CN109544241B

CN109544241B - Click rate estimation model construction method, click rate estimation method and device

Info

Publication number: CN109544241B
Application number: CN201811428618.2A
Authority: CN
Inventors: 陈晓爽; 郑胤; 马文晔; 黄俊洲
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2018-11-27
Filing date: 2018-11-27
Publication date: 2023-09-22
Anticipated expiration: 2038-11-27
Also published as: CN109544241A

Abstract

The embodiment of the application discloses a construction method of a click rate estimation model, a click rate estimation method and a device, wherein the click rate estimation model is provided with N different candidate dimensions, one candidate dimension is determined from the N different candidate dimensions to serve as a maximum dimension corresponding to a target feature, and i candidate dimensions smaller than the maximum dimension in the N different candidate dimensions are determined to serve as projection dimensions corresponding to the feature; the target feature has a corresponding maximum dimension and i corresponding projection dimensions, and i+1 feature vectors can be trained for the target feature, respectively. Therefore, the feature vector corresponding to the maximum dimension can reasonably reflect the information of the target feature in the training sample, and the problem of over-fitting or under-fitting cannot occur. Moreover, the inner product calculation of the feature vector is carried out on the target feature and other features with the maximum dimension smaller than that of the target feature, so that higher estimation accuracy is achieved.

Description

Click rate estimation model construction method, click rate estimation method and device

Technical Field

The present application relates to the field of data processing, and in particular, to a method for constructing a click rate estimation model, a click rate estimation method, and a related device.

Background

The click rate is the ratio of the number of times a certain content (news, advertisement or product) is clicked by the user to the number of times it is displayed on the client, i.e. the probability that the content is clicked by the user. In online application, the click rate of a user on a certain content is predicted, so that whether to recommend the information to the user is determined, and the method is an important mode for improving user experience. The model for estimating the click rate is called a click rate estimation model, and the model can estimate the probability of clicking a certain content by a user under a certain background according to the relevant information of the user and the content and the like through the click rate estimation model.

The resolver (Factorization Machine, FM) model is a commonly used click rate estimation model. In the FM model, a user, content and the like are respectively used as different features, each feature is allocated with a corresponding feature vector, and when the click rate of the user on certain content is estimated, the inner product between the feature vectors corresponding to the user and the content can be calculated through the FM model to obtain an estimated result.

The premise of calculating the inner product between feature vectors is that the dimensions of the feature vectors involved in the calculation are the same. Thus, in the conventional approach, in order to facilitate computation of the inner product between feature vectors, in the FM model, the feature vectors assigned to different features must have the same dimensions.

However, in real data, a large number of features have fewer non-zero samples, and only a small number of features have more non-zero samples. Taking the content, particularly a movie, as an example, the number of times (i.e., the number of samples) of watching a small number of hot movies is large, while the number of times of watching a large number of cold movies is relatively small. The content of the feature can be represented by adopting the feature vector with fewer dimensions for the feature with fewer non-zero samples, and the content of the feature can be represented by adopting the feature vector with more dimensions for the feature with more non-zero samples.

However, in order to calculate the inner product between feature vectors of different features, the click rate estimation models such as the FM model in the conventional manner have the same feature vector dimension for different features, so that feature vectors corresponding to certain features are over-fitted, for example, feature vectors of a cold film, and certain feature vectors are under-fitted, for example, feature vectors of a hot film, thereby affecting the estimation accuracy of the click rate.

Disclosure of Invention

In order to solve the technical problems, the application provides a click rate estimation model construction method, a click rate estimation method and a click rate estimation device, the constructed click rate estimation model can not have the problem of over fitting or under fitting, and the inner product calculation of the feature vector of the target feature and other features with the maximum dimension smaller than that of the target feature is not influenced, so that higher estimation precision is achieved.

The embodiment of the application discloses the following technical scheme:

in a first aspect, an embodiment of the present application provides a method for constructing a click rate prediction model, where N different candidate dimensions are set in the click rate prediction model, and N is a natural number greater than or equal to 2; the candidate dimension is used to identify the dimension of the feature vector, the method comprising:

determining one candidate dimension from the N different candidate dimensions as a maximum dimension corresponding to the target feature;

determining i candidate dimensions smaller than the maximum dimension in the N different candidate dimensions as projection dimensions corresponding to the target features; i is a natural number smaller than N and larger than or equal to 1;

and training i+1 feature vectors with different dimensions for the target feature according to the training sample corresponding to the target feature, wherein the dimension of any one feature vector in the i+1 feature vectors with different dimensions is one of the maximum dimension and the projection dimension corresponding to the target feature.

In a second aspect, an embodiment of the present application provides a click rate estimation method, where the method includes:

obtaining a sample to be estimated comprising a plurality of features, the plurality of features comprising at least a first feature and a second feature;

Calculating a click rate estimated value of the sample to be estimated through a click rate estimated model according to the feature vectors respectively corresponding to the plurality of features; the click rate estimation model is provided with N different candidate dimensions, and the candidate dimensions are used for identifying the dimensions of the feature vector; if the maximum dimension corresponding to a feature in the click rate estimation model is greater than i of the N different candidate dimensions, the feature has i+1 feature vectors of different dimensions; the maximum dimension corresponding to any one of the plurality of features is used for identifying the maximum dimension of the feature vector corresponding to the feature; n is a natural number greater than or equal to 2, i is a natural number less than N and greater than or equal to 1;

for the first feature and the second feature, the calculating the click rate estimated value of the sample to be estimated through the click rate estimation model includes:

if the first maximum dimension corresponding to the first feature in the click rate estimation model is different from the second maximum dimension corresponding to the second feature, the smallest dimension in the first maximum dimension and the second maximum dimension is used as the dimension to be calculated;

an inner product of the feature vector of the first feature having the dimension to be calculated and the feature vector of the second feature having the dimension to be calculated is calculated.

In a third aspect, an embodiment of the present application provides a device for constructing a click rate estimation model, where the click rate estimation model sets N different candidate dimensions, where the candidate dimensions are used to identify dimensions of feature vectors, and the device includes a first determining unit, a second determining unit, and a training unit:

the first determining unit is used for determining one candidate dimension from the N different candidate dimensions as a maximum dimension corresponding to the target feature;

the second determining unit is configured to determine i candidate dimensions smaller than the maximum dimension among the N different candidate dimensions as projection dimensions corresponding to the target feature; i is a natural number smaller than N and larger than or equal to 1;

the training unit is configured to train i+1 feature vectors with different dimensions for the target feature according to the training samples corresponding to the target feature, where the dimension of any one feature vector is one of the maximum dimension and the projection dimension corresponding to the target feature in the i+1 feature vectors with different dimensions.

In a fourth aspect, an embodiment of the present application provides a click rate estimating apparatus, where the apparatus includes an obtaining unit and a calculating unit:

The acquisition unit is used for acquiring a sample to be estimated, wherein the sample to be estimated comprises a plurality of characteristics, and the plurality of characteristics at least comprise a first characteristic and a second characteristic;

the calculating unit is used for calculating a click rate estimated value of the sample to be estimated through a click rate estimated model according to the feature vectors respectively corresponding to the plurality of features; the click rate estimation model is provided with N different candidate dimensions, and the candidate dimensions are used for identifying the dimensions of the feature vector; if the maximum dimension corresponding to a feature in the click rate estimation model is greater than i of the N different candidate dimensions, the feature has i+1 feature vectors of different dimensions; the maximum dimension corresponding to any one of the plurality of features is used for identifying the maximum dimension of the feature vector corresponding to the feature; n is a natural number greater than or equal to 2, i is a natural number less than N and greater than or equal to 1;

for the first and second features, the computing unit is further to:

In a fifth aspect, an embodiment of the present application provides a device for constructing a click rate estimation model, where the device includes a processor and a memory:

the memory is used for storing program codes and transmitting the program codes to the processor;

the processor is configured to execute the method for constructing the click rate estimation model according to the first aspect according to the instruction in the program code.

In a sixth aspect, an embodiment of the present application provides an apparatus for click rate estimation, where the apparatus includes a processor and a memory:

the processor is configured to execute the click rate estimation method according to the second aspect according to the instruction in the program code.

In a seventh aspect, an embodiment of the present application provides a computer readable storage medium, where the computer readable storage medium is configured to store program code for executing the method for constructing the click rate estimation model described in the first aspect, or for executing the method for estimating the click rate described in the second aspect.

According to the technical scheme, the click rate estimation model is provided with N different candidate dimensions, and each candidate dimension is used for identifying the dimension of the feature vector in the click rate estimation model. When the feature vector of the target feature is determined, one candidate dimension is determined from the N different candidate dimensions as a maximum dimension corresponding to the target feature, and i candidate dimensions smaller than the maximum dimension in the N different candidate dimensions are determined as projection dimensions corresponding to the feature; the target feature has a corresponding maximum dimension and i corresponding projection dimensions, and when feature vectors are found for the target feature according to training samples corresponding to the target feature, i+1 feature vectors can be trained respectively, and the dimension corresponding to different feature vectors is one of the maximum dimension and the projection dimension. Therefore, the feature vector corresponding to the maximum dimension can reasonably reflect the information of the target feature in the training sample, and the problem of over-fitting or under-fitting cannot occur. In order to realize the inner product of feature vectors among features with different maximum dimensions, the click rate estimation model trains i+1 feature vectors with different dimensions for the target features, so that the inner product calculation of the feature vectors is carried out on the target features and other features with the maximum dimensions smaller than the target features, and higher estimation accuracy is achieved.

Drawings

In order to more clearly illustrate the embodiments of the application or the technical solutions of the prior art, the drawings which are used in the description of the embodiments or the prior art will be briefly described, it being obvious that the drawings in the description below are only some embodiments of the application, and that other drawings can be obtained according to these drawings without inventive faculty for a person skilled in the art.

FIG. 1 is a schematic diagram of a click rate estimation system according to an embodiment of the present application;

FIG. 2 is a schematic diagram illustrating a comparison between feature vectors in a click rate estimation model and feature vectors in an FM model according to an embodiment of the present application;

FIG. 3 is a flowchart of a method for constructing a click rate estimation model according to an embodiment of the present application;

FIG. 4 is a flowchart of a click rate estimation method according to an embodiment of the present application;

FIG. 5 is a schematic diagram of calculating a click rate prediction value according to an embodiment of the present application;

FIG. 6 is a block diagram of a device for constructing a click rate estimation model according to an embodiment of the present application;

FIG. 7 is a block diagram of a click rate estimating apparatus according to an embodiment of the present application;

FIG. 8 is a schematic structural diagram of a device for constructing a click rate estimation model according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of a construction device for click rate estimation model according to an embodiment of the present application.

Detailed Description

In order to make the present application better understood by those skilled in the art, the following description will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings, and it is apparent that the described embodiments are only some embodiments of the present application, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

The terms "first," "second," "third," "fourth" and the like in the description and in the claims and in the above drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments of the application described herein may be implemented, for example, in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

Since the inner product between feature vectors needs to be calculated in the calculation of the estimated click rate, and the dimensions of the feature vectors for calculating the inner product must be the same, the conventional click rate estimation model trains feature vectors of the same dimension no matter how many non-zero samples correspond to the features when training the corresponding feature vectors for the features. This results in overfitting of feature vectors corresponding to certain features, such as those of a cold film, and underfilling of certain feature vectors, such as those of a hot film, thereby affecting the accuracy of the click rate estimate.

Therefore, the embodiment of the application provides a construction method of a click rate estimation model and a corresponding click rate estimation method, which are applied to a click rate estimation system.

FIG. 1 is a schematic diagram of a click rate estimation system.

The feature extraction module is used for extracting features from user data of a user, content data of a content platform and background data of a background database and sending the features to the training database. The training database sends the characteristics and the obtained actual click data of the user to the model training module, and the model training module trains the characteristic vectors corresponding to the characteristics according to the obtained characteristics and the actual click data of the user, so that a click rate estimation model is constructed. The training database can send the click rate estimation model obtained by construction to the estimation module, and the estimation module carries out click rate estimation on the sample carrying the characteristics according to the click rate estimation model. The estimating module sends the obtained click rate estimating result to the reordering module, the reordering module sorts the click rate estimating result and returns the obtained recommending result to the user.

The method for constructing the click rate estimation model provided by the embodiment of the application is mainly implemented by the model training module of fig. 1, and the corresponding click rate estimation method is mainly implemented by the estimation module of fig. 1.

The click rate estimation model provided by the embodiment of the application belongs to a variable parameter decomposer model (also called an integrated decomposer model Ensemble Factorization Machine and EFM). In the click rate estimation model, N different candidate dimensions are set, wherein N is a natural number greater than or equal to 2, and each candidate dimension is used for identifying the dimension of the feature vector in the click rate estimation model. That is, in the click rate estimation model, the dimensions of all feature vectors are one of a plurality of candidate dimensions.

Aiming at the target feature of the feature vector to be trained, one of the N different candidate dimensions can be determined as the maximum dimension corresponding to the target feature for the target feature, and the feature vector with the maximum dimension of the target feature can reasonably embody the information of the target feature in the training sample, so that the problem of over-fitting or under-fitting can not occur, and the higher estimation precision can be achieved.

In order to realize the inner product of feature vectors among features with different maximum dimensions, the click rate estimation model trains i+1 feature vectors with different dimensions for the target features, and i is a natural number smaller than N and larger than or equal to 1, so that the inner product calculation of the feature vectors between the target features and other features with the maximum dimensions smaller than the target features is not influenced, and the method has higher practicability.

For example, fig. 2 shows the difference between the click rate estimation model (EFM model) and the feature vector in the conventional click rate estimation model according to the embodiment of the present application, where part a in fig. 2 shows the case of the feature vector in the EFM model, part b shows the case of the feature vector in the conventional high-dimensional FM model, and part c shows the case of the feature vector in the conventional low-dimensional FM model.

In the example shown in fig. 2, for four features: user "Zhang San", user "Lifour", movie "A" and movie "B". It is assumed that in a training database including scores of two users "Zhang Sano" and "Lifour" for movies "A" and "B", there is a large amount of score data for Zhang Sano, and relatively few score data for Lifour. Meanwhile, movie a is watched in a smaller number than movie B.

In the click rate estimation model (EFM model) provided in the embodiment of the present application, two candidate dimensions are set, where d1=4 and d2=8 respectively. Since the training samples corresponding to the feature 'Zhang Sanling' are more, the training samples corresponding to the feature 'Liqu' are fewer, and the training samples corresponding to the feature 'B' are more, the training samples corresponding to the feature 'A' can be used as the maximum dimension of the feature 'Zhang Sanling', the feature 'Liqu' is used as the maximum dimension of the feature 'Liqu', the feature 'B' is used as the maximum dimension of the feature 'Liqu', and the feature 'A' is used as the maximum dimension of the feature 'A'. Since the maximum dimension D2 corresponding to the feature "Zhang San" and the feature "B" is larger than D1 in the candidate dimensions, D1 can be taken as the projection dimension corresponding to the feature "Zhang San" and the feature "B".

Thus, the feature "Zhang Sano" and feature "B" may be trained separately, each resulting in two feature trains, with dimensions D1 and D2, respectively, as shown in part a of FIG. 2. Training "Lifour" and feature "A" separately, each results in a feature training, dimension D1, for example as shown in section a of FIG. 2.

The feature 'Zhang San' and the feature 'B' have feature vectors which can reasonably embody information of target features in training samples, the dimension is D2, and the feature 'Li Si' and the feature 'A' have feature vectors which can reasonably embody information of target features in training samples, and the dimension is D1. None of these four features presents problems with over-fitting or under-fitting.

Meanwhile, in order to calculate the inner product between the feature vectors of the feature "Zhang Sano" and the feature "B" and the feature "Liqu" and the feature "A", the feature "Zhang Sano" and the feature "B" also have feature vectors with the dimension of D1, for example, if the click rate of the three-click movie A needs to be estimated, the inner product between the feature vectors with the dimension of D1 and the feature vectors with the dimension of D1 of the feature "Zhang Sano" can be calculated. If the click rate of three pairs of film B needs to be estimated, the dimension of the feature vector is D2, and the interaction term is the inner product of the feature vector and the interaction term. If the click rate of the four pairs of movies A needs to be estimated, the dimension of the feature vectors of the two pairs of movies A is D1, and the interaction term is the inner product of the two pairs of feature vectors.

In the above example, the conventional click rate estimation model may suffer from over-fitting problems, such as a uniformly arranged high-dimensional FM model with a high dimension (as shown in part b of fig. 2). The click rate estimation model assigns 8-dimensional feature vectors to all features, so that interaction between Lifour and film A can be over-fitted due to insufficient number of non-zero samples, and estimation accuracy is affected.

The conventional click rate estimation model also suffers from the problem of under fitting, such as a uniformly arranged low-dimensional FM model with a low dimension (as shown in part c of fig. 2). All features of the click rate estimation model are distributed with 4-dimensional feature vectors, so that the number of non-zero samples of Zhang three and film B is large, the number of dimension of the feature vectors (reflecting the expressive force of the model) is small, and the inner products between the feature vectors are under-fitted, so that the estimation accuracy is also influenced.

In real data, a large number of features have fewer non-zero samples, and a small number of features have more non-zero samples. For example, if some feature is the id of a movie, then a small number of popular movies will be watched much more times (i.e., number of samples) similar to movie B, while a large number of movies will be watched much less times similar to movie a. Therefore, compared with the traditional click rate estimation model, the EFM model provided by the embodiment of the application distributes low-dimensional vectors for a large number of features, and only distributes high-dimensional vectors and low-dimensional projections thereof for a small number of features. In this case, the number of parameters of the EFM model will be slightly greater than the case of assigning low-dimensional vectors to all features in the low-dimensional FM model shown in part c of fig. 2, while significantly less than the case of assigning high-dimensional vectors to all features in the high-dimensional FM model shown in part b of fig. 2. Therefore, the EFM model can maintain reasonable model scale while improving the prediction effect.

Next, first, the construction of the click rate estimation model in the embodiment of the present application will be described.

Fig. 3 is a flowchart of a method for constructing a click rate estimation model according to an embodiment of the present application. In the constructed click rate estimation model, N different candidate dimensions are set, wherein the candidate dimensions are used for identifying the dimensions of the feature vector.

The candidate dimension belongs to super parameters, and the size of the candidate dimension can be determined according to the scene and the problem actually corresponding to the click rate estimation model. The candidate dimensions may be pre-specified, assuming a total of m, noted as D ₁ ,D ₂ ，…,D _m Dimension of D ₁ <D ₂ <…<D _m . In the subsequent calculation, the dimension of the feature vector corresponding to each feature can only be selected from the candidate dimensions, namely, in the click rate estimation model, the dimension of all the feature vectors is one of a plurality of candidate dimensions.

The method comprises the following steps:

s301: and determining one candidate dimension from the N different candidate dimensions as a maximum dimension corresponding to the target feature.

The target feature may be a feature extracted by the feature extraction module shown in fig. 1, and the target feature may be a feature identifying relevant information of a user or a feature identifying relevant information of a clicked object such as content data. The determined maximum dimension is used for identifying the dimension of a feature vector of the target feature, and the feature vector is a feature vector which can reasonably embody information of the target feature in the training sample.

In one possible implementation, this step may be implemented by:

and determining one candidate dimension from the N different candidate dimensions as the maximum dimension corresponding to the target feature according to the number of non-zero training samples corresponding to the target feature.

For example, the candidate dimensions include two of 5 dimensions and 10 dimensions, and when there are more non-zero training samples of the target feature, 10 dimensions may be selected as the largest dimension of the target feature, and when there are fewer non-zero training samples of the target feature, 5 dimensions may be selected as the largest dimension of the target feature.

The embodiment of the application provides a specific alternative mode for determining the maximum dimension, which is shown in a formula (1):

（1）

wherein n is _i D for the number of non-zero training samples including the feature _k For any one of the candidate dimensions.

Equation (1) represents: the dimension of a feature should be as close as possible to the number of non-zero training samples to which the feature corresponds.

S302: and determining i candidate dimensions smaller than the maximum dimension in the N different candidate dimensions as projection dimensions corresponding to the target features.

Since the maximum dimension of the target feature determined in S301 is not the smallest of the candidate dimensions, i candidate dimensions smaller than the maximum dimension among the N different candidate dimensions may be determined as projection dimensions corresponding to the target feature, i being an integer of 1 or more.

For example, the candidate dimensions include four of 5 dimensions, 10 dimensions, 50 dimensions, and 100 dimensions, and when the maximum dimension corresponding to the target feature is 50 dimensions, 5 dimensions and 10 dimensions of the candidate dimensions are determined as projection dimensions of the target feature.

S303: and training i+1 feature vectors with different dimensions for the target feature according to the training sample corresponding to the target feature, wherein the dimension of any one feature vector in the i+1 feature vectors with different dimensions is one of the maximum dimension and the projection dimension corresponding to the target feature.

If the maximum dimension of the target feature is D in the previous example ₁ ,D ₂ ，…,D _m Any one D of _k It is assigned a dimension D _k Feature vector v of (2) ^(k) Called the feature vector of the feature. At the same time for all D ₁ -D _k-1 Any one D of _n Assigning dimension D to the target feature _n Feature vector v of (2) ⁽ⁿ⁾ Called feature vector v ^(k) Is of dimension D _n Is a projection of (a). So called asProjection is because these feature vectors are defined by v ^(k) And (3) determining. In the embodiments of the present application, either v ^(k) Or v ⁽ⁿ⁾ Are feature vectors corresponding to the target features. The specific values of all feature vectors are parameters, and can be calculated by a training algorithm in the training process of S303.

For example, when the maximum dimension corresponding to the target feature is 50 dimensions, 5 dimensions and 10 dimensions of the candidate dimensions are determined as projection dimensions of the target feature, which is equivalent to i=2, and training samples corresponding to the target feature can be used for training the target feature to obtain 1+2=3 feature vectors, including a 50-dimensional feature vector, a 10-dimensional feature vector and a 5-dimensional feature vector.

It can be seen that the click rate estimation model sets N different candidate dimensions, each of which is used to identify the dimensions of the feature vector in the click rate estimation model. When the feature vector of the target feature is determined, one candidate dimension is determined from the N different candidate dimensions as a maximum dimension corresponding to the target feature, and i candidate dimensions smaller than the maximum dimension in the N different candidate dimensions are determined as projection dimensions corresponding to the feature; thus, the target feature has a corresponding maximum dimension and i corresponding projection dimensions, and when the feature vector is found according to the training sample corresponding to the target feature as the target feature, i+1 feature vectors can be trained respectively, and the dimension corresponding to the different feature vectors is one of the maximum dimension and the projection dimension. Therefore, the feature vector corresponding to the maximum dimension can reasonably reflect the information of the target feature in the training sample, and the problem of over-fitting or under-fitting cannot occur. In order to realize the inner product of feature vectors among features with different maximum dimensions, the click rate estimation model trains i+1 feature vectors with different dimensions for the target features, so that the inner product calculation of the feature vectors is carried out on the target features and other features with the maximum dimensions smaller than the target features, and higher estimation accuracy is achieved.

Because the feature vector is different from one feature vector corresponding to one feature in the traditional model, in the embodiment of the application, the target feature has at least two feature vectors with different dimensions, so in order to train out the accurate feature vector which can reasonably embody the information of the target feature in the training sample, the embodiment of the application adopts a separate training mode, namely N feature vectors with different dimensions of the target feature are obtained by separate training.

Next, how any feature vector of the target feature is trained is described, and for S303, two calculation methods are provided in the embodiment of the present application.

The first calculation mode is as follows:

the first feature vector is any one of the i+1 feature vectors of different dimensions, that is, the first feature vector may be the feature vector having the largest dimension or the feature vector having the projected dimension.

For the first feature vector, in one possible implementation, S303 may include:

s3031: and acquiring a target training sample and a click rate corresponding to the target training sample from the training samples corresponding to the target characteristics.

The training samples corresponding to the target features are non-zero training samples comprising the target features, and the number of the target training samples can be one or a plurality of the target training samples. Being a training sample, there is a known click rate for that feature.

S3032: and calculating a first click rate estimated value of the target training sample through the click rate estimation model according to the first feature vector corresponding to the target feature.

On the first training, the first feature vector may be randomly initialized with the unknown parameters. In the following ith training, the first feature vector uses the parameters corrected by the ith-1 th training.

S3033: and calculating the gradient of the first eigenvector according to a loss function determined by the click rate and the first click rate predicted value.

In one possible implementation, the loss function may be calculated according to equation (2):

(2)

wherein, the liquid crystal display device comprises a liquid crystal display device,the dimension of the first eigenvector is D as a loss function _k ，/>And (3) a first click rate predicted value, and y is the click rate.

In one possible implementation, the gradient of the loss function to the first eigenvector may be calculated according to equation (3):

(3)

wherein the dimension of the first feature vector is D _k ， Dimension D for the ith feature in the target training sample _k The f-th component of the feature vector (including the first feature vector), the feature vector of the included feature,/-, the first feature vector, the second feature vector, the third feature vector, the fourth feature vector, the fifth feature vector, the sixth feature vector, the seventh feature vector, the eighth feature vector, the seventh feature vector, the>The dimension of the jth feature in the target training sample is greater than or equal to D _k F component of the eigenvector of +.>For the value of the ith feature, +.>For the value of the j-th feature +.>Is the learning rate.

The meaning of the formula is:should be directed toward more accurate prediction of yAnd updating the direction.

S3034: and correcting parameters of the first feature vector according to the gradient of the first feature vector.

After the parameters are corrected, S3031 may be re-executed to continue the next training until the algorithm converges or reaches a preset maximum training number.

The second calculation mode:

aiming at the feature vector which is not the largest dimension in the feature vectors corresponding to the target features, the embodiment of the application provides a mode that the high-dimensional feature vector is used as the training basis of the low-dimensional feature vector.

The second feature vector is any one of the i+1 different-dimension feature vectors, and the dimension of the second feature vector is not the maximum dimension. If the dimension of the first feature vector is the maximum dimension, in one possible implementation, S303 may include, for a second feature vector:

s3035: and acquiring a target training sample from the training samples corresponding to the target features.

S3036: and calculating a second click rate estimated value of the target training sample through the click rate estimated model according to a second feature vector corresponding to the target feature.

S3037: and calculating the gradient of the second feature vector according to a loss function determined by the first click rate pre-estimated value and the second click rate pre-estimated value.

(4)

Wherein, in particular, in the present formula,for dimension D _k+1 The click rate predictive value of the feature vector of (2) may be the first click rate predictive value,/-, etc.>A second click rate estimate is provided.

The meaning of the formula is:should be more accurately predicted +.>Is updated in the direction of the (c). The method can embody ∈ ->The reason meaning of the low-dimensional projection called higher-dimensional vector is because +.>Is calculated using a higher dimensional feature vector, so under this calculation method, it can be considered +.>Determined by the higher dimensional feature vectors.

S3038: and correcting parameters of the second eigenvector according to the gradient of the second eigenvector.

After the parameters are corrected, S3035 may be re-executed to continue the next training until the algorithm converges or the preset maximum training times are reached.

It is noted that the two calculation modes are not essentially different. This is because of the high-dimensional vector predictionUsually predicted +.>Closer to true y, thus let +.>Approximation y and let- >Approximation->Has similar effects.

The training mode provided by the embodiment of the application is different from the prior art in that: the embodiment of the application is used for calculatingIn the gradient of (2) is chosen by letting + ->Closer to y or->，/>It is understood that "no more than dimension D _k A click rate predictive value obtained by the feature vector of (a). This effectively reflects that feature vectors of different dimensions have different targets during training.

In the prior art, the same training objectives are used for all parameters. The training method of the embodiment of the application has the advantage that the click rate estimation model of the embodiment of the application is assumed to have 2-dimensional characteristics and 4-dimensional characteristics. By using the training method of the embodiment of the application, 4-dimensional features can make predictions in 4-dimensional space as close as possible to the true value, and 2-dimensional features can be close as possible to the true value in 2-dimensional space (the predicted values of the 4-dimensional features can also be approximated). Thus, the feature vector for each dimension number expresses the most important information for that dimension number. This is consistent with what we want to "2-dimensional features are projections of 4-dimensional features in 2-dimensional space" because the meaning of projections is to preserve the portion that matches the 2-dimensional space, removing the unmatched portion.

In contrast, if a unified training objective as in the prior art is directly adopted, the 2-dimensional features and the 4-dimensional features "cooperate" to give a predicted value, and the information they give may be feature information of the 6-dimensional space, instead of the features of the 4-dimensional space that we want, in which case over-fitting is easy to occur (i.e., training errors are small, but test errors become large).

Next, click rate estimation in the embodiment of the present application will be described.

Fig. 4 is a flowchart of a method for estimating click rate according to an embodiment of the present application, where the method includes:

s401: a sample to be estimated is obtained that includes a plurality of features.

The sample to be estimated belongs to a non-zero sample with the included characteristics, the plurality of characteristics included in the sample to be estimated at least comprise one characteristic related to a user and a characteristic related to a clicked object, and the probability of the user with the related characteristic clicking the clicked object with the related characteristic can be estimated through calculation of the click rate estimation model on the sample to be estimated.

The plurality of features at least comprises a first feature and a second feature, wherein the first feature can be a feature related to a user or a feature related to a clicked object, and the second feature can be a feature related to a user or a feature related to a clicked object.

S402: and calculating the click rate estimated value of the sample to be estimated through a click rate estimated model according to the feature vectors respectively corresponding to the features.

In this step, the click rate estimation model used is a click rate estimation model provided in the embodiment of the present application, for example, as illustrated in the embodiment corresponding to fig. 3, where N different candidate dimensions are set for the click rate estimation model, where the candidate dimensions are used to identify the dimensions of the feature vector. If the maximum dimension corresponding to a feature in the click rate estimation model is greater than i of the N different candidate dimensions, the feature has i+1 feature vectors of different dimensions; the maximum dimension corresponding to any one of the plurality of features is used to identify the maximum dimension of the feature vector corresponding to the feature.

That is, in the click rate estimation model, a feature may have N feature vectors with different dimensions, so that when the click rate estimation model is used to estimate the click rate estimated value of the sample to be estimated, a calculation mode different from that of the conventional model is adopted.

Since the calculation of the click rate estimated value includes the calculation of the inner product between feature vectors of different features, and the click rate estimated model provided by the embodiment of the present application includes feature vectors of different dimensions, how to calculate the inner product between feature vectors belongs to one of the core improvement points of the present application, and in order to clearly illustrate the specific manner of calculating the click rate estimated value provided by the embodiment of the present application, next, how to calculate the inner product between the feature vector of the first feature and the feature vector of the second feature is illustrated by taking the first feature and the second feature of the plurality of features as examples. For the first feature and the second feature, S402 may specifically include:

S4021: and if the first maximum dimension corresponding to the first feature in the click rate estimation model is different from the second maximum dimension corresponding to the second feature, taking the smallest of the first maximum dimension and the second maximum dimension as the dimension to be calculated.

That is, when the first maximum dimension corresponding to the first feature and the second maximum dimension corresponding to the second feature are different, in order to calculate the inner product between the feature vector of the first feature and the feature vector of the second feature, it is necessary to select the feature vector having the same dimension from the feature vector of the first feature and the feature vector of the second feature as the basis for calculating the inner product.

For example, the candidate dimensions are three, 5, 10 and 50, respectively, with a first maximum dimension of 50 and a second maximum dimension of 10. The first feature has three feature vectors, one 50-dimensional feature vector, one 10-dimensional feature vector and one 5-dimensional feature vector, respectively, and the second feature has two feature vectors, one 10-dimensional feature vector and one 5-dimensional feature vector, respectively. The feature vectors of the first feature and the second feature have the same dimension and are maximally 10 dimensions, i.e. the smallest dimension of the first maximum dimension and the second maximum dimension. Thus, 10 dimensions can be taken as the dimensions of the feature vector that calculates the inner product between the feature vectors of the first feature and the second feature, i.e. the dimensions to be calculated.

S4022: an inner product of the feature vector of the first feature having the dimension to be calculated and the feature vector of the second feature having the dimension to be calculated is calculated.

After the dimension to be calculated is determined, the calculation of the inner product of the feature vector of the first feature and the feature vector of the second feature can be implemented, wherein the feature vector adopted in the calculation is the feature vector of the first feature with the dimension to be calculated and the feature vector of the second feature with the dimension to be calculated.

The calculation of S4022 can be performed by formulas (5) and (6).

(5)

Wherein b is a constant term,is the i-th parameter value,/-, for example>Is the interaction term of the ith feature and the jth feature, and the calculation method is as follows: />

(6)

Wherein, the liquid crystal display device comprises a liquid crystal display device,for the maximum dimension (e.g. first maximum dimension) of the ith feature>Is the largest dimension (e.g., the second largest dimension) of the jth feature.

The expressions of formulas (5) and (6) mean that, assuming, k ₁ =8 dimensions, k ₂ =4 dimensions, then the feature vector of the 1 st feature is an 8-dimensional vector, and the feature vector of the 2 nd feature is a 4-dimensional vector. The interaction term of these two feature vectors is equal to an 8-dimensional vector"4-dimensional projection">And the 4-dimensional vector- >Is a product of the inner product of (a).

Therefore, the embodiment of the application provides a method for distributing different feature vector lengths for different features, and the method distributes a low-dimensional projection vector for the high-dimensional feature vector of each feature, so that the problem of interactive item calculation among the feature vectors with different dimension lengths is solved.

It should be noted that in the actual calculation, we need to calculate for every two combinations of input features according to formulas (5) and (6), assuming a total of n input features. And n features are combined in pairs to total n (n+1)/2. In practical applications, the number of features is typically in the order of millions to billions, and even if sparsity is considered, the number of actual calculations required is very large. In this case, the calculated amount of n (n+1)/2 times of calculation is relatively not small, so that the embodiment of the application provides an efficient calculation mode, and the requirement of the system on calculation performance is reduced.

If the dimension of the feature vector used for calculating the click rate pre-estimated value of the sample to be estimated is determined to be D with sequentially increasing dimensions according to the maximum dimension corresponding to the features ₁ To D _m K is any one of 1 to m, in one possible implementation, S402 may be implemented by:

According to D _k-1 The corresponding predicted value is added with the dimension D in the feature vector corresponding to the features _k The inner product of the feature vectors of the target feature is subtracted by the dimension D of the feature vector corresponding to the target feature _k-1 Inner product calculation D of eigenvectors of (C) _k A corresponding predicted value; the target feature is a plurality of features having a dimension D _k Is a feature of the feature vector of (a).

Will D _m The corresponding estimated value is used as the click rate estimation of the sample to be estimatedValues.

The specific calculation can be performed by formulas (7) and (8):

(7)

(8)

wherein, the liquid crystal display device comprises a liquid crystal display device,dimension D for the ith feature ₁ The dimension of the feature vector of (1) and the j-th feature is D ₁ Inner product between eigenvectors, +.>Dimension D for the ith feature _k The dimension of the feature vector of (1) and the j-th feature is D _k Is a product of the inner products between the eigenvectors of (a) and (b).

Equations (7) and (8) are equivalent variations from equations (5) and (6).

Calculated therefromThe click rate predicted value of the sample to be estimated can be used as +.>。

After the equations (7) and (8) are obtained by equivalently deforming the equations (5) and (6), the equations (7) and (8) contain three interactive terms, namely，/>，/>The main difference between these three interaction terms and formulas (5) and (6) is: formula [ (formula ] 5) The dimensions of the feature vectors are different for different i and j in (6), but the dimensions of the feature vectors corresponding to different i and j are the same for each interaction term in equations (7) and (8).

This condition allows for a further equivalent transformation of the three interaction terms described above in formulas (7) and (8) to yield formula (9):

(9)

according to this equation, the predicted value of EFM can be efficiently calculated. Taking the first expression of the formula (9) as an example, the right side of the formula has two terms respectivelyAnd->For each term, we only need to do n times of multiplication and addition operation to obtain the result. When n is large, n times of operation is much smaller than n (n+1)/2 times of operation of the foregoing formulas (5) and (6), and of course, a plurality of formulas (7) and (8) need to be calculated by the method of formula (9), so the final operation times are integer multiples of n. Even so, since this multiple is typically low (about 10 to 20), the algorithm given by equations (7), (8) and (9) is significantly improved over the direct calculation with equations (5) and (6).

Next, by illustrating the principle of formulas (7) and (8), as shown in fig. 5. The 4 features included in the sample to be estimated in fig. 5 are denoted as features 1, 2, 3, 4, and the candidate dimensions set in the efm model include two, a high-dimensional dimension and a low-dimensional dimension, respectively, wherein the maximum dimensions of feature 1 and feature 4 are larger than the maximum dimensions of feature 2 and feature 3. In this example, if it is desired to calculate interactions between the 4 features, summation is performed to estimate the click rate estimated value of the sample to be estimated.

In the calculation process, only low-dimensional vectors are considered first, and every two vectors are calculatedInteractive and add to obtain. At->In the above, the interaction between the feature 1 and the feature 4 adopts the inner product between the two low-dimensional projections, so long as the inner product between the original two high-dimensional vectors is replaced by the inner product. In this alternative, the inner product of the low-dimensional projections of feature 1 and feature 4 is subtracted first, plus the inner product of the high-dimensional vector, thus yielding +.>I.e. the final click rate estimate.

Based on the method for constructing the click rate estimation model provided in the foregoing embodiment, the present embodiment provides a device 600 for constructing the click rate estimation model, and referring to fig. 6, the device 600 includes a first determining unit 601, a second determining unit 602, and a training unit 603. For the description of each unit in the embodiment corresponding to fig. 6, reference may be made to the related description in the embodiment corresponding to fig. 3, which is not repeated here.

The first determining unit 601 is configured to determine one candidate dimension from the N different candidate dimensions as a maximum dimension corresponding to the target feature;

the second determining unit 602 is configured to determine i candidate dimensions smaller than the maximum dimension among the N different candidate dimensions as projection dimensions corresponding to the target feature;

The training unit 603 is configured to train i+1 feature vectors with different dimensions for the target feature according to a training sample corresponding to the target feature, where the dimension of any one feature vector is one of the maximum dimension and the projection dimension corresponding to the target feature in the i+1 feature vectors with different dimensions.

In a possible implementation, the first feature vector is any one of the i+1 feature vectors of different dimensions, for which the training unit 603 is further configured to:

acquiring a target training sample and a click rate corresponding to the target training sample from the training samples corresponding to the target characteristics;

calculating a first click rate estimated value of the target training sample through the click rate estimated model according to a first feature vector corresponding to the target feature;

calculating a gradient to the first feature vector according to a loss function determined by the click rate and the first click rate pre-estimated value;

and correcting parameters of the first feature vector according to the gradient of the first feature vector.

In one possible implementation, the second feature vector is any one of the i+1 different-dimension feature vectors, and the dimension of the second feature vector is not the maximum dimension; if the dimension of the first feature vector is the maximum dimension, the training unit 603 is further configured to, for the second feature vector:

Acquiring a target training sample from training samples corresponding to the target features;

calculating a second click rate estimated value of the target training sample through the click rate estimated model according to a second feature vector corresponding to the target feature;

calculating a gradient to the second feature vector according to a loss function determined by the first click rate pre-estimation value and the second click rate pre-estimation value;

and correcting parameters of the second eigenvector according to the gradient of the second eigenvector.

In a possible implementation manner, the first determining unit 601 is further configured to determine, according to the number of non-zero training samples corresponding to the target feature, one candidate dimension from the N different candidate dimensions as a maximum dimension corresponding to the target feature.

Based on the click rate estimation method provided in the foregoing embodiments, the present embodiment provides a click rate estimation device 700, referring to fig. 7, where the device 700 includes an obtaining unit 701 and a calculating unit 702. For the description of each unit in the embodiment corresponding to fig. 7, reference may be made to the related description in the embodiment corresponding to fig. 4, which is not repeated here.

The acquiring unit 701 is configured to acquire a sample to be estimated including a plurality of features, where the plurality of features includes at least a first feature and a second feature;

the calculating unit 702 is configured to calculate, according to the feature vectors corresponding to the features, a click rate estimated value of the sample to be estimated through a click rate estimated model; the click rate estimation model is provided with N different candidate dimensions, and the candidate dimensions are used for identifying the dimensions of the feature vector; if the maximum dimension corresponding to a feature in the click rate estimation model is greater than i of the N different candidate dimensions, the feature has i+1 feature vectors of different dimensions; the maximum dimension corresponding to any one of the plurality of features is used for identifying the maximum dimension of the feature vector corresponding to the feature;

For the first feature and the second feature, the computing unit 702 is further configured to:

In one possible implementation manner, if the dimension of the feature vector used for calculating the click rate estimated value of the sample to be estimated is determined to be D with sequentially increasing dimensions according to the maximum dimension corresponding to each of the plurality of features ₁ To D _m K is any one of 1 to m, and the computing unit 702 is further configured to:

according to D _k-1 The corresponding predicted value is added with the dimension D in the feature vector corresponding to the features _k The inner product of the feature vectors of the target feature is subtracted by the dimension D of the feature vector corresponding to the target feature _k-1 Inner product calculation D of eigenvectors of (C) _k A corresponding predicted value; the target feature is a plurality of features having a dimension D _k Features of the feature vector of (a);

will D _m And the corresponding predicted value is used as the click rate predicted value of the sample to be estimated.

The embodiment of the application also provides construction equipment for the click rate estimation model, and the construction equipment for the click rate estimation model is described below with reference to the accompanying drawings. Referring to fig. 8, an embodiment of the present application provides a device 800 for constructing a click-through rate estimation model, where the device 800 may be a server, may be relatively different due to configuration or performance, and may include one or more central processing units (Central Processing Units, abbreviated as CPUs) 822 (e.g., one or more processors) and a memory 832, and one or more storage media 830 (e.g., one or more mass storage devices) storing application programs 842 or data 844. Wherein the memory 832 and the storage medium 830 may be transitory or persistent. The program stored in the storage medium 830 may include one or more modules (not shown), each of which may include a series of instruction operations on a server. Still further, the central processor 822 may be configured to communicate with the storage medium 830 to execute a series of instruction operations in the storage medium 830 on the dynamic storage 800 for deep learning networks.

The dynamic storage device 800 for deep learning networks may also include one or more power supplies 826, one or more wired or wireless network interfaces 850, one or more input/output interfaces 858, and/or one or more operating systems 841, such as Windows ServerTM, mac OS XTM, unixTM, linuxTM, freeBSDTM, and the like.

The steps performed by the server in the above embodiments may be based on the server structure shown in fig. 8.

CPU 822, among other things, is configured to perform the following steps:

determining i candidate dimensions smaller than the maximum dimension in the N different candidate dimensions as projection dimensions corresponding to the target features;

The embodiment of the present application further provides a device for click rate estimation, and a structure of the device for click rate estimation may also be shown in fig. 8, where in the device, CPU 822 is configured to perform the following steps:

calculating a click rate estimated value of the sample to be estimated through a click rate estimated model according to the feature vectors respectively corresponding to the plurality of features; the click rate estimation model is provided with N different candidate dimensions, and the candidate dimensions are used for identifying the dimensions of the feature vector; if the maximum dimension corresponding to a feature in the click rate estimation model is greater than i of the N different candidate dimensions, the feature has i+1 feature vectors of different dimensions; the maximum dimension corresponding to any one of the plurality of features is used for identifying the maximum dimension of the feature vector corresponding to the feature;

Referring to fig. 9, an embodiment of the present application provides a device 900 for constructing a click rate estimation model, where the device 900 may also be a device for click rate estimation provided by the embodiment of the present application.

The device 900 may also be a terminal device, where the terminal device may be any terminal device including a mobile phone, a tablet computer, a personal digital assistant (Personal Digital Assistant, PDA for short), a Point of Sales (POS for short), a vehicle-mounted computer, and the like, and the terminal device is taken as an example of the mobile phone:

fig. 9 is a block diagram showing a part of the structure of a mobile phone related to a terminal device provided by an embodiment of the present application. Referring to fig. 9, the mobile phone includes: radio Frequency (RF) circuitry 910, memory 920, input unit 930, display unit 940, sensor 950, audio circuitry 960, wireless fidelity (wireless fidelity, wiFi) module 970, processor 980, and power source 990. It will be appreciated by those skilled in the art that the handset construction shown in fig. 9 is not limiting of the handset and may include more or fewer components than shown, or may combine certain components, or a different arrangement of components.

The following describes the components of the mobile phone in detail with reference to fig. 9:

the RF circuit 910 may be used for receiving and transmitting signals during a message or a call, and particularly, after receiving downlink information of a base station, the signal is processed by the processor 980; in addition, the data of the design uplink is sent to the base station. Generally, the RF circuitry 910 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low noise amplifier (Low Noise Amplifier, LNA for short), a duplexer, and the like. In addition, the RF circuitry 910 may also communicate with networks and other devices via wireless communications. The wireless communication may use any communication standard or protocol, including but not limited to global system for mobile communications (Global System of Mobile communication, GSM for short), general packet radio service (General Packet Radio Service, GPRS for short), code division multiple access (Code Division Multiple Access, CDMA for short), wideband code division multiple access (Wideband Code Division Multiple Access, WCDMA for short), long term evolution (Long Term Evolution, LTE for short), email, short message service (Short Messaging Service, SMS for short), and the like.

The memory 920 may be used to store software programs and modules, and the processor 980 performs various functional applications and data processing by operating the software programs and modules stored in the memory 920. The memory 920 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program (such as a sound playing function, an image playing function, etc.) required for at least one function, and the like; the storage data area may store data (such as audio data, phonebook, etc.) created according to the use of the handset, etc. In addition, memory 920 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage device.

The input unit 930 may be used to receive input numeric or character information and to generate key signal inputs related to user settings and function control of the handset. In particular, the input unit 930 may include a touch panel 931 and other input devices 932. The touch panel 931, also referred to as a touch screen, may collect touch operations thereon or thereabout by a user (such as operations of the user on the touch panel 931 or thereabout using any suitable object or accessory such as a finger, a stylus, or the like) and drive the corresponding connection device according to a predetermined program. Alternatively, the touch panel 931 may include two parts of a touch detection device and a touch controller. The touch detection device detects the touch azimuth of a user, detects a signal brought by touch operation and transmits the signal to the touch controller; the touch controller receives touch information from the touch detection device and converts it into touch point coordinates, which are then sent to the processor 980, and can receive commands from the processor 980 and execute them. In addition, the touch panel 931 may be implemented in various types such as resistive, capacitive, infrared, and surface acoustic wave. The input unit 930 may include other input devices 932 in addition to the touch panel 931. In particular, other input devices 932 may include, but are not limited to, one or more of a physical keyboard, function keys (e.g., volume control keys, switch keys, etc.), a trackball, mouse, joystick, etc.

The display unit 940 may be used to display information input by a user or information provided to the user and various menus of the mobile phone. The display unit 940 may include a display panel 941, and optionally, the display panel 941 may be configured in the form of a liquid crystal display (Liquid Crystal Display, LCD), an Organic Light-Emitting Diode (OLED), or the like. Further, the touch panel 931 may overlay the display panel 941, and when the touch panel 931 detects a touch operation thereon or thereabout, the touch operation is transferred to the processor 980 to determine a type of touch event, and then the processor 980 provides a corresponding visual output on the display panel 941 according to the type of touch event. Although in fig. 9, the touch panel 931 and the display panel 941 are implemented as two separate components for the input and output functions of the mobile phone, in some embodiments, the touch panel 931 may be integrated with the display panel 941 to implement the input and output functions of the mobile phone.

The handset may also include at least one sensor 950, such as a light sensor, motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor, wherein the ambient light sensor may adjust the brightness of the display panel 941 according to the brightness of ambient light, and the proximity sensor may turn off the display panel 941 and/or the backlight when the mobile phone moves to the ear. The accelerometer sensor can be used for detecting the acceleration in all directions (generally three axes), detecting the gravity and the direction when the accelerometer sensor is static, and can be used for identifying the gesture of a mobile phone (such as transverse and vertical screen switching, related games, magnetometer gesture calibration), vibration identification related functions (such as pedometer and knocking), and other sensors such as gyroscopes, barometers, hygrometers, thermometers, infrared sensors which are also configured by the mobile phone are not repeated herein.

Audio circuitry 960, speaker 961, microphone 962 may provide an audio interface between a user and a cell phone. Audio circuit 960 may transmit the received electrical signal converted from audio data to speaker 961, where it is converted to a sound signal by speaker 961 for output; on the other hand, microphone 962 converts the collected sound signals into electrical signals, which are received by audio circuit 960 and converted into audio data, which are processed by audio data output processor 980 for transmission to, for example, another cell phone via RF circuit 910 or for output to memory 920 for further processing.

WiFi belongs to a short-distance wireless transmission technology, and a mobile phone can help a user to send and receive emails, browse webpages, access streaming media and the like through a WiFi module 970, so that wireless broadband Internet access is provided for the user. Although fig. 9 shows a WiFi module 970, it is understood that it does not belong to the necessary constitution of the handset, and can be omitted entirely as needed within the scope of not changing the essence of the invention.

Processor 980 is a control center for the handset, connecting the various parts of the entire handset using various interfaces and lines, performing various functions and processing data for the handset by running or executing software programs and/or modules stored in memory 920, and invoking data stored in memory 920. Optionally, processor 980 may include one or more processing units; preferably, the processor 980 may integrate an application processor with a modem processor, wherein the application processor primarily handles operating systems, user interfaces, applications programs, etc., and the modem processor primarily handles wireless communications. It will be appreciated that the modem processor described above may not be integrated into the processor 980.

The handset further includes a power supply 990 (e.g., a battery) for powering the various components, which may be logically connected to the processor 980 by a power management system, such as for performing charge, discharge, and power management functions via the power management system.

Although not shown, the mobile phone may further include a camera, a bluetooth module, etc., which will not be described herein.

The embodiment of the application also provides a computer readable storage medium for storing program code, where the program code is configured to execute any one of the methods for constructing a click rate estimation model according to the foregoing embodiments, and may also be configured to execute any one of the methods for estimating a click rate according to the foregoing embodiments.

Those of ordinary skill in the art will appreciate that: all or part of the steps for implementing the above method embodiments may be implemented by hardware related to program instructions, where the above program may be stored in a computer readable storage medium, and when the program is executed, the program performs steps including the above method embodiments; and the aforementioned storage medium may be at least one of the following media: read-only memory (ROM), RAM, magnetic disk or optical disk, etc., which can store program codes.

It should be noted that, in the present specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment is mainly described in a different point from other embodiments. In particular, for the apparatus and system embodiments, since they are substantially similar to the method embodiments, the description is relatively simple, with reference to the description of the method embodiments in part. The apparatus and system embodiments described above are merely illustrative, in which elements illustrated as separate elements may or may not be physically separate, and elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present application without undue burden.

The foregoing is only one specific embodiment of the present application, but the scope of the present application is not limited thereto, and any changes or substitutions easily contemplated by those skilled in the art within the technical scope of the present application should be included in the scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. The utility model provides a click rate estimation model construction method, which is characterized in that the click rate estimation model is applied to the data processing field, wherein N different candidate dimensions are set in the click rate estimation model, N is a natural number greater than or equal to 2, and the candidate dimensions are used for identifying the dimension of a feature vector, and the method comprises the following steps:

determining one candidate dimension from the N different candidate dimensions as a maximum dimension corresponding to a target feature, wherein the target feature is a feature for identifying relevant information of a user or a feature for identifying relevant information of a clicked object, and the feature is a feature extracted from user data of the user, content data of a content platform or background data of a background database;

respectively training i+1 feature vectors with different dimensions for the target feature according to training samples corresponding to the target feature, wherein the dimension of any one feature vector is one of the maximum dimension and the projection dimension corresponding to the target feature, the feature vector is a feature vector reflecting information of the target feature in the training samples, the maximum dimension of the target feature is a candidate dimension with the minimum absolute value of the difference value of the number of non-zero training samples corresponding to the target feature in the N different candidate dimensions, the feature of the projection dimension is the projection of the feature of the maximum dimension in a projection dimension space, the projection is a part which is reserved to be matched with the projection dimension space, and the unmatched part is removed;

The first feature vector is any one of the i+1 feature vectors with different dimensions, for the first feature vector, the training samples corresponding to the target feature respectively train the i+1 feature vectors with different dimensions for the target feature, including:

acquiring a target training sample and a click rate corresponding to the target training sample from the training sample corresponding to the target feature, wherein the click rate is the ratio of the number of times a content is clicked to the number of times the content is displayed;

calculating a first click rate estimated value of the target training sample through the click rate estimated model according to a first feature vector corresponding to the target feature, wherein all unknown parameters in the first feature vector are randomly initialized during first training, and the first feature vector adopts parameters corrected by the i-1 th training during the i-th training;

and correcting the parameters of the first feature vector according to the gradient of the first feature vector, wherein after correcting the parameters, the next training is continued until the algorithm converges or the preset maximum training times are reached.

2. The method according to claim 1, wherein a second feature vector is any one of the i+1 different-dimension feature vectors, and the dimension of the second feature vector is not the maximum dimension; if the dimension of the first feature vector is the maximum dimension, for the second feature vector, training i+1 feature vectors with different dimensions for the target feature according to the training samples corresponding to the target feature, including:

3. The method according to claim 1 or 2, wherein said determining a candidate dimension from said N different candidate dimensions as the largest dimension corresponding to said target feature comprises:

4. The click rate estimation method is characterized by being applied to the field of data processing, and comprises the following steps:

calculating a click rate estimated value of the sample to be estimated through a click rate estimated model according to the feature vectors respectively corresponding to the plurality of features; the click rate estimation model is provided with N different candidate dimensions, and the candidate dimensions are used for identifying the dimensions of the feature vector; if the maximum dimension corresponding to a feature in the click rate estimation model is greater than i of the N different candidate dimensions, the feature has i+1 feature vectors of different dimensions; the maximum dimension corresponding to any one of the plurality of features is used for identifying the maximum dimension of the feature vector corresponding to the feature; n is a natural number greater than or equal to 2, i is a natural number less than N and greater than or equal to 1, and the click rate estimation model is the click rate estimation model as described in claim 1;

and calculating an inner product of the feature vector with the dimension to be calculated of the first feature and the feature vector with the dimension to be calculated of the second feature, and taking the inner product as a click rate pre-estimated value of the sample to be estimated.

5. The method according to claim 4, wherein if the dimension of the feature vector for calculating the click rate prediction value of the sample to be estimated is determined as sequentially increasing D according to the maximum dimension corresponding to each of the plurality of features ₁ To D _m K is 1 to mAccording to the feature vectors corresponding to the features, calculating the click rate estimated value of the sample to be estimated through a click rate estimated model, including:

6. The device for constructing the click rate estimation model is characterized by being applied to the field of data processing, the click rate estimation model is provided with N different candidate dimensions, the candidate dimensions are used for identifying the dimensions of a feature vector, and the device comprises a first determining unit, a second determining unit and a training unit:

the first determining unit is used for determining one candidate dimension from the N different candidate dimensions as a maximum dimension corresponding to the target feature; n is a natural number greater than or equal to 2, the target feature is a feature of relevant information of a mark user or a feature of relevant information of a clicked object, and the feature is a feature extracted from user data of the user, content data of a content platform or background data of a background database;

The training unit is configured to train, for the target feature, i+1 feature vectors with different dimensions according to training samples corresponding to the target feature, where the dimension of any one feature vector is one of the maximum dimension and the projection dimension corresponding to the target feature, where the feature vector is a feature vector that reflects information of the target feature in the training samples, the maximum dimension of the target feature is a candidate dimension that has a minimum absolute value of a difference value between the N different candidate dimensions and a number of non-zero training samples corresponding to the target feature, and where the feature of the projection dimension is a projection of the feature of the maximum dimension in a projection dimension space, where the projection is a portion that remains matching with the projection dimension space, and where a non-matching portion is removed;

wherein the first feature vector is any one of the i+1 feature vectors of different dimensions, and for the first feature vector, the training unit is further configured to:

7. The apparatus of claim 6, wherein a second feature vector is any one of the i+1 different-dimension feature vectors, and the dimension of the second feature vector is not the maximum dimension; if the dimension of the first feature vector is the maximum dimension, the training unit is further configured to, for the second feature vector:

8. The apparatus according to claim 6 or 7, wherein the first determining unit is further configured to determine, from the N different candidate dimensions, one candidate dimension as a maximum dimension corresponding to the target feature according to a number of non-zero training samples corresponding to the target feature.

9. The click rate estimating device is characterized by being applied to the field of data processing, and comprises an acquiring unit and a calculating unit:

the calculating unit is used for calculating a click rate estimated value of the sample to be estimated through a click rate estimated model according to the feature vectors respectively corresponding to the plurality of features; the click rate estimation model is provided with N different candidate dimensions, and the candidate dimensions are used for identifying the dimensions of the feature vector; if the maximum dimension corresponding to a feature in the click rate estimation model is greater than i of the N different candidate dimensions, the feature has i+1 feature vectors of different dimensions; the maximum dimension corresponding to any one of the plurality of features is used for identifying the maximum dimension of the feature vector corresponding to the feature; n is a natural number greater than or equal to 2, i is a natural number less than N and greater than or equal to 1, and the click rate estimation model is the click rate estimation model as described in claim 1;

For the first and second features, the computing unit is further to:

10. The apparatus of claim 9, wherein if the dimension of the feature vector used to calculate the click rate estimate of the sample to be estimated is determined to be successively higher-order D based on the largest dimension corresponding to each of the plurality of features ₁ To D _m K is any one of 1 to m, the computing unit being further configured to:

11. An electronic device comprising a memory and a processor, the memory storing instructions that are executed by the processor to implement the method of constructing a click rate estimation model according to any one of claims 1 to 3, or the click rate estimation method according to claim 4 or 5.

12. A computer-readable storage medium storing program code for execution by a processor to implement a method of constructing a click rate estimation model according to any one of claims 1 to 3.

13. A computer readable storage medium storing program code for execution by a processor to implement the click rate estimation method of claim 4 or 5.