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

Abstract

The present invention is applicable to the field of information technology, and provides a method and system for APP recommendation. The method includes: acquiring and recording the behavior log of the user, and the behavior log includes: the user downloads the APP record and is used to browse the APP record; The behavior log generates a user behavior matrix; according to a preset recommendation algorithm, the user behavior matrix is used to calculate the user's expectation of downloading an APP; and the APP with the high expectation is recommended to the user. In the embodiment of the present invention, the user's behavior log is obtained and recorded, a user behavior matrix is generated according to the behavior log, and according to a preset recommendation algorithm, the user behavior matrix is used to calculate the user's expectation of downloading an APP, and the APP with a high expectation is recommended to the user, so that The APP platform can use accurate algorithms to accurately recommend APPs to users based on the user's APP download records and APP preview records.

Description

A method and system for APP recommendation

technical field

The invention belongs to the field of information technology, and in particular relates to an APP recommendation method and system.

Background technique

Today, when mobile applications (application, app for short) greatly facilitate people's daily life, a large number of mobile applications also bring confusion to users in choosing. Users need an efficient way to help them choose the few interesting apps from hundreds of thousands of apps. Inspired by the widespread application of recommendation systems in the traditional Internet field, the industry has turned its attention to app recommendations.

At present, most app stores use categories and popular rankings to help users find the apps they need, but this cannot meet the individual needs of different users, because all users are presented with the same content , while ignoring differences such as gender, age, and culture of users. Even for the same user, his interests tend to change over time. The keyword-based search is based on the user's clear description of their own needs. Therefore, when the user cannot clearly describe their own needs, the presented recommendation results often appear blind. There are very few personalized app recommendation systems that are actually put into use in the industry.

Contents of the invention

The purpose of the embodiments of the present invention is to provide an APP recommendation method and system to solve the problem of no personalized APP recommendation in the prior art.

The embodiment of the present invention is implemented in this way, an APP recommendation method, the method includes:

Obtain and record the user's behavior log, the behavior log includes: the user downloads APP records, and uses to browse APP records;

Generate a user behavior matrix according to the behavior log;

According to the preset recommendation algorithm, use the user behavior matrix to calculate the expectation of the user to download the APP;

The APP with high expectation is recommended to the user.

Another object of the embodiments of the present invention is to provide an APP recommendation system, the system comprising:

The behavior log acquisition unit is used to obtain and record the user's behavior log, and the behavior log includes: the user downloads the APP record, and is used to browse the APP record;

a user behavior matrix generating unit, configured to generate a user behavior matrix according to the behavior logs acquired by the behavior log acquisition unit;

An expected calculation unit, configured to use the user behavior matrix generated by the user behavior matrix generating unit to calculate the user's expectation of downloading the APP according to a preset recommendation algorithm;

An APP recommendation unit, configured to recommend to the user an APP with a high expectation calculated by the expectation calculation unit.

In the embodiment of the present invention, the user's behavior log is obtained and recorded, a user behavior matrix is generated according to the behavior log, and according to a preset recommendation algorithm, the user behavior matrix is used to calculate the user's expectation of downloading an APP, and the APP with a high expectation is recommended to the user, so that The APP platform can use accurate algorithms to accurately recommend APPs to users based on the user's APP download records and APP preview records.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings that need to be used in the descriptions of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only of the present invention. For some embodiments, those of ordinary skill in the art can also obtain other drawings based on these drawings without paying creative efforts.

FIG. 1 is a flowchart of an APP recommendation method provided by an embodiment of the present invention;

Fig. 2 is a structural diagram of an APP recommendation system provided by an embodiment of the present invention.

Detailed ways

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

In order to illustrate the technical solutions of the present invention, specific examples are used below to illustrate.

Embodiment one

As shown in Figure 1, it is a flow chart of the APP recommendation method provided by the embodiment of the present invention, and the method includes the following steps:

Step S101 , acquire and record the user's behavior log, the behavior log includes: the user downloads APP records, and uses to browse APP records.

In the embodiment of the present invention, the APP recommendation system obtains the user's behavior log, and records the behavior log in the form of a data set, wherein, the data includes but not limited to: Download data set, Browse data set, and the data is specifically:

1. Each record in the Download dataset is the record of the app downloaded by the user at a certain time. The shape is (uid, aid, timestamp), where uid and aid are the encrypted IDs of the user and app respectively, and timestamp is the time when the download behavior occurs, such as 2000-01-01;

2. Each record in the Browse dataset is the record of the app that the user browsed at a certain time. The format is (uid, aid, timestamp), where uid and aid are the encrypted IDs of the user and app respectively, and timestamp is the time when the browsing behavior occurred, such as 2000-01-01.

Step S102, generating a user behavior matrix according to the behavior log.

In the embodiment of the present invention, after the APP recommendation system acquires the behavior log, it generates a user behavior matrix for subsequent algorithms according to the behavior log. The user behavior matrix can be a binary matrix of 0 and 1, or other values .

Step S103 , using the user behavior matrix to calculate the user's expectation of downloading the APP according to the preset recommendation algorithm.

In the embodiment of the present invention, after acquiring the user behavior matrix, the APP recommendation system invokes a preset recommendation algorithm, and uses the user behavior matrix to calculate the user's expectation of downloading a given APP. The recommendation algorithm includes:

1. Memory-based collaborative filtering (memory-based CF): Due to the large data size, the Jaccard formula is used to calculate the similarity between users or apps for the sake of computational efficiency. Take item-based collaborative filtering as an example:

${\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cup;</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

s _ij represents the similarity between app i and app j, and N(i) represents the set of all users who have downloaded app i.

For a given user u and app i, the following formula can be used to predict the possibility of u downloading i, which is described by r in the formula:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; ui</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; uj</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$

s _ij represents the similarity between app i and j in the similarity matrix, and r _uj represents the corresponding numerical information in the user behavior matrix R. Note that r, which is used to measure the possibility of user u downloading app i, is actually a value that can be greater than 1. For a given user, calculate user u's r _ui for all apps that have not been downloaded, sort these apps in descending order according to r, and recommend the top N apps for the user, which is the Top-N recommendation;

2. Time-Based Collaborative Filtering (Time-Based CF): Traditional collaborative filtering algorithms focus on how to connect users' interests and items, but ignore the important contextual information of time. Time has a profound impact on user interest, manifested in two aspects.

First, user interests change over time. For example, a user mainly focuses on game apps for a period of time, and when he gets bored, he may turn to social apps more. Another example is that users may want to download apps that assist work during work, but they tend to use apps for leisure and travel during vacations.

Second, apps are time-sensitive. For example, when the app game Angry Bird first became popular, it was installed on almost all mobile devices. And now it seems unnecessary to recommend this app, because users are generally familiar with it.

We do not use time to carry weights in the similarity processing of products. The reason is that, for example, in IBCF, if two users do not purchase two products at the same time, the calculated product similarity will be low, which is unreasonable. processing. Therefore, the Jaccard formula is still used to calculate the similarity between users or apps. The difference is that the following formula is used for prediction:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; ui</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; uj</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; *</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}$

a ranges from 0.1 to 0.9, t represents the time when the user downloads the app, and T represents the last time of the entire training set, in days. When a=0.1, the predicted score hardly declines over time, and the result is similar to the traditional Item-Based CF. When a=0.9, the predicted score declines significantly over time, that is, the closer the user is to the predicted time, the more likely the behavior will be. affect forecast results;

3. Collaborative filtering based on pseudo-rating (Pseudo-Rating-Based CF): After research on user behavior, it is found that the more popular a product is, the more users tend to click the download button to download it directly, instead of paying attention to the detailed information of the app page; for those apps that are relatively unpopular, users are more inclined to make a download decision after browsing the detailed information page. This research conclusion is in line with our intuitive feeling, because those popular app users are already familiar with it, and there is no need to view the detailed information page at all. More likely, the user is already looking for such an app, and the recommendation of the recommendation system is just to facilitate the user's search. Therefore, it is necessary to distinguish the behavior of browsing and downloading from the behavior of merely downloading.

For this reason, the R matrix used by the algorithm here is no longer 01 binary, but draws on the idea of scoring. The user writes 1 in the corresponding matrix unit only after browsing the app, and writes 1 in the corresponding matrix unit only after downloading the app. Write 4, the user has downloaded and browsed the app, then write 5 in the corresponding unit. Of course, there are other more reasonable scoring schemes. At the same time, this pseudo-scoring scheme also greatly alleviates the problem of too sparse data.

Since the calculation of score-based Pearson similarity in the traditional field is not practical in the actual data set tested by the present invention. The reason why Pearson similarity is not used is because it essentially describes the change trend between two vectors, so that it will have a better measure of the user's real rating vector. And if a user downloads an app and another user browses the app, they are actually very similar, and the Pearson similarity is slightly insufficient compared to Jaccard in measuring this situation. Therefore, based on the above considerations, this paper finally uses the Jaccard similarity.

Therefore, the integration of pseudo-scoring does not affect the calculation of similarity, and its role is mainly reflected in the calculation of prediction scores. For a formula like this:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; ui</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; uj</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$

The value of r _uj used in the formula is no longer 0 and 1, but replaced by a more accurate pseudo score. Moreover, the data support from the pseudo-scoring matrix is far more sufficient than the 0, 1 matrix, which is also confirmed in the actual model test;

4. Based on the Singular Value Decomposition (SVD) model: As an improvement to the traditional SVD model, the implicit factor model achieves dimensionality reduction by decomposing the user-item matrix R into a user factor matrix and an item factor matrix. These models are very common in rating prediction, by calculating the inner product of the corresponding vectors of the user factor matrix and the item factor matrix to predict the rating r _u,i of user u for item i:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; ui</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

μ represents the global bias term, which is the global average of all ratings in the training set. Before matrix decomposition, the user bias and item bias are generally removed by subtracting the average values of the rows and columns of the R matrix. The bias is finally added back to the inner product above to produce the final prediction. bu and bi denote user bias and item bias, respectively. The matrix R used by the SVD algorithm can be binary or based on pseudo-scoring. The specific effect depends on the actual data set characteristics.

Step S104, recommending the APP with high expectation to the user.

In the embodiment of the present invention, the user's behavior log is obtained and recorded, a user behavior matrix is generated according to the behavior log, and according to a preset recommendation algorithm, the user behavior matrix is used to calculate the user's expectation of downloading an APP, and the APP with a high expectation is recommended to the user, so that The APP platform can use accurate algorithms to accurately recommend APPs to users based on the user's APP download records and APP preview records.

As an optional embodiment of the present invention, before the step of recommending the highly expected APP to the user, the method further includes the following steps:

The number of recommended APPs is preset.

As another optional embodiment of the present invention, after the step of recommending the highly expected APP to the user, the method further includes the following steps:

The accuracy of downloading the APP with high expectation by the user is obtained.

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

Let R(u) represent the recommendation list shown to the user by the recommendation system, and T(u) represent the actual behavior list of the user on the test set. Then the accuracy of the recommendation result can be defined as the following form:

$\mathrm{<span\; class="notranslate">\; Precision</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

The corresponding recall rate can be defined as follows:

$\mathrm{<span\; class="notranslate">\; recall</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

In the TopN recommendation, each recommendation algorithm has a set of accuracy/recall rates according to different N values, and then the accuracy/recall rate curve (precision/recall curve) can be drawn.

A comprehensive metric is needed to compare precision/recall for each group. Among them, the most common is F-Measure:

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; PR</span>}}{\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; R</span>}}$

P and R are the accuracy rate and recall rate respectively. F1 comprehensively considers the accuracy rate and recall rate, so it can be considered that the higher the F1, the stronger the ability of the recommendation algorithm to predict user behavior.

The industry uses coverage to describe the ability of a recommendation system to mine the long tail of items. The simplest definition of coverage is the ratio of recommended items to all items, expressed as follows:

$\mathrm{<span\; class="notranslate">\; Coverage</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; \&cup;</span>}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; |</span>}$

U represents the set of all users in the system, I represents the set of all items in the system, and correspondingly R(u) represents the set of items recommended by the system for a certain user u.

If an app is very popular, it is very unlikely that the user’s download will be recommended and guided, and the benefits that our recommendation system can receive will be small; but if an app is relatively unpopular, then if it is pushed to the user and he accepts it, the recommendation system will get The benefits are even greater. Therefore, we redefine the weighted accuracy, recall, and F1 for specific problems, as shown below, which can be regarded as a comprehensive measure of prediction accuracy and diversity to a certain extent.

$\mathrm{<span\; class="notranslate">\; Precision</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

$\mathrm{<span\; class="notranslate">\; recall</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; PR</span>}}{\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; R</span>}}$

I=|R(u)∩T(u)|, R(u) represents the user recommendation list generated according to the training set, and T(u) represents the user's actual behavior list on the test set. All apps have a corresponding weight. In order to impose penalties on popular apps, the determined weight must satisfy that the weight of popular apps is smaller than that of unpopular apps. Accordingly, this paper proposes the weight calculation as follows:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}}$

C(i) represents the number of times app i appears in the training set. Generally, N is equal to 2.

Embodiment two

As shown in Figure 2, it is a structural diagram of the APP recommendation system provided by the embodiment of the present invention. For the convenience of description, only the parts related to the embodiment of the present invention are shown, including:

The behavior log obtaining unit 201 is configured to obtain and record the user's behavior log, the behavior log includes: the user downloads APP records, and uses to browse APP records.

In the embodiment of the present invention, the APP recommendation system obtains the user's behavior log, and records the behavior log in the form of a data set, wherein, the data includes but not limited to: Download data set, Browse data set, and the data is specifically:

1. Each record in the Download dataset is the record of the app downloaded by the user at a certain time. The shape is (uid, aid, timestamp), where uid and aid are the encrypted IDs of the user and app respectively, and timestamp is the time when the download behavior occurs, such as 2000-01-01;

2. Each record in the Browse dataset is the record of the app that the user browsed at a certain time. The format is (uid, aid, timestamp), where uid and aid are the encrypted IDs of the user and app respectively, and timestamp is the time when the browsing behavior occurred, such as 2000-01-01.

The user behavior matrix generation unit 202 is configured to generate a user behavior matrix according to the behavior log acquired by the behavior log acquisition unit 201 .

In the embodiment of the present invention, after the APP recommendation system acquires the behavior log, it generates a user behavior matrix for subsequent algorithms according to the behavior log. The user behavior matrix can be a binary matrix of 0 and 1, or other values .

The expectation calculation unit 203 is configured to use the user behavior matrix generated by the user behavior matrix generation unit 202 to calculate the user's expectation of downloading an APP according to a preset recommendation algorithm.

In the embodiment of the present invention, after acquiring the user behavior matrix, the APP recommendation system invokes a preset recommendation algorithm, and uses the user behavior matrix to calculate the user's expectation of downloading a given APP. The recommendation algorithm includes:

1. Memory-based collaborative filtering (memory-based CF): Due to the large data size, the Jaccard formula is used to calculate the similarity between users or apps for the sake of computational efficiency. Take item-based collaborative filtering as an example:

${\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cup;</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

s _ij represents the similarity between app i and app j, and N(i) represents the set of all users who have downloaded app i.

For a given user u and app i, the following formula can be used to predict the possibility of u downloading i, which is described by r in the formula:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; ui</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; uj</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$

s _ij represents the similarity between app i and j in the similarity matrix, and r _uj represents the corresponding numerical information in the user behavior matrix R. Note that r, which is used to measure the possibility of user u downloading app i, is actually a value that can be greater than 1. For a given user, calculate user u's r _ui for all apps that have not been downloaded, sort these apps in descending order according to r, and recommend the top N apps for the user, which is the Top-N recommendation;

2. Time-Based Collaborative Filtering (Time-Based CF): Traditional collaborative filtering algorithms focus on how to connect users' interests and items, but ignore the important contextual information of time. Time has a profound impact on user interest, manifested in two aspects.

First, user interests change over time. For example, a user mainly focuses on game apps for a period of time, and when he gets bored, he may turn to social apps more. Another example is that users may want to download apps that assist work during work, but they tend to use apps for leisure and travel during vacations.

Second, apps are time-sensitive. For example, when the app game Angry Bird first became popular, it was installed on almost all mobile devices. And now it seems unnecessary to recommend this app, because users are generally familiar with it.

We do not use time to carry weights in the similarity processing of products. The reason is that, for example, in IBCF, if two users do not purchase two products at the same time, the calculated product similarity will be low, which is unreasonable. processing. Therefore, the Jaccard formula is still used to calculate the similarity between users or apps. The difference is that the following formula is used for prediction:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; ui</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; uj</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; *</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}$

a ranges from 0.1 to 0.9, t represents the time when the user downloads the app, and T represents the last time of the entire training set, in days. When a=0.1, the predicted score hardly declines over time, and the result is similar to the traditional Item-Based CF. When a=0.9, the predicted score declines significantly over time, that is, the closer the user is to the predicted time, the more likely the behavior will be. affect forecast results;

3. Pseudo-Rating-Based Collaborative Filtering (Pseudo-Rating-Based CF): After research on user behavior, it is found that the more popular the product, the more users tend to click the download button to download directly, without paying attention to the detailed information of the app page; for those apps that are relatively unpopular, users are more inclined to make a download decision after browsing the detailed information page. This research conclusion is in line with our intuitive feeling, because those popular app users are already familiar with it, and there is no need to view the detailed information page at all. More likely, the user is already looking for such an app, and the recommendation of the recommendation system is just to facilitate the user's search. Therefore, it is necessary to distinguish the behavior of browsing and downloading from the behavior of merely downloading.

For this reason, the R matrix used by the algorithm here is no longer 01 binary, but draws on the idea of scoring. The user writes 1 in the corresponding matrix unit only after browsing the app, and writes 1 in the corresponding matrix unit only after downloading the app. Write 4, the user has downloaded and browsed the app, then write 5 in the corresponding unit. Of course, there are other more reasonable scoring schemes. At the same time, this pseudo-scoring scheme also greatly alleviates the problem of too sparse data.

Since the calculation of score-based Pearson similarity in the traditional field is not practical in the actual data set tested by the present invention. The reason why Pearson similarity is not used is that it essentially describes the change trend between two vectors, so that it will have a better measure of the user's real rating vector. And if a user downloads an app and another user browses the app, they are actually very similar, and the Pearson similarity is slightly insufficient compared to Jaccard in measuring this situation. Therefore, based on the above considerations, this paper finally uses the Jaccard similarity.

Therefore, the integration of pseudo-scoring does not affect the calculation of similarity, and its role is mainly reflected in the calculation of prediction scores. For a formula like this:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; ui</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; uj</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}$

The value of r _uj used in the formula is no longer 0 and 1, but replaced by a more accurate pseudo score. Moreover, the data support from the pseudo-scoring matrix is far more sufficient than the 0, 1 matrix, which is also confirmed in the actual model test;

4. Based on the Singular Value Decomposition (SVD) model: As an improvement to the traditional SVD model, the implicit factor model achieves dimensionality reduction by decomposing the user-item matrix R into a user factor matrix and an item factor matrix. These models are very common in rating prediction, by calculating the inner product of the corresponding vectors of the user factor matrix and the item factor matrix to predict the rating r _u,i of user u for item i:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; ui</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

μ represents the global bias term, which is the global average of all ratings in the training set. Before matrix decomposition, the user bias and item bias are generally removed by subtracting the average values of the rows and columns of the R matrix. The bias is finally added back to the inner product above to produce the final prediction. bu and bi denote user bias and item bias, respectively. The matrix R used by the SVD algorithm can be binary or based on pseudo-scoring. The specific effect depends on the actual data set characteristics.

The APP recommending unit 204 is configured to recommend, to the user, APPs with high expectations calculated by the expectation calculating unit 203 .

In the embodiment of the present invention, the user's behavior log is obtained and recorded, a user behavior matrix is generated according to the behavior log, and according to a preset recommendation algorithm, the user behavior matrix is used to calculate the user's expectation of downloading an APP, and the APP with a high expectation is recommended to the user, so that The APP platform can use accurate algorithms to accurately recommend APPs to users based on the user's APP download records and APP preview records.

As an optional embodiment of the present invention, before the APP recommendation unit 204 recommends, the system further includes:

APP quantity preset unit 205, configured to preset recommended quantity of APPs.

As another optional embodiment of the present invention, after the APP recommending unit 204 recommends, the system further includes:

The download accuracy obtaining unit 206 is configured to obtain the user's download accuracy of the expected high APP.

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

Let R(u) represent the recommendation list shown to the user by the recommendation system, and T(u) represent the actual behavior list of the user on the test set. Then the accuracy of the recommendation result can be defined as the following form:

$\mathrm{<span\; class="notranslate">\; Precision</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

The corresponding recall rate can be defined as follows:

$\mathrm{<span\; class="notranslate">\; recall</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&cap;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

In the TopN recommendation, each recommendation algorithm has a set of accuracy/recall rates according to different N values, and then the accuracy/recall rate curve (precision/recall curve) can be drawn.

A comprehensive metric is needed to compare precision/recall for each group. Among them, the most common is F-Measure:

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; PR</span>}}{\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; R</span>}}$

P and R are the accuracy rate and recall rate respectively. F1 comprehensively considers the accuracy rate and recall rate, so it can be considered that the higher the F1, the stronger the ability of the recommendation algorithm to predict user behavior.

The industry uses coverage to describe the ability of a recommendation system to mine the long tail of items. The simplest definition of coverage is the ratio of recommended items to all items, expressed as follows:

$\mathrm{<span\; class="notranslate">\; Coverage</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; \&cup;</span>}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; |</span>}$

U represents the set of all users in the system, I represents the set of all items in the system, and correspondingly R(u) represents the set of items recommended by the system for a certain user u.

If an app is very popular, it is very unlikely that the user’s download will be recommended and guided, and the benefits that our recommendation system can receive will be small; but if an app is relatively unpopular, then if it is pushed to the user and he accepts it, the recommendation system will get The benefits are even greater. Therefore, we redefine the weighted accuracy, recall, and F1 for specific problems, as shown below, which can be regarded as a comprehensive measure of prediction accuracy and diversity to a certain extent.

$\mathrm{<span\; class="notranslate">\; Precision</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

$\mathrm{<span\; class="notranslate">\; recall</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; u</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}$

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; PR</span>}}{\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; R</span>}}$

I=|R(u)∩T(u)|, R(u) represents the user recommendation list generated according to the training set, and T(u) represents the user's actual behavior list on the test set. All apps have a corresponding weight. In order to impose penalties on popular apps, the determined weight must satisfy that the weight of popular apps is smaller than that of unpopular apps. Accordingly, this paper proposes the weight calculation as follows:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}^{\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}}$

C(i) represents the number of times app i appears in the training set. Generally, N is equal to 2.

Those of ordinary skill in the art can understand that the various units included in the above-mentioned embodiment 2 are only divided according to functional logic, but are not limited to the above-mentioned division, as long as the corresponding functions can be realized; in addition, the specific functions of each functional unit The names are only for the convenience of distinguishing each other, and are not used to limit the protection scope of the present invention.

Those of ordinary skill in the art can also understand that all or part of the steps in the method of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, so The storage medium mentioned above includes ROM/RAM, magnetic disk, optical disk, etc.

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

Claims (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.