CN103198418A

CN103198418A - Application recommendation method and application recommendation system

Info

Publication number: CN103198418A
Application number: CN2013100845681A
Authority: CN
Inventors: 郑巍; 罗峰; 黄苏支; 李娜
Original assignee: BEIJING IZP TECHNOLOGIES Co Ltd
Current assignee: BEIJING IZP TECHNOLOGIES Co Ltd
Priority date: 2013-03-15
Filing date: 2013-03-15
Publication date: 2013-07-10

Abstract

The invention provides an application recommendation method and an application recommendation system. The application recommendation method includes the following steps: access behavior data of multiple reference users to preset multiple applications are obtained, the multiple reference users are subjected to grouping according to the access behavior data, and each of the multiple applications has a corresponding application class; a group which a target user belongs to is determined according to access behavior data of the target user to the multiple applications, access weighted values of the target user to different application classes are subjected to statistics, and the application classes with the access weighted values which are in a preset range are regarded as interest application classes; and the applications which belong to the interest application classes are extracted from the multiple applications which are accessed by all the reference users and are in the group which the target user belongs to, and then the applications are recommended to the target user. The application recommendation method and the application recommendation system can occupy fewer system resources when the applications are recommended.

Description

A kind of application recommend method and system

Technical field

The application relates to networking technology area, particularly relates to a kind of application recommend method, and, a kind of application recommendation apparatus.

Background technology

Along with the continuous increase of people to the network application demand, the personalized recommendation service has slowly entered in people's the life, and the personalized recommendation service refers to carry out pointed recommendation at different visitors.At present, many emerging websites adopt the personalized recommendation service to go to attract browsing of consumer one after another, particularly, analyze by the visit behavior to each visitor, give the recommendation of visitor's personalization respectively according to analysis result.

At present, recommend method relatively more commonly used mainly comprises following two kinds:

1, content-based filter method (CB, Content-Based Filtering)

This method is that the visit data according to targeted customer's (be video recommend object) screens associated information and recommends, for example, the relevant information of some videos that the extraction targeted customer browsed in the past filters out some videos according to relevant information and recommends.

2, collaborative filtering method (COL, Collaborative Filtering)

This method is based on the preference information that has other users of similar interests with the targeted customer and generates recommendation to the targeted customer, particularly, search other users that seen similar video with the targeted customer, find out that these users have seen, and candidate's video that the targeted customer has not seen, according to watching the higher user of video similarity to the marking of candidate's video with the targeted customer, or according to the marking of the higher video of candidate's video similarity, each candidate's video is given a mark, and candidate's video of giving a mark higher is recommended.

The problem that exists in the above background technology is:

When adopting content filtering method, because the relevant information of partial video is more limited, be difficult to find the video that to recommend; Collaborative filtering method does not need to judge according to the relevant information of video, can solve these problems of content filtering method, but the collaborative filtering method in the employing background technology, need filter out similarity higher other users or video, with the marking of predicting candidate video, these operations have taken a large amount of system resource, and, some video may not have associated user's marking, can't further predict the marking of this video, makes these videos recommended to go out.

Summary of the invention

The embodiment of the present application provides a kind of application recommend method, taking system resource when using recommendation to reduce.

The embodiment of the present application also provides a kind of application recommendation apparatus, in order to guarantee said method application and realization in practice.

In order to address the above problem, the embodiment of the present application discloses a kind of application recommend method, comprising:

Obtain a plurality ofly with reference to the visit behavioral data of user to a plurality of application of presetting, and a plurality ofly divide into groups with reference to the user to described according to described visit behavioral data, each application possesses corresponding applicating category;

According to the visit behavioral data of targeted customer to a plurality of application, determine to divide into groups under the targeted customer, and the statistics targeted customer satisfies the access weight value applicating category of preset range as the interest applicating category to the access weight value of different applicating categories;

In a plurality of application that each of dividing into groups under the targeted customer visited with reference to the user, extract the application that belongs to described interest applicating category and recommend the targeted customer.

Preferably, in a plurality of application that each of dividing into groups under the targeted customer visited with reference to the user, extract the application of described interest applicating category and recommend before the targeted customer, described method also comprises:

With reference to the visit behavioral data of user to a plurality of application, statistics satisfies the access weight value applicating category of preset range as the interest applicating category respectively with reference to the access weight value of user to different applicating categories according to each;

Calculate targeted customer's interest applicating category, with each similarity with reference to user's interest applicating category of affiliated grouping, and definite similarity is greater than the reference user of preset value.

Preferably, in a plurality of application that described each of dividing into groups under the targeted customer visited with reference to the user, extract the step that the application that belongs to described interest applicating category recommends the targeted customer and be:

Each of searching grouping under the targeted customer be with reference among the user, a plurality of application that similarity is visited greater than the reference user of preset value;

In a plurality of application of searching, extract the application that belongs to described interest applicating category and recommend the targeted customer.

Preferably, a plurality of groupings of described division with reference to the user are divided into high value grouping and low value and divide into groups according to the reference user's who comprises visit behavioral data;

At described low value grouping, also extract the application and the application that belongs to described interest applicating category of at least one default applicating category, together recommend the targeted customer.

Preferably, comprise the applicating category under access time and the access application in the described visit behavioral data, describedly according to the visit behavioral data a plurality of steps of dividing into groups with reference to the user comprised:

With reference to the user, the access time that comprises in the behavioral data according to visit and the applicating category under the access application calculate the number of current nearest access time, access frequency and the access application of distance at each;

Add up all mean values with reference to user's the nearest access time, the mean value of access frequency, and the mean value of the number of access application;

, with reference to the user number of nearest access time, access frequency and access application is compared with corresponding average respectively at each, will identical reference user be divided in the grouping according to the comparative result of three numerical value.

Preferably, described according to the visit behavioral data of targeted customer to a plurality of application, determine that the step of dividing into groups under the targeted customer comprises:

At the targeted customer, the application access time that comprises in the behavioral data according to visit and the applicating category under the access application calculate the number of current nearest access time, access frequency and the access application of distance;

Number with targeted customer's nearest access time, access frequency and access application, mean value with the number of the mean value of all mean values with reference to user's the nearest access time, access frequency and access application compares respectively, and the grouping that comparative result is identical is as dividing into groups under the targeted customer.

Preferably, described statistics targeted customer is to the access weight value of different applicating categories, and the applicating category that the access weight value is satisfied scope comprises as the step of interest applicating category:

At each applicating category that the targeted customer visits, according to the classification weighted value of preset algorithm and applicating category, add up the access weight value of each applicating category;

Remove the access weight value less than the applicating category of default weighted value, any two applicating categories are formed the binomial classification;

According to the classification weighted value of described preset algorithm and binomial classification, the access weight value of each binomial classification that statistics is visited;

Extract the access weight value greater than the binomial classification of default weighted value as the interest applicating category.

Preferably, described each applicating category of visiting at the targeted customer, according to the classification weighted value of preset algorithm and applicating category, the step of adding up the access weight value of each applicating category comprises:

Step 11 at each applicating category of targeted customer, is extracted current nearest j * N visit of distance behavioral data, and j and N are positive integer, and the initial value of j is 1;

Step 12 according to the definite access times to each applicating category of described visit behavioral data, according to the classification weighted value Wj of preset algorithm and applicating category, is upgraded the access weight value WSj of each applicating category, and the initial value of WSj is 0;

Step 13 is if described classification weighted value has been got then shut-down operation smaller or equal to default weighted value or described visit behavioral data; If described classification weighted value has not been got greater than default weighted value and described visit behavioral data, then make j=j+1, and return step 11.

Preferably, described classification weighted value according to preset algorithm and binomial classification, the step of the access weight value of each binomial classification that statistics is visited comprises:

Step 21 at each binomial classification of targeted customer, is extracted current nearest j * N visit of distance behavioral data, and j and N are positive integer, and the initial value of j is 1;

Step 22 according to the definite access times to each binomial classification of described visit behavioral data, according to the classification weighted value Wj ' of preset algorithm and binomial classification, is upgraded the access weight value WSj ' of each binomial classification, and the initial value of WSj ' is 0;

Step 23 is if described classification weighted value has been got then shut-down operation smaller or equal to default weighted value or described visit behavioral data; If described classification weighted value has not been got greater than default weighted value and described visit behavioral data, then make j=j+1, and return step 21.

Preferably, described calculating targeted customer's interest applicating category comprises with each step with reference to the similarity of user's interest applicating category of affiliated grouping:

At in the affiliated grouping each with reference to the user, according to targeted customer's interest applicating category with set up respectively with reference to user's interest applicating category corresponding regular vectorial;

Calculating is with reference to user's rule vector and similarity with reference to user's rule vector, as the similarity of reference user and targeted customer's interest applicating category.

Preferably, obtain a plurality of with reference to the step of user to the visit behavioral data of a plurality of application of presetting before, described method also comprises:

When the reference user conducted interviews to a plurality of application of presetting, record was with reference to user's visit behavioral data.

The embodiment of the invention also provides a kind of application recommendation apparatus, comprising:

Data acquisition module is used for obtaining a plurality of with reference to the visit behavioral data of user to a plurality of application of presetting;

The grouping determination module is used for a plurality ofly dividing into groups with reference to the user to described according to described visit behavioral data, and according to the visit behavioral data of targeted customer to a plurality of application, determines grouping under the targeted customer, and each application possesses corresponding applicating category;

Interest applicating category determination module is used for the statistics targeted customer to the access weight value of different applicating categories, the access weight value is satisfied the applicating category of preset range as the interest applicating category;

Use recommending module, a plurality of application that each that is used for dividing into groups under the targeted customer visited with reference to the user are extracted the application that belongs to described interest applicating category and are recommended the targeted customer.

Compare with background technology, the application comprises following advantage:

Than background technology when the targeted customer is recommended, need high reference user or the application of screening similarity earlier, the application is in advance by analyzing a plurality of visit behavioral datas with reference to the user, to divide into groups with reference to the user, when recommending at the targeted customer, grouping and the higher interest applicating category of access weight under only determining, in a plurality of application that the reference user who divides into groups under the targeted customer then visits, extraction meets the application of interest applicating category and recommends the user, simplify the step of recommending greatly, alleviated taking system resource.

The application is after grouping under definite targeted customer, can also be further from affiliated grouping, carry out similar gathering with the similarity that similar vector matrix is calculated between the user, further filter out the reference user higher with targeted customer's similarity, the user is recommended in the application that these users visit, use the accuracy of recommending thereby increased.

The application is when determining user's interest applicating category, weight increment method for digging joins in the Apriori algorithm, watch the mode of data volume to excavate nearest interest rule with the increment increase, and do not need whole historical datas are done analysis, so just can save assess the cost, the time.

Among the application, can will be divided into high value group and low value group with reference to the user according to the visit behavioral data, can extract the application of pre-set categories recommends for low value grouping, simultaneously, because the application is when using recommendation, need not the marking of user to using, thereby make each use all recommended users of giving of possibility.

Certainly, arbitrary product of enforcement the application not necessarily needs to reach simultaneously above-described all advantages.

Description of drawings

Fig. 1 is the described a kind of process flow diagram of using recommend method of the embodiment of the present application;

Fig. 2 is the described a kind of structured flowchart of using recommendation apparatus of the embodiment of the present application;

Fig. 3 uses the process synoptic diagram of recommending in the embodiment of the present application.

Embodiment

For above-mentioned purpose, the feature and advantage that make the application can become apparent more, below in conjunction with the drawings and specific embodiments the application is described in further detail.

Analysis mining need be carried out to active user's visit behavior in that personalized recommendation when service is provided in the website, and the result by data mining recommends some interested application to the user, improves the user with this and watch rate on the website.Recommend method commonly used comprises content-based filter method and collaborative filtering method.

Content-based filter method is based on the relevance of file and recommends, it is the extension to the content information that comprises of visit data, particularly, historical information (as the document of estimating, sharing, collected) structuring user's preference document according to the active user, calculated recommendation is used the similarity with the user preference document, and the active user is recommended in the most similar application.For example, in film is recommended, at first analyze marking that the active user seen than the general character (performer, director, style etc.) of higher film, recommend other films high with these user's interest movie contents similarities again.

Collaborative filtering method is based on the recommendation that the relevance between the user is carried out, be a kind of recommendation of carrying out based on the identical user of one group of interest or application, it produces recommendation list to the targeted customer according to neighbours user's (user similar to targeted customer's interest) preference information.Mainly be divided into based on user's collaborative filtering method with based on the collaborative filtering method of using.

Based on (User based) collaborative filtering method of user based on such hypothesis: if the marking that some users use a certain class is more approaching, then their marking that other class is used is also more approaching.The search similar users of watching similar video with the user at first according to similar users and the active user marking to the video watched jointly, filters out some marking similarities than higher similar users; The video that these users have been seen and video that the visitor has not seen are as designated then, and according to the scoring of these similar users to designated, the prediction active user feeds back to the active user to the marking of this designated with designated and marking situation.Working contents based on user's collaborative filtering method is similarity measurement and the prediction scoring of user or application.

Based on (Item based) collaborative filtering method of using based on such hypothesis: if most of user is more close to the marking of some project, then the active user also can be more approaching to these marking.At first according to each user respectively to the marking of each video, calculate the similarity between video and the video, the nearest application that ferret out is used, the scoring of the score information target of prediction of the nearest-neighbors of intended application being used according to the user produces the top n recommendation information at last then.

See table 1, provided content-based filter method and collaborative filtering method relatively:

Table 1

Than content-based filter method, collaborative filtering method can be recommended the unconspicuous application of some content characteristics (for example artwork, music etc.), but can't recommend for the application that does not have the user to estimate, and, when the user that the screening similarity is higher or application, can take a large amount of system resource.

When using recommendation, system resource is taken bigger problem, the method that the application provides a kind of application to recommend in order to solve collaborative filtering method.

Realization flow below by the described method of the application of embodiment is elaborated.

With reference to Fig. 1, it shows the method flow diagram that the described a kind of application of the embodiment of the present application is recommended.

Step 101, obtain a plurality ofly with reference to the visit behavioral data of user to a plurality of application of presetting, and a plurality ofly divide into groups with reference to the user to described according to described visit behavioral data.

In the embodiment of the present application, preset a plurality of application, described application can be video, audio frequency, Word message, info web, advertisement or application program etc., for example presets a video database, and the inside comprises a plurality of videos.

Refer to visit the user of wherein one or more application with reference to the user.In a preferred embodiment of the present application, obtain a plurality of with reference to the step of user to the visit behavioral data of a plurality of application of presetting before, can be when the reference user conducts interviews to a plurality of application of presetting, record is with reference to user's visit behavioral data, and be kept in the user's business database, the user that stores in the user's business database as the reference user, is used for the reference the when active user recommended.

The process that the user is recommended has comprised visiting the link that behavioral data excavates and foundation excavation result carries out personalized recommendation.

Data mining method has classification, sequential analysis, cluster and correlation rule etc. multiple.

Particularly, during data mining, from database or data warehouse (Data Warehousing), obtain form, summary table or record document as the data of excavating.Then for reducing data volume, at first the data of collecting are carried out data preparation (Data Cleaning), data integration (Data Integration) or data-switching (Data Transformation) etc., and guarantee the complete of user data information.Excavate multiple characteristics and the information that exists then from the data after the conversion, the knowledge of discovery can be used in aspects such as decision-making, flow process control, data management, searching and managing.

The in store historical data that does not have integration at random in the user's business database, in the embodiment of the present application, when user's visit behavioral data is analyzed, at first carry out data clusters (Data Clustering), divide into groups with reference to the user to a plurality of exactly, the similar user of visit behavior is divided into one group, so, user's similarity in the group is the highest, and the user's similarity between group and the group is minimum, can further be stored in classification integrated data storehouse according to affiliated grouping with reference to user and corresponding visit behavioral data thereof.

In the embodiment of the present application, according to the classification of RFM model to the reference user.The RFM model is made of three special key elements, and the last consumption time (Recency), consuming frequency (Frequency) and consumption amount of money data (Monetary) are in order to indicate this user's value information.

1, the last consumption time (Recency)

Time span when the last consumption time refers to that the user consumes distance analysis for the last time.Less when the Recency value, i.e. the time gap current time length of the last consumption of user hour can predict that the probability that this user can consume again is bigger, thereby it is higher in the last consumption time eigenwert.When the height with the last consumption time eigenwert decides user's principal characteristic degree, should consider the characteristic of this application, can not only rely on the significance level that the eigenwert height of borrowing the last consumption time just determines the user.

2, consuming frequency (Frequency)

Consuming frequency refers to that the user consumes the number of times of this product within a certain period of time.Generally speaking, get over for a long time when user's consumption number of times, can think that its user is worth and loyalty is higher.Otherwise, think that then its user is worth and loyalty is lower.

3, consumption amount of money data (Monetary Amount)

The consumption amount of money refers in a period of time the total charge of this product of customer consumption.Generally speaking, when user's the consumption amount of money was more high, it was more high to think that its user is worth.Yet, consider that new user exists that the consumption number of times is very few, the consumption amount of money low excessively situation, when handling the consumption amount of money, replaces with the average consumption amount of money usually.

Based on above RFM model, at first, the day of consumption recently is more little, just illustrates that the time gap of the last consumption of user is very near now, so be considered to the user that might consume again, user's height that this group user's eigenwert is longer than the last consumption time; Secondly, consuming frequency is more high, and then the user is just more high to the loyalty of product, and this user is also important relatively, and therefore, the user characteristics value of high consumption frequency is higher; At last, the user that amount of money summation is more high, it is also just more high with respect to system's importance; Otherwise summation is worth more low user characteristics value just relatively can be very not high.

By above three kinds of characteristic indexs are carried out comprehensively can obtaining each user to the value of system or product, can classify to the user according to each user's value then, obtain system or the different a plurality of groupings of value of the product.

By above-mentioned RFM model reference user's visit behavioral data is analyzed in the embodiment of the present application, so that the reference user is classified.In the embodiment of the present application, each application has set in advance corresponding applicating category, is example with the film, and " rifle king's king " belongs to feature film, and " handou sir " belongs to comedy.Recorded the user in the visit behavioral data to the access time of application and the applicating category under the access application.

Accordingly, in the embodiment of the present application, with time of the last access application of user as the last consumption time, access frequency in a period of time is used as consuming frequency, to be used as the consumption amount of money to total access times of each applicating category, each category file is exactly the unit amount of money, and the accessed number of times of a certain category file is more many, also representing the user can be on this category file spended time.

Further, the step that the reference user is divided into groups can comprise:

Substep S11, at each with reference to the user, the access time that comprises in the behavioral data according to visit and the applicating category under the access application calculate the number of current nearest access time, access frequency and the access application of distance;

Substep S12, add up all mean values with reference to user's the nearest access time, the mean value of access frequency, and the mean value of the number of access application;

Substep S13, at each with reference to the user, the number of nearest access time, access frequency and access application is compared with corresponding average respectively, will identical reference user be divided in the grouping according to the comparative result of three numerical value.

Be example with the video, with the user recently with this time of watching video as the last consumption time, will watch frequency as consuming frequency in a period of time, will be to all kinds of videos always watch number of times as the consumption amount of money.

At first, at each with reference to the user, according to the applicating category under the application of the access time that comprises in the visit behavioral data and visit, determine the number of current nearest access time, access frequency and the access application of distance, provided certain visit behavioral data with reference to user U1 as following table 2, in this segment record, user U1 is that the application of A and B is visited on Dec 12nd, 2011 to using classification, is that the application of A and D is visited on January 9th, 2012 to using classification.

Table 2, The history of user U1

Three desired values of RFM of calculating user U1 respectively are as follows:

The last video-see time (R): suppose that the present date is 2012/01/28, and last the watching of user U1 being 2012/01/09, is 2012/01/28-2012/01/09=20 at a distance of fate;

Watch frequency (F): 20;

Watch number (M): 42.

Identical with the statistical method of the RFM value of user U1, all users' RFM value can be obtained, and all can be further added up with reference to the mean value of user's nearest viewing time, watch the mean value of frequency, and the mean value of watching the number of application.Each is compared with reference to user's RFM value and all users' mean value, can each be divided into groups with reference to the user according to result relatively, particularly:

With ↑ expression value greater than population mean, and ↓ represent less than population mean, the comparative result of RFM value is divided into eight types: ↑ ↑ ↑, ↑ ↑ ↓, ↑ ↓ ↑, ↑ ↓ ↓, ↓ ↑ ↑, ↓ ↑ ↓, ↓ ↓ ↑ and ↓ ↓ ↓, and according to the classification of comparative result, corresponding is divided into eight groups with a plurality of with reference to the user, each is done a comparison with reference to user's RFM value and mean value, can find out each thus with reference to user's group type, with each with reference to user grouping in the group that meets.

Provided a plurality of group result with reference to the user as following table 3:

Table 3, The expression of RFM value by groups

According to the RFM value each is divided into groups with reference to the user, obtained at the different a plurality of groupings of a plurality of using values that preset.

Step 102, according to the visit behavioral data of targeted customer to a plurality of application, determine to divide into groups under the targeted customer, and the statistics targeted customer satisfies the access weight value applicating category of preset range as the interest applicating category to the access weight value of different applicating categories.

Be divided in a plurality of groupings with reference to the user a plurality of according to the RFM model in the step 101, at targeted customer to be recommended, determined to belong to which grouping with reference to the user earlier, and definite targeted customer's interest applicating category.

Concrete, according to the visit behavioral data of targeted customer to a plurality of application, determine that the step of dividing into groups under the targeted customer can comprise:

Substep S21, at the targeted customer, the application access time that comprises in the behavioral data according to visit and the applicating category under the access application calculate the number of current nearest access time, access frequency and the access application of distance;

Substep S22, with the number of targeted customer's nearest access time, access frequency and access application, mean value with the number of the mean value of all mean values with reference to user's the nearest access time, access frequency and access application compares respectively, and the grouping that comparative result is identical is as dividing into groups under the targeted customer.

According to method same in the step 101, can from the user's business database, extract the visit behavioral data of targeted customer's correspondence, according to the access time that comprises in the visit behavioral data and the applicating category of visit, can determine the number of current nearest access time, access frequency and the access application of targeted customer distance, RFM data and all RFM mean values with reference to the user with the targeted customer compare then, according to comparative result the targeted customer are divided in the corresponding grouping and go.For example, three of the RFM of user U9 desired values are as follows:

The last video-see time (R): suppose that the present date is 2012/01/28, and last the watching of user U9 being 2011/12/25, is 2012/01/28-2011/12/25=34 at a distance of fate;

Watch frequency (F): 40;

Watch number (M): 65.

Be respectively with reference to user RFM mean value: R=26.75, F=25, M=40, contrast as can be known targeted customer and all comparative results with reference to user's mean value can be expressed as R ↑ F ↑ M ↑, can be divided in the grouping of U7.

In the embodiment of the present application, also need to determine the interest applicating category at the targeted customer.Concrete, the statistics targeted customer is to the access weight value of different applicating categories, and the applicating category that the access weight value is satisfied scope can comprise as the step of interest applicating category:

Substep S31, at each applicating category that the targeted customer visits, according to the classification weighted value of preset algorithm and applicating category, add up the access weight value of each applicating category;

Substep S32, removal access weight value are formed the binomial classification less than the applicating category of presetting weighted value with any two applicating categories;

Substep S33, according to the classification weighted value of described preset algorithm and binomial classification, the access weight value of each binomial classification that statistics is visited;

Substep S34, extract the access weight value greater than the binomial classification of default weighted value as the interest applicating category.

The interest applicating category refers in a plurality of different application classifications of user visit, and the applicating category that the access weight value is higher adds the access weight value and can obtain an accurate category of interest, rather than approximately or increase other category item.When determining targeted customer's interest applicating category, the access weight value of at first adding up each applicating category is removed access weight value smaller applications classification, keeps the higher applicating category of access weight value.

In the embodiment of the present application, after obtaining the higher applicating category of access weight value, also need further applicating category to be formed the binomial classification in twos, add up the access weight value of each binomial classification then, the binomial classification that weighted value is bigger is as targeted customer's interest applicating category.Two project categories are formed a binomial classification, can compare the similarity of each project category, the higher binomial classification of access weight value also is the higher project category of similarity.

When the access weight value of the single applicating category of statistics and binomial classification, preestablished corresponding classification weighted value at each applicating category, also preestablish corresponding classification weighted value at each binomial classification, when calculating the access weight value, the statistics targeted customer calculates corresponding access weight value to using the access times of classification or binomial classification according to preset algorithm and classification weighted value.

In the embodiment of the present application, preset algorithm can be the Apriori algorithm, and the Apriori algorithm is the classic algorithm of finding the correlation rule field.

The Apriori algorithm finds that the process of correlation rule is divided into two steps: the first step retrieves all frequent item sets in the transaction database by iteration, and namely support is not less than the item collection of user's preset threshold; Second step utilized frequent item set to construct to satisfy the rule of user's the minimum confident degree.Specific practice is exactly: at first find out frequent 1-item collection, be designated as L1; Utilize L1 to produce candidate C2 then, the item among the C2 is judged excavated L2, be i.e. frequent 2-item collection; Constantly so circulation is gone down till can't finding more frequent k-item collection.Every excavation one deck Lk just needs scanning entire database one time.

In the embodiment of the present application, the step of adding up the access weight value of each applicating category can comprise:

Substep S31-1 at each applicating category of targeted customer, extracts current nearest j * N visit of distance behavioral data, and j and N are positive integer, and the initial value of j is 1;

Substep S31-2 according to the definite access times to each applicating category of described visit behavioral data, according to the classification weighted value Wj of preset algorithm and applicating category, upgrades the access weight value WSj of each applicating category, and the initial value of WSj is 0;

Substep S31-3 is if described classification weighted value has been got then shut-down operation smaller or equal to default weighted value or described visit behavioral data; If described classification weighted value has not been got greater than default weighted value and described visit behavioral data, then make j=j+1, and return substep S31-1.

Comprise all visit behavioral datas of targeted customer in the user's business database, when data volume is bigger, if the whole visit behavioral data of statistics can cause the increase of system's execution time and cost, influenced the instantaneity function of ecommerce; And the user not necessarily can be centered around identical classification style to the selection of using recently always.Therefore, in a preferred embodiment of the present application, can add up near present visit behavioral data, though may there be the last and last visit situation too of a specified duration of being separated by, but but these visits also are within user's interest, can't change because of the time.

Further, the embodiment of the present application also adds the Apriori algorithm with the thought of IMW (Incremental Mining based on Weight, weight increment).The speed that early time data increases is slow, after carrying out a data mining, excavates the result and can use for a long time; Nowadays the growth rate of data is very fast, after the data mining, excavates the standing state that the result often can not represent data, by the mode of weight increment, is exactly on original pattern, excavates again in conjunction with newly-increased data.

More specifically, the weight increment refers to not satisfy certain when pre-conditioned the data of once extracting being carried out statistics, extracts another batch data again and together adds up until the result with the last data of extracting and reach pre-conditioned.Adopt the Apriori algorithm after improving can find out useful nearest interest applicating category, namely customary rule (Recent behavior rules-Rbr) is recommended accurately recently.

Utilize increment to excavate to save the time of whole excavations and can excavate dynamically and watch custom recently.The difference of increment number of times, resulting result arranges also can be different, thus number and degree of accuracy that influence is recommended, so pre-conditioned arranging is extremely important.

When the access weight value of each applicating category of adding up targeted customer's visit, at first extract current nearest j * N visit of distance behavioral data, wherein j and N are positive integer, the initial value of j is 1, that is, in first time during incremental computations, extract current nearest N of distance earlier and visit behavioral data; In second time during incremental computations, extract current nearest 2N of distance and visit behavioral data, by that analogy.

Foundation visit behavioral data can be determined the access times to each applicating category, be example with table 2, be the visit situation of user U1 in certain accrual accounting to using, the access times of the A of user U1, B, C, D, four applicating categories of E are 42, be 12 to other access times of category-A, can determine other access weight value of category-A according to access times and other classification weighted value of category-A, computing formula is as follows:

{WS}_{i}^{j} = {WS}_{i}^{j - 1} + ({Count}_{i}^{j} \times W_{j})

J is the number of times that increment excavates,

The access weight value that is i applicating category in the j time incremental computations, W _jBe the applicating category weight,

W _j=β ^J-1, j=1,2 ... n, wherein, β is a constant, and β＜1,

It is the occurrence number summation that i category file appears at the j time increment transaction.

In the embodiment of the present application, with classification weighted value W _jSize pre-conditioned as incremental computations, classification weighted value W _jChange along with the increment number of times, and be the trend that reduces gradually.After an incremental computations, judge that whether this classification weighted value in calculating is smaller or equal to default weighted value, if, then do not carry out incremental computations next time, directly carry out substep S32, perhaps the used data of this incremental computations have been all visit behavioral datas of targeted customer, and the visit behavioral data of namely working as the targeted customer has extracted and has been over, the incremental computations of next time can't be carried out, also substep S32 can be directly carried out; Otherwise, if this classification weighted value in calculating is greater than default weighted value, or visits behavioral data and also do not got, then can further extract more visit behavioral data, carry out incremental computations next time.

In the embodiment of the present application, the step of the access weight value of statistics each binomial classification of visiting can comprise:

Substep S33-1 at each binomial classification of targeted customer, extracts current nearest j * N visit of distance behavioral data, and j and N are positive integer, and the initial value of j is 1;

Substep S33-2 according to the definite access times to each binomial classification of described visit behavioral data, according to the classification weighted value Wj ' of preset algorithm and binomial classification, upgrades the access weight value WSj ' of each binomial classification, and the initial value of WSj ' is 0;

Substep S33-3 is if described classification weighted value has been got then shut-down operation smaller or equal to default weighted value or described visit behavioral data; If described classification weighted value has not been got greater than default weighted value and described visit behavioral data, then make j=j+1, and return substep S33-1.

The binomial classification is made of in twos applicating category, for example, in film is recommended, screening through step S31 and step S32, the applicating category that obtains comprises feature film, horror film and comedy, and the binomial classification that obtains of combination can comprise { feature film-＞horror film }, { feature film-＞comedy } and { horror film-＞comedy } in twos.

According to the same computing method of step S31, to form the summation of access times of two applicating categories of binomial classification as the access times of binomial classification, according to the Apriori algorithm, calculate the access weight value of each binomial classification according to the classification weighted value of each binomial classification, when the classification weighted value when having got, then finishes incremental computations smaller or equal to default weighted value or visit behavioral data; When the classification weighted value when not got, then continues next incremental computations greater than default weighted value or visit behavioral data.

In order to make those skilled in the art understand the scheme of the embodiment of the present application better, provide the process of determining targeted customer's interest applicating category in the step 102 herein:

1. in the user's business database, extract targeted customer's nearest n pen visit behavioral data, and calculate the access times of each applicating category i, when fetching data for the first time, j=1;

2. calculate the access weight value of each applicating category And judgement W _jValue β ^J-1Value whether smaller or equal to default weighted value, or the visit behavioral data got;

3. if the condition that does not meet in 2 is just got next n pen visit behavioral data again, and recomputates again

Value is up to W _jValue got smaller or equal to default weighted value or visit behavioral data, stop to calculate;

4. deletion Be zero or less than the applicating category of default weighted value, and remaining applicating category is combined into the binomial candidate collection, repeat the access weight value that above-mentioned mode is calculated each binomial collection classification i, until W _jValue β ^J-1Value smaller or equal to default weighted value, or the visit behavioral data got;

5. deletion

Be zero or less than the binomial classification of default weighted value;

6. last remaining binomial classification will be considered as user's interest applicating category.

At the step of above screening interest applicating category, provide a concrete example herein and illustrate.

Suppose the increment amount of watching n=2, β=0.4, default weighted value Min-Support=1.5, work as β ^J-1Smaller or equal to 0.1 o'clock, stop incremental computations.

Interest applicating category Rbr can be expressed as:

Rbr = {[m &RightArrow; n] | Min - Support \leq {WS}_{i}^{j}

([m→n])，β ^j-1＜0.1}

Be that interest applicating category [m → n] satisfies the access weight value greater than default weighted value (default minimal weight), increment number of times j satisfies β ^J-1＜0.1.

Classification m and n are respectively A, B, C, D, E is representative.By last table 1 as can be known U1 20 user access activity data (namely watching video data) are arranged as historical record, at first excavate the higher applicating category of access weight value, suppose to get earlier at the beginning last two and watch video data (n=2), i.e. T20, T19.Calculate the number of times of each applicating category and calculate the access weight value by formula

And suppose to also have undrawn visit behavioral data.

Table 4, weight incremental computations for the first time

The first calculation of incremental weight

After first time incremental computations, because β ^J-1Greater than 0.1, therefore, must add other two affairs project T18, T17 again, carry out the incremental computations second time, as shown in table 5.

Table 5, weight incremental computations for the second time

The second calculation of incremental weight

After second time incremental computations, because β ^J-1Still greater than 0.1, must add other two affairs project T15, T16 again, carry out incremental computations for the third time, as shown in table 6.

Table 6, weight incremental computations for the third time

Table4-6The third calculation of incremental weight

Table 6 is the result of weight increment for the third time, wherein, and β ^J-1The value next iteration when calculating just smaller or equal to preset threshold, therefore stop incremental computations, because the access weight value of B, E applicating category equals 0, therefore with B, the deletion of E applicating category, residue A, C, D applicating category are the higher applicating category of access weight value.

With A, C, D category item be combined into the binomial classification A-＞C}, A-＞D}, C-＞D}, as shown in table 7.And the calculation times of binomial classification is to have done the incremental computations of how many times altogether for stopping according to (j=3) according to applicating category, as long as increment just stops so be calculated to for the third time, purpose is consistent with the increment of a collection, avoids in the incremental process of binomial collection, because of β ^jBe worth too little and do not have the meaning calculated, and constantly carry out incremental computations, also can waste unnecessary time and cost.

Table 7, binomial classification be the weight incremental computations for the first time

The first calculation of incremental weight of 2-itemsets

Table 8 binomial classification is the weight incremental computations for the second time

The second calculation of incremental weight of 2-itemsets

Table 9 binomial classification is the weight incremental computations for the third time

The third calculation of incremental weight of 2-itemsets

Can find out calculation process by table 4-8, table 4-9, remaining binomial classification be { A-＞C}, { A-＞D} and { three rules of C-＞D}, and { this is low excessively because of support for C-＞D}, therefore deletes.{ A-＞C}, { A-＞D} is as user's interest applicating category, and wherein, { the access weight value maximum of A-＞D} can be used as the applicating category of preferably recommending the targeted customer with the binomial classification at last.

Draw at last rule A-＞C}, A-＞D}, and therefore after recommendation process in, { A-＞D} related category product will recommend more number to the user.Excavate than traditional increment, the embodiment of the present application is after incremental computations each time, remove access weight value smaller applications classification or binomial classification, thereby the data volume that can solve the incremental mode excavation is bigger, expend long shortcoming of system resource and excavation time, wherein the β value can be adjusted according to experience or through repeatedly calculating, thereby obtains excavating more accurately the result.

In step 103, each a plurality of application of visiting with reference to the user of under the targeted customer, dividing into groups, extract the application that belongs to described interest applicating category and recommend the targeted customer.

In step 102, determined grouping and the interest applicating category under the targeted customer, when recommending, can further under the user, extract this its application of visiting with reference to the user of dividing into groups in the grouping, in then these being used, the user is recommended in the application that belongs to the user interest applicating category.

As above routine, the interest applicating category that draws the targeted customer is { A-＞C}, { A-＞D}, under be grouped into the grouping of U7, comprise U7 in the grouping of U7, two of U11 are with reference to the user, the applicating category that U7 visits comprises B and C, the applicating category that U11 visits comprises D and E, therefore, can recommend to meet C and the D of targeted customer's interest applicating category, in the application that U7 and U11 are visited, the targeted customer is recommended in the application that meets these two applicating categories, because binomial classification { A-＞C} that the A that applicating category D and targeted customer visited forms, { the access weight value of A-＞C} is big, therefore, can preferably recommend the application of D applicating category to recommend the targeted customer for the binomial classification that the A that visited than C and targeted customer forms.In concrete realization, can not do recommendation for the application that the targeted customer had visited.

For example, in video is recommended, suppose U ₁Other similar people having the same habits user U on the same group ₂, U ₃, three user-selected videos are numbered vt, represent t video in the video database.U ₁Interest video classification be A, B and C, U ₂The video classification of watching comprises A, B, C and D, U ₃The video classification of watching comprises A, B, D and E.For U ₂, A, B, three kinds of video classifications of C are recommended U ₁, specifically comprise { A:v1, v2, B:v6, C:v9}; For U ₃, two kinds of video classifications of A, B are recommended U ₁, specifically comprise { A:v2, v4, B:v6, v8}, comprehensive U ₂With U ₃Both results, the unduplicated U that recommends ₁, i.e. { A:v1, v2, v4, B:v6, v8, C:v9}.

In concrete the recommendation, can be with the application recommended and relative information displaying in client, make things convenient for the user to select and visit.Be example to recommend video, with the form of recommending the content tabs such as length, style, video name, countries and regions, age of video and video with tabulation, interface in client is showed, browse by the interface, the user learns the details of video, and the user can check list of videos and watch the video of oneself liking.

In a preferred embodiment of the present application, before step 103, described method can also comprise:

In the embodiment of the present application, in can also the grouping under the targeted customer, further determine that interest applicating category similarity with the targeted customer is than higher reference user, find out group more similar between the user, be equivalent to the reference user is carried out again cluster analysis, similarity can think to possess with the targeted customer the similar user of common hobby greater than this class user of preset value.

Particularly, at under the targeted customer grouping in each with reference to the user, can according to step 102 in the statistics targeted customer interest application class method for distinguishing, count the interest applicating category with reference to the user, described in the concrete step such as step 102, repeat no more herein.

After counting each interest applicating category with reference to the user, targeted customer's interest applicating category and interest applicating category with reference to the user are compared, add up both similarities.

Concrete, calculate targeted customer's interest applicating category, can comprise with each step with reference to the similarity of user's interest applicating category of affiliated grouping:

Can make up corresponding rule vector according to targeted customer's interest applicating category with reference to user's interest applicating category, in this step, what at first will carry out is the structure of similar matrix, particularly, is user U with SM (X) _xThe interest applicating category of excavating by the Apriori algorithm (is nearest customary rule, Rbr, Recent behavior rules) launches the similar matrix (Similar Matrix) be combined into, it is 0 that the interior i of matrix is listed as the capable element of j, Rbr is the set of nearest customary rule [i → j], shown in the following formula:

SM(X)＝[smij]m×n；(m，n＝1，2，3……，p)

Where

{sm}_{ij} = \{\begin{matrix} 1, & if [i &RightArrow; j] &Element; Rbr \\ 0, & Otherwise \end{matrix}

Suppose user U _xThe applicating category of visit comprises A, B, C, D and E, and user U _xThe interest applicating category is { A → C}, { A → D}.So in matrix A, C}, A, and D} inserts 1, and other still is 0.As shown in table 10 below, be user U _XSimilar matrix represent.

Table 10 user U _XSimilar matrix is represented

The Similar Matrix of U _X

Next carry out the calculating of similar vector.Similar matrix converts similar vector (Similar vector) to, and similar vector is defined as follows:

SV(X)＝[sm ₁₂ sm ₁₃…sm _1n sm ₂₃ sm ₂₄ sm _2n……sm _(m-1)n]

Via top conversion, the user U after table 10 conversion _XSimilar vector be 0,110 000 00 0}, and try to achieve similar vector after, just can carry out the similarity contrast between the user.

{Similar}_{xy} = (X_{1} &CirclePlus; Y_{1}) + (X_{2} &CirclePlus; Y_{2}) + . . . . . . + (X_{1} &CirclePlus; Y_{1}) - - - (1 - 1)

Similarset(X)＝{Uy|Similarxy≤θ} (1-2)

Formula (1-1) is to calculate user U _XWith user U _YBetween similarity (Similar _Xy), X wherein ₁And Y ₁Refer to the 1st element of similar vectorial X or Y here, Be mutual exclusion (be all 0 mutually, inequality is 1).Formula (1-2) is set classification thresholds, that is to say preset value θ, and the size of classification thresholds influences the change of similar users, can cause the difference of recommendation results, in concrete realization, can set up on their own according to the demand of applied environment.

With the user U of similar value smaller or equal to threshold value _YBe classified as user U _XSimilar set (Similarset (X)), namely with the reference user set of targeted customer's similarity greater than preset value, assemble via secondary classification, thereby further find the reference user higher with targeted customer's similarity, further improve the accuracy of recommending.

Below by concrete example the calculation of similarity degree process is described:

Suppose that category item i is respectively A, B, C, D and E, threshold value θ=2.User U _XThe interest applicating category be that { A-＞C}, { A-＞D} comprises with reference to the user U in the affiliated grouping _Y, U _ZAnd U _WAs following table 11, according to U _XSimilar vector { 0,110 000 00 0}, Gou Zao the U after the same method of interest applicating category structure _Y, U _ZAnd U _WSimilar vector be respectively 0,110 001 00 0,0,110 001 101 and 0,110 000 01 1.

Utilize formula (1-1) then, with U _XBe benchmark, contrast and calculate similar value with all in twos with reference to the user.See Table 11, obtain similar value after, distinguish by classification thresholds θ=2, similarity greater than preset value be considered as similar.Therefore know and find out U _XWith U _Y, U _WFor similar similar, and U _ZThen be considered as dissimilar class.

Table 11, U _XAnd the similar value between all users

The similar of U _X and other users

Further, described step 103 can comprise:

Substep S41, search grouping under the targeted customer each with reference among the user, a plurality of application that similarity is visited greater than the reference user of preset value;

Substep S42, in a plurality of application of searching, extract the application belong to described interest applicating category and recommend the targeted customer.

Find out the higher user of similarity among the reference user who under the targeted customer, divides into groups, when using recommendation, this part can be recommended the targeted customer with reference to the application that the user visits, concrete, at the application that the higher reference user of similarity visits, the application fetches that will meet targeted customer's interest applicating category is come out, and recommends intended application, so can further improve the accuracy of recommendation, reach the effect of the sharing information of real collaborative filtering method.

In a preferred embodiment of the present application, with reference to a plurality of groupings of user's division visit behavioral data according to the reference user who comprises, can be divided into high value grouping and low value and divide into groups.

For example, user U3 watch the number of video many, watch the frequency height, recently viewing time is near, show that user U3 probably watches video and higher to the access frequency of this video again, be a high-value user, the high-value user often visits, the application of recommending according to its historical visit data relatively conforms to its interest preference, recommends accuracy higher; And user U2, can find out be one to using the less user of visit, because it watches the number of times of video less, and the time of watching recently, early the video number of watching was also fewer, and this class user may not find more interested video, can not react its interest preference comparatively accurately according to its historical visit data, little with reference to property, the accuracy of recommendation is lower, is a low value user.

In a preferred embodiment of the present application, at high-value user and low value user, can adopt different recommendation strategies, for example user U3 belongs to the high-value user, can be according to its interest applicating category of visit behavior data analysis of its history, the targeted customer is recommended in the application that the reference user of affiliated grouping is met the interest applicating category.And user U2 belongs to the low value user, also unreliable according to the interest applicating category that its visit behavioral data draws, therefore, on the recommendation strategy, will recommend with a large amount of and multi-class mixing, for example, at described low value grouping, the application that meets the interest applicating category except the reference user who divides into groups under extracting is recommended, can also extract the application of at least one default applicating category, together recommend the targeted customer.

For example, user U1 belongs to high value group, and it excavates the result be that { A-＞C}, { A-＞D} two kinds of rules can be recommended user U1 with the reference user's of affiliated grouping relevant A, C, the video of D, reach accurate individualized video recommendation; If user U1 belongs in the low value group, then not only recommend relevant video, also to increase other classification video, make the user can obtain diversified recommendation, find own interested applicating category, from the angle of website, can allow the user of this group can have an opportunity to land use more and increase historical record data, for the user provides better personalized recommendation service and better visit experience.

Divide for 8 groupings with reference to the user according to the prize of RFM model in the step 101, therefore, in concrete realization, can at these 8 groupings corresponding recommendation strategy be set respectively according to practical application request, the application does not do restriction to this.

In the embodiment of the present application, visit behavioral data in conjunction with RFM model and user, value or the identical reference user of behavior are classified as same grouping, weight increment method for digging is joined the interest applicating category that comes digging user in the Apriori algorithm, use the method for collaborative filtering recommending at last, determine the grouping under the targeted customer, the targeted customer is recommended in the application that the reference user of affiliated grouping visits.

Need to prove, for aforesaid method embodiment, for simple description, so it all is expressed as a series of combination of actions, but those skilled in the art should know, the application is not subjected to the restriction of described sequence of movement, because according to the application, some step can adopt other orders or carry out simultaneously.Secondly, those skilled in the art also should know, the embodiment described in the instructions all belongs to preferred embodiment, and related action might not be that the application is necessary.

Based on the explanation of said method embodiment, the application also provides corresponding application recommendation apparatus embodiment, realizes the described content of said method embodiment.

With reference to Fig. 2, it shows the apparatus structure block diagram that the described a kind of application of the embodiment of the present application is recommended.

Data acquisition module 201 is used for obtaining a plurality of with reference to the visit behavioral data of user to a plurality of application of presetting, and a plurality ofly divides into groups with reference to the user to described according to described visit behavioral data, and each application possesses corresponding applicating category;

Grouping determination module 202 is used for according to the visit behavioral data of targeted customer to a plurality of application, determines to divide into groups under the targeted customer;

Interest applicating category determination module 203 is used for the statistics targeted customer to the access weight value of different applicating categories, the access weight value is satisfied the applicating category of preset range as the interest applicating category;

Use recommending module 204, a plurality of application that each that is used for dividing into groups under the targeted customer visited with reference to the user are extracted the application that belongs to described interest applicating category and are recommended the targeted customer.

In a preferred embodiment of the present application, described device can also comprise:

With reference to the user interest statistical module, be used for according to each with reference to the visit behavioral data of user to a plurality of application, statistics each with reference to the access weight value of user to different applicating categories, the access weight value is satisfied the applicating category of preset range as the interest applicating category;

Similarity calculation module is for the interest applicating category that calculates the targeted customer, with each similarity with reference to user's interest applicating category of affiliated grouping;

With reference to user's determination module, be used for determining that similarity is greater than the reference user of preset value.

In a preferred embodiment of the present application, described application recommending module 305 comprises:

Application fetches submodule, each that is used for searching grouping under the targeted customer be with reference to the user, a plurality of application that similarity is visited greater than the reference user of preset value;

Recommend submodule, be used in a plurality of application of searching, extract the application that belongs to described interest applicating category and recommend the targeted customer.

In a preferred embodiment of the present application, a plurality of groupings of described division with reference to the user are divided into high value grouping and low value and divide into groups according to the reference user's who comprises visit behavioral data;

At described low value grouping, described application recommending module also is used for:

Extract the application and the application that belongs to described interest applicating category of at least one default applicating category, together recommend the targeted customer.

In a preferred embodiment of the present application, comprise the applicating category that access time and access application are affiliated in the described visit behavioral data, described grouping determination module comprises:

Add up submodule with reference to the user, be used at each with reference to the user, according to the access time that comprises in the visit behavioral data and the applicating category under the access application, calculate the number of current nearest access time, access frequency and the access application of distance;

Mean value statistics submodule is used for all mean values with reference to user's the nearest access time of statistics, the mean value of access frequency, and the mean value of the number of access application;

The first contrast submodule, be used at each with reference to the user, the number of nearest access time, access frequency and access application is compared with corresponding average respectively, will identical reference user be divided in the grouping according to the comparative result of three numerical value.

In a preferred embodiment of the present application, described grouping determination module comprises:

The targeted customer adds up submodule, is used at the targeted customer, according to the application access time that comprises in the visit behavioral data and the applicating category under the access application, calculates the number of current nearest access time, access frequency and the access application of distance;

The second contrast submodule, be used for the number with targeted customer's nearest access time, access frequency and access application, mean value with the number of the mean value of all mean values with reference to user's the nearest access time, access frequency and access application compares respectively, and the grouping that comparative result is identical is as dividing into groups under the targeted customer.

In a preferred embodiment of the present application, described interest applicating category determination module comprises:

First weight statistics submodule for each applicating category of visiting at the targeted customer, according to the classification weighted value of preset algorithm and applicating category, is added up the access weight value of each applicating category;

The binomial classification is formed submodule, is used for removing the access weight value less than the applicating category of default weighted value, and any two applicating categories are formed the binomial classification;

Second weight statistics submodule is used for the classification weighted value according to described preset algorithm and binomial classification, the access weight value of each binomial classification that statistics is visited;

The binomial classification is extracted submodule, be used for extracting the access weight value greater than the binomial classification of default weighted value as the interest applicating category.

In a preferred embodiment of the present application, described first weight statistics submodule comprises:

First subelement is used for each applicating category at the targeted customer, extracts current nearest j * N visit of distance behavioral data, and j and N are positive integer, and the initial value of j is 1;

Second subelement is used for according to the classification weighted value Wj of preset algorithm and applicating category, upgrading the access weight value WSj of each applicating category according to the definite access times to each applicating category of described visit behavioral data, and the initial value of WSj is 0;

The 3rd subelement is used for if described classification weighted value has been got then shut-down operation smaller or equal to default weighted value or described visit behavioral data; If described classification weighted value has not been got greater than default weighted value and described visit behavioral data, then make j=j+1, and return first subelement.

In a preferred embodiment of the present application, described second weight statistics submodule comprises:

The 4th subelement is used for each the binomial classification at the targeted customer, extracts current nearest j * N visit of distance behavioral data, and j and N are positive integer, and the initial value of j is 1;

The 5th subelement is used for according to the classification weighted value Wj ' of preset algorithm and binomial classification, upgrading the access weight value WSj ' of each binomial classification according to the definite access times to each binomial classification of described visit behavioral data, and the initial value of WSj ' is 0;

The 6th subelement is used for if described classification weighted value has been got then shut-down operation smaller or equal to default weighted value or described visit behavioral data; If described classification weighted value has not been got greater than default weighted value and described visit behavioral data, then make j=j+1, and return step 21.

In a preferred embodiment of the present application, described similarity calculation module comprises:

The rule vector is set up submodule, be used at each of affiliated grouping with reference to the user, according to targeted customer's interest applicating category with set up respectively with reference to user's interest applicating category corresponding regular vectorial;

Vector similarity calculating sub module is used for calculating with reference to user's rule vector and similarity with reference to user's rule vector, as the similarity of reference user and targeted customer's interest applicating category.

In a preferred embodiment of the present application, described device also comprises:

Data recordin module is used for when the reference user conducts interviews to a plurality of application of presetting, and record is with reference to user's visit behavioral data.

With reference to figure 3, show and use the process synoptic diagram of recommending in the embodiment of the present application.

The method that the application's application is recommended can be used for the application of webpage recommends, and when user's browsing page, shows the application of recommending at webpage; Also the application that can be used on the program interface is recommended, and when user's visit can be used the program of recommendation, or shows the application of recommending on the program interface.

Before using recommendation, preset user's business database and similar vectorial comparison database, the user's business database is used for each user's of record visit behavioral data, and similar vectorial comparison database is used for the similarity of two vectors of statistics.

The process of use recommending comprises the four-stage of carrying out successively as lower module or database:

User's business database 301:

The visit behavioral data that is used for each user of record.

User grouping module 302:

User grouping module is used for extracting a plurality of with reference to user's visit behavioral data in the past from the user's business database, it is user's historical data, and according to the RFM model, by adding up these visit behavioral datas, be divided into a plurality of groupings with a plurality of with reference to the user, and the result that will divide into groups deposits in the user and is worth in the integrated data storehouse, reaches data and integrates and distribute different resources on one's body the dissimilar users.

Customer documentation module 303:

The customer documentation module is used for extracting the targeted customer and with reference to user's historical data of user from the user's business database, by the weight increment being excavated the historical data in conjunction with Apriori algorithm statistics user, obtain each user's nearest interest, i.e. the interest applicating category.

Similar vectorial comparison database 304:

Similar vectorial comparison database is for the similarity of the reference user's of comparison targeted customer and affiliated grouping interest applicating category.

Collaborative filtering recommending module 305:

The collaborative filtering recommending module is used for searching the affiliated grouping of targeted customer, and the user is recommended in the application that the reference user in will dividing into groups then visited.When recommending, can carry out the collaborative filtering recommending that the user is worth on the one hand, namely different grouping is corresponding to possessing the different price value, according to the corresponding recommendation strategy of affiliated grouping the targeted customer is used recommendation; Can carry out the collaborative filtering recommending of similar users on the other hand, the reference user higher with targeted customer's similarity in the grouping recommends the user with similarity than the application that higher reference user visits under further finding out.

For above-mentioned application recommendation apparatus embodiment, because it is similar substantially to method embodiment, so description is fairly simple, relevant part gets final product referring to the part explanation of method embodiment shown in Figure 1.

Each embodiment in this instructions all adopts the mode of going forward one by one to describe, and what each embodiment stressed is and the difference of other embodiment that identical similar part is mutually referring to getting final product between each embodiment.

What those skilled in the art were easy to expect is: it all is feasible that the combination in any of above-mentioned each embodiment is used, so the combination in any between above-mentioned each embodiment all is the application's embodiment, but this instructions has not just described in detail one by one at this as space is limited.

Claims

1. use recommend method for one kind, it is characterized in that, comprising:

2. method according to claim 1 is characterized in that, in a plurality of application that each of dividing into groups under the targeted customer visited with reference to the user, extracts the application of described interest applicating category and recommends before the targeted customer, and described method also comprises:

3. method according to claim 2 is characterized in that, in a plurality of application that described each of dividing into groups under the targeted customer visited with reference to the user, extracts the step that the application that belongs to described interest applicating category recommends the targeted customer and is:

4. according to claim 1 or 3 described methods, it is characterized in that a plurality of groupings of described division with reference to the user are divided into high value grouping and low value and divide into groups according to the reference user's who comprises visit behavioral data;

5. method according to claim 1 is characterized in that, comprises the applicating category under access time and the access application in the described visit behavioral data, describedly according to the visit behavioral data a plurality of steps of dividing into groups with reference to the user is comprised:

6. method according to claim 5 is characterized in that, and is described according to the visit behavioral data of targeted customer to a plurality of application, determines that the step of dividing into groups under the targeted customer comprises:

7. method according to claim 1 is characterized in that, described statistics targeted customer is to the access weight value of different applicating categories, and the applicating category that the access weight value is satisfied scope comprises as the step of interest applicating category:

8. method according to claim 7 is characterized in that, described each applicating category of visiting at the targeted customer, and according to the classification weighted value of preset algorithm and applicating category, the step of adding up the access weight value of each applicating category comprises:

9. method according to claim 7 is characterized in that, described classification weighted value according to preset algorithm and binomial classification, and the step of the access weight value of each binomial classification that statistics is visited comprises:

10. method according to claim 2 is characterized in that, described calculating targeted customer's interest applicating category comprises with each step with reference to the similarity of user's interest applicating category of affiliated grouping:

11. method according to claim 1 is characterized in that, obtain a plurality of with reference to the step of user to the visit behavioral data of a plurality of application of presetting before, described method also comprises:

12. use recommendation apparatus for one kind, it is characterized in that, comprising: