CN104503973A - Recommendation method based on singular value decomposition and classifier combination - Google Patents

Abstract

The invention discloses a fusion recommendation method based on singular value decomposition and classifier. Through data preprocessing, the average score and probability distribution of items are calculated, and the singular value decomposition model is trained by stochastic gradient descent method. Item collection, calculate the user's entropy set on the item category through the calculation method of entropy value, and determine the critical value of item uncertainty, determine whether to use the classifier by comparing the uncertainty and critical value of the predicted item, and use Top-N The method recommends the highest rated N items among all unrated items of the user. The present invention adopts the historical score data of users, analyzes it, and finally generates personalized recommendations; uses the singular value decomposition algorithm to obtain the predicted score of the specified item i, and determines whether to classify by calculating the information entropy of the item on each user. Finally, the final prediction score of the item is obtained through the classifier, thereby improving the accuracy of the recommendation method.

Description

A Fusion Recommendation Method Based on Singular Value Decomposition and Classifier

technical field

The invention belongs to the technical field of recommendation algorithms, in particular to a method for fusion recommendation based on singular value decomposition and classifiers.

Background technique

Whether it is in the field of e-commerce or online video sites such as music and movies, information overload is one of the problems people face in the world today, and the recommendation system is the way to solve this problem. It mines massive data, analyzes users' historical browsing records, and scores historical items to obtain users' preferences in a certain field, thereby realizing personalized recommendations.

Among the commonly used recommendation algorithms, there are currently three categories, namely, content-based recommendation algorithms, collaborative filtering algorithms, and hybrid recommendation algorithms. The content-based recommendation algorithm performs feature extraction, modeling and comparison on the user's historical items to recommend items, but there are complex attributes that are difficult to deal with; collaborative filtering is based on similar users who have the same hobbies or users who give similar items the same score However, it performs poorly in terms of new users, sparsity and scalability; while hybrid recommendation algorithms mostly mix content-based and system filtering algorithms, and achieve better results under appropriate fusion conditions.

Contents of the invention

The purpose of the embodiments of the present invention is to provide a method for fusion recommendation based on singular value decomposition and classifier, which aims to solve the problems of poor performance of new users, low accuracy of system recommendation and complex attributes existing in existing recommendation algorithms.

The embodiment of the present invention is implemented in this way, a method based on singular value decomposition and classifier fusion recommendation, the method based on singular value decomposition and classifier fusion recommendation includes the following steps:

Step 1, by preprocessing the data, calculate the average score and probability distribution of the item;

Step 2. For each user, calculate the entropy value of all rated items in the category of the item. For each item, there is a category, and the user's preference for the item is set as the target variable, and the value range is determined. According to the user A collection of rated items for , using the formula Calculate the entropy value of the user on the category, where {0,...,k} is the value range of the target variable, P _k,u is the probability of user u’s preference value for item i on a certain category when the user u’s preference value is k, If the entropy value of a certain category does not exist in user u, then the maximum value of entropy of all categories is the entropy value of this category. The greater the uncertainty of the user's rating of this type of item;

From the entropy set E={E ₁ , E ₂ ,…,E _N } of user u on the item category, for any item i of user u, according to the entropy set E of user u, the uncertainty of the item is obtained as where n is the number of categories in item i; through user u's item rating subset Calculate the uncertainty of all elements in the set, and find out the critical value e _u =min(e _i,u ) for the low score of user u; for all users, obtain the set and critical value of their item category entropy respectively;

Step 3: Determine whether to use the classifier by comparing the uncertainty and critical value of the predicted items, and use the Top-N method to recommend the N items with the highest ratings among all the unrated items of the user.

Further, the data preprocessing in step 1 specifically includes that the data set consists of user set U, item set I, and scoring matrix R, where the value range of scoring matrix R is {1, 2, 3, 4, 5}, giving After setting the dimension as f, there is a vector q _i for each item i∈I, which is used to measure the degree to which the item has these influence factors; for each user u∈U, there is a vector P _u , which is used to calculate the user's influence on the item. The value of these influence factors is similar to the value in q _i , both positive and negative; the inner product of q _i ^T and Pu captures the relationship between user _u and item i; the singular value decomposition algorithm and the most basic scoring Compared with the formula R=μ+b _u +b _i +q _i ^T p _u , remove the average score μ of all item ratings and replace it with the average score a _i of each item rating; before starting the learning algorithm of singular value decomposition , the average score for each item must be calculated.

Further, the specific steps for calculating the average score of each item are as follows:

In the first step, for items with user rating records, count the item ratings, calculate the average score, and calculate the distribution probability on the rating range [1,2,3,4,5];

In the second step, for items without user rating records, the average score of all ratings is used as the average score of the item, and the distribution probability on the rating range is set to 0.2.

Further, to calculate the prediction score r _ui , it is necessary to calculate $\underset{p,q,b}{\min }\underset{(u,i)\∈R}{\mathrm{\Σ}}{(r_{\mathrm{ui}}-b_u-b_i-a_i-p_u^T{q}_i)}^2+\mathrm{\λ}({\left|\right|p_u\left|\right|}^2+{\left|\right|q_i\left|\right|}^2+{\left|\right|b_u\left|\right|}^2+{\left|\right|b_i\left|\right|}^2)$ Among them, λ is a regularization parameter, and the square of the absolute values of _pu , q _i , b _i and b _u is added to prevent overfitting. Iteratively learn the values of b _u , b _i , q _i , and p _u through stochastic gradient descent algorithm or cross least square method, and adopt stochastic gradient descent algorithm.

Further, Step 3 specifically includes, according to the probability distribution of all items obtained in Step 1 on [1,2,3,4,5], and the item category entropy set and critical value obtained in Step 2, for any user u, compare the uncertainty e _i,u of item i under the premise of user u and the critical value e _u of user u, if e _i,u ≥ e _u , it means that the uncertainty of the item is large, then it will not be classified and directly Use the singular value decomposition to calculate the predicted score of the unrated item i; otherwise, according to the probability distribution of the item and the SVD predicted score of the item, the probability of rounding up and down is obtained, and the value with the highest probability is selected as the user u The predicted score of the item i; finally, calculate the score of user u for each unbalanced item, and recommend the N items with the highest score to the user.

Further, the specific implementation steps of personalized recommendation based on singular value decomposition and classifier fusion recommendation method are as follows:

Step 1, firstly, preprocess the user data, known user set U={u ₁ ,u ₂ ,u ₃ }, item set I={i ₁ ,i ₂ ,i ₃ ,i ₄ }, category set C={C1,C2,C3,C4}; Calculate the average score μ of all items, and calculate the average score a _i of each item and the probability distribution of the item on the scoring range [1,2,3,4,5] , use μ to fill items without scoring records, get μ=3.78, a _i ={3.67,3.67,4,4}, the probability distribution is {P ₁ ={0,0.33,0,0.33,0.33},P ₂ ={0,0.33,0,0.33,0.33}, P ₃ ={0,0,0,1,0},P ₄ ={0,0,0.5,0,0.5}};

Step 2. After the given dimension f, learning rate and number of iterations, use the user's historical rating data and loss equation $\underset{p,q,b}{\min }\underset{(u,i)\∈R}{\mathrm{\Σ}}{(r_{\mathrm{ui}}-b_u-b_i-a_i-p_u^T{q}_i)}^2+\mathrm{\λ}({\left|\right|p_u\left|\right|}^2+{\left|\right|q_i\left|\right|}^2+{\left|\right|b_u\left|\right|}^2+{\left|\right|b_i\left|\right|}^2)$ Calculate the value of b _u , b _i , _{pu , q i by the method of stochastic gradient descent; the steps are as follows, in the case of given parameters b u , b i , p u , q i first calculate the partial} value of each parameter for the loss equation Derivation, such as b _u ←b _u +α*(rr _ui -λ*b _u ), where r is the real score scored by user u on i, α is the learning rate, and λ is the regularization parameter; use the partial derivative to update each parameter , and finally get the SVD model R＝a _i +b _u +b _i +q _i ^T p _u ;

Step 3, use the information entropy formula Calculate the entropy set of each user on each item category {C1,C2,C3,C4}; such as user u ₁ , the value range of the target variable is {-1,0,1}, where -1 means the score is less than or equal to 2 and less than the average score of the item, that is, dislike; 0 means that the score is equal to 3, that is, general; 1 means that the score is greater than 3, that is, like; thus, Ε(C _u )={0,1.43,0.60, 0}, through E(C _u ) and the uncertainty calculation formula of the project Get the uncertainty critical point e _u of the item is the item rating subset of user u The minimum value of uncertainty of all items in , which is 0.68; thus, the critical value of all users is obtained;

To predict the scoring situation of u ₁ to i ₃ , first obtain the preliminary prediction score through the trained singular value decomposition model, and then get the item uncertainty of item i ₃ when user u ₁ is 0<e _u , then it needs to be classified device; if the initial prediction score is 3.21, get the i ₃ score up to 4, and down to 3, calculate the probability P(3)=(4-3.21)*0=0 when the score is 3, calculate When the score is 4, the probability P(4)=(3.21-3)*1=0.21; P(4)>P(3), so the predicted score is classified as 4, and finally the score of u1 to _i3 is 4 , and will be recommended to the user (in this example, N=1).

The method for recommendation based on singular value decomposition and classifier fusion provided by the present invention adopts the historical score data of users, analyzes it, and finally produces personalized recommendation; uses the singular value decomposition algorithm to obtain the predicted score of the specified item i, and calculates The information entropy of the item on each user determines whether to classify according to the fluctuation degree of the item rating (the size of the information entropy), and finally obtains the final predicted score of the item through the classifier. The use of information entropy in the present invention effectively eliminates the impact on the classifier when the item rating fluctuates greatly, improves the classification effect to improve the scoring accuracy of the predicted rating, thereby improving the accuracy of the recommendation method.

Description of drawings

FIG. 1 is a flow chart of a method for recommendation based on singular value decomposition and classifier fusion provided by an embodiment of the present invention;

Fig. 2 is a schematic flowchart of Embodiment 1 of the method for recommendation based on singular value decomposition and classifier fusion provided by an embodiment of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with the examples. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The application principle of the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

As shown in Figure 1, the method for recommendation based on singular value decomposition and classifier fusion recommendation in the embodiment of the present invention includes the following steps:

S101: Calculate the average score and probability distribution of the items by preprocessing the data.

S102: Train the singular value decomposition model by the stochastic gradient descent method, calculate the entropy set of the user on the item category by the calculation method of the entropy value for the user's scored item set, and determine the item uncertainty critical value;

S103: Determine whether to use a classifier by comparing the uncertainty of the predicted items with the critical value, and use the Top-N method to recommend N items with the highest scores among all unrated items of the user.

Concrete steps of the present invention are: with reference to Fig. 2:

The first step is data preprocessing. The data set consists of a user set U, an item set I, and a scoring matrix R, where the value range of the scoring matrix R is {1,2,3,4,5}, and the given dimension is After f, there is a vector q _i for each item i∈I, which is used to measure the degree to which the item has these influence factors; for each user u∈U, there is a vector P _u , which is used to calculate the influence factors of the user on the item. The interest on , its value is similar to the value in q _i , both can be positive or negative. The inner product of q _i ^T and P _u can capture the relationship between user u and item i; considering that users in real life will be guided by the mode and average score in the rating, the singular value decomposition algorithm and the most Compared with the basic scoring formula R=μ+b _u +b _i +q _i ^T p _u , the average score μ of all item ratings is removed and replaced by the average score _ai of each item rating. Therefore, before starting the learning algorithm of singular value decomposition, the average score of each item must be calculated; the specific steps are as follows:

Step 1, for an item with user rating records, count the item rating, calculate its average score, and calculate its distribution probability on the rating range [1,2,3,4,5];

Step 2, for items without user rating records, use the average score of all ratings as the average score of the item, and set its distribution probability on the rating range to 0.2;

The second step is to calculate the predicted score r _ui , which needs to be calculated $\underset{p,q,b}{\min }\underset{(u,i)\&Element;R}{\mathrm{\&Sigma;}}{(r_{\mathrm{ui}}-b_u-b_i-a_i-p_u^T{q}_i)}^2+\mathrm{\&lambda;}({\left|\right|p_u\left|\right|}^2+{\left|\right|q_i\left|\right|}^2+{\left|\right|b_u\left|\right|}^2+{\left|\right|b_i\left|\right|}^2)$ Among them, λ is a regularization parameter, and the square of the absolute values of _pu , q _i , b _i and b _u is added to prevent overfitting. Iteratively learn the values of b _u , b _i , q _i , and p _u through the stochastic gradient descent (SGD) algorithm or crossed least squares (ALS). Since the SGD method is superior to ALS in time, the stochastic gradient descent algorithm is used;

In the third step, for each user u, calculate the entropy value of all its rated items in the category to which the item belongs. For each item i∈I, there is its category, and its set is C={C ₁ ,C ₂ ,C ₃ ,...,C _N }, where N is the number of categories. The user's preference for this item is set as the target variable, and its value range is determined. According to the user u's evaluated item set I _u , use the formula Calculate the entropy value of the user on N categories, where {0,...,k} is the value range of the target variable, and P _k,u is the user u's preference value for item i on a certain category when the value is k Probability, if the entropy value of a certain category does not exist in user u, then the maximum value of entropy value of all categories is the entropy value of this category. Larger, the greater the uncertainty of the user's rating of this type of item;

The fourth step, from the third step, the entropy set E={E ₁ , E ₂ ,…,E _N } of user u on the item category can be obtained. For any item i of user u, according to the entropy set E of user u , the uncertainty of the project is obtained as where n is the number of categories in item i. Subset of item ratings by user u Calculate the uncertainty of all elements in the set, and find out the critical value e _u =min(e _i,u ) for the low score of user u. For all users, get the set and critical value of their item category entropy respectively;

The fifth step, according to the probability distribution of all items obtained in the first step on [1,2,3,4,5], and the item category entropy set and critical value obtained in 4), for any user u, Compare the uncertainty e _i,u of item i under the premise of user u and the critical value e _u of user u, if e _i,u ≥ e _u , it means that the uncertainty of the item is large, then do not classify, and directly use SVD Calculate the predicted score of the unrated item i; otherwise, according to the probability distribution of the item and the SVD predicted score of the item, the probability of rounding up and rounding down is obtained, and the value with a higher probability is selected as the value of the item i for user u. i's prediction score. Finally, calculate the ratings of user u on each unbalanced item, and recommend the N items with the highest ratings to the user.

The stochastic gradient descent algorithm is used, and the specific steps are as follows:

Step 1, use the formula b _u ←b _u +α*(rr _ui -λ*b _u ) to fit the value of b _u in the scoring equation R=a _i +b _u +b _i +q _i ^T p _u , b _i , q _i and p _u are fitted in the same way, given the learning rate α, the regularization parameter λ, the maximum number of iterations and the initial b _u , b _i , q _i , p _u and the initial root mean square error RSME value;

Step 2, import the training data, get the predicted score according to the scoring equation R=a _i +b _u + _bi +q _i ^T p _u , and use the partial derivative equations of bu, bi, qi and pu to update bu, bi, qi and pu value;

Step 3, import the test data, and according to the formula Calculate the value of RMSE, where u, i belong to the test data, R _ui is the prediction score, is the real rating, and nui is the number of ratings.

Step 4, compare the initial RMSE and the current RMSE value, if the current RMSE is greater than the initial or previous RMSE value or the number of iteration steps has reached the maximum number of iteration steps, then exit the iteration; otherwise, use the current RMSE as the RMSE of the next iteration value, return to step 2.

The resulting probability distribution of all items on [1,2,3,4,5], and the resulting item category entropy set and critical value, for any user u, compare the uncertainty of item i under the premise of user u e _i , _u and the critical value e _u of user u, if e _i,u ≥ e _u , it means that the item has a large uncertainty, then it will not be classified, and the singular value decomposition formula R=a _i +b _u +b will be used directly _i +q _i ^T p _u Calculate the predicted score of unrated item i; otherwise, according to the probability distribution of the item and the SVD predicted score of the item, the probability of rounding up and down is obtained, and the value with the highest probability is selected As user u's prediction score for item i; finally, calculate user u's score for each unbalanced item, and recommend the N items with the highest score to the user.

Using a recommendation algorithm based on singular value decomposition and classifier fusion provided by the present invention can realize personalized recommendation. Taking the scoring matrix and item category description table in Table 1 as an example, the specific implementation steps are as follows: (a) Scoring Matrix; (b) item category description table.

Step 1. First, preprocess the user data. The scoring matrix is shown in Table 1(a); the known user set U={u ₁ ,u ₂ ,u ₃ }, and the item set I={i ₁ ,i ₂ ,i ₃ , i ₄ }, the set of categories C={C1, C2, C3, C4}. According to the first step of the claim, calculate the average score μ of all items, and calculate the average score a _i of each item and the probability distribution of the item on the scoring range [1,2,3,4,5], for no scoring record The item of is filled with μ, it can be obtained that μ=3.78, a _i ={3.67,3.67,4,4}, and its probability distribution is {P ₁ ={0,0.33,0,0.33,0.33},P ₂ ={0 ,0.33,0,0.33,0.33},P ₃ ={0,0,0,1,0},P ₄ ={0,0,0.5,0,0.5}};

Step 2. After the given dimension f, learning rate and number of iterations, use the user's historical rating data and loss equation $\underset{p,q,b}{\min }\underset{(u,i)\&Element;R}{\mathrm{\&Sigma;}}{(r_{\mathrm{ui}}-b_u-b_i-a_i-p_u^T{q}_i)}^2+\mathrm{\&lambda;}({\left|\right|p_u\left|\right|}^2+{\left|\right|q_i\left|\right|}^2+{\left|\right|b_u\left|\right|}^2+{\left|\right|b_i\left|\right|}^2)$ Calculate the values of b _u , b _i , p _u , q _i by stochastic gradient descent method. The main steps are as follows. In the case of given parameters b _u , b _i , p _u , q _i , first calculate the partial derivative of each parameter of the loss equation, such as b _u ← b _u +α*(rr _ui -λ* b _u ), where r is the real score scored by user u on i, α is the learning rate, and λ is the regularization parameter; use the partial derivative to update each parameter, and finally get the SVD model R=a _i +b _u + _bi + q _i ^T p _u ;

Step 3, according to the third step of table 1 (b) and the concrete steps of the present invention, utilize information entropy formula Compute the set of entropies for each user on each item category {C1,C2,C3,C4}. For example, user u ₁ , in this example, the value range of the target variable is {-1,0,1}, where -1 means that the score is less than or equal to 2 and less than the average score of the item, that is, dislike; 0 means that the score is equal to 3 In the case of , it is general; 1 means that the score is greater than 3, that is, like it. Thus, Ε(C _u )={0,1.43,0.60,0} can be obtained, through E(C _u ) and the uncertainty calculation formula of the project The uncertainty critical point e _u of the item can be obtained as the item rating subset of user u The minimum value of uncertainty of all items in , which is 0.68. Therefore, the critical value of all users can be obtained;

Table 1

i ₁ i ₂ i ₃ i ₄ u ₁ 5 2 0 3 u ₂ 4 5 4 0 u ₃ 2 4 0 5

(a)

C i ₁ {C1,C2} i ₂ {C2,C3,C4} i ₃ {C4} i ₄ {C2,C3}

(b)

To predict the rating of u ₁ to i ₃ , first obtain the preliminary prediction score through the trained SVD model, and then get the item uncertainty of item i ₃ when user u ₁ is 0<e _u , then it needs to go through a classifier. If the preliminary prediction score is 3.21, the i ₃ score is rounded up to 4 and down to 3, and the probability P(3)=(4-3.21)*0=0 when the calculated score is 3, the calculated score is The probability of 4 is P(4)=(3.21-3)*1=0.21. P(4)>P(3), so the predicted score is classified as 4, and finally u1 scores 4 for i3, and recommends it to the user (in this example, N=1).

The invention uses the singular value decomposition algorithm to obtain the predicted score of the specified item i, and determines whether to classify by calculating the information entropy of the item on each user, and finally obtains the final predicted score of the item through the classifier, thereby improving the accuracy of the recommendation method .

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention should be included in the protection of the present invention. within range.

Claims (6)

1. based on the method that svd and Multiple Classifier Fusion are recommended, it is characterized in that, the method should recommended based on svd and Multiple Classifier Fusion comprises the following steps:

Step one, by the pre-service to data, the average mark of computational item and probability distribution;

Step 2, for each user, calculates and has allly commented the entropy of project on project generic, for each project, have generic, and user is set to target variable for the preference of project, determine span, comment project set according to user, use formula calculate user at classificatory entropy, wherein 0 ..., k} is the span of target variable, P _k,ufor user certain classification on user u be k to the preference value of project i time probability, if certain classification entropy is not present in user u, then the maximal value of all categories entropy is such entropy;

By the entropy set E={E of user on project category ₁, E ₂..., E _n, for the project i of any one user, according to the entropy set E of user, the uncertainty obtaining project is wherein n is the number of classification in project i; By the project scoring subclass of user calculate the uncertainty of this set all elements, find out user and to give a mark critical value e on the low side _u=min (e _i,u); For all users, obtain set and the critical value of their project category entropy respectively;

Step 3, determines whether use sorter by the uncertainty of comparison prediction project and critical value, and using Top-N method to recommend out, user is all does not comment in project the highest N number of project of marking.

2. the method for recommending based on svd and Multiple Classifier Fusion as claimed in claim 1, it is characterized in that, the data prediction of step one specifically comprises, data set gathers U by user, project set I, and rating matrix R forms, wherein the span of rating matrix R is { 1,2,3,4,5}, given dimension is after f, for each project i ∈ I directed quantity q _i, being used for measurement project has the degree of factor of influence; For each user u ∈ U directed quantity P _u, be used for calculate user on project these impact because of on interest, be worth same q _iin value similar, be all positive and negative; q _i ^twith P _uinner product catch relation between user u and project i; Singular value decomposition algorithm and most basic evaluate formula R=μ+b _u+ b _i+ q _i ^tp _ucompare, remove the average mark μ of all items scoring, and with the average mark a that each project is marked _ireplace; Before the learning algorithm starting svd, calculate the average mark of each project.

3. the method for recommending based on svd and Multiple Classifier Fusion as claimed in claim 2, it is characterized in that, the concrete steps calculating the average mark of each project are as follows:

The first step, for have user mark record project, add up this project scoring, according to calculate average mark, wherein n _itotal number of scoring user on project i, r _jbe each concrete scoring of user to project i of having marked, and calculate the distribution probability in scoring scope [1,2,3,4,5];

Second step, for do not have user mark record project, use the average mark of average mark as this project of all scorings, and the distribution probability be located in scoring scope is 0.2.

4. the method for recommending based on svd and Multiple Classifier Fusion as claimed in claim 1, is characterized in that, computational prediction scoring r _ui, need to calculate $\underset{p,q,b}{\min }\underset{(u,i)\&Element;R}{\mathrm{\&Sigma;}}{(r_{\mathrm{ui}}-b_u-b_i-a_i-p_u^T{q}_i)}^2+\mathrm{\&lambda;}({\left|\right|p_u\left|\right|}^2+{\left|\right|q_i\left|\right|}^2+{\left|\right|b_u\left|\right|}^2+{\left|\right|b_i\left|\right|}^2)$ Wherein λ is regularization parameter, by stochastic gradient descent algorithm or the least square method iterative learning b that intersects _u, b _i, q _i, p _uvalue, adopt stochastic gradient descent algorithm, concrete steps are as follows:

Step one, with the formula b _u← b _u+ α * (r-r _ui-λ * b _u) carry out matching scoring equation R=a _i+ b _u+ b _i+ q _i ^tp _umiddle b _uvalue, b _i, q _iand p _uwith same manner matching, given learning rate α, regularization parameter λ, greatest iteration step number and initial b _u, b _i, q _i, p _uand the value of initial root mean square error RSME;

Step 2, imports training data, according to scoring equation R=a _i+ b _u+ b _i+ q _i ^tp _uobtain prediction scoring, and upgrade the value of bu, bi, qi and pu with the local derviation equation of bu, bi, qi and pu;

Step 3, imports test data, and according to formula calculate the value of RMSE, wherein u, i belong to test data, R _uifor prediction scoring, for true scoring, nui is scoring number.

Step 4, compares the value of Initial R MSE and current RMSE, if current RMSE is greater than the value of an initial or front RMSE or iterative steps has reached greatest iteration step number, then exits iteration; Otherwise, using the value of current RMSE as the RMSE of next iteration, return step 2.

5. the method for recommending based on svd and Multiple Classifier Fusion as claimed in claim 1 or 2, it is characterized in that, step 3 specifically comprises, according to all items of gained in step one [1,2,3,4,5] probability distribution on, and the project category entropy set of gained and critical value in step 2, for any one user u, item compared i uncertain e under user u prerequisite _i,uwith the critical value e of user u _uif, e _i,u>=e _u, descriptive item object uncertainty is not then classified comparatively greatly, directly uses svd formula R=a _i+ b _u+ b _i+ q _i ^tp _ucalculate the prediction score value not commenting project i; Otherwise, probability when obtaining rounding up according to the probability distribution of project and the SVD prediction point of this project and round downwards, the value that select probability is large is divided the prediction of this project i as user u; Finally, user u is calculated to the scoring of each not flat project, by N number of project recommendation the highest for scoring to user.

6. the method for recommending based on svd and Multiple Classifier Fusion as claimed in claim 1 or 2, is characterized in that, the concrete implementation step that the method should recommended based on svd and Multiple Classifier Fusion realizes personalized recommendation is as follows:

Step one, first, carries out pre-service to user data, known users set U={u ₁, u ₂, u ₃, project set I={i ₁, i ₂, i ₃, i ₄, the set C={C1 of classification, C2, C3, C4}; Calculate the average mark μ of all items, and calculate the average mark a of each project _iand the probability distribution of project in scoring scope [1,2,3,4,5], the project for record of not marking uses μ to fill, and obtains μ=3.78, a _i={ 3.67,3.67,4,4}, probability distribution is { P ₁={ 0,0.33,0,0.33,0.33}, P ₂={ 0,0.33,0,0.33,0.33}, P ₃={ 0,0,0,1,0}, P ₄={ 0,0,0.5,0,0.5}};

Step 2, after given dimension f, learning rate and iterations, utilizes user's history score data and loss equation $\underset{p,q,b}{\min }\underset{(u,i)\&Element;R}{\mathrm{\&Sigma;}}{(r_{\mathrm{ui}}-b_u-b_i-a_i-p_u^T{q}_i)}^2+\mathrm{\&lambda;}({\left|\right|p_u\left|\right|}^2+{\left|\right|q_i\left|\right|}^2+{\left|\right|b_u\left|\right|}^2+{\left|\right|b_i\left|\right|}^2)$ B is calculated by stochastic gradient descent method _u, b _i, p _u, q _ivalue; Step is as follows, at given parameters b _u, b _i, p _u, q _iwhen first loss equation is asked to the local derviation of parameters, as b _u← b _u+ α * (r-r _ui-λ * b _u), wherein r is the true score that user u gives a mark to i, and α is learning rate, and λ is regularization parameter; Utilize partial derivative to upgrade parameters, and finally obtain SVD model R=a _i+ b _u+ b _i+ q _i ^tp _u;

Step 3, utilizes information entropy formula calculate each user in { the entropy set on C1, C2, C3, C4} of each project category; As user u ₁, the span of target variable be-1,0,1}, wherein-1 represents that scoring is less than or equal to 2 and is less than project average mark, does not namely like; 0 represents the situation that scoring equals 3, namely generally; 1 represents the situation that scoring is greater than 3, namely likes; Thus, Ε (C is obtained _u)={ 0,1.43,0.60,0}, by E (C _u) and the indeterminacy of calculation formula of project obtain the uncertain critical point e of project _ufor the project scoring subclass of user u the probabilistic minimum value of middle all items, namely 0.68; Therefore, the critical value of all users is obtained;

Prediction u ₁to i ₃scoring situation, first obtain tentative prediction scoring by the svd model after training, then obtain project i ₃at user u ₁time project uncertainty be 0<e _u, then need through sorter; If tentative prediction scoring is 3.21, obtain i ₃rounding up of scoring is 4, and rounding downwards is 3, probability P (3)=(4-3.21) * 0=0 when calculating scoring is 3, probability P (4)=(3.21-3) * 1=0.21 when calculating scoring is 4; P (4) >P (3), so, prediction scoring is classified as 4, last u ₁to i ₃scoring be 4, and will user be recommended, N=1.