CN112288471A - Advertisement click rate prediction method based on user historical behavior sequence - Google Patents

Advertisement click rate prediction method based on user historical behavior sequence Download PDF

Info

Publication number
CN112288471A
CN112288471A CN202011154969.6A CN202011154969A CN112288471A CN 112288471 A CN112288471 A CN 112288471A CN 202011154969 A CN202011154969 A CN 202011154969A CN 112288471 A CN112288471 A CN 112288471A
Authority
CN
China
Prior art keywords
advertisement
user
vector
sequence
historical behavior
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011154969.6A
Other languages
Chinese (zh)
Inventor
周仁杰
钱豪
万健
张纪林
殷昱煜
蒋从锋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Dianzi University
Original Assignee
Hangzhou Dianzi University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Dianzi University filed Critical Hangzhou Dianzi University
Priority to CN202011154969.6A priority Critical patent/CN112288471A/en
Publication of CN112288471A publication Critical patent/CN112288471A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0241—Advertisements
    • G06Q30/0242—Determining effectiveness of advertisements
    • G06Q30/0245—Surveys
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0241—Advertisements
    • G06Q30/0251—Targeted advertisements
    • G06Q30/0255—Targeted advertisements based on user history
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0241—Advertisements
    • G06Q30/0277—Online advertisement

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an advertisement click rate prediction method based on a user historical behavior sequence, which comprises the following steps: acquiring target user characteristic information, advertisement information and user historical behavior information, constructing an advertisement related graph, performing enhanced representation on user historical behavior information data by using a graph embedding method, and learning to obtain a representation vector capable of fully expressing advertisement characteristics; a self-attention mechanism layer is added in the click rate estimation model, so that the internal correlation among the historical behaviors of the user can be learned more directly, and the dependence on external information is reduced. Experiments on a plurality of public amazon commodity click data sets and a traditto company advertisement click data set show that the technical scheme provided by the invention improves the accuracy of click rate prediction, so that advertisements which are interested in the user can be displayed to the user more accurately.

Description

Advertisement click rate prediction method based on user historical behavior sequence
Technical Field
The invention relates to an advertisement click rate prediction method, in particular to an advertisement click rate prediction method based on a user historical behavior sequence.
Background
The internet advertisement is a novel advertisement operation mode emerging under the wave of network development, and with the gradual maturity of the internet industry and the continuous expansion of user groups in recent years, the income brought by the internet advertisement is continuously increased. Internet advertising has many natural advantages over traditional media advertising. First, instant interactivity. The advertisement information release of the advertiser and the advertisement information receiving of the consumer are synchronized in real time, and the consumer can browse the preferred advertisement content; second, the popularity is advertised. The worldwide coverage of the Internet ensures that the Internet has wider publicity space; third, category diversity. The internet advertisement has a plurality of categories in size, technical means and user reading form, which cannot be achieved by the traditional media advertisement; fourth, it is easy to be statistical. The relevant data such as internet advertisement display, browsing, clicking and the like are very easy to count by using a computer and are more accurate than the traditional media; and fifthly, putting pertinence. The internet advertisement is a natural characteristic with absolute advantages, and can deeply understand user preference and accurately put in advertisement information matched with the user preference through long-term accumulation and deep analysis of browsing history, personal information and purchasing behavior of each consumer.
The advertisement click rate prediction algorithm is a core means for accurately putting online advertisements in a large scale, and the accuracy of advertisement click rate prediction not only relates to whether a user can obtain good experience, but also relates to whether advertisers and media merchants can obtain more economic benefits. The essence of the click rate estimation problem is to judge whether a user clicks an advertisement in a certain scene, so that the click rate estimation problem can be regarded as two types of division of 'clicking' and 'not clicking' on a sample. For the two-classification problem, the current common solutions are mainly classified into solutions based on the traditional machine learning and the deep learning. The traditional machine learning model comprises methods such as logistic regression, gradient decision tree, factorization machine and the like, and has the advantages of simple solution and strong interpretability, but the characteristic engineering needs to be carried out manually. The Deep learning models are typically Wide & Deep model proposed by google, Deep & Cross, Deep fm model proposed by hua corporation, and the like. The Wide & Deep model is formed by connecting a linear structure and a Deep neural network structure in parallel, the linear model of the Wide part of the model receives the characteristics extracted manually and is used for improving the memory capacity of the model to a specific rule, the high-order abstract characteristics extracted by the Deep neural network of the Deep part of the model Deep part; the Deep neural network and the Deep Cross network are combined in the Deep & Cross model, the Deep part of the model is used for extracting high-order abstract features, the Cross part of the model is used for extracting high-order Cross features with better interpretability, and the two high-order feature extraction modes are combined, so that a good effect is achieved on interpretability and generalization capability. The Deep FM model is provided by combining the FM model on the basis of the Wide & Deep model, compared with the Wide & Deep model, the Deep FM model increases the introduction of the second-order interaction characteristics of the FM model, does not need to carry out additional artificial characteristic engineering, and simplifies the work of characteristic processing.
Although the workload of feature engineering is greatly reduced by the existing deep neural network model, some problems also exist: the existing model only carries out implicit interactive modeling on the input original characteristics and does not consider the historical behavior characteristics of the user. And a great deal of user interest information is often contained in the user historical behaviors, and for a click rate prediction model, it is necessary to discover potential user interests through user behavior data. The method mainly aims at modeling the historical behavior information of the user, constructs the advertisement correlation diagram and trains to obtain the time sequence vector capable of fully expressing the advertisement characteristics through the diagram embedding method. And a click rate estimation model added into a self-attention layer is provided, and internal correlation among historical behaviors of the user is directly learned through a self-attention mechanism, so that the accuracy of advertisement click rate estimation is improved.
Disclosure of Invention
In order to solve the above problems, the present invention provides an advertisement click rate prediction method based on user historical behaviors from the perspective of a user.
The technical scheme adopted by the invention is as follows:
a click rate prediction method based on a user historical behavior sequence is realized by the following steps:
step 1: acquiring target user characteristic information, advertisement information and user historical behavior information;
step 2: constructing an advertisement related graph, performing enhanced representation on user historical behavior information data by using a graph embedding method, and constructing a data set by combining user characteristic information and advertisement information codes for model training;
and step 3: training a click rate estimation model HDSAN (Hybrid Deep Self-attack Neural Network) based on a Self-Attention mechanism layer and a multilayer fully-connected Neural Network by using a data set;
and 4, step 4: and performing polling training on the data through a forward propagation algorithm and a backward propagation algorithm to update the weight parameters to obtain an optimal parameter model.
An apparatus characterized by at least one processor, and a memory in communication with the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executable by the at least one processor to cause the at least one processor to perform one of the above-described advertisement click rate prediction methods based on a sequence of historical behaviors of a user.
The technical scheme provided by the invention has the following beneficial effects:
1. according to the method, the historical behavior information of the user is collected, the advertisement related graph is constructed, the graph embedding method is used for carrying out enhanced representation on the historical behavior information data of the user, the time sequence vector of the advertisement is obtained, and the defect that the existing model does not consider the advertisement time sequence information is overcome;
2. the click rate estimation model added with the self-attention mechanism layer is provided, and the internal correlation among the user historical behaviors is learned more directly through the self-attention mechanism, so that the dependence on external information is reduced.
Drawings
FIG. 1 is a flow chart according to the present invention;
FIG. 2 is a diagram of a process for training an advertisement timing vector according to the present invention;
FIG. 3 is a diagram of a click through rate prediction model based on the self-attention mechanism.
Detailed Description
Embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. The specific steps are described as shown in fig. 1, wherein:
step 1: and collecting characteristic information of the target user, advertisement information and historical behavior information of the user.
The target user characteristic information comprises user gender, age, address, occupation, interest and hobby;
the advertisement information comprises advertisement types (such as pictures, flash, pictures and texts), advertisement industry, advertisement positions, advertisement position sizes and advertisement audiences;
the historical behavior information of the user comprises historical clicks of the user and un-clicked advertisements, and time sequence information of behavior generation;
step 2: and constructing an advertisement related graph, performing enhanced representation on the historical behavior information data of the user by using a graph embedding method, and constructing a data set by combining the user characteristic information and the advertisement information codes for HDSAN model training.
2.1 user historical behavior information graph embedding enhancement representation
2.1.1 constructing advertisement correlation graphs corresponding to all users according to the time sequence information in the historical behavior information of the users. The advertisement correlation graph is generated by using advertisements as nodes and clicking the advertisements by the user, and the weight of the edge is the number of times of the clicking sequence of the user appearing in the data set, as shown in FIG. 2;
2.1.2 on the advertisement correlation diagram, randomly walking along the edge with an advertisement node as a starting point every time to generate a random advertisement sequence, and performing x rounds of random walks in total to obtain x random advertisement sequences, wherein x is a positive integer, as shown in fig. 2;
2.1.3 training the random advertisement sequence generated in the last step by a Word embedding method Word2vec to obtainTo each advertisement A in the user's historical behavior sequenceiTiming vector of
Figure BDA0002742425060000031
As shown in fig. 2; the obtained advertisement timing vector is used as the input of the HDSAN model self-attention mechanism layer, as shown in FIG. 3;
2.2, carrying out one-hot coding on the user characteristic information and the advertisement information to obtain a user characteristic vector and an advertisement information vector which are used as the input of an HDSAN model embedding layer; the method comprises the following steps:
2.2.1 obtaining user feature vectors
And (3) splicing the unique hot codes of m1 user features to obtain user feature vectors, wherein m1 user features are used: hU＝[U1，U2，…，Um1]
Wherein U isiIs a one-hot encoding of user characteristic i.
2.2.2 obtaining advertisement information vectors
The total number of the advertisement information is m2, and the unique hot codes of m2 advertisement information are spliced to obtain an advertisement information vector:
HT＝[T1，T2，…，Tm2]
wherein T isiIs a one-hot encoding of the advertisement information i.
2.3 constructing output item label of data set according to whether advertisement is clicked, if clicked, label is 1, otherwise, label is 0. Each sample in the data set consists of a user historical click sequence, user characteristic information, advertisement information, and a label of whether the advertisement was clicked. And finally, dividing the training set and the test set according to the ratio of 7:3 for the data set.
And step 3: a click rate estimation model HDSAN (Hybrid Deep Self-attack Neural Network) based on a Self-Attention mechanism layer and a multilayer fully-connected Neural Network is trained by using a data set, as shown in FIG. 3.
The HDSAN model comprises an embedding layer, a self-attention mechanism layer, a full connection layer and an output layer.
The embedding layer is used for embedding the user feature vector GUAdvertisement information vector HTMapping to a low-dimensional embedded vector to obtain a user characteristic embedded vector eUAdvertisement feature embedding vector eT；
The self-attention mechanism layer calculates the correlation among advertisements according to the time sequence vector of each advertisement in the user historical behavior sequence, and performs linear transformation on the advertisement time sequence vector according to the correlation weight to finally obtain the expression vector of the user historical behavior, specifically:
1) for a user historical behavior sequence [ A ]1，A2，…，An]The timing vector corresponding to each advertisement in the sequence is
Figure BDA0002742425060000045
The timing vector matrix is:
Figure BDA0002742425060000041
where Q, K, V represents a matrix of timing vectors of historical behavior of the user, each row of the matrix being a timing vector of one advertisement in the sequence; q ═ K ═ V, Q ∈ Rn×d，K∈Rn×d，V∈Rn×dN is the number of advertisements in the sequence and d is the dimension of the advertisement vector.
2) Attentions are carried out on a time sequence vector of each advertisement in a user historical behavior sequence and time sequence vectors of other advertisements in the sequence, wherein the attentions are carried out in a mode of solving for vector inner products, the similarity between the vectors is obtained by calculating the inner products, in order to prevent the similarity result from being too large, scaling is needed, namely, the scaling is divided by the dimension d of the vectors, and finally, the similarity is normalized through softmax to obtain a final Attention weight WijThe calculation formula is as follows:
Figure BDA0002742425060000042
wherein WijRepresenting advertisement AiTiming vector of
Figure BDA0002742425060000043
And advertisement AjTiming vector of
Figure BDA0002742425060000044
The similarity of (2), namely the attention weight;
3) forming an attention matrix W by the attention weights among all advertisement time sequence vectors in the user historical behavior sequenceatt；
Figure BDA0002742425060000051
Wherein n is the number of advertisements in the user historical behavior sequence, WijRepresenting advertisement AiTiming vector of
Figure BDA0002742425060000055
And advertisement AjTiming vector of
Figure BDA0002742425060000056
The similarity of (c).
4) The attention weight value of the attention weight value matrix and the time sequence vectors of all advertisements are used for weighting and summing to serve as a single advertisement AiIs represented by vector eiThe vector contains the correlation information between the advertisements in the sequence, and the calculation formula is as follows:
Figure BDA0002742425060000052
where n is the number of advertisements in the user's historical behavior sequence.
5) Summing the representation vectors of all advertisements in the user historical behavior sequence to obtain a representation vector e of the whole user historical behavior sequenceSThe calculation formula is as follows:
Figure BDA0002742425060000053
where n is the number of advertisements in the user's historical behavior sequence.
The third layer and the fourth layer are full connection layers and represent a user historical behavior vector eSEmbedding vector e with user characteristicsUAdvertisement feature embedding vector eTSplicing, learning high-order combination characteristics through a feedforward neural network, and calculating the output of the fully-connected neural network:
a(0)＝[eS，eU，eT]
a(l+1)＝σ(W(l)a(l)+b(l))
where l is the current layer depth, σ is the activation function, W(l)Is a weight matrix of the l-th layer, b(l)Is the bias term for the l-th layer. The final output of the feedforward neural network is:
yMLP＝σ(Wh+1ah+bh+1)
the fifth layer is an output layer, the output of the feedforward neural network is input into the sigmoid function, and finally, the click rate estimated value between 0 and 1 is output
Figure BDA0002742425060000057
Figure BDA0002742425060000058
And 4, step 4: the HDSAN model carries out polling training on data through a forward propagation algorithm and a backward propagation algorithm, updating of weight parameters is achieved, and an optimal parameter model is obtained. In the backward propagation, the present invention uses loglos as a loss function, which is expressed by the following formula:
Figure BDA0002742425060000054
in the above formula, n is the number of samples, i represents the ith sample, and its true label is y(i)The model prediction probability is hθ(xi)。
The experimental results are as follows:
the effect of the model of the invention was compared with other models on three data sets:
1) amazon Electronic data set: the data set contained review data from Amazon electronic products containing 192403 users, 63.001 goods, 801 categories, 1689188 samples.
2) Amazon Books dataset: the data set contained review data from Amazon book-like products, containing 603668 users, 367982 goods, 1600 categories, 603668 samples.
3) Set of stone advertising data: the discai information technology limited is a leading enterprise in china that is dedicated to advertising. In the data set, one training sample contains the characteristics of the advertisement (advertisement targeting gender, advertisement targeting industry, advertisement targeting age) and the characteristics of the advertisement position on the medium website (website visit amount, website affiliated industry, website user gender ratio).
The invention adopts AUC as the measurement standard for evaluating the click rate estimation model effect. AUC is a widely used metric in the field of advertisement click-through rate prediction, and refers to the probability that a positive sample A and a negative sample B are randomly drawn and the positive sample is arranged in front of the negative sample in the model prediction result. The calculation formula is as follows:
Figure BDA0002742425060000061
wherein M is the number of positive samples in the data set, N is the number of negative samples in the data set,
Figure BDA0002742425060000063
for the M × N pairs of samples, the prediction probability of the positive samples is greater than the number of prediction probabilities of the negative samples.
TABLE 1 results of the experiment
Figure BDA0002742425060000062
Experimental results show that the advertisement click rate prediction method based on the user historical behavior sequence, provided by the invention, has a great improvement in recommendation accuracy compared with other advanced models.

Claims (8)

1. An advertisement click rate prediction method based on a user historical behavior sequence is characterized by comprising the following steps:
step 1: collecting target user characteristic information, advertisement information and user historical behavior information; the historical behavior information of the user comprises historical clicks of the user, untapped advertisements and time sequence information of behavior generation;
step 2: constructing an advertisement related graph, performing enhanced representation on user historical behavior information data by using a graph embedding method, and constructing a data set by combining user characteristic information and advertisement information codes;
2.1 user historical behavior information graph embedding enhancement representation
2.1.1, constructing advertisement correlation graphs corresponding to all users according to time sequence information in the historical behavior information of the users; the advertisement correlation graph is generated by using advertisements as nodes and clicking the advertisements by a user in sequence, and the weight of the edges is the number of times of the clicking sequence of the user in the data set;
2.1.2 on the advertisement correlation graph, taking any advertisement node as a starting point, and randomly walking along the edge to generate a random advertisement sequence, and finally obtaining x random advertisement sequences, wherein x is a positive integer;
2.1.3 training the random advertisement sequence generated in the last step by a Word embedding method Word2vec to obtain each advertisement A in the user historical behavior sequenceiTiming vector of
Figure FDA0002742425050000011
2.2, carrying out one-hot coding on the user characteristic information and the advertisement information to obtain a user characteristic vector and an advertisement information vector which are used as the input of an HDSAN model embedding layer;
2.3 constructing output item labels of the data set according to whether the advertisement is clicked or not;
and step 3: training a click rate estimation model HDSAN based on a self-attention mechanism layer and a multilayer fully-connected neural network by using a data set;
the HDSAN model comprises an embedding layer, a self-attention mechanism layer, a full connection layer and an output layer;
the embedding layer is used for embedding the user feature vector HUAdvertisement information vector HTMapping to a low-dimensional embedded vector to obtain a user characteristic embedded vector eUAdvertisement information embedding vector eT；
The self-attention mechanism layer calculates the correlation among the advertisements according to the time sequence vector of each advertisement in the user historical behavior sequence, linearly transforms and sums the advertisement time sequence vectors according to the correlation weight value, and finally obtains the expression vector e of the user historical behaviorS；
The third layer and the fourth layer are full connection layers and represent a user historical behavior vector eSEmbedding vector e with user characteristicsUAdvertisement information embedding vector eTSplicing, learning high-order combination characteristics through a feedforward neural network, and calculating the output of the fully-connected neural network:
a(0)＝[eS，eU，eT]
a(l+1)＝σ(W(l)a(l)+b(l))
where l is the current layer depth, σ is the activation function, W(l)Is a weight matrix of the l-th layer, b(l)Is the bias term for the l-th layer; the final output of the feedforward neural network is:
yMLP＝σ(Wh+1ah+bh+1)
the fifth layer is an output layer, the output of the feedforward neural network is input into the sigmoid function, and finally, the click rate estimated value between 0 and 1 is output
Figure FDA0002742425050000021
Figure FDA0002742425050000022
2. The method of claim 1, wherein the target user characteristic information comprises user gender, age, address, occupation, interest, and hobby.
3. The method as claimed in claim 1, wherein the advertisement information includes advertisement type, industry of advertisement, advertisement location, advertisement slot size, and advertisement audience.
4. The method as claimed in claim 1, wherein in step 2.2, there are m1 user features, and the unique hot codes of m1 user features are concatenated to obtain the user feature vector: hU＝[U1，U2，…，Um1]
Wherein U isiIs a one-hot encoding of user characteristic i.
5. The method as claimed in claim 1, wherein the step 2.2 is implemented by concatenating m2 unique hot codes of the advertisement information to obtain the advertisement information vector, wherein m2 advertisement information are obtained in total according to the advertisement click rate prediction method based on the user historical behavior sequence:
HT＝[T1，T2，…，Tm2]
wherein T isiIs a one-hot encoding of the advertisement information i.
6. The method for predicting advertisement click-through rate based on user historical behavior sequence as claimed in claim 1, wherein the self-attention mechanism layer is specifically:
for a user historical behavior sequence [ A ]1，A2，…，An]The timing vector corresponding to each advertisement in the sequence is
Figure FDA0002742425050000023
The timing vector matrix is:
Figure FDA0002742425050000031
where Q, K, V represents a matrix of timing vectors of historical behavior of the user, each row of the matrix being a timing vector of one advertisement in the sequence; wherein Q ═ K ═ V, Q ∈ Rn×d，K∈Rn×d，V∈Rn×dN is the number of advertisements in the sequence, d is the dimension of the advertisement vector;
the time sequence vector of each advertisement in the user historical behavior sequence and the time sequence vectors of other advertisements in the sequence are subjected to Attention in a way of solving vector inner products, and the similarity W between the vectors is obtained by calculating the inner productsijI.e. attention weight, the calculation formula is as follows:
Figure FDA0002742425050000032
wherein WijRepresenting advertisement AiTiming vector of
Figure FDA0002742425050000033
And advertisement AjTiming vector of
Figure FDA0002742425050000034
The similarity of (2);
the attention weights among all advertisement time sequence vectors in the sequence form an attention matrix Watt：
Figure FDA0002742425050000035
Wherein n is the number of advertisements in the user historical behavior sequence;
weight matrix W according to attentionattTime of all advertisements in the sequenceThe order vectors are weighted and summed as the final expression vector e of the single advertisementiThe calculation formula is as follows:
Figure FDA0002742425050000036
finally, summing the representation vectors of all advertisements in the user historical behavior sequence to obtain a representation vector e of the whole user historical behavior sequenceSThe calculation formula is as follows:
Figure FDA0002742425050000037
7. the method for predicting advertisement click rate based on user historical behavior sequence as claimed in claim 1, wherein the HDSAN model performs polling training on data through a forward propagation algorithm and a backward propagation algorithm to update weight parameters and obtain an optimal parameter model; in the backward propagation, logloss is used as a loss function, and the formula is as follows:
Figure FDA0002742425050000038
in the above formula, n is the number of samples, i represents the ith sample, and its true label is y(i)The model prediction probability is hθ(xi)。
8. An apparatus characterized by at least one processor, and a memory in communication with the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executable by the at least one processor to cause the at least one processor to perform a method of predicting advertisement click-through rates based on a sequence of historical behaviors of a user according to any one of claims 1 to 7.
CN202011154969.6A 2020-10-26 2020-10-26 Advertisement click rate prediction method based on user historical behavior sequence Pending CN112288471A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011154969.6A CN112288471A (en) 2020-10-26 2020-10-26 Advertisement click rate prediction method based on user historical behavior sequence

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011154969.6A CN112288471A (en) 2020-10-26 2020-10-26 Advertisement click rate prediction method based on user historical behavior sequence

Publications (1)

Publication Number Publication Date
CN112288471A true CN112288471A (en) 2021-01-29

Family

ID=74372240

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011154969.6A Pending CN112288471A (en) 2020-10-26 2020-10-26 Advertisement click rate prediction method based on user historical behavior sequence

Country Status (1)

Country Link
CN (1) CN112288471A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112801719A (en) * 2021-03-01 2021-05-14 深圳市欢太科技有限公司 User behavior prediction method, user behavior prediction device, storage medium, and apparatus
CN112836081A (en) * 2021-03-01 2021-05-25 腾讯音乐娱乐科技（深圳）有限公司 Neural network model training method, information recommendation method and storage medium
CN113034196A (en) * 2021-04-07 2021-06-25 西北工业大学 Click rate prediction method based on core interest network
CN113468867A (en) * 2021-06-04 2021-10-01 淮阴工学院 Reference citation validity prediction method based on Attention mechanism
CN114491342A (en) * 2022-01-26 2022-05-13 阿里巴巴（中国）有限公司 Training method of personalized model, information display method and equipment

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106997549A (en) * 2017-02-14 2017-08-01 火烈鸟网络(广州)股份有限公司 The method for pushing and system of a kind of advertising message
CN111325579A (en) * 2020-02-25 2020-06-23 华南师范大学 Advertisement click rate prediction method

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106997549A (en) * 2017-02-14 2017-08-01 火烈鸟网络(广州)股份有限公司 The method for pushing and system of a kind of advertising message
CN111325579A (en) * 2020-02-25 2020-06-23 华南师范大学 Advertisement click rate prediction method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
李强: "基于图嵌入的广告推荐模型", 中国优秀硕士学位论文全文数据库信息科技辑, pages 0021 - 0036 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112801719A (en) * 2021-03-01 2021-05-14 深圳市欢太科技有限公司 User behavior prediction method, user behavior prediction device, storage medium, and apparatus
CN112836081A (en) * 2021-03-01 2021-05-25 腾讯音乐娱乐科技（深圳）有限公司 Neural network model training method, information recommendation method and storage medium
CN113034196A (en) * 2021-04-07 2021-06-25 西北工业大学 Click rate prediction method based on core interest network
CN113468867A (en) * 2021-06-04 2021-10-01 淮阴工学院 Reference citation validity prediction method based on Attention mechanism
CN114491342A (en) * 2022-01-26 2022-05-13 阿里巴巴（中国）有限公司 Training method of personalized model, information display method and equipment
CN114491342B (en) * 2022-01-26 2023-09-22 阿里巴巴（中国）有限公司 Training method of personalized model, information display method and equipment

Similar Documents

Publication Publication Date Title
CN111538912B (en) Content recommendation method, device, equipment and readable storage medium
CN111339415B (en) Click rate prediction method and device based on multi-interactive attention network
CN112288471A (en) Advertisement click rate prediction method based on user historical behavior sequence
CN111222332A (en) Commodity recommendation method combining attention network and user emotion
CN112288042B (en) Updating method and device of behavior prediction system, storage medium and computing equipment
CN112258260A (en) Page display method, device, medium and electronic equipment based on user characteristics
CN113379449B (en) Multimedia resource recall method and device, electronic equipment and storage medium
CN111241394A (en) Data processing method and device, computer readable storage medium and electronic equipment
CN112287238B (en) User characteristic determination method and device, storage medium and electronic equipment
CN115495654A (en) Click rate estimation method and device based on subspace projection neural network
CN116541607A (en) Intelligent recommendation method based on commodity retrieval data analysis
CN111047009A (en) Event trigger probability pre-estimation model training method and event trigger probability pre-estimation method
US20230316106A1 (en) Method and apparatus for training content recommendation model, device, and storage medium
WO2020211616A1 (en) Method and device for processing user interaction information
CN110555719A (en) commodity click rate prediction method based on deep learning
CN114330519A (en) Data determination method and device, electronic equipment and storage medium
CN114841765A (en) Sequence recommendation method based on meta-path neighborhood target generalization
CN114519600A (en) Graph neural network CTR estimation algorithm fusing adjacent node variances
CN112118486B (en) Content item delivery method and device, computer equipment and storage medium
CN114781625B (en) Network model training and push content determining method and device
Di Deep interest network for taobao advertising data click-through rate prediction
CN113888238B (en) Advertisement click rate prediction method and device and computer equipment
Wen et al. Cat2Vec: Learning distributed representation of multi-field categorical data
Zhou Understanding and Improving Recommender Systems’ Performance in the Presence of Practical User-, Item-, and Marketing-Oriented Considerations
Yin et al. Clr: coupled logistic regression model for ctr prediction

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination