance indicates the strength of the relationship between items. This method defines the degree of relevance as the possibility that the user _browses another item after browsing one item. The number of users co-viewed by i and project j is calculated by the ratio of the number of users browsed by project i, as follows:

Among them, N _i , N _j represent the number of users who rated item i and item j respectively, and the item correlation matrix G=nxn can be established according to the item correlation degree, where r _ij =(i,j=1,2...n) is the correlation degree between project j and project i, the main diagonal elements of the matrix are all 0, and the correlation degree between projects is generally different, and the size of the correlation degree indicates the strength of the relationship between projects, and the project correlation matrix is generally is asymmetric, that is, r _ij ≠ r _ji ;

The correlation degree is introduced into the selection process of candidate item sets, and the item correlation matrix is used instead of the similarity matrix to generate candidate item sets;

S4. Aiming at the problem of poor similarity accuracy in the case of sparse data, a modified Pearson correlation coefficient is proposed, a user-rating matrix is established, and the matrix is optimized using the forgetting function to calculate the score between the target user and other users. Similarity, according to the order from high score to low score, ranked by the similarity of the score, select the top users as the nearest neighbors of the target user, and calculate the score prediction of the target user for the unrated item according to the score information of the neighbors value, and then select the N items with the highest predicted value to recommend to the user;

When selecting co-rating data, the number of users who co-rated should also be considered. If the proportion of co-rating users among the two item scoring users is higher, the two items will be more similar to a certain extent. For example, item i ₁ is compared with 10 item i ₂ is rated by 12 users, item i ₃ is rated by 50 users, if items i ₁ and i ₂ , and items i ₁ and i ₃ have 8 common rating users, then item i ₁ is considered It is more similar to i _2. Similarly, the project i ₁ and i ₂ have 8 co-rated users, which means that the two items are more similar than the case of 5 co-rated users. Therefore, a modified Pearson correlation coefficient is proposed, and the formula is as follows :

分别为项目i和项目j的平均评分；Among them, U _i , U _j , U represent the user sets who rate item i, item j and both item i and item j respectively; |U _i |, |U _j |, |U| represent sets U _i , U _j , the number of elements in U, r _ui , r _uj are user u’s ratings on item i and item j respectively,

are the average ratings of item i and item j respectively;

The algorithm is as follows:

(1) environment setting

Assuming that user u is a new user to the current system, the historical behavior information is empty, but he already has historical behavior information in other systems (such as social networking sites, business sites, video sites) (can be authorized or the current system has cross-site cookie);

(2) Construction of user relationship network

Build a relationship network between users based on their historical information in other systems, and divide them into communities, and use a similarity calculation formula when building a relationship network based on user purchases or rating information. The formula is:

和

分别表示r_u和r_v中非零元素的标准差[20]；Among them, r _u and r _v represent the rating vectors of users u and v respectively, covar(r _u , r _v ) represents the covariance of r _u and r _v ,

and

Respectively represent the standard deviation of the non-zero elements in r _u and r _v [20];

It is worth noting that a similarity of 0 means that no edge should be added between the corresponding two users, and the overlapping elements between the two vectors should be no less than three, otherwise the similarity between the corresponding two users will be considered to be 0;

If a user has friends in a social networking site, that is, has a circle of friends, then it can be directly divided into communities;

(3) Community division

The community division adopts the method of principal module maximization (PMM). The basic principle is to maximize the modularity index Q. Q measures the community division of the real network by measuring the difference between the random network with the same degree distribution and the interconnection within the community. The strength of , is defined as:

其中

in

S is the affiliation matrix of node u and community C, if node u belongs to community C, then S _uC is 1, otherwise it is 0, B is the module matrix, A is the adjacency matrix of node relationship, d is the degree sequence of all nodes, m is the total number of edges in the node relationship network, and Q less than 0 means that the division effect is very poor;

Considering the multi-dimensional relationship between users, including the user's rating of the project, the friendship between users, etc., the PMM divides the community according to the multi-dimensional relationship, and the elements in the division result S represent the user u. How high is the degree of belonging to community C;

(4) Changes in the strategy of selecting similar neighbors

Previously, among all users, the k neighbors with the highest similarity were selected, but this method only selects users in the same community as the target user u, and the target user u’s predicted score for item j:

瓦为用户u的平均评分,a为常数，C为用户v所在社区，r_V,j为用户v对项目j的评分，

氏为用户v的平均评分。in,

W is the average rating of user u, a is a constant, C is the community of user v, r _V,j is the rating of user v on item j,

is the average rating of user v.

Simulation:

The data set in this experiment is from the imhonet website, which is a multi-faceted social networking site. The data set used in this method includes the user's online connections and their ratings for books and movies. Among them, the user has at least one Contact, 48,000 users have 900,000 movie ratings for 50,000 movies, 13,000 users have 1.2 million book reviews for 14,000 books, and the records of each rating in the data set include user ID, item ID and the ratings between them Value (0-10, where 0 means not rated). From this data set, users who had only 20 ratings on books and movies were selected. The resulting experimental data set included 2,363,394 ratings on 12,621 movies from 6,089 users and 17,907 books from these users. 1,138,401 ratings.

20% of the users in the data set are used as the test set, and the remaining 80% are used as the training set. In order to simulate the cold start problem, the book reviews of all users in the test set are deleted, so as to simulate that all users in the book test set are new users and predict ratings for books for them.

This experiment uses the comparison method. The comparison experiment is a traditional collaborative filtering algorithm, and N nearest neighbors are selected for prediction. The calculation method for calculating the similarity between users here uses the Pcarson correlation coefficient. Compare by modifying the number of nearest neighbors. The predicted neighbors are incremented from 10 to 100, and the increment is 10. The evaluation standard still adopts MAE, and the accuracy of the prediction is measured by calculating the deviation between the predicted user rating and the actual rating. The smaller the MAE, the higher the quality of the recommendation algorithm.

The experimental results are shown in Figure 1. The square dotted line represents the collaborative filtering score prediction based on traditional Pearson correlation, and the rectangular dotted line represents the score prediction based on user community division.

Experimental results

The experimental results show that the score prediction based on user community division is better than the traditional collaborative filtering, especially when the number of nearest neighbors is greater than 40, the traditional collaborative filtering recommendation quality tends to be stable, while the community-based prediction score can still improve the recommendation quality, In the data set, the cold start problem in the recommendation system is simulated by deleting the user's rating for the book. The experiment proves that the prediction accuracy of the former is better than that of the latter under the condition that the number of neighbors increases continuously, that is to say, the The algorithm can overcome the cold start problem and improve the recommendation quality.

As shown in Figure 2 and Figure 3, the comparison between the item-based collaborative filtering algorithm (IBCF) and the content-based collaborative filtering algorithm (UBCF) on the MovicLens dataset, MAE and Precision. In the experiment, the data set is divided into 80%-20% training set and test set, and the 5-fold crossover method is used for experiments. At the same time, this experiment uses cosine similarity to calculate the similarity between users and items. In the experiment of MAE comparison, the number of neighbors knear increased from 5 to 50 with an interval of 5; in the experiment of Precision comparison, the number of recommendations topN increased from 2 to 20 with an interval of 2. At the same time, when generating candidate itemsets, For each rated item or user, take the 5 most similar items or users (the number of candidate neighbors is 5), for the user-based collaborative filtering algorithm, the neighbors take 30, and for the item-based collaborative filtering algorithm, the neighbors take 15 (the reason why the values are different , because the two algorithms can already obtain stable top measurement accuracy under the above values). As can be seen from the following two figures, when the data is quite sparse, the collaborative filtering algorithm based on the base and the item has better prediction accuracy (MAE value) than the collaborative filtering algorithm based on the user, and it does alleviate the problem to a certain extent. The impact of data sparsity, but its recommendation accuracy (Precision value) is not as good as the user-based collaborative filtering algorithm. In fact, the MAE value of the item-based collaborative filtering algorithm is better than that of the user-based collaborative filtering algorithm in most cases, not only in the case of sparse data. In addition, at this time, the size of the average candidate itemsets for each user of the two algorithms is different, and the average user candidate itemsets of the user-based collaborative filtering algorithm is larger, and the average user candidate itemsets can be reduced by reducing the number of candidate neighbors, and If the average user candidate item sets of the two algorithms become smaller; and if the average user candidate item sets of the two algorithms are similar in size (UBCF takes 2, IBCF takes 5), the recommendation accuracy of the user-based collaborative filtering algorithm should be far superior to the latter.

This method analyzes the core process of the traditional item-based collaborative filtering algorithm: first, generate candidate itemsets; then predict active users' ratings on the items in the candidate itemsets. The size and accuracy of the candidate set will affect the final recommendation accuracy. The item-based collaborative filtering algorithm needs to operate on all rated items i∈I _u of user u, and read k nearest neighbor sets N _i ={i ₁ ,i ₂ ,…,i _k } for each rated item , merge all N _i and delete the scored items in I _u from them, and get the candidate item set C.

Figure 4 shows the change curve of the size of the user’s average candidate item set with the k value of the number of neighbors in the candidate set obtained according to the project-based collaborative filtering algorithm on the MovieLens data set. In the experiment, the cosine similarity was used to calculate the relationship between items. similarity. On the way, the horizontal axis is the k value of the number of neighbors in the candidate set, increasing from 2 to 20, with an interval of 2. The vertical axis is the average size of the user's candidate item set, that is, the number of candidate items in the candidate item set. It can be seen from the figure that when k=2, the size of the average candidate item set has exceeded 50, and as the value of k increases, the size of the candidate item set also increases sharply. When k=20, the average size of each user The size of the candidate set has exceeded 300.

Figure 5 shows the proportion of the items that the user is interested in in the candidate item set in the total candidate set under the same experimental conditions. The horizontal axis in the figure is the k value of the number of neighbors in the candidate set, and the vertical axis is the percentage of the items that the user is interested in. It can be seen from the figure that when k=2, there is a maximum value of about 12.2%. With the increase of the k value of the nearest neighbors in the similar selection set, the proportion of the user's interested item in the candidate item set gradually decreases. When k=20 , the proportion has dropped to around 4.8%. Obviously, based on the project-based collaborative filtering algorithm, the proportion of the project that the user is interested in in the candidate project set is very small. It can be seen that item-based collaborative filtering does not consider the differences between users, which can easily cause items that users are not interested in to be included in the candidate item set, so the recommendation accuracy is relatively poor, and the candidate item set is too large, which will lead to the next step The time to predict scores for items in the candidate set increases, directly affecting the scalability of the system.

The following two tables show the performance of several similarity calculation methods on the MAE value using the MovieLens data set and the Jester data set respectively. The table lists the optimal value, worst value, and The average value and the value of the number of neighbors when the optimal value is obtained. The previous table shows the experimental results on the MovieLens dataset. In the experiment, the number of neighbors knear was increased from 5 to 50 with an interval of 5. The latter table shows the experimental results of the Jester dataset. In the experiment, the number of neighbors knear was increased from 2 to 30 with an interval of 2. It can be seen from the previous table that the performance of cosine similarity is very good no matter from the optimal value or the average value, but the performance of Pearson correlation coefficient is not so good when the data is sparse, because the common scoring between items The number of users is rare, and it is easy to cause the calculation results to be either too large or too small. From the following table, it can be seen that the effect of the Pearson correlation coefficient has increased significantly, which is better than other traditional algorithms. This is because as the sparsity decreases, the Pearson correlation coefficient gets enough common scoring users. In addition, it can also be seen that the proposed modified Pearson correlation coefficient shows good performance no matter in the case of sparse data or in the case of improvement, especially in the case of sparse data, its effect is far more than that of Pearson correlation coefficient. As the sparsity gradually decreases, the gap between the effects of the modified Pearson correlation coefficient and the Pearson correlation coefficient gradually decreases. At this time, the effects of both are very good. At the same time, it can also be seen from the table that when the data is sparse, the Pearson correlation coefficient converges slowly. When the number of neighbors k=50, the optimal value is obtained. As the sparsity decreases, the convergence is significantly improved. . The convergence of the modified Pearson correlation coefficient is always good when the data is sparse and not sparse, and when k=15 and k=10 respectively, the optimal value has been achieved.

Table 1 Comparison of various similarity calculation methods-1

Table 2 Comparison of various similarity calculation methods-2

This experiment uses the comparison method to verify the effectiveness of the proposed improved algorithm by comparing it with several other recommendation algorithms.

The experimental data of this experiment is a 100K data set collected on the MovieLens website, including 1682 movies and 943 users and 100,000 rating data of these movies by these users. The scores scored by users are divided into 1,2 , 3, 4, 5 such 5 grades, each user has rated at least 20 movies. In the data set, there are m users U={u ₁ ,u ₂ ,u ₃ ,...u _m }, n movies M={m ₁ ,m ₂ ,m ₃ ...... m _n }, then a mxn matrix R can be expressed as the user's rating of the movie. r _ui is user u's rating for movie i, if there is no rating, then r _ui =0.

Firstly, the similarity between users is evaluated, which is calculated by Pearson correlation, vector cosine and modified vector cosine. Then use the improved algorithm proposed in this experiment to calculate the similarity between users. The similarity is calculated based on the data in the training set, and then the user's rating is predicted based on the data in the prediction set. This method has done two experiments. In the first one, when calculating the neighbors of the target user, the Top-N recommendation is selected. In the experiment, in order to verify the feasibility and effectiveness of the proposed algorithm, different numbers of neighbors are selected. The second is to change the similarity threshold one by one, adding a closed value of 0.1 each time, and then verify the accuracy of the prediction, and gradually increase the threshold from 0.1 to 0.9.

Recommended System Performance Evaluation Criteria

The mean absolute error ((MeanAbsluteErrorMAE) is used to measure the accuracy of the prediction results in the recommendation system. The accuracy of the prediction is measured by comparing the deviation between the predicted user rating and the actual rating. The smaller the MAE, the better the recommendation algorithm The higher the quality of recommendation. Let the user's predicted score set be {t ₁ ,t ₂ ,...t _n }, and the corresponding actual score set be {p ₁ ,p ₂ .....p _n }, Then the MAE is:

Recall (Recall) and precision (Precise) are two metrics widely used in the field of statistical classification and information retrieval, and they are applied to the recommendation system to evaluate the quality of the results. The recall rate is the probability that a recommendation list generated by the system contains resources rated by users in the recommendation system. It can reflect the completeness of user interests. Let the recommendation list generated by the system for user u _i be L _i , and the set of resources that user u _i is interested in in the test set is T _i, whose calculation formula is:

Accuracy is also called accuracy rate. As the name suggests, it is the ratio of how many resources in the generated recommendation list are the resources that users are really interested in, that is, the accuracy of predicting user interests:

Experimental data preparation and design

Data preparation, as shown in Figures 6 and 7, the experimental data of this experiment is collected from the MovieLens website, with a size of 1,000,000 data sets, including 943 users and 100,000 records of these users on 1682 movies. Rating data, the scores scored by users are divided into 5 levels of 1, 2, 3, 4, and 5, and each user has rated at least 20 movies. The experiment selects 80% of the collected data set as the training set (base set), and the remaining 20% as the prediction set (test set).

As shown in FIG. 8 , the user-rating table includes four fields, which are user ID (userid), item ID (itemid), rating (rating), and timestamp (timestamp). The time is calculated from January 1, 1970, UTC time in the Unix system.

experimental design

In this experiment, a total of two sets of experiments were done. In the first set of experiments, Pearson and vector cosine, modified vector cosine similarity calculation method 1 was compared with the similarity calculation method based on the forgetting curve, and verified by changing the number of neighbors The effectiveness of the proposed algorithm is measured by MAE to measure the final experimental results; in the second set of experiments, Pearson similarity and similarity calculation based on forgetting curve are used to verify the effectiveness of the proposed algorithm by changing the similarity threshold . As mentioned before, the accuracy of the predicted rating is measured by calculating the deviation between the predicted user rating and the actual rating, and the lower the MAE value, the better the recommendation quality.

experiment one

In this set of experiments, comparisons are made by modifying the number of nearest neighbors. The predicted neighbors are incremented from 10 to 100, and the increment is 10. The result is shown in Figure 9, the experimental results of the traditional Pearson similarity and the improved similarity algorithm (where the red silk solid line is the traditional algorithm, and the blue dotted line is the improved algorithm, the following are the same). It can be seen from Figure 8 that when the number of neighbors is equal to 30, the MAE values of the two algorithms are roughly equal, but when the number of neighbors continues to increase to 100, the MAE value of the improved similarity algorithm is smaller, indicating that its effect is better .

Figure 10 is the experimental results of the traditional vector cosine and vector cosine similarity algorithm based on forgetting curve. It is not difficult to see from the figure that with the continuous increase in the number of neighbors, the MAE value of the improved algorithm is smaller than that of the traditional vector cosine, indicating that the prediction is also closer to accuracy.

Figure 11 is the experimental results of the traditional modified vector cosine and the improved similarity algorithm based on forgetting curve. It can be seen from the figure that in the process of increasing the number of neighbors from 10 to 100, the MAE value of the improved algorithm is always smaller than that of the traditional modified vector cosine, which shows that when using the modified vector cosine to calculate the similarity, the algorithm based on the forgetting curve can make forecasts more accurate.

The effectiveness of the proposed algorithm has been verified by the above three experiments. When calculating the similarity between users, the change of users' ratings of resources due to natural forgetting should be considered. User interests will change over time. changed over time. By simulating the forgetting process of human beings, the accuracy of prediction is improved.

Experiment 2

Using Pearson similarity and similarity calculation based on forgetting curve, the effectiveness of the proposed algorithm is verified by changing the similarity threshold. As mentioned earlier, the deviation between the predicted user rating and the actual rating is calculated to measure the accuracy of the predicted rating. The lower the MAE value, the better the recommendation quality. The threshold value ranges from 0.1 to 0.9, and the increment is 0.1. By setting the threshold value, the neighbor users used are all similarities greater than this closed value, and then the experiment is carried out separately. The results are shown in Figure 12. It can be seen from the figure that when the threshold value is greater than 0.4, the MAE value of the traditional Pearson correlation coefficient is significantly higher than the MAE value of the improved algorithm based on the forgetting curve, and when it is less than 0.4, the traditional method is better. In other words, when the threshold is greater than 0.4, the improved algorithm predicts more accurately.

To sum up, according to the results of two groups of experiments, generally speaking, the calculation method based on the similarity of the forgetting curve is better than the traditional algorithm. Then in the recommendation system, combined with the laws of nature, by using the law of forgetting expressed by the Einstein's forgetting curve to attenuate the user's score, the accuracy of the system's prediction can be significantly improved. This also shows that in the recommendation system, human cognitive laws can play a very important role.

Claims (1)

1. A collaborative filtering algorithm of a neural network recommendation system is characterized in that: the method comprises the following steps:

s1, firstly, establishing a convolutional neural network model for some previous behaviors of a user and the attributes of items by adopting a convolutional neural network, so that more accurate recommendation is provided for the user;

the data set mainly comprises two static objects, namely a user and a project, wherein the related attributes of the user comprise a user ID, a gender, an occupation and an age, and the project attributes comprise a project ID, a project type and a project name, firstly, the data are coded digitally, discrete variables are converted into continuous value vectors, category fields in the data, such as the project category attributes, are represented by characters, and the common processing mode is to convert the category fields into numerical values by adopting one-hot coding;

s2, extracting characteristics of the user and the project attributes of the user, and fitting the scores in a full-connection mode;

firstly, converting attribute values into numbers for representation, taking the converted numbers as indexes N of an embedded matrix, then allocating potential factors with fixed sizes to each index, namely the size of a vector, usually selecting a criminal, wherein the size of the matrix dimension is (N) 32, obtaining user characteristic vectors and project characteristic vectors after data related attributes are operated, and fitting score values;

s3, introducing an association rule idea, generating a project association matrix according to the project association relation, and generating a candidate project set by using the project association matrix;

the relevance represents the strength of the relation between the items, the relevance is defined as the possibility that a user browses one item and then another item, the relevance is represented by r, and r _ij The association degree of the item j to the item i is represented, and the ratio of the co-browsing user ratio of the item i and the item j to the browsing user ratio of the item i is defined to be calculated as follows:

wherein, N _i ,N _j Respectively representing the number of users scoring the item i and the item j, and establishing an item association matrix G = nxn according to the item association degree, wherein r _ij N, the main diagonal elements of the matrix are all 0, the degree of association between items is generally different, the magnitude of the degree of association indicates the strength of the relationship between items, and the item association matrix is generally asymmetric, i.e., r is _ij ≠r _ji ；

The relevance is introduced into the selection process of the candidate item set, and the item relevance matrix is used for replacing a similarity matrix to generate the candidate item set;

s4, aiming at the problem of poor similarity accuracy under the condition of sparse data, providing a modified Pearson correlation coefficient, establishing a user-scoring matrix, optimizing the matrix by using a forgetting function, calculating scoring similarity between a target user and other users, ranking according to the sequence from high score to low score and the scoring similarity, selecting a plurality of users arranged in the front as nearest neighbors of the target user, calculating a scoring predicted value of the target user for an unscored item according to the scoring information of the neighbors, and then selecting N items with the ranking predicted values being close to the front to recommend the users;

when selecting the common scoring data, the number of users scoring together is also considered, and if the proportion of the users scoring together is higher in the two items of scoring users, the two items are similar to each other to a certain extent, for example, the item i ₁ Is scored by 10 users, item i ₂ Is scored by 12 users, item i ₃ Is scored by 50 users if item i ₁ And i ₂ And item i ₁ And i ₃ All have 8 common scoring users, then consider project i ₁ And i ₂ More similarly, item i ₁ And i ₂ There are 8 users who score together and represent two projects more similar than the case of 5 users who score together, for this reason, a modified Pearson correlation coefficient is proposed, the formula is as follows:

wherein, U _i ,U _j U represents the set of users scoring project i, project j, and both project i and project j, respectively, | U _i |,|U _j Respectively, representing the set U _i ,U _j Number of elements in U, r _ui ,r _uj Item i and item j are scored separately for user u,

average scores for item i and item j, respectively;

the algorithm is as follows:

(1) Environmental settings

Let user u be a new user to the current system, and the historical behavior information is null, but he already has the historical behavior information in other systems (such as social network sites, business network sites, video network sites) (either through authorization or the current system has a cross-site cookie);

(2) Construction of a user relationship network

The method comprises the following steps of constructing a relationship network among users according to historical information of the users in other systems, carrying out community division on the users, and using a similarity calculation formula when constructing the relationship network according to purchasing or grading information of the users, wherein the formula is as follows:

wherein r is _u And r _v Score vectors, covar (r), representing users u and v, respectively _u ,r _v ) Is represented by r _u And r _v The covariance of (a) is determined,

and

respectively represent r _u And r _v Standard deviation of medium non-zero elements [ 20%]；

It should be noted that, a similarity of 0 indicates that no connecting edge should be added between the two corresponding users, and the number of the overlapped elements between the two vectors should be not less than three, otherwise, the similarity between the two corresponding users is considered as 0;

if a user has friends in a certain social network site, namely has a friend circle, the user can directly perform community division on the friends;

(3) Community partitioning

The community division adopts a method of maximizing a master module (PMM), the basic principle is to maximize a modularity index Q, and the Q measures the intensity of community division on a real network by measuring the difference between a random network with same degree distribution and the interconnection in a community, and is defined as follows:

wherein

S is an attribution relationship matrix of the node u and the community C, if the node u belongs to the community C, S _uC If the number is 1, otherwise, the number is 0, B is a module matrix, A is an adjacent matrix of the node relationship, d is a degree sequence of all nodes, m is the total number of edges in the node relationship network, and Q is less than 0, which indicates that the partitioning effect is very poor;

because the multi-dimensional relationship among the users is considered due to the fact that the multi-dimensional relationship among the users comprises the scores of the users for the items, the friend relationships among the users and the like, the PMM performs community division according to the multi-dimensional relationship, and the element in the division result S shows how high the affiliation degree of the user u to the community C is;

(4) Change of selection similar neighbor policy

Selecting k neighbors with the maximum similarity from all users, selecting only the users in the same community with the target user u according to the method, and scoring the target user u for the prediction of the item j:

wherein,

w is the average score of user u, a is a constant, C is the community in which user v is located, r _V,j For the user v to score the item j,

and "is the average score of user v.

Images