CN108363804A - Local model weighted fusion Top-N movie recommendation method based on user clustering - Google Patents

Abstract

A local model weighted fusion Top-N movie recommendation method based on user clustering comprises the steps of 1, preprocessing data, cleaning data of an inactive user and a movie with low popularity, constructing a user movie label document, converting explicit grading information into implicit feedback information, constructing a user-movie implicit feedback matrix A, 2, clustering users, training a L DA topic model by utilizing movie label information to obtain a user characteristic vector, realizing user clustering by using a spectral clustering algorithm, 3, determining a local recommendation model and performing global recommendation model training, 4, performing a model weighted fusion recommendation stage, and 5, proving the effectiveness of the model by leave-one-out cross verification.

Description

Local model weighted fusion Top-N movie recommendation method based on user clustering

technical field

The invention relates to a method for recommending movies on the network.

Background technique

With the rapid development of information technology and social networks, the data generated by the Internet has recently skyrocketed, and the era of big data is coming. With the increase of the amount of data, it becomes more and more difficult for people to find the information they really want from the massive data. At this time, the recommendation system can exert its maximum application value. Based on user profiles, item information, and user historical behavior data, the recommendation algorithm can accurately predict user preferences and recommend items that may be of interest to users in a personalized manner, greatly reducing the cost for users to discover target information.

Recommendation algorithms can be divided into content-based recommendation and collaborative filtering recommendation. Modern recommendation systems mainly have two tasks, one is score prediction, and the other is Top-N recommendation, which is most widely used in real business scenarios. The Top-N recommendation algorithm allows users to choose what they are interested in by recommending a ranked and n-sized item list to users. Top-N recommendation models are mainly divided into two types, namely neighborhood-based collaborative filtering and model-based collaborative filtering. The former can be subdivided into user-based neighborhood model (UserKNN) and item-based neighborhood model (ItemKNN), and the latter is represented by latent factor model.

As the saying goes, "Things of a feather flock together and people are divided into groups." Different user groups often form their own unique behavior patterns, which makes the similarity of two identical items change in different groups of people. However, single recommendation algorithm models often fail to capture these local similarity differences. They believe that the similarity of two identical items in any scene is the same. These models cannot accurately capture the real preferences of users, which reduces the personalized recommendation the quality of. The recommendation algorithm that improves the overall recommendation effect by training multiple local recommendation models and then fusing the local models can solve the above problems to a certain extent, but these algorithms often do not make full use of the data provided by the recommendation scene, and the data used is relatively single. The recommended effect is also average.

Contents of the invention

In order to overcome the problem that a single model in the prior art cannot accurately capture user preferences and the multi-model fusion algorithm uses a single training data, the present invention provides a new local model weighted fusion movie recommendation algorithm based on user clustering to realize the Top- N personalized recommendation.

The present invention utilizes the text content information of movies to calculate semantic level user feature vectors through the LDA topic model, and based on this, realizes user clustering through spectral clustering algorithm and constructs local groups of people. The present invention further uses the user's rating information on movies, constructs a local recommendation model and a global recommendation model through a sparse linear model, and realizes the final Top-N personalized recommendation of movies through the linear weighted fusion of the local model and the global model.

The local model weighted fusion Top-N movie recommendation method based on user clustering, the overall process is shown in Figure 1, including the following steps:

Step 1: Data preprocessing stage. Perform data cleaning on some inactive users and movies with low popularity; construct user movie label documents; convert explicit rating information into implicit feedback information, and construct user-movie implicit feedback matrix A;

1.1 Carry out data cleaning work on the original data set, remove users who watched less than 20 movies, and remove movies rated less than 20 times to obtain a new training data set;

1.2 Count all the tags that users put on movies in the new data set to generate a tag dictionary, and use the tags of all the movies that the user has watched to represent the current user. TF-IDF values in the corpus. The calculation formulas of term frequency TF, inverse document frequency IDF and term frequency-inverse document frequency TF-IDF are shown in formula (1)(2)(3);

TFIDF _i,j =TF _i,j ×IDF _i (3)

Where TF _i,j represents the word frequency of word t _i in document d _j , ni _,j represents the number of times word t _i appears in document d _j , ∑ _k n _k,j represents the number of occurrences of all words in document d _j Sum. IDF _i represents the inverse document frequency of term t _i , |D| represents the total number of documents in the corpus, and |{j:t _i ∈ d _j }| represents the number of documents containing term t _i . TFIDF _{i, j} represents the word frequency inverse document frequency of word t _i in document d _j ;

1.3 Convert explicit scoring information such as 1-5 points into implicit feedback information represented by 0-1. If the current user has rated the current movie too much, it will be recorded as 1, and the movie that has not rated too much is the movie to be recommended. is 0, and an n×m user-movie implicit feedback matrix is obtained, the number of users is n, and the number of movies is m;

Step 2: User clustering stage. Using movie label information, user feature vectors are obtained through LDA topic model training, and user clustering is realized by spectral clustering algorithm;

2.1 The LDA topic model is a document-topic-word three-layer Bayesian network. Given a corpus, the model can analyze the topic distribution of each document in the corpus and the word distribution of each topic. Its joint probability is shown in formula (4);

θ represents the topic distribution of a document, z represents a topic, w represents a document, α represents the Dirichlet prior parameter of the multinomial distribution of topics under each document, and β represents the Dirichlet of the multinomial distribution of words under each topic A priori parameter, N represents the number of documents in the corpus, z _n represents the topic of the nth word in a document, w _n represents the nth word in a document;

Each movie has multiple tags given to it by users. A movie tag is mapped to a word w _n , a set of tags of all movies a user has watched is mapped to a document w, and a user-preferred Class-specific movie genres are mapped to a topic z. If there are n users in the data set, a corpus containing n documents and a dictionary can be generated. Each document in the corpus is represented by a vector of the length of the dictionary, and each value in the vector is the corresponding label in the dictionary. TF-IDF values in documents and corpora;

In order to be able to distinguish more unique user groups, the more differences between the different topics, the better. In order to determine the optimal number of topics, train multiple LDA models by setting multiple topic numbers, calculate the average similarity between the topic vectors obtained by each LDA model training, and take the topic number corresponding to the model with the smallest average similarity of topic vectors The optimal number of topics to use as a model. Through LDA model training, the topic distribution θ of each document is obtained, and it is used to represent the feature vector of each user;

2.2 Utilize the user feature vectors (n in total) obtained by the above steps, and use the spectral clustering algorithm to realize the clustering of users;

Before clustering, it is first necessary to determine the number of clusters. Because each dimension of each user vector obtained through training represents the membership degree of the user belonging to the corresponding topic, so in order to determine the importance of each topic in the current user group, all user feature vectors are accumulated by dimension and then averaged , get a topic strength vector representing the whole, and determine the optimal number of clusters by observing the value distribution of the topic strength vector. For example, in an LDA training process with a topic number of 10, a 10-dimensional topic strength vector is obtained by the above method, and the visualization is shown in Figure 2 (the vertical axis represents the topic strength, and the horizontal axis is the topic). Through observation, we can see Topics 2, 9, 3, 8, and 6 are the most intense in the current data set, indicating that the most people like to watch these types of movies. Therefore, it is more appropriate to use the spectral clustering algorithm to cluster users into 5 categories in the current situation. The specific steps of the spectral clustering algorithm are as follows:

(1) Calculate n×n similarity matrix W and degree matrix D;

(2) Calculate the Laplacian matrix L=D-W;

(3) Calculate the first k eigenvectors t ₁ , t ₂ ,...,t _k of L;

(4) Composing k column vectors t ₁ , t ₂ ,...,t _k into a matrix T, T∈R ^n×k ;

(5) For i=1,...,n, let y _i ∈ R ^k be the i-th row vector of T;

(6) Use the K-Means algorithm to cluster users (y _i ) _{i=1, 2,..., n} into clusters C ₁ , C ₂ ,..., C _k ;

For each user cluster, the user row vectors in the original implicit feedback training matrix A that do not belong to the cluster are set to 0, and each cluster generates a corresponding local implicit feedback training matrix P _u represents the cluster number, and P _u ∈ {1,…,k};

Step 3 determines the local recommendation model and performs global recommendation model training. The loss function of the sparse linear model SLIM is shown in formula (5);

Among them, A represents the original user-movie implicit feedback matrix, α and ρ control the weights of L1 and L2 norms, and a sparse matrix W of movie similarity of size m×m can be obtained by minimizing the loss function. In this model, the L1 norm controls the sparsity of W, and the L2 norm controls the complexity of the model to prevent the model from overfitting. The model uses the random gradient descent method to train each column w _j of the W matrix in parallel to obtain the final W matrix, as shown in formula (6);

Among them, a _j represents the jth column in the matrix A. User i's predicted recommendation for movie j The calculation formula is shown in formula (7);

Use the sparse linear model SLIM as the basic recommendation model to build a global recommendation model and a local recommendation model, use the global implicit feedback training matrix A to train the global movie similarity matrix W, and use the local implicit feedback training matrix Train to get the local movie similarity matrix corresponding to each cluster

Step 4: Model weighted fusion recommendation stage. The calculation formula of local model weighted fusion recommendation degree is shown in formula (8);

in Represents the weighted fusion recommendation degree of movie j for user u, R _u is the set of all movies that have interacted with user u, w _lj is the similarity between movie l and movie j in the global model, is the similarity between movie l and movie j in the local model corresponding to the cluster P _u to which user u belongs, and the parameter g is the weight parameter of the global model. The weight ratio of the global model and the local model in the fusion model is controlled by adjusting the parameter g, and the best recommendation effect of the fusion model is obtained by determining the optimal weight parameter g. The optimal global model weight parameters in the current dataset can be determined experimentally. After determining all the parameters in the model, by calculating the weighted fusion recommendation of all movies for the current user u, sort them from large to small, and delete the movies that have interacted with the current user, and take the top N positions movies recommended to the current user;

Step 5. The recommended method can prove the validity of the model by leave-one-out cross-validation. A movie can be randomly selected from each user's movie rating set and put into the test set, and other movies are used as the training set of the model. Then use the trained model to recommend a Top-N movie list for each user, and observe whether the corresponding movie of the user in the test set appears in the recommended list and its specific position p _i in the list. Finally, two indicators, hit rate (HR) and average ranking hit rate (ARHR), can be used to measure the recommendation quality of the model, where #hits represents the number of recommended hits, and #users represents the total number of users. Their definitions are as in formula (9), (10);

This concludes the recommended method flow steps.

The present invention combines the above technologies and proposes a Top-N movie recommendation algorithm based on user clustering based on local model weighted fusion. In order to solve the problem that the traditional single recommendation model cannot accurately estimate the local differences of items, resulting in the inability to accurately capture user preferences, a global recommendation model and a local recommendation model based on user clustering are proposed to be trained separately, through linear weighted fusion between the models. Realize the Top-N recommendation of movies. In addition, in order to make full use of the data in the movie recommendation scene and improve the quality of recommendation from multiple dimensions, the present invention uses movie label information to realize the calculation of the feature vector of the user at the semantic level through the LDA topic model, and realizes the user's semantic level. The division of hierarchical groups.

The advantages of the present invention are: (1) The algorithm idea is novel. Use the sparse linear model as the basic recommendation model, train the global recommendation model and the local recommendation model based on user clustering respectively, and finally generate the final fusion model through linear weighted fusion. This idea can deal with the difference in similarity between different groups of people. , which effectively overcomes the problem that a single model cannot accurately capture user preferences. (2) Improve recommendation quality in multiple dimensions. In addition to using traditional scoring data to train the recommendation model, in the user clustering stage, the present invention introduces movie label data, uses the LDA topic model to analyze the topic attributes of the crowd at the semantic level, obtains user feature vectors, and uses the spectral clustering algorithm to realize the crowd Clustering further improves the quality of recommendations. (3) The algorithm implementation is simple and fast. In the training phase of the local model and the global model, since each model is independent of each other, each column of the similarity matrix of each model is also independent of each other, so the method of parallel training can be adopted, which greatly reduces the training time of the model and improves the The efficiency of model training. (4) The recommendation quality is better. The local model weighted fusion recommendation algorithm proposed by the present invention is an effective combination of content recommendation, neighborhood-based collaborative filtering, and model-based collaborative filtering. It makes full use of the advantages of each recommendation algorithm and makes up for the shortcomings of each other. , compared with a single recommendation algorithm, the recommendation quality has been greatly improved.

Description of drawings

Fig. 1 is the general flowchart of the inventive method;

Figure 2 is a graph of the subject intensity distribution of the method of the present invention.

Detailed ways

Referring to the overall flow chart of the technical solution in Figure 1, the present invention has four stages, namely: data preprocessing stage, user clustering stage, global recommendation model and local recommendation model training stage, and recommendation model linear weighted fusion stage. The data preprocessing stage is to clean the data set, remove some inactive users and unpopular movies, construct a corpus for LDA topic model training and a user movie implicit feedback training matrix for sparse linear model training; user clustering stage , use the user corpus obtained in the first stage to train the LDA topic model to obtain the user feature vector, and realize the clustering of users through the spectral clustering algorithm, and each cluster generates a local implicit feedback training matrix; the global recommendation model and the local In the recommendation model training stage, the original implicit feedback matrix and the local implicit feedback matrix are used to obtain the global model and local model through sparse linear model training respectively; in the model linear weighted fusion recommendation stage, the global model and local model obtained in the previous step are linearly weighted way to get the final recommendation model.

The input of the present invention is the rating data of the user watching movies and the label data of the movie, and the output is the Top-N personalized movie recommendation list for the user.

Specific steps are as follows:

Step 1: Data preprocessing stage. Perform data cleaning on some inactive users and movies with low popularity; construct user movie label documents; convert explicit rating information into implicit feedback information, and construct user-movie implicit feedback matrix A;

1.1 Carry out data cleaning work on the original data set, remove users who watched less than 20 movies, and remove movies rated less than 20 times to obtain a new training data set;

1.2 Count all the tags that users put on movies in the new data set to generate a tag dictionary, and use the tags of all the movies that the user has watched to represent the current user. TF-IDF values in the corpus. The calculation formulas of TF (term frequency), IDF (inverse document frequency) and TF-IDF (term frequency-inverse document frequency) are shown in formula (1)(2)(3);

TFIDF _i,j =TF _i,j ×IDF _i (3)

Where TF _i,j represents the word frequency of word t _i in document d _j , ni _,j represents the number of times word t _i appears in document d _j , ∑ _k n _k,j represents the number of occurrences of all words in document d _j Sum. IDF _i represents the inverse document frequency of term t _i , |D| represents the total number of documents in the corpus, and |{j:t _i ∈ d _j }| represents the number of documents containing term t _i . TFIDF _{i, j} represents the word frequency inverse document frequency of word t _i in document d _j ;

1.3 Convert explicit scoring information such as 1-5 points into implicit feedback information represented by 0-1. If the current user has rated the current movie too much, it will be recorded as 1, and the movie that has not rated too much is the movie to be recommended. is 0, and an n×m user-movie implicit feedback matrix is obtained, the number of users is n, and the number of movies is m;

Step 2: User clustering stage. Using movie label information, user feature vectors are obtained through LDA topic model training, and user clustering is realized by spectral clustering algorithm;

2.1 The LDA topic model is a document-topic-word three-layer Bayesian network. Given a corpus, the model can analyze the topic distribution of each document in the corpus and the word distribution of each topic. Its joint probability is shown in formula (4);

θ represents the topic distribution of a document, z represents a topic, w represents a document, α represents the Dirichlet prior parameter of the multinomial distribution of topics under each document, and β represents the Dirichlet of the multinomial distribution of words under each topic A priori parameter, N represents the number of documents in the corpus, z _n represents the topic of the nth word in a document, w _n represents the nth word in a document;

Each movie has multiple tags given to it by users. A movie tag is mapped to a word w _n , a set of tags of all movies a user has watched is mapped to a document w, and a user-preferred Class-specific movie genres are mapped to a topic z. If there are n users in the data set, a corpus containing n documents and a dictionary can be generated. Each document in the corpus is represented by a vector of the length of the dictionary, and each value in the vector is the corresponding label in the dictionary. TF-IDF values in documents and corpora;

In order to be able to distinguish more unique user groups, the more differences between the different topics, the better. In order to determine the optimal number of topics, train multiple LDA models by setting multiple topic numbers, calculate the average similarity between the topic vectors obtained by each LDA model training, and take the topic number corresponding to the model with the smallest average similarity of topic vectors The optimal number of topics to use as a model. Through LDA model training, the topic distribution θ of each document is obtained, and it is used to represent the feature vector of each user;

2.2 Utilize the user feature vectors (n in total) obtained by the above steps, and use the spectral clustering algorithm to realize the clustering of users;

Before clustering, it is first necessary to determine the number of clusters. Because each dimension of each user vector obtained through training represents the membership degree of the user belonging to the corresponding topic, so in order to determine the importance of each topic in the current user group, all user feature vectors are accumulated by dimension and then averaged , get a topic strength vector representing the whole, and determine the optimal number of clusters by observing the value distribution of the topic strength vector. For example, in an LDA training process with a topic number of 10, a 10-dimensional topic strength vector is obtained by the above method, and the visualization is shown in Figure 2 (the vertical axis represents the topic strength, and the horizontal axis is the topic). Through observation, we can see Topics 2, 9, 3, 8, and 6 are the most intense in the current data set, indicating that the most people like to watch these types of movies. Therefore, it is more appropriate to use the spectral clustering algorithm to cluster users into 5 categories in the current situation. The specific steps of the spectral clustering algorithm are as follows:

(1) Calculate n×n similarity matrix W and degree matrix D;

(2) Calculate the Laplacian matrix L=D-W;

(3) Calculate the first k eigenvectors t ₁ , t ₂ ,...,t _k of L;

(4) Composing k column vectors t ₁ , t ₂ ,...,t _k into a matrix T, T∈R ^n×k ;

(5) For i=1,...,n, let y _i ∈ R ^k be the i-th row vector of T;

(6) Use the K-Means algorithm to cluster users (y _i ) _{i=1, 2,..., n} into clusters C ₁ , C ₂ ,..., C _k ;

For each user cluster, the user row vectors in the original implicit feedback training matrix A that do not belong to the cluster are set to 0, and each cluster generates a corresponding local implicit feedback training matrix P _u represents the cluster number, and P _u ∈ {1,…,k};

Step 3 determines the local recommendation model and performs global recommendation model training. The loss function of the sparse linear model SLIM is shown in formula (5);

Among them, A represents the original user-movie implicit feedback matrix, α and ρ control the weights of L1 and L2 norms, and a sparse matrix W of movie similarity of size m×m can be obtained by minimizing the loss function. In this model, the L1 norm controls the sparsity of W, and the L2 norm controls the complexity of the model to prevent the model from overfitting. The model uses the random gradient descent method to train each column w _j of the W matrix in parallel to obtain the final W matrix, as shown in formula (6);

Among them, a _j represents the jth column in the matrix A. User i's predicted recommendation for movie j The calculation formula is shown in formula (7);

Use the sparse linear model SLIM as the basic recommendation model to build a global recommendation model and a local recommendation model, use the global implicit feedback training matrix A to train the global movie similarity matrix W, and use the local implicit feedback training matrix Train to get the local movie similarity matrix corresponding to each cluster

Step 4: Model weighted fusion recommendation stage. The calculation formula of local model weighted fusion recommendation degree is shown in formula (8);

in Represents the weighted fusion recommendation degree of movie j for user u, R _u is the set of all movies that have interacted with user u, w _lj is the similarity between movie l and movie j in the global model, is the similarity between movie l and movie j in the local model corresponding to the cluster P _u to which user u belongs, and the parameter g is the weight parameter of the global model. The weight ratio of the global model and the local model in the fusion model is controlled by adjusting the parameter g, and the best recommendation effect of the fusion model is obtained by determining the optimal weight parameter g. The optimal global model weight parameters in the current dataset can be determined experimentally. After determining all the parameters in the model, by calculating the weighted fusion recommendation of all movies for the current user u, sort them from large to small, and delete the movies that have interacted with the current user, and take the top N positions movies recommended to the current user;

Step 5. The recommended method can prove the validity of the model by leave-one-out cross-validation. A movie can be randomly selected from each user's movie rating set and put into the test set, and other movies are used as the training set of the model. Then use the trained model to recommend a Top-N movie list for each user, and observe whether the corresponding movie of the user in the test set appears in the recommended list and its specific position p _i in the list. Finally, two indicators, hit rate (HR) and average ranking hit rate (ARHR), can be used to measure the recommendation quality of the model, where #hits represents the number of recommended hits, and #users represents the total number of users. Their definitions are as in formula (9), (10);

This concludes the recommended method flow steps.

The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. Equivalent technical means that a person can think of based on the concept of the present invention.

Claims (1)

1. the partial model Weighted Fusion Top-N films based on user clustering recommend method, specifically comprise the following steps：

Step 1：Data prediction；Data cleansing is carried out to the film of inactive user and popularity very little；Structuring user's electricity Shadow lagged document；Explicit score information is converted into implicit feedback information, structuring user's-film implicit feedback matrix A；

1.1 pairs of raw data sets carry out data cleansing work, reject the user that viewing number is less than 20 films, while rejecting and being commented Gradation number is less than 20 films, obtains new training dataset；

All users generate a label dictionary to the label that film is beaten in 1.2 statistics new data sets, user have been seen all The document of the label composition of film indicates active user, and the document of all users forms a corpus, calculate every in document TF-IDF value of a word in corpus；Word frequency TF, the calculation formula of inverse document frequency IDF and term frequency-inverse document frequency TF-IDF As shown in formula (1) (2) (3)；

TFIDF_i,j=TF_i,j×IDF_i (3)

Wherein TF_{I, j}Indicate word t_iIn document d_jIn word frequency, n_{I, j}Indicate word t_iIn document d_jThe number of middle appearance, ∑_kn_{K, j} Indicate document d_jIn all words the sum of occurrence number；IDF_iIndicate word t_iInverse document frequency, | D | indicate corpus in document Sum, | { j：t_i∈d_j| it indicates to include word t_iNumber of documents；TFIDF_{I, j}Indicate document d_jMiddle word t_iThe inverse text of word frequency Shelves frequency；

1.3 such as 1-5 points of explicit score informations, are converted into the implicit feedback information indicated with 0-1, if active user is to working as Preceding film, which is beaten, is excessively then denoted as 1, does not beat excessive film film i.e. to be recommended and is denoted as 0, obtains user-film of a n × m Implicit feedback matrix, number of users n, film number are m；

Step 2：User clustering；Using film label information, train to obtain user characteristics vector by LDA topic models, with spectrum Clustering algorithm realizes user clustering；

2.1 LDA topic models are three layers of Bayesian networks of a document-theme-word, give a corpus, LDA master The theme distribution of every document in the topic model analysis corpus, and the word of each theme are distributed；The connection of the word distribution of theme It closes shown in probability such as formula (4)；

θ indicates that the theme distribution of a document, z indicate that a theme, w indicate that a document, α indicate theme under every document The Dirichlet Study firsts of multinomial distribution, β indicate the Dirichlet Study firsts of the multinomial distribution of word under each theme, N Indicate the number of files in corpus, z_nIndicate the theme of n-th of word in a document, w_nIndicate n-th of word of a document；

Every film has the label that multiple users assign to it, a film label mapping at a word w_n, a use The compound mapping of the label composition for all films that family has been seen is at a document w, the preferred specific film of one kind of user Type mapping is at a theme z；If sharing n user in data set, producible one corpus containing n documents and One dictionary, every document in corpus indicate that each value of vector is that corresponding dictionary is got the bid with the vector of dictionary length Sign the TF-IDF values in the customer documentation and corpus；

In order to distinguish more unique user group, the otherness between different themes is the bigger the better；It is best in order to determine Theme number trains multiple LDA models by the way that multiple numbers of topics are arranged, calculates the theme vector that each LDA model trainings obtain Between average similarity, take the corresponding number of topics of model of theme vector average similarity minimum as model best theme Number；By LDA model trainings, the theme distribution θ of each document is obtained, the feature vector of each user is indicated with it；

The 2.2 n user characteristics vectors obtained using above step, the cluster to user is realized using spectral clustering；

Before cluster number is clustered firstly the need of determining；It should because each dimension for each user vector that training obtains indicates User belongs to the degree of membership of corresponding theme, therefore in order to determine importance of each theme in active user group, institute is useful Family feature vector by dimension do it is cumulative after be averaged again, obtain one and represent whole theme intensity vector, pass through observation theme The Distribution value of intensity vector most preferably clusters number to determine；；Spectral clustering is as follows：

(1) the similarity matrix W and degree matrix D of n × n are calculated；

(2) Laplacian Matrix L=D-W is calculated；

(3) the preceding k feature vector t of L is calculated₁,t₂,…,t_k；

(4) by k column vector t₁,t₂,…,t_kForm matrix T, T ∈ R^n×k；

(5) for i=1 ..., n enables y_i∈R^kIt is the i-th row vector of T；

(6) use K-Means algorithms by user (y_i)_{I=1,2 ..., n}Cluster cluster C₁,C₂,…,C_k；

For each user clustering, user's row vector that the cluster is not belonging in original implicit feedback training matrix A is all set to 0, each cluster generates a corresponding local implicit feedback training matrixP_uIndicate cluster number, and P_u∈ { 1 ..., k }；

Step 3 determines local recommended models and carries out global recommended models training；The loss function of sparse linear model SLIM is such as Shown in formula (5)；

Wherein, A indicates that original user-film implicit feedback matrix, the weight of α and ρ control L1 and L2 norms pass through minimum The loss function can obtain the film similarity sparse matrix W that a size is m × m；L1 Norm Controls W is sparse in the model Degree, the complexity of L2 Norm Control models, prevents model over-fitting；The model passes through stochastic gradient descent method, parallel training W Each row w of matrix_jFinal W matrixes are obtained, as shown in the formula (6)；

Wherein, a_jJth row in representing matrix A；Prediction recommendations of the user i about film jShown in calculation formula such as formula (7)；

Using sparse linear model SLIM as basic recommendation model construction overall situation recommended models and local recommended models, using complete Office implicit feedback training matrix A trains to obtain global film similarity matrix W, utilizes local implicit feedback training matrixInstruction Get the corresponding local film similarity matrix of each cluster

Step 4 model-weight merges the recommendation stage；Shown in partial model Weighted Fusion recommendation calculation formula such as formula (8)；

WhereinIndicate film j for the Weighted Fusion recommendation of user u, R_uFor all films interacted occurred with user u Set, w_ljFor the similarity of film l and film j in world model,For the cluster of film l and film j belonging to user u P_uSimilarity in corresponding partial model, parameter g are the weight parameter of world model；The overall situation is controlled by adjustment parameter g The weight proportion of model and partial model in Fusion Model, by determining that optimal weights parameter g obtains the best of Fusion Model Recommendation effect；Best world model's weight parameter is concentrated in current data to determine by testing；In model is determined After all parameters, the Weighted Fusion recommendation by all films of calculating about active user u presses sequence from big to small, The film for occurring to interact with active user is deleted simultaneously, and the film for coming top N is taken to recommend active user；

Step 5. proves the validity of model by leave one cross validation；It is random from the film of each user scoring set It extracts a film to be put into test set, other films are used as the training set of model；Then it is every with trained model Whether a user recommends the movie listings of a Top-N, observe that film of the correspondence of the user in test set and appear in and push away It recommends in list and it appears in specific location p in list_i；Finally, with hit rate HR and average ranking hit rate ARHR two A index weighs the recommendation quality of model, and wherein #hits indicates that hits, #users is recommended to indicate total number of users, such as formula (9), shown in (10)；