CN104572962A - APP (Application) recommendation method and system - Google Patents

Abstract

The invention is applicable to the technical field of information, and provides an APP (Application) recommendation method and system. The method comprises the following steps: acquiring and recording a behavior log of a user, wherein the behavior log comprises a record for the user to download the APP and a record for browsing the APP; generating a user behavior matrix according to the behavior log; calculating the expectation for the user to download the APP by using the user behavior matrix according to a preset recommendation algorithm; recommending the APP with high expectation to the user. According to the embodiment of the invention, the behavior log of the user is acquired and recorded, the user behavior matrix is generated according to the behavior log, the expectation for the user to download the APP is calculated by using the user behavior matrix according to the preset recommendation algorithm, and the APP with high expectation is recommended to the user, so an APP platform can call an accurate algorithm to accurately recommend the APP to the user according to an APP download record and an APP preview record of the user.

Description

The method and system that a kind of APP recommends

Technical field

The invention belongs to areas of information technology, particularly relate to the method and system that a kind of APP recommends.

Background technology

At mobile applications, (application, is called for short: today app) greatly facilitating people's daily life, and the mobile applications of substantial amounts is also with giving user the puzzlement of selection.User needs a kind of efficient approach and helps them select interested small part from tens of thousands app.Be subject to the inspiration of commending system in the widespread use of conventional internet field, industry has turned one's attention to the recommendation of app.

The scheme that current most of app shop have employed split catalog and popular ranking list helps user to find required app, but this is not well positioned to meet the individual demand of different user, because being presented in face of all users is all the same content, and have ignored the differences such as the sex of user, age and culture.Even if for same user, his interest is often along with time variations.Search based on key word is then based upon user on the clearly description of self-demand, and therefore when user clearly can not describe self-demand, the recommendation results presented often seems blindly.The personalized app commending system that industry really comes into operation is little.

Summary of the invention

The object of the embodiment of the present invention is the method and system providing a kind of APP to recommend, to solve the problem that prior art APP with no personalization recommends.

The embodiment of the present invention is achieved in that a kind of method that APP recommends, and described method comprises:

Obtain and the user behaviors log of recording user, described user behaviors log comprises: user downloads APP record, for browsing APP record;

User behavior matrix is generated according to described user behaviors log;

According to the proposed algorithm preset, described user behavior matrix computations user is used to download the expection of APP;

To the APP that user recommends described expection high.

Another object of the embodiment of the present invention is the system providing a kind of APP to recommend, and described system comprises:

User behaviors log acquiring unit, for obtaining and the user behaviors log of recording user, described user behaviors log comprises: user downloads APP record, for browsing APP record;

User behavior matrix generation unit, generates user behavior matrix for the user behaviors log obtained according to described user behaviors log acquiring unit;

Expection computing unit, for the proposed algorithm that basis is preset, the user behavior matrix computations user using described user behavior matrix generation unit to generate downloads the expection of APP;

APP recommendation unit, the APP that the expection for recommending described expection computing unit to calculate to user is high.

The embodiment of the present invention, obtain and the user behaviors log of recording user, user behavior matrix is generated according to user behaviors log, according to the proposed algorithm preset, user behavior matrix computations user is used to download the expection of APP, to the APP that user recommends described expection high, APP platform according to the APP Download History of user and APP preview record, can be called algorithm accurately and recommends APP to user accurately.

Accompanying drawing explanation

In order to be illustrated more clearly in the technical scheme in the embodiment of the present invention, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skill in the art, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.

Fig. 1 is the process flow diagram of the APP recommend method that the embodiment of the present invention provides;

Fig. 2 is the structural drawing of the APP commending system that the embodiment of the present invention provides.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

In order to technical solutions according to the invention are described, be described below by specific embodiment.

Embodiment one

Be illustrated in figure 1 the process flow diagram of the APP recommend method that the embodiment of the present invention provides, said method comprising the steps of:

Step S101, obtain and the user behaviors log of recording user, described user behaviors log comprises: user downloads APP record, for browsing APP record.

In embodiments of the present invention, APP commending system obtains the user behaviors log of user, and passes through the form record behavior daily record of data set, and wherein, data include but not limited to: Download data set, Browse data set, and data are specially:

1, each of Download data centralization be recorded as user certain time download the record of app.Shape as (uid, aid, timestamp), wherein uid, aid be respectively user, app encryption after ID, timestamp be download behavior occur time, as 2000-01-01;

2, each of Browse data centralization be recorded as user certain time browse the record of app.Shape as (uid, aid, timestamp), wherein uid, aid be respectively user, app encryption after ID, timestamp be navigation patterns occur time, as 2000-01-01.

Step S102, generate user behavior matrix according to described user behaviors log.

In embodiments of the present invention, APP commending system is after obtaining user behaviors log, and generate the user behavior matrix for subsequent algorithm according to behavior daily record, this user behavior matrix can be 0,1 two values matrix, also can be other numerical value.

The proposed algorithm that step S103, basis are preset, uses described user behavior matrix computations user to download the expection of APP.

In embodiments of the present invention, APP commending system, after obtaining user behavior matrix, calls default proposed algorithm, uses this user behavior matrix computations user to download the expection of given APP.Described proposed algorithm comprises:

1, based on the collaborative filtering (memory-based CF) of memory: factor data scale is excessive, consider to adopt the similarity between Jaccard formulae discovery user or app for counting yield.Collaborative filtering for based on article:

$s_{\mathrm{ij}}=\frac{|N\left(i\right)\∩N\left(j\right)|}{|N\left(i\right)\∪N\left(j\right)|}$

S _ijrepresent the similarity of app i and app j, N (i) represents all user's set of downloading app i.

For given user u and app i, can predict that u downloads the possibility of i with following formula, this possibility describes with r in formula:

$r_{\mathrm{ui}}=\underset{j\∈R\left(u\right)}{\mathrm{\Σ}}{r}_{\mathrm{uj}}*s_{\mathrm{ij}}$

S _ijrepresent the similarity of app i and j in similarity matrix, r _ujrepresent corresponding numerical information in user behavior matrix R.Notice that the r downloading app i possibility for weighing user u is the numerical value that can be greater than 1 in fact.For given user, calculate the r of user u to all app also do not downloaded _ui, carry out descending sort according to r to these app, for user recommends top n app, Here it is, and Top-N recommends;

2, time-based collaborative filtering (Time-Based CF): energy is concentrated on the interest and article of studying how contact user by traditional collaborative filtering, but have ignored this important contextual information of time.Time, for the impact of user interest own profound, shows two aspects.

One, user interest is along with time variations.Such as user's section of having time is mainly absorbed in game class app, may turn to social class app more after he is weary of.User may want the app downloading back work during operation for another example, then tends to the app of amusement and travel class vacation.

Its two, app have ageing.Such as when this app of Angry Bird game wind in upper space wastes time, nearly all mobile device has installation.And recommend to seem to this app more now and there is no need, because user is generally very familiar to it for a long time.

We do cum rights without the time in the similarity process of commodity, and reason is such as in IBCF, and two users are if not have purchased two commodity simultaneously, and the commodity similarity so calculated will be lower, and this is irrational process.Therefore, still adopt the similarity between Jaccard formulae discovery user or app, unlike, with as shown in the formula form predict:

$r_{\mathrm{ui}}=\underset{j\∈R\left(u\right)}{\mathrm{\Σ}}{r}_{\mathrm{uj}}*s_{\mathrm{ij}}*\frac{1}{1+a(T-t)}$

A gets 0.1-0.9, and t represents that user downloads the time of app, and T represents the time that whole training set is last, in units of sky.Work as a=0.1, the prediction that obtains scoring is not almost failed in time, and result is similar to traditional Item-Based CF, works as a=0.9, and significant decline appears in prediction scoring in time, namely user from predicted time more close to behavior more can impact prediction result;

3, based on the collaborative filtering (Pseudo-Rating-Based CF) of puppet scoring: through finding the research of user behavior, popular commodity, user tends to directly click download button to download more, and does not go the details page paying close attention to app; And for those relatively unfashionable app, user then tends to the decision making download after having browsed details page more.This research conclusion meets our direct feel because those popular app user what's frequently heard can be repeated in detail already, do not need to check details page.More likely, user is natively finding such app, and the recommendation of commending system just facilitates searching of user.Therefore, be necessary the behavior of browsing and downloading and the behavior of only downloading to distinguish.

For this reason, the R matrix that algorithm uses herein is no longer 01 two-value, but uses for reference the thought of scoring, user has only browsed app just in corresponding matrix unit write 1, only downloaded app just in corresponding matrix unit write 4, user had not only downloaded but also had browsed app, just write 5 in corresponding units.Certainly also there is other more reasonably marking schemes.This pseudo-marking scheme also alleviates the too sparse problem of data greatly simultaneously.

Concentrated and impracticable owing to calculating the real data of testing in the present invention based on Pearson's similarity of marking in traditional field.Pearson's similarity is not used to be that it has reasonable tolerance for the scoring vector that user is real like this because it is portray the variation tendency between two vectors in essence.And if user has downloaded an app and another one user has browsed this app, in fact they or closely similar, and Pearson's similarity shows slightly not enough relative to Jaccard in this situation of tolerance.So based on above-mentioned consideration, final or employing Jaccard similarity herein.

Therefore pseudo-incorporating of scoring does not affect Similarity Measure link, and its effect is mainly reflected in prediction score calculation.Formula for as follows:

$r_{\mathrm{ui}}=\underset{j\∈R\left(u\right)}{\mathrm{\Σ}}{r}_{\mathrm{uj}}*s_{\mathrm{ij}}$

The r used in formula _ujvalue be no longer 0 and 1, the substitute is more accurate pseudo-scoring.And from the Data support of pseudo-rating matrix far away than 0,1 matrix is abundant, has also confirmed this point in actual model measurement;

4, based on svd (Singular Value Decomposition, SVD) model: implicit expression factor model as the improvement to traditional SVD model, by by user-article matrix R resolves into user's factor matrix and article factor matrix realizes dimensionality reduction.These models are very common in score in predicting, predict the scoring r of user u to article i by the inner product calculating user's factor matrix and the corresponding vector of article factor matrix _u,i:

$r_{\mathrm{ui}}=\mathrm{\μ}+b_u+b_i+p_u^T*q_i$

μ represents global bias item, i.e. the overall average of all scorings in training set.Before matrix decomposition, it is biased biased with article that the mean value generally by deducting R matrix row and column removes user.Be biased and finally can be added back to above-mentioned inner product to produce last prediction.Bu and bi represents user's bias term and article bias term respectively.Svd algorithm matrix R used can be two-value, also can mark based on puppet.Concrete effect is depending on actual data set characteristic.

Step S104, the APP recommending described expection high to user.

The embodiment of the present invention, obtain and the user behaviors log of recording user, user behavior matrix is generated according to user behaviors log, according to the proposed algorithm preset, user behavior matrix computations user is used to download the expection of APP, to the APP that user recommends described expection high, APP platform according to the APP Download History of user and APP preview record, can be called algorithm accurately and recommends APP to user accurately.

As an embodiment of the present invention, before the step of the described APP recommending described expection high to user, described method is further comprising the steps of:

Preset APP recommended amount.

As another alternative embodiment of the invention, after the step of the described APP recommending described expection high to user, described method is further comprising the steps of:

Obtain the accuracy that user downloads the high APP of described expection.

In embodiments of the present invention, accuracy rate/recall rate (precision/recall, pr) is generally used to measure the prediction accuracy of TopN recommendation.

If commending system shows the recommendation list of user to make R (u) represent, and T (u) represents the behavior list that user is actual on test set.So the accuracy rate of recommendation results can be defined as form as follows:

$\mathrm{Precision}=\frac{{\mathrm{\Σ}}_{u\∈U}|R\left(u\right)\∩T\left(u\right)|}{{\mathrm{\Σ}}_{u\∈U}\left|R\left(u\right)\right|}$

Corresponding recall rate can be defined as the form be shown below:

$\mathrm{Recall}=\frac{{\mathrm{\Σ}}_{u\∈U}|R\left(u\right)\∩T\left(u\right)|}{{\mathrm{\Σ}}_{u\∈U}\left|T\left(u\right)\right|}$

During TopN recommends, each proposed algorithm has one group of accuracy rate/recall rate according to different N values, and then can draw accuracy rate recall rate curve (precision/recall curve).

Need a kind of comprehensive index to compare each group accuracy rate/recall rate.Wherein, modal is exactly F-Measure:

$F_1=\frac{2\mathrm{PR}}{P+R}$

P, R are accuracy rate and recall rate respectively, and F1 has considered accuracy rate and recall rate, so can think that F1 is higher, the ability of proposed algorithm prediction user behavior is stronger.

Industry coverage rate (coverage) describes a commending system to the mining ability of article long-tail.It is exactly that recommended article account for the ratio of all article, under statement that coverage rate the most simply defines:

$\mathrm{Coverage}=\frac{\left|{\∪}_{u\∈U}R\left(u\right)\right|}{\left|I\right|}$

U represents all user's set of system, and I represents all article set of system, and correspondingly R (u) expression system is the article set that certain user u recommends.

If an app is very popular, so the download of user recommends the possibility of guiding very little, and the income that our commending system can receive is just less; If but an app compares unexpected winner, so pushing away to user him, to receive the income that commending system obtains just larger.Therefore we redefine the accuracy rate of cum rights, recall rate and F1 for particular problem, as follows, can be regarded as prediction accuracy and multifarious comprehensive measurement to a certain extent.

$\mathrm{Precision}=\frac{{\mathrm{\Σ}}_{u\∈U}{\mathrm{\Σ}}_{i\∈I}{w}_i}{{\mathrm{\Σ}}_{u\∈U}\left|R\left(u\right)\right|}$

$\mathrm{Recall}=\frac{{\mathrm{\Σ}}_{u\∈U}{\mathrm{\Σ}}_{i\∈I}{w}_i}{{\mathrm{\Σ}}_{u\∈U}\left|T\left(u\right)\right|}$

$F_1=\frac{2\mathrm{PR}}{P+R}$

I=|R (u) ∩ T (u) |, R (u) represents the user's recommendation list produced according to training set, and T (u) represents the behavior list that user is actual on test set.All app have a corresponding weight, and in order to the app of pop applies punishment, the weight determined must meet popular app weight and be less than unexpected winner app.Accordingly, herein propose weight calculation such as formula under:

$w_i=\frac{1}{{\log }_N^{C\left(i\right)}}$

C (i) represents the number of times that app i occurs in training set.Generally get N and equal 2.

Embodiment two

Be illustrated in figure 2 the structural drawing of the APP commending system that the embodiment of the present invention provides, for convenience of explanation, the part relevant to the embodiment of the present invention be only shown, comprise:

User behaviors log acquiring unit 201, for obtaining and the user behaviors log of recording user, described user behaviors log comprises: user downloads APP record, for browsing APP record.

In embodiments of the present invention, APP commending system obtains the user behaviors log of user, and passes through the form record behavior daily record of data set, and wherein, data include but not limited to: Download data set, Browse data set, and data are specially:

1, each of Download data centralization be recorded as user certain time download the record of app.Shape as (uid, aid, timestamp), wherein uid, aid be respectively user, app encryption after ID, timestamp be download behavior occur time, as 2000-01-01;

2, each of Browse data centralization be recorded as user certain time browse the record of app.Shape as (uid, aid, timestamp), wherein uid, aid be respectively user, app encryption after ID, timestamp be navigation patterns occur time, as 2000-01-01.

User behavior matrix generation unit 202, generates user behavior matrix for the user behaviors log obtained according to described user behaviors log acquiring unit 201.

In embodiments of the present invention, APP commending system is after obtaining user behaviors log, and generate the user behavior matrix for subsequent algorithm according to behavior daily record, this user behavior matrix can be 0,1 two values matrix, also can be other numerical value.

Expection computing unit 203, for the proposed algorithm that basis is preset, the user behavior matrix computations user using described user behavior matrix generation unit 202 to generate downloads the expection of APP.

In embodiments of the present invention, APP commending system, after obtaining user behavior matrix, calls default proposed algorithm, uses this user behavior matrix computations user to download the expection of given APP.Described proposed algorithm comprises:

1, based on the collaborative filtering (memory-based CF) of memory: factor data scale is excessive, consider to adopt the similarity between Jaccard formulae discovery user or app for counting yield.Collaborative filtering for based on article:

$s_{\mathrm{ij}}=\frac{|N\left(i\right)\∩N\left(j\right)|}{|N\left(i\right)\∪N\left(j\right)|}$

S _ijrepresent the similarity of app i and app j, N (i) represents all user's set of downloading app i.

For given user u and app i, can predict that u downloads the possibility of i with following formula, this possibility describes with r in formula:

$r_{\mathrm{ui}}=\underset{j\∈R\left(u\right)}{\mathrm{\Σ}}{r}_{\mathrm{uj}}*s_{\mathrm{ij}}$

S _ijrepresent the similarity of app i and j in similarity matrix, r _ujrepresent corresponding numerical information in user behavior matrix R.Notice that the r downloading app i possibility for weighing user u is the numerical value that can be greater than 1 in fact.For given user, calculate the r of user u to all app also do not downloaded _ui, carry out descending sort according to r to these app, for user recommends top n app, Here it is, and Top-N recommends;

2, time-based collaborative filtering (Time-Based CF): energy is concentrated on the interest and article of studying how contact user by traditional collaborative filtering, but have ignored this important contextual information of time.Time, for the impact of user interest own profound, shows two aspects.

One, user interest is along with time variations.Such as user's section of having time is mainly absorbed in game class app, may turn to social class app more after he is weary of.User may want the app downloading back work during operation for another example, then tends to the app of amusement and travel class vacation.

Its two, app have ageing.Such as when this app of Angry Bird game wind in upper space wastes time, nearly all mobile device has installation.And recommend to seem to this app more now and there is no need, because user is generally very familiar to it for a long time.

We do cum rights without the time in the similarity process of commodity, and reason is such as in IBCF, and two users are if not have purchased two commodity simultaneously, and the commodity similarity so calculated will be lower, and this is irrational process.Therefore, still adopt the similarity between Jaccard formulae discovery user or app, unlike, with as shown in the formula form predict:

$r_{\mathrm{ui}}=\underset{j\∈R\left(u\right)}{\mathrm{\Σ}}{r}_{\mathrm{uj}}*s_{\mathrm{ij}}*\frac{1}{1+a(T-t)}$

A gets 0.1-0.9, and t represents that user downloads the time of app, and T represents the time that whole training set is last, in units of sky.Work as a=0.1, the prediction that obtains scoring is not almost failed in time, and result is similar to traditional Item-Based CF, works as a=0.9, and significant decline appears in prediction scoring in time, namely user from predicted time more close to behavior more can impact prediction result;

3, based on the collaborative filtering (Pseudo-Rating-Based CF) of puppet scoring: through finding the research of user behavior, popular commodity, user tends to directly click download button to download more, and does not go the details page paying close attention to app; And for those relatively unfashionable app, user then tends to the decision making download after having browsed details page more.This research conclusion meets our direct feel because those popular app user what's frequently heard can be repeated in detail already, do not need to check details page.More likely, user is natively finding such app, and the recommendation of commending system just facilitates searching of user.Therefore, be necessary the behavior of browsing and downloading and the behavior of only downloading to distinguish.

For this reason, the R matrix that algorithm uses herein is no longer 01 two-value, but uses for reference the thought of scoring, user has only browsed app just in corresponding matrix unit write 1, only downloaded app just in corresponding matrix unit write 4, user had not only downloaded but also had browsed app, just write 5 in corresponding units.Certainly also there is other more reasonably marking schemes.This pseudo-marking scheme also alleviates the too sparse problem of data greatly simultaneously.

Concentrated and impracticable owing to calculating the real data of testing in the present invention based on Pearson's similarity of marking in traditional field.Pearson's similarity is not used to be that it has reasonable tolerance for the scoring vector that user is real like this because it is portray the variation tendency between two vectors in essence.And if user has downloaded an app and another one user has browsed this app, in fact they or closely similar, and Pearson's similarity shows slightly not enough relative to Jaccard in this situation of tolerance.So based on above-mentioned consideration, final or employing Jaccard similarity herein.

Therefore pseudo-incorporating of scoring does not affect Similarity Measure link, and its effect is mainly reflected in prediction score calculation.Formula for as follows:

$r_{\mathrm{ui}}=\underset{j\∈R\left(u\right)}{\mathrm{\Σ}}{r}_{\mathrm{uj}}*s_{\mathrm{ij}}$

The r used in formula _ujvalue be no longer 0 and 1, the substitute is more accurate pseudo-scoring.And from the Data support of pseudo-rating matrix far away than 0,1 matrix is abundant, has also confirmed this point in actual model measurement;

4, based on svd (Singular Value Decomposition, SVD) model: implicit expression factor model as the improvement to traditional SVD model, by by user-article matrix R resolves into user's factor matrix and article factor matrix realizes dimensionality reduction.These models are very common in score in predicting, predict the scoring r of user u to article i by the inner product calculating user's factor matrix and the corresponding vector of article factor matrix _u,i:

$r_{\mathrm{ui}}=\mathrm{\μ}+b_u+b_i+p_u^T*q_i$

μ represents global bias item, i.e. the overall average of all scorings in training set.Before matrix decomposition, it is biased biased with article that the mean value generally by deducting R matrix row and column removes user.Be biased and finally can be added back to above-mentioned inner product to produce last prediction.Bu and bi represents user's bias term and article bias term respectively.Svd algorithm matrix R used can be two-value, also can mark based on puppet.Concrete effect is depending on actual data set characteristic.

APP recommendation unit 204, the APP that the expection for recommending described expection computing unit 203 to calculate to user is high.

The embodiment of the present invention, obtain and the user behaviors log of recording user, user behavior matrix is generated according to user behaviors log, according to the proposed algorithm preset, user behavior matrix computations user is used to download the expection of APP, to the APP that user recommends described expection high, APP platform according to the APP Download History of user and APP preview record, can be called algorithm accurately and recommends APP to user accurately.

As an embodiment of the present invention, before described APP recommendation unit 204 is recommended, described system also comprises:

APP quantity presets unit 205, for default APP recommended amount.

As another alternative embodiment of the invention, after described APP recommendation unit 204 is recommended, described system also comprises:

Downloading accuracy acquiring unit 206, downloading the accuracy of the high APP of described expection for obtaining user.

In embodiments of the present invention, accuracy rate/recall rate (precision/recall, pr) is generally used to measure the prediction accuracy of TopN recommendation.

If commending system shows the recommendation list of user to make R (u) represent, and T (u) represents the behavior list that user is actual on test set.So the accuracy rate of recommendation results can be defined as form as follows:

$\mathrm{Precision}=\frac{{\mathrm{\Σ}}_{u\∈U}|R\left(u\right)\∩T\left(u\right)|}{{\mathrm{\Σ}}_{u\∈U}\left|R\left(u\right)\right|}$

Corresponding recall rate can be defined as the form be shown below:

$\mathrm{Recall}=\frac{{\mathrm{\Σ}}_{u\∈U}|R\left(u\right)\∩T\left(u\right)|}{{\mathrm{\Σ}}_{u\∈U}\left|T\left(u\right)\right|}$

During TopN recommends, each proposed algorithm has one group of accuracy rate/recall rate according to different N values, and then can draw accuracy rate recall rate curve (precision/recall curve).

Need a kind of comprehensive index to compare each group accuracy rate/recall rate.Wherein, modal is exactly F-Measure:

$F_1=\frac{2\mathrm{PR}}{P+R}$

P, R are accuracy rate and recall rate respectively, and F1 has considered accuracy rate and recall rate, so can think that F1 is higher, the ability of proposed algorithm prediction user behavior is stronger.

Industry coverage rate (coverage) describes a commending system to the mining ability of article long-tail.It is exactly that recommended article account for the ratio of all article, under statement that coverage rate the most simply defines:

$\mathrm{Coverage}=\frac{\left|{\∪}_{u\∈U}R\left(u\right)\right|}{\left|I\right|}$

U represents all user's set of system, and I represents all article set of system, and correspondingly R (u) expression system is the article set that certain user u recommends.

If an app is very popular, so the download of user recommends the possibility of guiding very little, and the income that our commending system can receive is just less; If but an app compares unexpected winner, so pushing away to user him, to receive the income that commending system obtains just larger.Therefore we redefine the accuracy rate of cum rights, recall rate and F1 for particular problem, as follows, can be regarded as prediction accuracy and multifarious comprehensive measurement to a certain extent.

$\mathrm{Precision}=\frac{{\mathrm{\Σ}}_{u\∈U}{\mathrm{\Σ}}_{i\∈I}{w}_i}{{\mathrm{\Σ}}_{u\∈U}\left|R\left(u\right)\right|}$

$\mathrm{Recall}=\frac{{\mathrm{\Σ}}_{u\∈U}{\mathrm{\Σ}}_{i\∈I}{w}_i}{{\mathrm{\Σ}}_{u\∈U}\left|T\left(u\right)\right|}$

$F_1=\frac{2\mathrm{PR}}{P+R}$

I=|R (u) ∩ T (u) |, R (u) represents the user's recommendation list produced according to training set, and T (u) represents the behavior list that user is actual on test set.All app have a corresponding weight, and in order to the app of pop applies punishment, the weight determined must meet popular app weight and be less than unexpected winner app.Accordingly, herein propose weight calculation such as formula under:

$w_i=\frac{1}{{\log }_N^{C\left(i\right)}}$

C (i) represents the number of times that app i occurs in training set.Generally get N and equal 2.

One of ordinary skill in the art will appreciate that the unit included by above-described embodiment two is carry out dividing according to function logic, but be not limited to above-mentioned division, as long as corresponding function can be realized; In addition, the concrete title of each functional unit, also just for the ease of mutual differentiation, is not limited to protection scope of the present invention.

Those of ordinary skill in the art it is also understood that, the all or part of step realized in above-described embodiment method is that the hardware that can carry out instruction relevant by program has come, described program can be stored in a computer read/write memory medium, described storage medium, comprises ROM/RAM, disk, CD etc.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, all any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within protection scope of the present invention.

Claims (8)

1. a method for APP recommendation, it is characterized in that, described method comprises:

Obtain and the user behaviors log of recording user, described user behaviors log comprises: user downloads APP record, for browsing APP record;

User behavior matrix is generated according to described user behaviors log;

According to the proposed algorithm preset, described user behavior matrix computations user is used to download the expection of APP;

To the APP that user recommends described expection high.

2. the method for claim 1, is characterized in that, before the step of the described APP recommending described expection high to user, described method is further comprising the steps of:

Preset APP recommended amount.

3. the method for claim 1, after the step of the described APP recommending described expection high to user, described method is further comprising the steps of:

Obtain the accuracy that user downloads the high APP of described expection.

4. the method as described in any one of claims 1 to 3, is characterized in that, described proposed algorithm comprises: based on memory collaborative filtering, time-based collaborative filtering, based on puppet scoring collaborative filtering.

5. a system for APP recommendation, it is characterized in that, described system comprises:

User behaviors log acquiring unit, for obtaining and the user behaviors log of recording user, described user behaviors log comprises: user downloads APP record, for browsing APP record;

User behavior matrix generation unit, generates user behavior matrix for the user behaviors log obtained according to described user behaviors log acquiring unit;

Expection computing unit, for the proposed algorithm that basis is preset, the user behavior matrix computations user using described user behavior matrix generation unit to generate downloads the expection of APP;

APP recommendation unit, the APP that the expection for recommending described expection computing unit to calculate to user is high.

6. system as claimed in claim 5, is characterized in that, before described APP recommendation unit is recommended, described system also comprises:

APP quantity presets unit, for default APP recommended amount.

7. system as claimed in claim 5, is characterized in that, after described APP recommendation unit is recommended, described system also comprises:

Downloading accuracy acquiring unit, downloading the accuracy of the high APP of described expection for obtaining user.

8. the system as described in any one of claim 5 ~ 7, is characterized in that, described proposed algorithm comprises: based on memory collaborative filtering, time-based collaborative filtering, based on puppet scoring collaborative filtering.