CN103823908A - Method and server for content recommendation on basis of user preferences - Google Patents

Abstract

The invention provides a method and server for content recommendation on the basis of user preferences. The method comprises the steps that (1) preference index vectors of a user are calculated on the basis of historical preference information of the user; (2) content items to be recommended are selected from a recommendation list for selection on the basis of the preference index vectors so that a final recommendation list can be formed; (3) the final commendation list is provided for the user.

Description

Content recommendation method based on user preference and server

Technical field

The present invention relates to commending contents field, relate more specifically to content recommendation method and server based on user preference.

Background technology

In recent years, along with popularizing of the mobile terminal such as smart mobile phone, panel computer, mobile Internet has become our an indispensable part in producing, living.Mobile terminal has been no longer the terminal that a basic communication and information are transmitted, but becomes the entertainment applications terminal that people carry.This variation, has expedited the emergence of huge mobile application market industry, and people are the various application (for example, app) on a large amount of use mobile terminal in the time of daily study, work, rest.For example, in the time of study, can obtain various knowledge by educational app; In the time of work, can obtain various important Financial Information, news etc. or apply and obtain location/navigation feature etc. by map class by such as financial class app; In the time of rest, can appreciate film, TV play, music, books etc. by media class app.

Meanwhile, mobile terminal has also changed user's consumption pattern, consumption habit and consumer behavior.Pc user and mobile phone users have in time remarkable difference in their service of buying: the consumer on mobile terminal is generally more out of patience, always wish the thing that just can find them to want at once, pc user is relatively more patient, the removal search or wait for more cheap but good commidities and/or service of being ready to take time.

The exemplary of this respect is 82% the user who utilizes the predetermined hotel room of mobile terminal decision completing in 24 hours.This is almost just equivalent to user and has arrived destination just with mobile phone Lai Ding hotel.Compared with the user of predetermined hotel room on computers, these user's time spents want short many.This " impulse buying ", " instantaneity purchase " behavior of mobile phone users are that the one of the business model relatively leisurely to conventional internet is overturned in fact.For this new variation, enterprise need to help user to find them may interested commodity and recommend within the extremely short time, to capture the first chance of mobile marketing.Therefore just need a kind of content recommendation system that for example, carrys out the commodity/service that predictive user may pay close attention to according to existing data (, sales data).

In existing commending contents scheme, collaborative filtering is one of important commending contents algorithm.GroupLens has proposed collaborative filtering (referred to as CF-U, the i.e. Collaborative Filtering User-based) algorithm based on user in 1994.This algorithm is the collaborative filtering being applied the earliest, and it is mainly divided into three steps:

1) data statement.Conventionally first obtain user-project rating matrix R of a m × n, the capable representative of consumer number of m, n row represent item number, matrix element r _{i, j}represent the score value of user i to project j;

2) calculate user's similarity according to user-project rating matrix, according to similarity from big to small for active user tries to achieve an arest neighbors set N; And

3) produce recommending data collection.Obtain after k arest neighbors for active user, can, based on this k neighbour to the scoring of Arbitrary Term object, predict the scoring of active user to this project.Then according to the height of prediction scoring, from all projects of predicting scoring, select one or more projects as last recommendation results.

In this process, in order to obtain the k neighbour the most similar to active user's scoring, must calculate the similarity between user.Traditional method for measuring similarity is generally cosine similarity.User's scoring is treated as n-dimensional space vector.If user does not mark to project, user is made as 0 to the scoring of this project.Similarity between user a and b is measured as follows and is provided by the cosine angle between their scoring vectors:

$\mathrm{sim}(u_a,u_b)=\cos (u_a,u_b)=\frac{u_a*u_b}{\left|u_a\right|*\left|u_b\right|}=\frac{{\mathrm{\&Sigma;}}_{j=1}^n{r}_{a,j}*r_{b,j}}{\sqrt{{\mathrm{\&Sigma;}}_{j=1}^n{r}_{a,\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="HDA0000480172300000022.png"\; state="\{\{state\}\}">j</figure-callout>}}^2}*\sqrt{{\mathrm{\&Sigma;}}_{j=1}^n{r}_{b,\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="HDA0000480172300000022.png"\; state="\{\{state\}\}">j</figure-callout>}}^2}}$

Wherein, u _arepresent the n dimension scoring vector of user a, u _brepresent the n dimension scoring vector of user b, sim (u _a, u _b) expression u _aand u _bbetween similarity, cos (u _a, u _b) expression u _aand u _bbetween cosine similarity, || represent vectorial norm, r _{a, j}represent the scoring of user a to content item j.By traditional Forecasting Methodology, the scoring of predictive user a to project j, wherein N is the arest neighbors set of user a, the average score of user a,

the average score of user b:

$\mathrm{pred}(a,j^{\&prime;})=\stackrel{\&OverBar;}{r_a}+\frac{{\mathrm{\&Sigma;}}_{b\&Element;N}(\mathrm{sim}(a,b)*(r_{{b,j}^{\&prime;}}-\stackrel{\&OverBar;}{r_b}))}{{\mathrm{\&Sigma;}}_{b\&Element;N}\mathrm{sim}(a,b)}$

Wherein, pred (a, j ') represents the user a that the dopes scoring to project j '.

But along with the data volume in system is day by day huge, cause some current recommended technologies can not effectively make real-time recommendation.Meanwhile, a problem that is perplexing all the time commending system is local data's sparse property problem.Although the conceptual data amount of system is very big, to divide to each user, really very little, this has just caused cannot accomplishing accurate and effective in the problem of calculating user's similarity the ratio that its commodity of browsing and/or buying account in system total commodity number.

For example, in doing personalized recommendation to user a, according to for example above-mentioned algorithm, the personal behavior feature of user a is all embodied by his arest neighbors.In the situation that having adequate data, the large probability of arest neighbors ground can embody the interest preference of user a.But in the time that data are too sparse, the arest neighbors of the user a finding according to above-mentioned algorithm and the similarity of user a are in fact very low, the confidence level of carrying out the behavioural characteristic of representative of consumer a by arest neighbors is also just had a greatly reduced quality naturally.

In addition, above-mentioned algorithm is also usually ignored user to this important information of the preference of variety classes commodity, thereby has also had a strong impact on the validity of recommendation results.

At present, in commending system field, the application of comparative maturity comprises the commending system of Amazon and Netflix.Amazon commending system relates to e-commerce field, has used mixing proposed algorithm, and one is improved project-based collaborative filtering, and another kind is according to good friend's relation in user social contact network, the article of liking to user's commending friends on Amazon.Netflix commending system relates to online film lease, employing be a kind of project-based collaborative filtering of improveing after user behavior pattern that combines equally.These two is in the problem all having in varying degrees aspect above.And with respect to these two fields, in the field of mobile application market, because its concern being subject to is relatively less, therefore the achievement in research of user behavior feature is wherein also few, and Sparse Problem is more serious.

Summary of the invention item

In order to address the above problem, provide according to content recommendation method and the server based on user preference of the present invention.

According to a first aspect of the invention, provide a kind of content recommendation method based on user preference.The method comprises: (a) the historical preference information based on described user, calculates described user's preference function vector; (b), based on described preference function vector, from recommendation list to be selected, select the content item that will recommend, to form final recommendation list; And (c) provide described final recommendation list to described user.

In certain embodiments, described historical preference information is to obtain by the content item that in statistics predetermined amount of time, described user downloads.

In certain embodiments, described predetermined amount of time is 1 month.

In certain embodiments, step (a) also comprises: the content item of (a1) user who counts being downloaded is divided into one or more classifications; (a2), for the each classification in described one or more classifications, in the content item that calculating user downloads, belong to the number of such other content item and the total ratio of the content item that user downloads; And (a3) according to the corresponding ratio of all categories in described one or more classifications, form the preference function vector of described user in described predetermined amount of time.

In certain embodiments, described one or more classification comprises following at least one: audio frequency, video, books, theme, game and tool software.

In certain embodiments, described recommendation list to be selected generates for described user by proposed algorithm, and described proposed algorithm is following at least one: content-based recommendation; Based on the recommendation of collaborative filtering; Based on the recommendation of correlation rule; Based on the recommendation of effectiveness; Based on the recommendation of knowledge; And above-mentioned every combination in any.

In certain embodiments, the described recommendation based on collaborative filtering comprises following at least one: the collaborative filtering recommending (CF-U) based on user; Project-based collaborative filtering recommending (CF-I); Collaborative filtering recommending (CF-M) based on model; And above every combination in any.

In certain embodiments, step (b) comprising: (b1), according to the classification of each content item in described recommendation list to be selected, described recommendation list to be selected is divided into the sub-list of one or more classification; (b2), based on described preference function vector, from the sub-list of corresponding classification, select the forward content item of rank, to insert described final recommendation list; And (b3) based on non-selected content item in described recommendation list to be selected, adjust described final recommendation list.

In certain embodiments, step (b2) comprising: (b21) preference function based on corresponding with each classification in described preference function vector and the length of described final recommendation list, calculate the shared expection number of each classification in described final recommendation list; And (b22) for each classification, from the sub-list of corresponding classification, select the forward content item of corresponding expected numbers object rank, to insert described final recommendation list.

In certain embodiments, selection in step (b22) is one of following two kinds of modes: in a sub-list of classification, select a current content item ranking the first at every turn, and loop for the sub-list of all classification, wherein, skip the sub-list of classification of selected sky and the number of selected content item and reach the sub-list of classifying of corresponding expected numbers object; Or disposablely in the sub-list of each classification select corresponding expected numbers object content item, wherein, if the number of the content item in sub-list of classifying is less than corresponding expection number, all the elements item in sub-this classification list is all selected.

In certain embodiments, step (b3) comprising: if (b31) in the sub-list of described classification the number of content item be less than corresponding expection number, from described recommendation list to be selected, in non-selected residue content item, select the forward content item of rank, with by described final recommendation list completion, (b32) judge whether the difference between the scoring of content item the last in the scoring of the content item ranking the first in non-selected residue content item in described recommendation list to be selected and described final recommendation list is greater than predetermined threshold, if be greater than described predetermined threshold, exchange these two and repeating step (b32) until in the scoring of the content item ranking the first in non-selected residue content item in described recommendation list to be selected and described final recommendation list the difference between the scoring of the last content item be less than or equal to described predetermined threshold, to obtain described final recommendation list.

In certain embodiments, the content item in described final recommendation list is to arrange from high to low by the marking order of content item.

In certain embodiments, described content item at least comprises the following information: content name, content type and content marking.

According to a second aspect of the invention, provide a kind of commending contents server based on user preference.Described server comprises: preference function computing unit, for the historical preference information based on described user, calculates described user's preference function vector; Content recommendation selected cell for based on described preference function vector, is selected the content item that will recommend, to form final recommendation list from recommendation list to be selected; And recommendation list provides unit, for described final recommendation list is provided to described user.

In certain embodiments, described historical preference information is to obtain by the content item that in statistics predetermined amount of time, described user downloads.

In certain embodiments, described predetermined amount of time is 1 month.

In certain embodiments, described preference function computing unit also for: the content item of (a1) user that counts being downloaded is divided into one or more classifications; (a2), for the each classification in described one or more classifications, in the content item that calculating user downloads, belong to the number of such other content item and the total ratio of the content item that user downloads; And (a3) according to the corresponding ratio of all categories in described one or more classifications, form the preference function vector of described user in described predetermined amount of time.

In certain embodiments, described one or more classification comprises following at least one: audio frequency, video, books, theme, game and tool software.

In certain embodiments, described recommendation list to be selected generates for described user by proposed algorithm, and described proposed algorithm is following at least one: content-based recommendation; Based on the recommendation of collaborative filtering; Based on the recommendation of correlation rule; Based on the recommendation of effectiveness; Based on the recommendation of knowledge; And above-mentioned every combination in any.

In certain embodiments, the described recommendation based on collaborative filtering comprises following at least one: the collaborative filtering recommending (CF-U) based on user; Project-based collaborative filtering recommending (CF-I); Collaborative filtering recommending (CF-M) based on model; And above every combination in any.

In certain embodiments, described content recommendation selected cell also for: (b1) according to the classification of described each content item of recommendation list to be selected, described recommendation list to be selected is divided into the sub-list of one or more classification; (b2), based on described preference function vector, from the sub-list of corresponding classification, select the forward content item of rank, to insert described final recommendation list; And (b3) based on non-selected content item in described recommendation list to be selected, adjust described final recommendation list.

In certain embodiments, described content recommendation selected cell also for: (b21) based on the described preference function vector preference function corresponding with each classification and the length of described final recommendation list, calculate in described final recommendation list the shared expection number of each classification; And (b22) for each classification, from the sub-list of corresponding classification, select the forward content item of corresponding expected numbers object rank, to insert described final recommendation list.

In certain embodiments, it is one of following two kinds of modes that described content recommendation selected cell is selected from the sub-list of classifying: in a sub-list of classification, select a current content item ranking the first at every turn, and loop for the sub-list of all classification, wherein, skip the sub-list of classification of selected sky and the number of selected content item and reach the sub-list of classifying of corresponding expected numbers object; Or disposablely in the sub-list of each classification select corresponding expected numbers object content item, wherein, if the number of the content item in sub-list of classifying is less than corresponding expection number, all the elements item in sub-this classification list is all selected.

In certain embodiments, described content recommendation selected cell also for: if (b31) number of the sub-list content item of described classification is less than corresponding expection number, from described recommendation list to be selected, in non-selected residue content item, select the forward content item of rank, with by described final recommendation list completion, (b32) judge whether the difference between the scoring of content item the last in the scoring of the content item ranking the first in non-selected residue content item in described recommendation list to be selected and described final recommendation list is greater than predetermined threshold, if be greater than described predetermined threshold, exchange these two and repeating step (b32) until in the scoring of the content item ranking the first in non-selected residue content item in described recommendation list to be selected and described final recommendation list the difference between the scoring of the last content item be less than or equal to described predetermined threshold, to obtain described final recommendation list.

In certain embodiments, the content item in described final recommendation list is to arrange from high to low by the marking order of content item.

In certain embodiments, described content item at least comprises the following information: content name, content type and content marking.

The method of the application of the invention and respective server can be introduced user interest preference information in commending system algorithm, make last recommendation list make personalized recommendation for the interest preference difference of different user.In addition, target of the present invention is on the basis of existing or proposed algorithm that develop in the future, and a kind of method of adjustment and server of many classification (being generally more than or equal to 3) recommendation list of the commending system for mobile application market is provided.It can, for commending system algorithm service existing or that develop in the future, increase under the prerequisite of commending system added burden hardly, makes recommendation results as far as possible consistent with each other with user's interest tendency.

Particularly, the method according to this invention and server, have following some advantages:

(1) high divergence.Recommend method/server of the present invention goes for any one commending system model based on scoring.As long as can provide the mark about commodity (content item), no matter how concrete technology realizes, recommend method/server of the present invention can improve result by list adjustment algorithm.Therefore it is well positioned to meet the demand of various businessmans, and does not need system existing to businessman or that purchase in the future to do extensive adjustment;

(2) quick.In recommend method/server of the present invention, can adopt offline mode to carry out compute user preferences (for example, preference function vector), thereby improve the formation speed of Recommendations lists, be applicable to large-scale businessman to recommending the demand of efficiency;

(3) user interest is changed to reflection sensitive.Client's interest preference is for example, in the relatively short time period (, one or several month) upper stable, and for example, can change on the relatively long time period (, 1 year).Native system adopts timing to calculate the mode of user interest preference, along with passage of time, constantly adjusts user interest preference value, the demand constantly changing to meet client;

(4) improve precision.The recommendation list after adjusting is conducive to eliminate owing to having introduced the index vector of user interest preference the impact that Sparse brings, so will have larger raising than the recommendation list before adjusting in accuracy.

Accompanying drawing explanation

By below in conjunction with accompanying drawing explanation the preferred embodiments of the present invention, will make of the present invention above-mentioned and other objects, features and advantages are clearer, wherein:

Fig. 1 shows according to the schematic diagram of the example application scene of content recommendation system of the present invention.

Fig. 2 shows according to the example flow diagram of the commending contents based on user preference of the present invention.

Fig. 3 shows the example of adjusting according to the recommendation list of the commending contents based on user preference of the present invention.

Fig. 4 shows another example of adjusting according to the recommendation list of the commending contents based on user preference of the present invention.

Fig. 5 shows according to the process flow diagram of the exemplary contents recommend method based on user preference of carrying out at server place of the embodiment of the present invention.

Fig. 6 shows according to the block diagram of the example server for the method shown in execution graph 5 of the embodiment of the present invention.

Embodiment

Before specifically introducing design of the present invention, will first simply introduce the term that may occur herein, do not have in this article in the situation of clear and definite contrary indication, these terms should be understood with the implication illustrating herein.

Content:

In commending system field, the content of recommendation can be the interested books of user, audio frequency, video, theme, game, tool software; Or more generally, can be any commodity and/or service etc., no matter it is charge or free, tangible or invisible, in esse or virtual.

Mobile application market:

Assemble all kinds of mobile phone application developer and outstanding application thereof, met the comprehensive on-line mall of dissimilar cellphone subscriber's real-time experience, download and order demand.It provides the one-stop services such as tool software, game, theme, video, audio frequency, books by cell-phone customer terminal, wap and www website etc. for user.Famous mobile application market has: Google Play, Apple Store etc.

Categorised content:

The classification of carrying out for content, total number of categories is L (the application relates generally to 3 classes and situations more than 3 classes, and certainly, the application is applicable to the situation of 1 class or 2 classes too).For example: the application take audio plays as its main propagating contents is classified as audio class application (for example, audio player); The application of games is classified as game class application (for example, the bird of indignation) etc.The method of classification can artificially define, and also can divide by mathematical method, and the present invention is not limited to this.In other words, the present invention is applicable to classification and any mode classification of any number.

Recommendation list:

When proposed algorithm is recommended for certain user, meeting marking to each different content based on other similar users, predicts the marking of each different content this user.In the recommended content of all energy, select N content wherein to recommend to user together as net result, such set is recommendation list.N is list length, and N is natural number.The recommendation order of recommendation list inside conventionally in no particular order, can certainly sort by certain order, for example, by the height etc. of marking.

Redundancy recommendation list (or recommendation list to be selected):

In the present invention, before the final recommendation list that is N in generation length, can generate a length by various proposed algorithms existing or that develop is in the future M (M >=N, and M is natural number) redundancy recommendation list, final recommendation list produces after adjusting redundancy recommendation list by related algorithm disclosed in this invention.

Classification loop list:

Initial length is L (equaling total number of categories), be used for that record hereinafter describes in detail in the circular list that generates optional classification while carrying out cyclic pac king in the process of final recommendation list.In the time that a certain classification does not meet packing part, this classification is removed from circular list.L is natural number.

Final recommendation list:

List final lengths is N, and its initial content is empty, is filled by the content item that meets regularization condition in redundancy recommendation list.After list adjustment completes, the N of a final recommendation list content item is presented to user.

Collaborative filtering (CF, i.e. Collaborative Filtering):

A kind of algorithms most in use in commending system field.This algorithm has two layers of meaning: compared with the implication of narrow sense and compared with the implication of broad sense.In a broad sense, CF uses working in coordination with between multiple mechanisms, viewpoint, data source etc. to filter information or pattern.The application of collaborative filtering is usually directed to very large data set.It is for mine locating, large-area environmental monitoring, finance data, ecommerce and network application etc.

Narrowly say, CF is the method for predicting (filtration) user's focus from a lot of users' preference or taste information (working in coordination with) by collecting.The logic of its bottom is: if user a has identical suggestion in a certain content with user b, for another content x, user a more likely has identical suggestion with user b than another (having random opinion on content x) user c.These predictions are specific for user, but in fact they are according to generating from a lot of other users' feedback.Please note: this is different from the recommendation of based on all users, the average suggestion of a certain content being made simply.

Particularly, this Algorithm Analysis user interest, in customer group, find the several users (arest neighbors) the highest with the similarity degree of designated user, and the comprehensively marking evaluation of these similar users to a certain content, the marking to this content (fancy grade) prediction to this designated user with formation system.Conventionally the higher representative of consumer of prediction mark is more liked.According to the method, the recommended content of all energy is all given a mark, and the sequence of giving a mark according to prediction, recommend one or more content to user.The marking information that marking evaluation is collected is not necessarily confined to interested especially, and the record of the content of loseing interest in is especially also quite important, for example uninterested especially user content is got rid of outside recommendation list.

Collaborative filtering (CF-U, i.e. Collaborative Filtering User-based) based on user:

It is based on such hypothesis " you also probably like to like with you thing that similar people likes ".So the main task of collaborative filtering based on user is exactly to find out user's nearest-neighbors, thereby makes the score in predicting of this user for unknown term according to the hobby of nearest-neighbors.

Project-based collaborative filtering (CF-I, i.e. Collaborative Filtering Item-based):

By user, the scoring of different content project (item) is carried out to the similarity between evaluation item, the similarity based between project is made recommendation.Have a basic hypothesis " can cause the project of user's interest, must be similar to the high project of marking before it " take project as basic collaborative filtering method, the similarity seeing through between computational item replaces the similarity between user.

Interest preference:

In a period of time, mobile Internet user can tend to use a certain class application.This and individual hobby, habits and customs have close contacting.

Preference function vector:

Describe the fancy grade of designated user to a certain kind content with preference function quantitatively, and by a certain user, the preference function of all L class contents (for multiple classification, L is generally more than or equal to 3) is formed this user's preference function vector.In this application, be applied as example with six classes, i.e. L=6, the preference function of user to game class application, is in downloaded all application game class application proportion.In like manner, to the preference function of tool software, books, audio frequency, theme, video class application, be in downloaded all application respective classes application proportion.Formed again user's preference function vector by the preference function of these six large classes.In one example, preference function vector=(tool software, game, books, audio frequency, theme, video).Such as, for example, in 10 application that user A downloads within a period of time (, one month), the application of tool software class has 3, and accounting is 0.3; Game class application has 5, and accounting is 0.5; The application of books class has 0, and accounting is 0.0; Audio class application has 1, and accounting is 0.1; Theme class application has 0, and accounting is 0.0; And video class application has 1, accounting is 0.1.This user's preference function vector is (0.3,0.5,0.0,0.1,0.0,0.1).But in the present invention, the number of content (application) classification is not limited to 6, also can be still less or more, the dimension of preference function vector is not limited to 6 dimensions, can be also still less or more.In addition, in preference function vector, the order of each dimension (, the order of content type) is also not limited to (tool software, game, books, audio frequency, theme, video) order, and can be any order, for example (theme, video, audio frequency, tool software, game, books) etc.

Tool software:

General reference other application except audio frequency, video, books, game and theme class in application.For example, can comprise map class application, the application of financing class, address list management application, the application of shopping class etc.Sometimes also by it referred to as " software ".

To a preferred embodiment of the present invention will be described in detail, in description process, having omitted is unnecessary details and function for the present invention with reference to the accompanying drawings, obscures to prevent that the understanding of the present invention from causing.Below, the scene that is applied to mobile radio system take the present invention is example, and the present invention be have been described in detail.But the present invention is not limited thereto, the present invention also can be applied to fixed communications, wired communication system, or is applied to any mixed structure of mobile radio system, fixed communications, wired communication system etc.With regard to mobile communication system, the present invention is not limited to the concrete communication protocol of each related mobile communication terminal, can include, but is not limited to 2G, 3G, 4G, 5G network, WCDMA, CDMA2000, TDSCDMA system etc., different mobile terminals can adopt identical communication protocol, also can adopt different communication protocol.In addition, the present invention is not limited to the specific operating system of mobile terminal, can include, but is not limited to iOS, Windows Phone, Symbian (Saipan), Android (Android) etc., different mobile terminals can adopt identical operating system, also can adopt different operating system.In addition, the present invention is not limited to the specific operating system of server, can include, but is not limited to Windows, Unix, Linux, FreeBSD, Solaris of various version etc., different servers can adopt identical operating system, also can adopt different operating system.

Fig. 1 shows according to the schematic diagram of the application scenarios of content recommendation system 1000 of the present invention.As shown in Figure 1, system 1000 can comprise terminal 100 and server 200.For the sake of clarity, in figure, only show a terminal 100, a server 200, but the present invention is not limited thereto, can comprise the terminal of two or more numbers and/or server etc.Terminal 100 and server 200 can communicate by communication network 300.The example of communication network 300 can include, but is not limited to: internet, mobile communications network, permanent haulage line (as xDSL, optical fiber etc.) etc.

In the embodiment shown in fig. 1, in order on server 200, content to be recommended, commending contents server end 250 (being designated hereinafter simply as server end 250) is arranged on server 200 according to an embodiment of the invention.Server end 250 can be arranged in server 200 with the form of software voluntarily by service provider, or can be arranged in server 200 with the form of hardware or firmware by production firm.In certain embodiments, server end 250 can be the application software of the present invention that is specifically designed to of for example having downloaded from network after service provider has bought server 200.In further embodiments, server end 250 can be to be for example arranged in advance the application program in server 200 by production firm with firmware or example, in hardware.In other embodiment, server end 250 can be that hardware module or the server 200 produced by production firm are own.

In this embodiment, related concrete content recommendation is the various application in mobile application market.Certainly the invention is not restricted to this, content can be also other commodity and/or service, and such as service is payed on behalf in food, medicine, household electrical appliances, charges for water and electricity, fine is payed on behalf service etc.Server 200 can be recommended above-mentioned extensive stock and/or service to user.

In this embodiment, a certain user's application download historical information can be collected or otherwise be obtained to server 200 first.A certain content (application) the classification quantity of downloading according to this user who records in this historical information accounts for the percentage of total download and recently describes interest preference, and calculates preference function vector.Meanwhile or before this or afterwards, recommend for specific user by traditional collaborative filtering, wherein, this tradition collaborative filtering carries out user's marking prediction to each application, select the highest front M item application program of marking prediction, to form redundancy recommendation list (, recommendation list to be selected).For example, by adjustment of the present invention (, cyclic pac king, supplementary replacement etc.) scheduling algorithm, generate final recommendation list.And then make the types of applications ratio in final recommendation list as far as possible can be consistent with user preference index.After adjustment operation all finishes, final recommendation list is recommended to user as net result.

In one embodiment, for example, if user's application program that (, 1 month) downloads within a period of time is 10, wherein the application of tool software class has 3, accounting 0.3; Game class application has 5, accounting 0.5; The application of e-book class has 0, accounting 0.0; Audio class application has 1, accounting 0.1; Theme class application has 0, accounting 0.0; And video class application has 1, accounting 0.1.This user's preference function vector is (0.3,0.5,0.0,0.1,0.0,0.1).This index vector has also just reflected the interest preference of user in this month.The object of the invention is to wish finally presenting identical or at least close interest preference to the ratio of types of applications in user's recommendation list.In the situation that for example final recommendation list length N equals 10, final recommendation list should be made up of 3 software class application of giving a mark the highest, 5 game class application, 0 books application, 1 audio class application, 0 theme class application and 1 video class application as far as possible.And traditional collaborative filtering cannot reach such requirement substantially, even if end product meets this ratio, also belong to coincidence, but not intention so.So we need to adjust recommendation list.

Next, we specifically describe the flow process for commending contents with reference to Fig. 2.

In one embodiment, related content recommendation is the various application in mobile application market, and flow process is roughly divided into following four basic execution steps:

Step 1: generate preference function vector.In one embodiment, utilize user's application to download historical record, calculate each user within a period of time recently, for the preference function vector of all categories application.Each user can obtain self a preference function vector.

Step 2: produce redundancy recommendation list.Utilize proposed algorithm existing or that develop in the future to recommend about this user, and produce redundancy recommendation list, operable proposed algorithm is not limited to above-mentioned collaborative filtering, and it can at least use one of the following: content-based recommendation, the recommendation based on collaborative filtering, the recommendation based on correlation rule, the recommendation based on effectiveness, recommendation and aforementioned every combination in any based on knowledge.In addition, if use collaborative filtering, it is also not limited to the collaborative filtering based on user, and it can also use the collaborative filtering recommending (CF-U) based on user, project-based collaborative filtering recommending (CF-I), collaborative filtering recommending (CF-M) and aforementioned every combination in any based on model.In a word, in the time generating redundancy recommendation list, the present invention does not limit the initial recommendation algorithm of concrete use.

In addition, the content item in redundancy recommendation list generally includes the information of three aspects: the prediction marking in commending system of content item (application) title, content item (application) type and content item (application).Redundancy recommendation list can be divided into L the sub-list of classification by content type.

Please note: the execution sequence of step 1 and step 2 might not be first to perform step one herein, in execution step two, in fact also can the two simultaneously (or simultaneously basic) carry out, or first perform step two and perform step again one, the invention is not restricted to this.

Step 3: generate final recommendation list (list adjustment).According to the user preference index vector calculating in step 1, fill and adjust in final recommendation list in the mode of hereinafter describing in detail, make the distribution of content of all categories in final recommendation list consistent or at least as far as possible consistent to the preference degree of variety classes content with user.

The method of aforementioned adjustment can be divided into two steps:

(a) cyclic pac king is chosen successively Section 1 (i.e. the highest of scoring) and insert the filling of order method in final recommendation list, and constantly between L class, circulates from the each classification of redundancy recommendation list.In one embodiment, this circulation can be safeguarded with a classification loop list (length is L).When in the sub-list of a certain classification of redundancy recommendation list without project, when the number of maybe filling the content item of (selecting) in the sub-list of this classification has reached " the preference function * N " of this classification (length that N is final recommendation list), this classification is removed from classification loop list, to avoid again filling the content item of this classification in final recommendation list.In the time that circular list is empty (that is, all classification are all removed from circular list), cyclic pac king process finishes.

(b) supplement replace, after cyclic pac king finishes, by whole the residue in redundancy recommendation list projects according to scoring descending sequence.When final recommendation list less than time, successively the Section 1 in redundancy list is filled in final recommendation list until fill up (completion) by slotting order method.After final recommendation list is filled up, difference between the scoring of the last item of the scoring of calculating the scoring Section 1 in current residual content item in redundancy recommendation list and final recommendation list (because slotting order method is filled, finally current scoring lowest term in recommendation list).In the time that this difference is greater than pre-set threshold value, the two is replaced.Replacement method is final recommendation list last item to be removed, and Section 1 in redundancy list is filled in final recommendation list by slotting order method.This difference of double counting afterwards is also replaced, until this difference is not more than threshold values, adjustment process completes, and final recommendation list is now final recommendation list.

In another embodiment, also can not use " circulation " to fill.For example, can be for each sub-list of classifying, the individual content item of disposable extraction from each sub-list of classifying " preference function * N " is inserted final recommendation list.If certain content item number of classifying in sub-list is less than " preference function * N ", can be equally follow-up " supplement and replace " in the stage by final recommendation list completion.

Step 4: present final recommendation list to user.For example, by this final recommendation list is provided from server 200 to terminal 100, and shown to provide to user by terminal 100.

In addition,, in above-mentioned steps one, can also calculate according to other data relevant to user and content item user's preference function vector.For example, can be according to user's movable a series of information that produce in mobile application market, comprise download, mark, browse, the record such as word comment, calculate preference function vector.Certainly, user behavior data is at this moment of a great variety, and is not that each user profile is all that we need.Therefore, can extract garbled data by behavioural characteristic.Finally, in conjunction with UAD, change by behavioural characteristic, convert concrete external data to computing machine understandable behavioural characteristic vector.Consider user behavior real-time change, therefore, this part operation can be carried out conventionally in real time.

In addition, the external data relevant to application can comprise application recommendation tables, and it comprises the information that a series of users such as Apply Names, affiliated classification, price pay close attention to.Application recommendation tables bonding behavior proper vector has just formed feature-content item associated recommendation.Add owing to having at set intervals new a collection of application, therefore this part can regularly be upgraded conventionally.

In general, redundancy recommendation list can be produced by traditional proposed algorithm, and obtains last recommendation results after recommendation list adjustment.We need to make an explanation to content recommendation, and present to user together with recommendation results.

So far, in conjunction with Fig. 2 detailed according to the commending contents flow process of the embodiment of the present invention.

Next, describe respectively according to the example generative process of the final recommendation list for user A and user B of the embodiment of the present invention in conjunction with Fig. 3 and Fig. 4.Fig. 3 and Fig. 4 all show example redundancy recommendation list, the final recommendation list of example and example adjustment process wherein.

User A:

First suppose that redundancy recommendation list (Fig. 3 bottom) length M being generated for user A by traditional collaborative filtering is 20, and we expect that the final recommendation list length N generating is 10 (Fig. 3 tops), total number of categories L is 6, the user's who for example calculates as mentioned above preference function vector is for (0.3,0.5,0.0,0.1,0.0,0.1), corresponding (tool software, game, books respectively, audio frequency, theme, video), it is 0.10 that threshold value (marking poor) is adjusted in list.

(1) cyclic pac king:

Initialization classification loop list: [tool software, game, books, audio frequency, theme, video].

Because user's books class and the preference function of theme class are 0.0, therefore books class and theme class are removed to (0.0*10=0) from classification loop list, remove in rear classification loop list only residue: [tool software, game, audio frequency, video].No. 18, No. 1, selection tool software class Section 1, No. 6, game class Section 1, No. 14, audio class Section 1 and video class Section 1 successively, is filled in final recommendation list by slotting order method.

The number that audio class after filling has been filled content item is 1, and it is more than or equal to 0.1 (audio-preferences index) * 10 (final recommendation list length)=1.It is 1 that video class has been filled item number, and it is more than or equal to 0.1*10=1.Therefore, audio class and video class are removed from classification loop list, only remain: [software, game].Select successively software class Section 1 No. 2 (now No. 1 not in redundancy recommendation list, other in like manner), No. 7, game class Section 1, be filled in final recommendation list by slotting order method.Then, No. 8, No. 3, game class Section 1 of selecting successively software class Section 1, are filled in final recommendation list by slotting order method.

Now, it is 3 that the rear software class of filling has been filled item number, and it is more than or equal to 0.3*10=3.Software class is removed from classification loop list, only remain: [game].Select No. 9, game class Section 1, be filled in final recommendation list by slotting order method.Select No. 10, game class Section 1, be filled in final recommendation list by slotting order method.Now, fill after game class to have filled item number be 5 to be more than or equal to 0.5*10=5.Game class is removed from classification loop list, only remain: [].Now classification loop list is empty, and cyclic pac king process finishes.

Now, can see in the redundancy recommendation list of Fig. 3 middle and lower part and being received in the final recommendation list on top with 10 content items that indicate without diagonal line hatches.

(2) supplement and replace:

Remaining whole projects in redundancy recommendation list are pressed to the descending sequence of scoring.Because now final recommendation list is full, without the process of supplementing.

Then, it is 4.00-3.66=0.34 that the last item of choosing No. 16, the Section 1 that remains in redundancy recommendation list in content item (non-selected content item) and final recommendation list calculates scoring differences for No. 18, and it is greater than 0.10 (threshold values).Now remove in final recommendation list last item No. 18, redundancy recommendation list Section 1 is inserted to final recommendation list by slotting order method No. 16.Repeat said process, that is, No. 14 calculated difference of last item of choosing No. 11, the current Section 1 of redundancy recommendation list and final recommendation list are 3.99-3.79=0.20, and it is greater than 0.10.Now remove in final recommendation list last item No. 14, redundancy recommendation list Section 1 is inserted to final recommendation list by slotting order method No. 11.Repeat said process, that is, No. 10 calculated difference of last item of choosing No. 17, the current Section 1 of redundancy recommendation list and final recommendation list are 3.88-3.81=0.07, and it is less than 0.10, and therefore without replacing, replacement process finishes.

Now, can see that 2 content items that indicate with diagonal line hatches in the redundancy recommendation list of Fig. 3 middle and lower part have been replaced in the final recommendation list on top, and indicate in the middle diagonal line hatches of two lists two content items that are replaced out.

After all adjustment processes finish, final recommendation list is now exactly the net result that will recommend to user.

User B:

Referring to Fig. 4, redundancy recommendation list (Fig. 4 bottom) length M that same supposition is generated for user B by traditional collaborative filtering is 20, and we expect that the final recommendation list length N generating is 10 (Fig. 4 tops), total number of categories L is 6, the user's who for example calculates as mentioned above preference function vector is for (0.1,0.2,0.4,0.1,0.1,0.1), respectively corresponding (tool software, game, books, audio frequency, theme, video), it is 0.10 that threshold value (marking poor) is adjusted in list.

(1) cyclic pac king:

Initialization classification loop list: [tool software, game, books, audio frequency, theme, video].

Because the preference function of all categories is not all 0.0, therefore No. 18, No. 1, selection tool software class Section 1, No. 6, game class Section 1, No. 11, books class Section 1, No. 14, audio class Section 1, No. 16, theme class Section 1 and video class Section 1 successively, is filled in final recommendation list by slotting order method.

The number of filling content item of software class, audio class, theme class and video class is 1, and it is more than or equal to 0.1 (preference function) * 10 (final recommendation list length)=1.Therefore, software class, audio class, theme class and video class are removed from classification loop list, only remain: [game, books].Select successively game class Section 1 No. 7 (now No. 6 not in redundancy recommendation list, other in like manner), No. 12, books class Section 1, be filled in final recommendation list by slotting order method.

Now, it is 2 that the rear game class of filling has been filled item number, and it is more than or equal to 0.2*10=2.Game class is removed from classification loop list, only remain: [books].Select No. 13, books classes Section 1, be filled in final recommendation list by slotting order method.Now, fill after books class to have filled item number be 3 to be less than 0.4*10=4.But because books class content item has all been selected into final recommendation list, therefore books class is removed from classification loop list, only remain: [].Now classification loop list is empty, and cyclic pac king process finishes.

Now, can see in the redundancy recommendation list of Fig. 4 middle and lower part and being received in the final recommendation list on top with 9 content items that indicate without diagonal line hatches.

(2) supplement and replace:

Remaining whole projects in redundancy recommendation list are pressed to the descending sequence of scoring.Due to final recommendation list now less than, therefore need the process of supplementing.In the process of supplementing, from redundancy recommendation list, remain in content item and select the highest content item of scoring, i.e. No. 2, the highest item of the current scoring of software class, is used slotting order method to be filled in final recommendation list.Now, final recommendation list is by completion, and the process of supplementing finishes.Now, can see that the content item in Fig. 4 bottom with diagonal line hatches is added in final recommendation list for No. 2.

Then, it is 3.96-3.66=0.30 that the last item of choosing No. 8, the Section 1 that remains in redundancy recommendation list in content item (non-selected content item) and final recommendation list calculates scoring differences for No. 18, and it is greater than 0.10 (threshold values).Now remove in final recommendation list last item No. 18, redundancy recommendation list Section 1 is inserted to final recommendation list by slotting order method No. 8.Repeat said process, that is, No. 13 calculated difference of last item of choosing No. 3, the current Section 1 of redundancy recommendation list and final recommendation list are 3.89-3.68=0.21, and it is greater than 0.10.Now remove in final recommendation list last item No. 13, redundancy recommendation list Section 1 is inserted to final recommendation list by slotting order method No. 3.Repeat said process, that is, No. 14 calculated difference of last item of choosing No. 17, the current Section 1 of redundancy recommendation list and final recommendation list are 3.88-3.79=0.09, and it is less than 0.10, and therefore without replacing, replacement process finishes.

Now, can see that 2 content items that indicate with diagonal line hatches in the redundancy recommendation list of Fig. 4 middle and lower part have been replaced to for No. 3 and No. 8 in the final recommendation list on top, and indicate in the middle diagonal line hatches of two lists two content items that are replaced out.

After all adjustment processes finish, final recommendation list is now exactly the net result that will recommend to user.

So far, describe the concrete example of carrying out commending contents for user A and user B in detail in conjunction with Fig. 3 and Fig. 4.

Fig. 5 shows according to the process flow diagram of the content recommendation method 400 of carrying out in server 200 of the embodiment of the present invention.As shown in Figure 5, method 400 can comprise step S410, S420 and S430.According to the present invention, execution can be carried out separately or combine to some steps of method 400, and can executed in parallel or order carry out, be not limited to the concrete operations order shown in Fig. 5.In certain embodiments, the server 200 that method 400 can be as shown in Figure 1 or the server end 250 on it are carried out.

Fig. 6 shows according to the block diagram of the example server 200 of the embodiment of the present invention.As shown in Figure 6, server 200 can comprise: preference function computing unit 210, content recommendation selected cell 220 and recommendation list provide unit 230.

Preference function computing unit 210 can, for the historical preference information based on user, calculate this user's preference function vector.Preference function computing unit 210 can be CPU (central processing unit) (CPU), digital signal processor (DSP), microprocessor, microcontroller of server 200 etc., its can with the communications portion of server 200 (for example, radio receiving-transmitting unit, Ethernet card, xDSL modulator-demodular unit etc.) and/or storage area is (for example, RAM, SD card etc.) match, based on the user's that receives and/or store historical preference information, calculate this user's preference function vector.

Content recommendation selected cell 220 can, for based on preference function vector, be selected the content item that will recommend, to form final recommendation list from recommendation list to be selected.Content recommendation selected cell 220 can be CPU (central processing unit) (CPU), digital signal processor (DSP), microprocessor, microcontroller of server 200 etc., its can with the communications portion of server 200 (for example, radio receiving-transmitting unit, Ethernet card, xDSL modulator-demodular unit etc.) and/or storage area is (for example, RAM, SD card etc.) match, based on preference function vector, from select the content item that will recommend the recommendation list to be selected that receives and/or store, to form final recommendation list.

Recommendation list provides the unit 230 can be for final recommendation list is provided to user.It can be CPU (central processing unit) (CPU), digital signal processor (DSP), microprocessor, microcontroller of server 200 etc. that recommendation list provides unit 230, its can with the output of server 200 (for example, display, printer etc.) match, the final recommendation list being generated by content recommendation selected cell 220 is provided to user.

Below with reference to Fig. 5 and Fig. 6, content recommendation method 400 and server 200 for carrying out at server 200 places according to the embodiment of the present invention are described in detail.

Method 400 starts from step S410, in step S410, can be by the preference function computing unit 210 of server 200 the historical preference information based on user, calculate this user's preference function vector.

In step S420, can, by the content recommendation selected cell 220 of server 200 based on this preference function vector, from recommendation list to be selected, select the content item that will recommend, to form final recommendation list.

In step S430, can provide unit 230 to provide this final recommendation list to user by the recommendation list of server 200.

In certain embodiments, historical preference information can be to obtain by the content item that in statistics predetermined amount of time, user downloads.

In certain embodiments, predetermined amount of time can be 1 month.

In certain embodiments, step S410 can also comprise that the content item of the user who counts being downloaded by the preference function computing unit 210 of server 200 is divided into one or more classifications (S412); For the each classification in one or more classifications, in the content item that calculating user downloads, belong to the number of such other content item and the total ratio (S414) of the content item that user downloads; And according to the corresponding ratio of all categories in one or more classifications, form user's preference function vector (S416) within a predetermined period of time.

In certain embodiments, one or more classifications can comprise following at least one: audio frequency, video, books, theme, game and tool software.

In certain embodiments, recommendation list to be selected can generate for user by proposed algorithm, and proposed algorithm can be following at least one: content-based recommendation; Based on the recommendation of collaborative filtering; Based on the recommendation of correlation rule; Based on the recommendation of effectiveness; Based on the recommendation of knowledge; And above-mentioned every combination in any.

In certain embodiments, the recommendation based on collaborative filtering can comprise following at least one: the collaborative filtering recommending (CF-U) based on user; Project-based collaborative filtering recommending (CF-I); Collaborative filtering recommending (CF-M) based on model; And above every combination in any.

In certain embodiments, step S420 can comprise that content recommendation selected cell 220 by server 200, according to the classification of each content item in recommendation list to be selected, is divided into the sub-list of one or more classification (S422) by recommendation list to be selected; Based on preference function vector, from the sub-list of corresponding classification, select the forward content item of rank, to insert final recommendation list (S424); And based on non-selected content item in recommendation list to be selected, adjust final recommendation list (S426).

In certain embodiments, step S424 can comprise by the preference function of content recommendation selected cell 220 based on corresponding with each classification in preference function vector and the length of final recommendation list of server 200, calculate the shared expection number (S424-1) of each classification in final recommendation list; And for each classification, from the sub-list of corresponding classification, select the forward content item of corresponding expected numbers object rank, to insert final recommendation list (S424-2).

In certain embodiments, selection in step S424-2 can be one of following two kinds of modes: in a sub-list of classification, select a current content item ranking the first at every turn, and loop for the sub-list of all classification, wherein, skip the sub-list of classification of selected sky and the number of selected content item and reach the sub-list of classifying of corresponding expected numbers object; Or disposablely in the sub-list of each classification select corresponding expected numbers object content item, wherein, if the number of the content item in sub-list of classifying is less than corresponding expection number, all the elements item in sub-this classification list is all selected.

In certain embodiments, if classified, in sub-list, the number of content item is less than corresponding expection number, step S426 can comprise by the content recommendation selected cell 220 of server 200 and from recommendation list to be selected, in non-selected residue content item, selects the forward content item of rank

With by final recommendation list completion: (S426-1); Judge whether the difference between the scoring of content item the last in the scoring of the content item ranking the first in non-selected residue content item in recommendation list to be selected and final recommendation list is greater than predetermined threshold, if be greater than predetermined threshold, exchange these two and repeating step (S426-1) until in the scoring of the content item ranking the first in non-selected residue content item in recommendation list to be selected and final recommendation list the difference between the scoring of the last content item be less than or equal to predetermined threshold, to obtain final recommendation list (S426-2).

In certain embodiments, the content item in final recommendation list can be to arrange from high to low by the marking order of content item.

In certain embodiments, content item can at least comprise the following information: content name, content type and content marking.

So far invention has been described in conjunction with the preferred embodiments.Should be appreciated that, those skilled in the art without departing from the spirit and scope of the present invention, can carry out various other change, replacement and interpolations.Therefore, scope of the present invention is not limited to above-mentioned specific embodiment, and should be limited by claims.

Claims (26)

1. the content recommendation method based on user preference, comprising:

(a) the historical preference information based on described user, calculates described user's preference function vector;

(b), based on described preference function vector, from recommendation list to be selected, select the content item that will recommend, to form final recommendation list; And

(c) provide described final recommendation list to described user.

2. method according to claim 1, wherein, described historical preference information is to obtain by the content item that in statistics predetermined amount of time, described user downloads.

3. method according to claim 2, wherein, described predetermined amount of time is 1 month.

4. method according to claim 3, wherein, step (a) also comprises:

(a1) content item of the user who counts being downloaded is divided into one or more classifications;

(a2), for the each classification in described one or more classifications, in the content item that calculating user downloads, belong to the number of such other content item and the total ratio of the content item that user downloads; And

(a3), according to the corresponding ratio of all categories in described one or more classifications, form the preference function vector of described user in described predetermined amount of time.

5. method according to claim 4, wherein, described one or more classifications comprise following at least one: audio frequency, video, books, theme, game and tool software.

6. method according to claim 1, wherein, described recommendation list to be selected generates for described user by proposed algorithm, and described proposed algorithm is following at least one:

Content-based recommendation;

Based on the recommendation of collaborative filtering;

Based on the recommendation of correlation rule;

Based on the recommendation of effectiveness;

Based on the recommendation of knowledge; And

Above-mentioned every combination in any.

7. method according to claim 6, wherein, the described recommendation based on collaborative filtering comprises following at least one:

Collaborative filtering recommending (CF-U) based on user;

Project-based collaborative filtering recommending (CF-I);

Collaborative filtering recommending (CF-M) based on model; And

Every combination in any above.

8. method according to claim 1, wherein, step (b) comprising:

(b1) according to the classification of each content item in described recommendation list to be selected, described recommendation list to be selected is divided into the sub-list of one or more classification;

(b2), based on described preference function vector, from the sub-list of corresponding classification, select the forward content item of rank, to insert described final recommendation list; And

(b3) based on non-selected content item in described recommendation list to be selected, adjust described final recommendation list.

9. method according to claim 8, wherein, step (b2) comprising:

(b21) preference function based on corresponding with each classification in described preference function vector and the length of described final recommendation list, calculates the shared expection number of each classification in described final recommendation list; And

(b22), for each classification, from the sub-list of corresponding classification, select the forward content item of corresponding expected numbers object rank, to insert described final recommendation list.

10. method according to claim 9, wherein, the selection in step (b22) is one of following two kinds of modes:

At every turn in a sub-list of classification, select a current content item ranking the first, and loop for the sub-list of all classification, wherein, skip the sub-list of classification of selected sky and the number of selected content item and reach the sub-list of classifying of corresponding expected numbers object; Or

Disposablely in the sub-list of each classification select corresponding expected numbers object content item, wherein, if the number of the content item in sub-list of classifying is less than corresponding expection number, all the elements item in sub-this classification list is all selected.

11. methods according to claim 9, wherein, step (b3) comprising:

(b31), if the number of content item is less than corresponding expection number in the sub-list of described classification, from described recommendation list to be selected, in non-selected residue content item, select the forward content item of rank, with by described final recommendation list completion;

(b32) judge whether the difference between the scoring of content item the last in the scoring of the content item ranking the first in non-selected residue content item in described recommendation list to be selected and described final recommendation list is greater than predetermined threshold, if be greater than described predetermined threshold, exchange these two and repeating step (b32) until in the scoring of the content item ranking the first in non-selected residue content item in described recommendation list to be selected and described final recommendation list the difference between the scoring of the last content item be less than or equal to described predetermined threshold, to obtain described final recommendation list.

12. methods according to claim 1, wherein, the content item in described final recommendation list is to arrange from high to low by the marking order of content item.

13. methods according to claim 1, wherein, described content item at least comprises the following information: content name, content type and content marking.

14. 1 kinds of commending contents servers based on user preference, comprising:

Preference function computing unit, for the historical preference information based on described user, calculates described user's preference function vector;

Content recommendation selected cell for based on described preference function vector, is selected the content item that will recommend, to form final recommendation list from recommendation list to be selected; And

Recommendation list provides unit, for described final recommendation list is provided to described user.

15. servers according to claim 14, wherein, described historical preference information is to obtain by the content item that in statistics predetermined amount of time, described user downloads.

16. servers according to claim 15, wherein, described predetermined amount of time is 1 month.

17. servers according to claim 16, wherein, described preference function computing unit also for:

(a1) content item of the user who counts being downloaded is divided into one or more classifications;

(a2), for the each classification in described one or more classifications, in the content item that calculating user downloads, belong to the number of such other content item and the total ratio of the content item that user downloads; And

(a3), according to the corresponding ratio of all categories in described one or more classifications, form the preference function vector of described user in described predetermined amount of time.

18. servers according to claim 17, wherein, described one or more classifications comprise following at least one: audio frequency, video, books, theme, game and tool software.

19. servers according to claim 14, wherein, described recommendation list to be selected generates for described user by proposed algorithm, and described proposed algorithm is following at least one:

Content-based recommendation;

Based on the recommendation of collaborative filtering;

Based on the recommendation of correlation rule;

Based on the recommendation of effectiveness;

Based on the recommendation of knowledge; And

Above-mentioned every combination in any.

20. servers according to claim 19, wherein, the described recommendation based on collaborative filtering comprises following at least one:

Collaborative filtering recommending (CF-U) based on user;

Project-based collaborative filtering recommending (CF-I);

Collaborative filtering recommending (CF-M) based on model; And

Every combination in any above.

21. servers according to claim 14, wherein, described content recommendation selected cell also for:

(b1) according to the classification of each content item in described recommendation list to be selected, described recommendation list to be selected is divided into the sub-list of one or more classification;

(b2), based on described preference function vector, from the sub-list of corresponding classification, select the forward content item of rank, to insert described final recommendation list; And

(b3) based on non-selected content item in described recommendation list to be selected, adjust described final recommendation list.

22. servers according to claim 21, wherein, described content recommendation selected cell also for:

(b21) preference function based on corresponding with each classification in described preference function vector and the length of described final recommendation list, calculates the shared expection number of each classification in described final recommendation list; And

(b22), for each classification, from the sub-list of corresponding classification, select the forward content item of corresponding expected numbers object rank, to insert described final recommendation list.

23. methods according to claim 22, wherein, it is one of following two kinds of modes that described content recommendation selected cell is selected from the sub-list of classifying:

At every turn in a sub-list of classification, select a current content item ranking the first, and loop for the sub-list of all classification, wherein, skip the sub-list of classification of selected sky and the number of selected content item and reach the sub-list of classifying of corresponding expected numbers object; Or

Disposablely in the sub-list of each classification select corresponding expected numbers object content item, wherein, if the number of the content item in sub-list of classifying is less than corresponding expection number, all the elements item in sub-this classification list is all selected.

24. servers according to claim 22, wherein, described content recommendation selected cell also for:

(b31), if the number of content item is less than corresponding expection number in the sub-list of described classification, from described recommendation list to be selected, in non-selected residue content item, select the forward content item of rank, with by described final recommendation list completion;

(b32) judge whether the difference between the scoring of content item the last in the scoring of the content item ranking the first in non-selected residue content item in described recommendation list to be selected and described final recommendation list is greater than predetermined threshold, if be greater than described predetermined threshold, exchange these two and repeating step (b32) until in the scoring of the content item ranking the first in non-selected residue content item in described recommendation list to be selected and described final recommendation list the difference between the scoring of the last content item be less than or equal to described predetermined threshold, to obtain described final recommendation list.

25. servers according to claim 14, wherein, the content item in described final recommendation list is to arrange from high to low by the marking order of content item.

26. servers according to claim 14, wherein, described content item at least comprises the following information: content name, content type and content marking.

Images