CN102508907B - Dynamic recommendation method based on training set optimization for recommendation system - Google Patents

Abstract

The invention discloses a dynamic recommendation method based on training set optimization for a recommendation system, which specifically includes: (1) establishing a preliminary recommendation portion: generating an original recommendation model according to original user grading data; (2) performing AdaBoost trainings: utilizing the original recommendation model as a classifying and judging basis to classify the data and adjust learning times of samples by means of multiple iterative learning training data; (3) screening incorrect samples: data of selected difficult samples are removed as the incorrect samples after multiple AdaBoost trainings so as to construct a new training data collection; (4) reconstructing a recommendation model: combining training results to regenerate the recommendation model based on the new training data; and (5) generating recommendation results: utilizing the new recommendation model to generate the recommendation results. The method is capable of removing the data without referential meaning in recommended service by the aid of great relevance of original training set data in content, so that validity of the training data and precision of the final recommendation model are improved.

Description

A kind of dynamic recommendation method of the commending system based on training set optimization

Technical field

The present invention relates to the technical field of user's commending system, particularly a kind of dynamic recommendation method of the commending system based on training set optimization.

Background technology

The personalized recommendation service is customer-centric, to understand user preferences as basis, is the method for service that the user provides customized customized information to present, and is also to solve from the magnanimity Internet resources, extracting a kind of effective way of user's information needed.With the generic services pattern, compare the personalized recommendation service following characteristics are arranged: at first, the personalized recommendation service can be by the user from rescue the predicament of information overload out, make the user can have an opportunity to enjoy the really hommization network information service of rich and varied, convenient appropriateness, greatly promote the user and experience and satisfaction; Secondly, the personalized recommendation service can fully improve service quality and the access efficiency of Web website, can also find the point of interest that the user is potential simultaneously, thereby excavate potential commercial value, for Internet service provider provides considerable economy return.

Since the commending system based on collaborative filtering is born, the proposition of the hidden vectorial recommended models of especially decomposing based on normalized matrix, the personalized recommendation technology has had quite high lifting in the recommendation precision of theoretical aspect.As the original score data of recommending important evidence, final recommendation results is had to decisive influence, apparent one group of data with high accuracy can obtain good recommendation effect in final recommendation.

The data set that general personalized recommendation service to the user all is based on existing historical accumulation carries out, and the data volume of this data set is very huge.Data set scale huge is difficult to avoid exist in the collection of data irrational data, such as user's mistake scoring or non-user replaces the phenomenons such as scoring.Itself does not have referential these data, and these data should not adopted in the recommendation service to the user.Therefore, to processing and the selection of original score data, can help to a great extent to improve the recommendation precision.Adopting the training set of determination methods screening comparatively accurately, and carrying out on this basis the foundation of recommended models, the recommended models that obtains so can recommend, on precision, more remarkable must the lifting arranged.

Summary of the invention

The technical problem to be solved in the present invention is: overcome the deficiencies in the prior art, a kind of dynamic recommendation method of the commending system based on training set optimization is provided, the method can be screened by the original training data to as the personalized recommendation foundation, and take the new training set of removing error sample as according to obtaining to have the more recommended models of high accurancy and precision, the accuracy that has improved personalized recommendation.

The technical scheme that the present invention solves the problems of the technologies described above is: a kind of dynamic recommendation method of the commending system based on training set optimization, and the method concrete steps are as follows:

Step (1) is set up preliminary recommended models: according to original user's score data, utilize based on the modeling method in normalized matrix factorization recommended models and generate initial recommended models;

Step (2) AdaBoost training: the foundation of utilizing the middle recommended models that generates of step (1) to judge as primary classification builds sorter, the recommendation that calculates according to recommended models and the classification of the difference condition decision data between raw value, utilize AdaBoost Algorithm Learning original training sample, and take turns and finish the rear new sorter that generates at each;

Step (3) screening error sample: each takes turns training in the training process that utilizes the AdaBoost algorithm all needs to filter out difficult sample, the division of difficult sample can adopt the otherness between predicted value and actual value to judge in the method, namely when this species diversity, judges difficult sample during greater than a certain threshold value.After the AdaBoost training that warp is too much taken turns, repeatedly the data of chosen difficult sample can be used as the error sample removal, thereby construct for the required training data set of next iteration;

Step (4) reconstruct recommended models: take the middle training data that obtains of step (3) as basis, in conjunction with the AdaBoost training data, regenerate recommended models.

Step (5) produces recommendation results: using the user characteristics vector as input, the recommended models of utilizing step (4) to obtain calculates recommendation results and returns to the user.

To the AdaBoost cluster training of raw data set, specific as follows in described step (2):

1. step revises normalized matrix factorization recommended models, no longer original score data set T is divided into to two subset T ₁, T ₂, wherein, data set T ₁For learning, data set T ₂For judging that study stops wherein, but all data in data set T are all learnt, set the AdaBoost training iteration wheel number I, often take turns the number of times R of study, the error range errPermission of permission, and initialization feature vector set;

2. step utilizes normalized matrix factorization recommended models learning training data R time in the iteration of the first round, the estimated value of respective user to the scoring of project in the calculation training data on the set of eigenvectors that training obtains

And obtain itself and actual value r _{U, i}Absolute error, i.e. absolute error

The AbsE value that 3. step calculates in 2. when step is judged the difficult sample of secondary data during greater than errPermission, travel through the total errCount of whole difficult samples of training data acquisition ⁿAnd by following formula, calculated the error rate ε of sample ⁿ, wherein | T| represents the number of samples in training set;

${\mathrm{\ϵ}}^n=\frac{{\mathrm{errCount}}^n}{\left|T\right|}$ (1) formula;

In formula: ε ⁿExpression sample error rate, | T| represents the number of samples in training set, errCount ⁿThe whole training data of expression traversal obtains the sum of difficult sample;

The error rate ε that 4. step is calculated in 3. according to step ⁿAdjust the study number of times of training sample in the next round iteration, be specially when the AbsE of training sample data value during less than errPermission, the study number of times of this sample in the next round iteration is trainTime ^N+1=trainTime ⁿ* ε ⁿIf (tramTime ^N+1Get 1 for<1), when the AbsE of training sample data value determined that it is difficult sample in (2) during greater than errPermission, the study number of times of this sample in the next round iteration was ${\mathrm{trainTime}}^{n+1}=\frac{{\mathrm{trainTime}}^n}{{\mathrm{\&epsiv;}}^n},$ Available as shown in the formula expression:

${\mathrm{trainTime}}^{n+1}=\left\{\begin{array}{cc}\frac{{\mathrm{trainTime}}^n}{{\mathrm{\&epsiv;}}^n},& (\mathrm{AbsE}\&GreaterEqual;\mathrm{errPermission})\\ \max ({\mathrm{trainTime}}^n*{\mathrm{\&epsiv;}}^n,1),& (\mathrm{AbsE}<\mathrm{errPermission})\end{array}\right.$ (2) formula;

In formula: trainTime ⁿThe study number of times of sample in n next round iteration, trainTime ^N+1The study number of times of sample in n+1 next round iteration, ε ⁿThe error rate of calculating in expression step (3), the absolute error that the sub-step of AbsE value representation step (2) is 2. calculated, errPermission represents the error range that allows;

Step 5. by errPermission with after fixed proportion declineRate reduction, start the iteration of a new round, and the study number of times of each sample calculates and carries out in the iteration that this is taken turns in 4. according to step.

The method of the screening error sample in described step (3), specific as follows:

In steps A, step (2), each is taken turns training and all needs to filter out difficult sample go forward side by side row labels and statistics in the training process that utilizes the AdaBoost algorithm;

After step B, the training of the AdaBoost iteration through too much taking turns, training data is traveled through to the number of times that each sample of statistics is determined into difficult sample;

Step C, according to clearance delRate, will judge into sample that difficult sample number of times is higher from training set, removing, thereby obtain new training set.

The present invention's advantage compared with prior art is:

The present invention can carry out take available data as basis automatic cluster, through repeatedly iterative learning, can filter out the misdata that is unfavorable for improving the recommended models precision, thereby remove misdata and obtain high-quality training data, the higher recommendation precision of final acquisition.Than classic method, only for recommended models, do not process original training data is filtered, this method can more effectively improve the recommendation precision.

The accompanying drawing explanation

Fig. 1 is summary workflow diagram of the present invention;

Fig. 2 is detailed operation process flow diagram of the present invention.

Embodiment

Existing accompanying drawings embodiments of the invention.

As shown in Figure 2, the present invention includes four key steps: set up recommended models, AdaBoost training, screening error sample and reconstruct recommended models.

Step (1) is set up recommended models: read original user's score data and test data, and according to the Customs Assigned Number UserID of maximum in two data and the dimension that bullets ItemID determines user characteristics vector sum item feature vector, based on the modeling method in normalized matrix factorization recommended models, newly-built and random initializtion user characteristics vector sum item feature vector;

Step (2) the AdaBoost training stage: the foundation structure sorter that utilizes the middle recommended models that generates of step (1) to judge as classification, the recommendation that calculates according to recommended models and the classification of the difference condition decision data between raw value, concrete steps are:

1. step revises normalized matrix factorization recommended models, no longer original score data set T is divided into to two subset T ₁, T ₂(utilize data set T ₁Learn, utilize T ₂Judge that study stops), but all data in data set T are all learnt, set the AdaBoost training iteration wheel number I, often take turns the number of times R of study, the error range errPermission of permission, and initialization feature vector set;

2. step is utilized above-mentioned amended based on the modeling in normalized matrix factorization recommended models and training method learning training data R time in AdaBoost every takes turns iteration.After each study, calculating the RMSE of resulting user characteristics vector sum product features vector model on training data, the RMSE value that reaches this wheel hour namely when RMSE value can finish in advance this when large and takes turns iteration than learning value afterwards last time.This is taken turns after iteration and obtains that new recommended models---this takes turns final user characteristics vector sum product features vector.Utilize the estimated value of respective user to the scoring of project in proper vector calculation training data

And obtain itself and actual value r _{U, i}Absolute error, i.e. absolute error

The AbsE value that 3. step calculates in 2. when step judges that the difficult sample of secondary data, the mistake that these data are corresponding judge that number of times increases by 1 during greater than errPermission, travel through the total errCount of whole difficult samples of training data acquisition ⁿAnd by (1) formula, calculated the error rate ε of sample ⁿ

${\mathrm{\&epsiv;}}^n=\frac{{\mathrm{errCount}}^n}{\left|T\right|}$ (1) formula;

In formula: ε ⁿExpression sample error rate, | T| represents the number of samples in training set, errCount ⁿThe whole training data of expression traversal obtains the sum of difficult sample;

The error rate ε that 4. step is calculated in 3. according to step ⁿ, utilize (2) formula to adjust the study number of times of training sample in the next round iteration;

${\mathrm{trainTime}}^{n+1}=\left\{\begin{array}{cc}\frac{{\mathrm{trainTime}}^n}{{\mathrm{\&epsiv;}}^n},& (\mathrm{AbsE}\&GreaterEqual;\mathrm{errPermission})\\ \max ({\mathrm{trainTime}}^n*{\mathrm{\&epsiv;}}^n,1),& (\mathrm{AbsE}<\mathrm{errPermission})\end{array}\right.$ (2) formula;

In formula: trainTime ⁿThe study number of times of sample in n next round iteration, trainTime ^N+1The study number of times of sample in n+1 next round iteration, ε ⁿThe error rate of calculating in expression step (3), the absolute error that the sub-step of AbsE value representation step (2) is 2. calculated, errPermission represents the error range that allows;

5. step reduces errPermission with fixed proportion declineRate after, start the iteration of a new round, and the study number of times of each sample calculates in 4. and carries out according to step in the iteration that this is taken turns, if completed the iteration of pre-determined number finished this stage;

Step (3) screening error sample: each takes turns training in the training process that utilizes the AdaBoost algorithm all needs to filter out difficult sample, the division of difficult sample can adopt the otherness between predicted value and actual value to judge in the method, namely when this species diversity, judges difficult sample during greater than a certain threshold value.After the AdaBoost training that warp is too much taken turns, repeatedly the data of chosen difficult sample can be used as the error sample removal, thereby construct for the required training data set of next iteration;

In steps A, step (2), each is taken turns training and all needs to filter out difficult sample go forward side by side row labels and statistics in the training process that utilizes the AdaBoost algorithm;

After step B, the training of the AdaBoost iteration through too much taking turns, training data is traveled through to the number of times that each sample of statistics is determined into difficult sample;

Step C, statistics calculates the maximum errors number that allows according to clearance del_rate, total sample number and 2., by judging into difficult sample number of times, higher than the sample of this value, at training concentrated setting frequency of training, is 0, thereby obtains new training set;

Step (4) reconstruct recommended models: take the middle training data that obtains of step (3) as basis, in conjunction with the AdaBoost training data, regenerate recommended models based on the normalized matrix factorization method;

Step (5) produces recommendation results: using the product of user characteristics vector sum item feature vector as the estimation score value of specific user to specific project, will estimate that project recommendation that score value is higher is to this user.

The present invention can carry out take available data as basis automatic cluster, through repeatedly iterative learning, can filter out the misdata that is unfavorable for improving the recommended models precision, thereby remove misdata and obtain high-quality training data, the higher recommendation precision of final acquisition.Than classic method, only for recommended models, do not process original training data is filtered, this method can more effectively improve commending system and recommend precision, as shown in table 1 below, when not adopting screening technique of the present invention, the RMSE of RMF method on test set is 0.792933, and several groups of data in table 1 have all obtained less RMSE value.

The RMSE value that table 1 the present invention obtains

The part that the present invention does not elaborate belongs to techniques well known.Above embodiment is only in order to technical scheme of the present invention to be described but not be limited in the scope of embodiment; to those skilled in the art; as long as various variations claim limit and the spirit and scope of the present invention determined in; these variations are apparent, and all utilize innovation and creation that the present invention conceives all at the row of protection.

Claims (2)

1. the dynamic recommendation method of a commending system of optimizing based on training set, it is characterized in that: the method concrete steps are as follows:

Step (1), set up preliminary recommended models: according to original user's score data, utilize based on the modeling method in normalized matrix factorization recommended models and generate preliminary recommended models;

Step (2), AdaBoost training: the foundation of utilizing the middle recommended models that generates of step (1) to judge as primary classification builds sorter, the recommendation that calculates according to recommended models and the classification of the difference condition decision data between raw value, utilize AdaBoost Algorithm Learning original training sample, and take turns and finish the rear new sorter that generates at each;

Wherein, to the AdaBoost cluster training of raw data set, specific as follows in described step (2):

Step 1., revise normalized matrix factorization recommended models, no longer original score data set T is divided into to two subset T ₁, T ₂, wherein, data set T ₁For learning, data set T ₂For judging that study stops, but all data in data set T all being learnt, set the AdaBoost training iteration wheel number I, often take turns the number of times R of study, the error range errPermission of permission, and initialization feature vector set;

Step 2., in the iteration of the first round, utilize normalized matrix factorization recommended models learning training data R time, the estimated value of respective user to the scoring of project in the calculation training data on the set of eigenvectors that training obtains

And obtain itself and actual value r _u,iAbsolute error, i.e. absolute error

Step 3., the AbsE value that calculates in 2. when step judges the difficult sample of these data during greater than errPermission, travels through the total errCount that whole training datas obtains difficult sample ⁿAnd by following formula, calculated the error rate ε of sample ⁿ, wherein | T| represents the number of samples in training set;

Formula;

In formula: ε ⁿExpression sample error rate, | T| represents the number of samples in training set, errCount ⁿThe whole training data of expression traversal obtains the sum of difficult sample;

Step 4., the error rate ε that calculates in 3. according to step ⁿAdjust the study number of times of training sample in the next round iteration, be specially: during less than errPermission, the study number of times of this sample in the next round iteration is trainTime when the AbsE of training sample data value ^N+1=trainTime ⁿ* ε ⁿIf trainTime wherein ^N+1Get 1 for<1, when the AbsE of training sample data value determined that it is difficult sample in (2) formula during greater than errPermission, the study number of times of this sample in the next round iteration was

Available as shown in the formula expression:

Formula;

In formula: trainTime ⁿThe study number of times of expression sample in n next round iteration, trainTime ^N+1The study number of times of expression sample in n+1 next round iteration, ε ⁿRepresent during step is 3. the error rate of calculating, the absolute error that the sub-step of AbsE value representation step (2) is 2. calculated, errPermission represents the error range that allows;

Step 5., by errPermission with after fixed proportion declineRate reduction, start the iteration of a new round, and the carrying out that the study number of times of each sample calculates in 4. according to step in the iteration that this is taken turns;

Step (3), screening error sample: each takes turns training in the training process that utilizes the AdaBoost algorithm all needs to filter out difficult sample, the division of difficult sample can adopt the otherness between predicted value and actual value to judge, namely when this species diversity, judges difficult sample during greater than a certain threshold value; After the AdaBoost training that warp is too much taken turns, repeatedly the data of chosen difficult sample can be used as the error sample removal, thereby construct for the required training data set of next iteration;

Step (4), reconstruct recommended models: with the training dataset that obtains in step (3), be combined into basis, in conjunction with the AdaBoost algorithm, train, regenerate recommended models;

Step (5), generation recommendation results: using the user characteristics vector as input, the recommended models of utilizing step (4) to obtain calculates recommendation results and returns to the user.

2. the dynamic recommendation method of a kind of commending system of optimizing based on training set according to claim 1 is characterized in that: the method for the screening error sample in described step (3), specific as follows:

Steps A, in step (2), utilizing in the training process of AdaBoost algorithm each to take turns training all to need to filter out difficult sample go forward side by side row labels and statistics;

After step B, the training of the AdaBoost iteration through too much taking turns, training data is traveled through to the number of times that each sample of statistics is determined into difficult sample;

Step C, according to clearance delRate, will judge into sample that difficult sample number of times is higher from training set, removing, thereby obtain new training set.

Images