CN106202515A - A kind of Mobile solution based on sequence study recommends method and commending system thereof - Google Patents

Abstract

The present invention discloses a kind of Mobile solution based on sequence study and recommends method and commending system thereof, and this recommendation method recommends that APP be modeled as a sequencing problem, compares and traditionally recommendation is regarded as scoring problem, has higher recommendation accuracy rate.The sort recommendations algorithm that commending system of the present invention uses, combines the high efficiency and by the high accuracy predicting sort algorithm and algorithm idea is skeletonisation that the algorithm that sorts top to bottom calculates, can have high expansion in conjunction with different proposed algorithms.

Description

A kind of Mobile solution based on sequence study recommends method and commending system thereof

Technical field

The invention belongs to technical field of electronic commerce, be specifically related to a kind of Mobile solution based on sequence study and recommend method And commending system.

Background technology

Today of people's daily life, substantial amounts is greatly facilitated at mobile applications (Application, APP) APP also band give user select puzzlement.User needs a kind of efficient approach to help them from millions of APP selects small part interested.Recommended system is in the wide variety of inspiration in conventional internet field, and industry is by sight Invest the recommendation of APP.

The most most of APP shop have employed the scheme help of classified catalogue and popular ranking list and the user discover that required APP, but this is not well positioned to meet the individual demand of different user, because being presented on being just as in face of all users Content, and have ignored the differences such as the sex of user, age and culture.For same user, his interest is past Toward changing over time.Search based on keyword is then set up on self-demand is expressly recited by user, therefore In the case of user can not be expressly recited self-demand, the recommendation results presented often seems blindly.Industry really puts into and makes Personalized APP commending system little.

Summary of the invention

Present invention aim at providing a kind of Mobile solution based on sequence study to recommend method and commending system thereof, to solve The problem certainly lacking of the sort recommendations algorithm frame of APP at present.

A kind of Mobile solution based on sequence study recommends method, comprises the steps:

(1) collecting user's behavioral data about APP, described behavioral data comprises user and downloads and browse APP's Historical record；

(2) described behavioral data is carried out pretreatment and therefrom extracts user's characteristic vector for each APP；Then Select a kind of score in predicting model, and then calculate user's scoring for each APP according to characteristic vector；

(3) use gradient descent method that each model parameter in described score in predicting model is carried out more according to above-mentioned scoring Newly；

(4) according to the score in predicting model established after updating, user's commenting for each APP not downloaded is calculated Point, and by this scoring, APP is ranked up, extract several the highest APP that mark and recommend user.

Described step (2) carries out pretreatment to behavioral data, implements and include missing values is passed through statistical simulation It is filled with and exceptional value carries out screening removing.

The score in predicting model selected in described step (2) can use logistic regression algorithm (LR) or gradient to promote back Return tree algorithm (GBRT).

In described step (3), loss function C based on intersection information entropy uses gradient descent method to score in predicting model In each model parameter be updated.

The expression formula of described loss function C is as follows:

$C=-\stackrel{\‾}{P_{ij}}\lg P_{ij}-(1-\stackrel{\‾}{P_{ij}})\lg (1-P_{ij})=\frac{1}{2}(1-S_{ij})(s_i-s_j)+\lg (1+e^{-(s_i-s_j)})$

Wherein: P_ijFor user, the preference of i-th APP is higher than the prediction probability to jth APP,For P_ijCorresponding True probability, s_iAnd s_jIt is respectively user for i-th APP and the scoring of jth APP, S_ij=0 or ± 1, if s_i=s_j, then S_ij =0；If s_i＞ s_j, then S_ij=1；If s_i＜ s_j, then S_ij=-1；I and j is natural number and 1≤i≤n, 1≤j≤n, i ≠ j, n Quantity for APP.

Described prediction probability P_ijExpression formula as follows:

$P_{ij}=\frac{1}{1+e^{-(s_i-s_j)}}$

Described scoring S_i=f (x_i), s_j=f (x_j), x_iAnd x_jIt is respectively i-th APP and the characteristic vector of jth APP, F () represents the score in predicting model before updating.

For the arbitrary model parameter in score in predicting model in described step (3), its update algorithm is as follows:

$w_k^{*}=w_k-\mathrm{\η}\frac{\∂C}{\∂w_k}=w_k-\mathrm{\η}(\frac{\∂C}{\∂s_i}\frac{\∂s_i}{\∂w_k}+\frac{\∂C}{\∂s_j}\frac{\∂s_j}{\∂w_k})$

Wherein: w_kWithBeing respectively the kth model parameter in score in predicting model before and after updating, η is for updating coefficient, k For natural number and 1≤k≤K, K is the quantity of model parameter in score in predicting model.

A kind of Mobile solution commending system based on sequence study, including:

User behavior data collection module, for collecting user's behavioral data about APP；

Training data generation module, for carrying out pretreatment and therefrom extracting user for respectively to described behavioral data The characteristic vector of APP；And then utilize selected score in predicting model to calculate user according to characteristic vector each APP is commented Point；

Parameter calculating module, for using gradient descent method to each mould in described score in predicting model according to above-mentioned scoring Shape parameter is updated；

Recommending module, for according to the score in predicting model established after updating, calculating user and not downloading for each The scoring of APP, and by this scoring, APP is ranked up, extracts several the highest APP that mark and recommend user.

Advantages of the present invention is as follows:

(1) present invention recommends that APP be modeled as a sequencing problem, compares and traditionally recommendation is regarded as scoring problem, tool There is higher recommendation accuracy rate.

(2) the sort recommendations algorithm that the present invention uses, combines the high efficiency of the algorithm calculating that sorts top to bottom and by sequence The high accuracy of algorithm predicts, and algorithm idea is skeletonisation, can have high expansion in conjunction with different proposed algorithms.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet that the present invention recommends method.

Fig. 2 is the structural representation of commending system of the present invention.

Detailed description of the invention

In order to more specifically describe the present invention, below in conjunction with the accompanying drawings and detailed description of the invention is to technical scheme It is described in detail.

Present invention Mobile solution based on sequence study recommends method, first carries out the variable in algorithm frame and formula Definition:

The prediction of APP i is marked by formula 1. active user u:

s_i=f (x)

Wherein: x is the related data vector of i, and concrete calculation f can be any calculation that can carry out score in predicting Method, such as logistic regression (LR), gradient promotes regression tree (GBRT) etc..

Formula 2. user u is to probability higher than j of the preference of APP i:

$P_{ij}=\frac{1}{1+e^{-(s_i-s_j)}}$

Wherein: e is the truth of a matter of natural logrithm.

Formula 3. loss function based on intersection information entropy:

$\begin{array}{c}C=-\stackrel{\‾}{P_{ij}}\lg P_{ij}-(1-\stackrel{\‾}{P_{ij}})\lg (1-P_{ij})\\ =\frac{1}{2}(1-S_{ij})(s_i-s_j)+\lg (1+e^{-(s_i-s_j)})\end{array}$

Wherein:Represent prediction probability P_ijCorresponding true probability.S_ij∈ 0, ± 1}, S_ij=0 represents APP i and APP The scoring of j is consistent, S_ij=1 represents that the scoring of i is higher than j, otherwise then less than j.

Formula 4. model parameter updates:

$w_k=w_k-\mathrm{\η}\frac{\∂C}{\∂w_k}=w_k-\mathrm{\η}(\frac{\∂C}{\∂s_i}\frac{\∂s_i}{\∂w_k}+\frac{\∂C}{\∂s_j}\frac{\∂s_j}{\∂w_k})$

$\frac{\∂C}{\∂s_i}\frac{\∂s_i}{\∂w_k}+\frac{\∂C}{\∂s_j}\frac{\∂s_j}{\∂w_k}=\left(\frac{1}{2}\right(1-S_{ij})-\frac{1}{1+e^{(s_i-s_j)}})(\frac{\∂s_i}{\∂w_k}-\frac{\∂s_j}{\∂w_k})={\mathrm{\λ}}_{ij}(\frac{\∂s_i}{\∂w_k}-\frac{\∂s_j}{\∂w_k})$

Wherein: w_kRepresent and calculate any one parameter comprised in the model f of prediction scoring s.Therefore, determine f it After i.e. can determine thatAnd then calculateAnd the mode using gradient to decline solves any one parameter w_k, k=1 ～K (K represents model parameter number).

The NDCG index in formula 5. conventional IR field:

$DCG=\underset{i=1}{\overset{T}{\Σ}}\frac{2^{l_i}-1}{log1+i}$

$NDCG=\frac{DCG}{\max DCG}$

Wherein: l_i{ 0,1,2,3,4,5} in conventional IR field, represents that active user u is in recommendation list for ∈ The interest level of i-th article.T represents the commodity number of recommendation.MaxDCG is expressed as user u and recommends the maximum of T commodity DCG, T the commodity i.e. recommended sort from high to low according to degree of user interest.

Formula 6. introduces the parameter update mode of score information:

${\mathrm{\λ}}_{ij}=\frac{-1}{1+e^{(s_i-s_j)}}\left|\mathrm{\Δ}NDCG\right|$

The present invention with user, the scoring of APP replaces tradition NDCG calculate in l_i, Δ NDCG represents current u Speech, after the sorting position of APP i and j exchanges, the knots modification of NDCG；Specifically calculate process as shown in Figure 1:

(1) model f is determined.Such as f represents Logic Regression Models.

(2) parameter more new formula is determined.With the λ in formula 6_ijReplace the λ in formula 4_ij, in combination with concrete f, really Determine the concrete form that formula 4 is current.

(3) parameter is calculated.According to the formula 4 in previous step, input training data gradient declines and carries out parameter calculating.

As shown in Figure 2, it is achieved the commending system of above-mentioned algorithm frame includes:

User behavior data collection module, the various actions data of real-time collecting user, it is deposited in background data base. Source data as various algorithms input data.

Training data generation module, the various algorithms used by framework produce concrete required input data.F can be not Same score calculation algorithm, its required input data of algorithms of different are different.The input number of the sorting algorithms such as such as logistic regression According to being: ID, application ID, score value, feature 1 ... feature n；And its required input data of matrix decomposition class algorithm are: use Family ID, application ID, score value.It is thus desirable to training data generation module carries out data prediction, generate corresponding for algorithms of different Input data.

Parameter calculating module, using the output data of training data generation module as input, its concrete internal process such as Fig. 1 Shown in.Training data according to the score in predicting algorithm f determined and corresponding parameter more new formula and needs carries out parameter meter Calculate, obtain final f.

Recommending module, according to final score in predicting algorithm f, it was predicted that user u comments the APP's that each was not downloaded Point, it is recommended that mark the highest front T APP.

The above-mentioned description to embodiment is to be understood that for ease of those skilled in the art and apply the present invention. Above-described embodiment obviously easily can be made various amendment by person skilled in the art, and described herein typically Principle is applied in other embodiments without through performing creative labour.Therefore, the invention is not restricted to above-described embodiment, ability Field technique personnel should be in protection scope of the present invention according to the announcement of the present invention, the improvement made for the present invention and amendment Within.

Claims (8)

1. Mobile solution based on sequence study recommends a method, comprises the steps:

(1) collecting user's behavioral data about APP, described behavioral data comprises the history that user downloads and browses APP Record；

(2) described behavioral data is carried out pretreatment and therefrom extracts user's characteristic vector for each APP；Then select A kind of score in predicting model, and then calculate user's scoring for each APP according to characteristic vector；

(3) use gradient descent method that each model parameter in described score in predicting model is updated according to above-mentioned scoring；

(4) according to the score in predicting model established after updating, user's scoring for each APP not downloaded is calculated, and By this scoring, APP is ranked up, extracts several the highest APP that mark and recommend user.

Mobile solution the most according to claim 1 recommends method, it is characterised in that: to behavior number in described step (2) According to carrying out pretreatment, implement and include being filled with and exceptional value is carried out screening going by statistical simulation to missing values Remove.

Mobile solution the most according to claim 1 recommends method, it is characterised in that: that selects in described step (2) comments Forecast model is divided to use logistic regression algorithm or gradient to promote regression tree algorithm.

Mobile solution the most according to claim 1 recommends method, it is characterised in that: based on intersection in described step (3) The loss function C of comentropy uses gradient descent method to be updated each model parameter in score in predicting model.

Mobile solution the most according to claim 4 recommends method, it is characterised in that: the expression formula of described loss function C is such as Under:

$C=-\stackrel{\‾}{P_{ij}}\lg P_{ij}-(1-\stackrel{\‾}{P_{ij}})\lg (1-P_{ij})=\frac{1}{2}(1-S_{ij})(s_i-s_j)+\lg (1+e^{-(s_i-s_j)})$

Wherein: P_ijFor user, the preference of i-th APP is higher than the prediction probability to jth APP,For P_ijCorresponding is true general Rate, s_iAnd s_jIt is respectively user for i-th APP and the scoring of jth APP, S_ij=0 or ± 1, if S_i=S_j, then S_ij=0；If s_i＞ s_j, then S_ij=1；If s_i＜ s_j, then S_ij=-1；I and j is natural number and 1≤i≤n, 1≤j≤n, and i ≠ j, n are APP's Quantity.

Mobile solution the most according to claim 5 recommends method, it is characterised in that: described prediction probability P_ijExpression formula such as Under:

$P_{ij}=\frac{1}{1+e^{-(s_i-s_j)}}.$

Mobile solution the most according to claim 5 recommends method, it is characterised in that: described scoring s_i=f (x_i), s_j=f (x_j), x_iAnd x_jBeing respectively i-th APP and the characteristic vector of jth APP, f () represents the score in predicting model before updating.

Mobile solution the most according to claim 5 recommends method, it is characterised in that: for scoring in described step (3) Arbitrary model parameter in forecast model, its update algorithm is as follows:

$w_k^{*}=w_k-\mathrm{\η}\frac{\∂C}{\∂w_k}=w_k-\mathrm{\η}(\frac{\∂C}{\∂s_i}\frac{\∂s_i}{\∂w_k}+\frac{\∂C}{\∂s_j}\frac{\∂s_j}{\∂w_k})$

Wherein: w_kWithBeing respectively the kth model parameter in score in predicting model before and after updating, η is for updating coefficient, and k is certainly So number and 1≤k≤K, K are the quantity of model parameter in score in predicting model.