CN113888238B - Advertisement click rate prediction method and device and computer equipment - Google Patents
Advertisement click rate prediction method and device and computer equipment Download PDFInfo
- Publication number
- CN113888238B CN113888238B CN202111243596.4A CN202111243596A CN113888238B CN 113888238 B CN113888238 B CN 113888238B CN 202111243596 A CN202111243596 A CN 202111243596A CN 113888238 B CN113888238 B CN 113888238B
- Authority
- CN
- China
- Prior art keywords
- user
- advertisement
- interest
- time
- data
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 44
- 239000013598 vector Substances 0.000 claims abstract description 56
- 238000012549 training Methods 0.000 claims abstract description 25
- 239000013604 expression vector Substances 0.000 claims abstract description 22
- 230000007246 mechanism Effects 0.000 claims abstract description 16
- 238000013528 artificial neural network Methods 0.000 claims abstract description 14
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 230000006399 behavior Effects 0.000 claims description 91
- 230000006870 function Effects 0.000 claims description 26
- 230000009471 action Effects 0.000 claims description 13
- 238000000605 extraction Methods 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 13
- 230000003068 static effect Effects 0.000 claims description 12
- 238000012545 processing Methods 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 9
- 239000011159 matrix material Substances 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 5
- 238000009499 grossing Methods 0.000 claims description 2
- 238000012163 sequencing technique Methods 0.000 claims description 2
- 230000006872 improvement Effects 0.000 claims 1
- 230000001502 supplementing effect Effects 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 4
- 239000010410 layer Substances 0.000 description 28
- 238000010586 diagram Methods 0.000 description 9
- 230000003993 interaction Effects 0.000 description 5
- 238000004088 simulation Methods 0.000 description 5
- 238000011161 development Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000000717 retained effect Effects 0.000 description 2
- 230000004075 alteration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000006397 emotional response Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009469 supplementation Effects 0.000 description 1
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- General Business, Economics & Management (AREA)
- Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Marketing (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to the field of electronic commerce big data recommendation, and relates to an advertisement click rate prediction method, an advertisement click rate prediction device and computer equipment; the method comprises the steps of obtaining user behavior data, user portrait data and advertisement data of an electronic commerce platform; preprocessing user behavior data to form a user behavior sequence; coding and representing the user behavior sequence, the user portrait data and the advertisement data respectively to obtain embedded vectors of corresponding features; extracting interest expression vectors of users by adopting a deep neural network based on an attention mechanism; extracting invisible relation vectors between portrait features and advertisement features of a user by adopting a stack type automatic encoder; inputting interest expression vectors and invisible relation vectors of users into a multi-layer perceptron for joint training to obtain a prediction result of the advertisement click rate; the invention can effectively improve the click rate of the e-commerce platform advertisement and realize the effects of accurate marketing and recommendation.
Description
Technical Field
The invention belongs to the field of electronic commerce big data recommendation, and particularly relates to an advertisement click rate prediction method, an advertisement click rate prediction device and computer equipment based on user interests and time sequence behaviors.
Background
With the development of information technology, many domestic and foreign internet electronic commerce platforms are increasingly focusing on the profit effect of an online advertisement system and on realizing personalized and accurate marketing strategies. The advertisement click rate (CTR, click Through Rate) is one of the most core indexes in the e-commerce platform system, and is important in the fields of advertisement recommendation, web page search, sponsored recommendation and the like. The accuracy of click rate prediction not only affects the benefits of the e-commerce platform, but also affects the satisfaction and consumption experience of the user.
In current e-commerce platforms, although marketers want to know the response of network visitors, it is almost impossible to quantify the emotional response to a website and the impact of the website on the company brands using current technology. However, click through rates are readily available. Click-through rate measures the ratio of the number of page visitors to the page after clicking on the commercial advertisement and redirecting it to the visitor of another page where they can purchase the commercial or learn more about the product or service. In general, a higher click rate indicates that the advertised product is more commercially valuable or that the marketing campaign is more attractive. Most e-commerce websites aim to adjust the display of homepage commodity advertisements through click rate for personalized recommendation.
At present, many domestic and foreign scholars develop intensive researches on the CTR model, and research results mainly show the following aspects: on the one hand, with the development of deep learning technology, the deep CTR model gradually replaces the LR and other machine learning-based CTR models requiring artificial feature engineering. On the other hand, some depth CTR models focus on compression and interaction of features. In addition, there are models that focus on the extraction of user behavior sequence features. However, the click rate of the advertisement at the current stage still has the following defects:
1. timeliness of the user's historical behavior sequence. The traditional time sequence model ignores the influence of time intervals between sequential behaviors on the expression of the user interests, and the traditional RNN can well capture the dependency relationship between the sequential relationships in the behavior sequence, but the user behaviors are not just the sequential relationships, the time intervals of the behaviors, the characteristics of the behaviors and the like contain more prior information, and the information is critical to the expression of the user interests.
2. Generalization and complexity of user interests. The interests of users have diversity and have a trend of variation, the preference of users has concentration in a certain period of time, each interest has own evolution trend, and different kinds of interests have little interaction.
3. The dimension of the data features is high, and the implicit information quantity is large. Besides the user behavior sequence characteristics, the relationship between the characteristics such as the context characteristics, the advertisement characteristics and the like also influences the accuracy of click rate estimation. These features are high in dimension and large in hidden information, and it becomes difficult to obtain the relationship between them.
Disclosure of Invention
In order to overcome the defects, the invention provides a method and a device for predicting the click rate of an advertisement and computer equipment for solving the problem of predicting the click rate of the advertisement.
In a first aspect of the present invention, the present invention provides an advertisement click rate prediction method, the method comprising:
acquiring user behavior data of an e-commerce platform, user portrait data and advertisement data;
preprocessing user behavior data of an e-commerce platform and forming a user behavior sequence;
coding and representing the user behavior sequence, the user portrait data and the advertisement data respectively to obtain embedded vectors of corresponding features;
inputting user behavior sequence characteristics, and outputting interest expression vectors of users by adopting a Time-GRU (Time-based neural network);
inputting an interest expression vector of a user, adopting a deep neural network of an AT-GRU based on an attention mechanism, simulating an interest updating process, and outputting an interest updating vector of the user;
inputting user portrait features and advertisement features, and extracting invisible relation vectors between the user portrait features and the advertisement features by adopting a stack type automatic encoder;
and inputting the interest expression vector of the user and the invisible relation vector between the portrait characteristic and the advertisement characteristic of the user into a multi-layer perceptron for joint training to obtain the prediction result of the advertisement click rate.
In a second aspect of the present invention, the present invention further provides an advertisement click rate prediction apparatus, the apparatus comprising:
the acquisition module is used for acquiring user behavior data, user portrait data and advertisement data of the electronic commerce platform;
the processing module is used for preprocessing the user behavior data of the e-commerce platform and forming a user behavior sequence;
the embedding module is used for respectively coding and representing the user behavior sequence, the user portrait data and the advertisement data to obtain an embedding vector of the corresponding feature;
the first feature extraction module is used for inputting user behavior sequence features, adopting a Time-GRU (Time-based group) deep neural network based on a Time factor, and outputting interest expression vectors of users;
the second feature extraction module is used for inputting interest expression vectors of users, simulating an interest updating process by adopting an AT-GRU deep neural network based on an attention mechanism, and outputting interest updating vectors of the users;
the third feature extraction module is used for inputting the portrait features and the advertisement features of the user and extracting the recessive relation vectors between the portrait features and the advertisement features of the user by adopting a stack type automatic encoder;
and the advertisement click rate prediction module is used for inputting the interest expression vector of the user and the invisible relation vector between the portrait characteristic of the user and the advertisement characteristic into the multi-layer perceptron for joint training to obtain a prediction result of the advertisement click rate.
In a third aspect of the invention there is also provided a computer device comprising a memory and a processor, the memory having stored therein a computer program which, when executed by the processor, causes the processor to perform the steps of the method according to the first aspect of the invention.
The invention has the beneficial effects that:
according to the invention, the user behavior data and the target advertisement data of the e-commerce platform are utilized, the interest of the user hidden behind the time sequence behavior sequence is updated and modeled to obtain the interest representation, and the invisible correlation between other non-time sequence features is combined to predict the advertisement click rate.
Drawings
FIG. 1 is a diagram of an advertisement click rate prediction framework in an embodiment of the present invention;
FIG. 2 is a flowchart of a method for predicting advertisement click rate according to an embodiment of the present invention;
FIG. 3 is a diagram of a simulated user interest feature of the present invention in constructing a time series model;
FIG. 4 is a diagram of an interest update model for constructing an attention mechanism in accordance with the present invention;
FIG. 5 is a diagram of an unsupervised feature extraction method introduced in the present invention;
fig. 6 is a block diagram of a private advertisement click rate prediction apparatus according to the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Fig. 1 is a frame diagram of advertisement click rate prediction in an embodiment of the present invention, as shown in fig. 1, in the prediction frame of the present embodiment, the frame mainly includes four parts, firstly, user behavior data, user portrait data and advertisement data are collected; processing the data to obtain a user behavior sequence, user portrait characteristics and advertisement characteristics; secondly, constructing an interest updating model by using a user behavior sequence in a data characteristic processing mode, obtaining an interest expression vector of a user by using the interest simulation model, and simultaneously processing user portrait data and advertisement data by using a characteristic interaction model to extract a recessive relation vector; then constructing a click rate prediction model by combining the final interest expression of the user and the non-time sequence characteristic invisible relation; the click rate prediction model can be used for predicting the click rate of the advertisement, and accurate advertisement pushing can be completed according to the click rate of the advertisement.
Fig. 2 is a flowchart of an advertisement click rate prediction method in an embodiment of the present invention, as shown in fig. 2, where the method includes:
s1, acquiring user behavior data, user portrait data and advertisement data of an electronic commerce platform;
in the embodiment of the invention, some basic data of the e-commerce platform can be acquired, including user historical behavior data, user portrait data and advertisement data. Taking the online advertisement display data set of the Taobao E-commerce platform as an example, the data set mainly records the browsing/clicking record of the Taobao user on the E-commerce platform, and comprises three parts, namely user behavior history, user portrait and advertisement basic information. The user behavior history data comprises fields such as a user ID, an advertisement ID, time, clicking or not, and the like, is time sequence characteristic data showing implicit interests of a user, and can obtain a user behavior sequence through processing the user behavior data. The user portrait data includes feature information such as user ID, age, sex, shopping depth, etc., and reflects basic feature information of the user. The advertisement basic data comprises characteristic information such as advertisement ID, commodity category ID, commodity brand ID, price and the like, and the advertisement data shows some basic characteristics of the advertisement to be recommended and is an important component of non-time sequence data.
In the embodiment of the invention, for the method for acquiring the data, the original data can be obtained by the modes of data sources provided by the e-commerce platform or directly downloading the existing public data sources.
The raw data that is typically acquired is unstructured and cannot be used directly for data analysis. Most unstructured data can be structured by simple data cleansing. For example, delete duplicates, clean up invalid nodes such as portions of guest data, and the like.
S2, preprocessing user behavior data of the e-commerce platform and forming a user behavior sequence;
in the embodiment of the invention, invalid user behavior data can be removed; for example, some user behavior that is too short of a browsing time can affect the effectiveness of the data, and the present invention defines a browsing threshold of 25 seconds for advertisement click effectiveness. When the user browses at an advertisement page above the threshold, the data is considered valid, otherwise, the data is invalid; these invalid data are deleted.
In the embodiment of the invention, the statistics of the data can be carried out according to the number of the users, and the original user behavior data can be spliced according to the ID, the browse advertisement and the time stamp information of each user; in this way a series of user behavior data is formed.
In the embodiment of the invention, the statistical user behavior data can be complemented by adopting a multiple interpolation method; for example, when the missing rate exceeds 15%, the data is directly removed, and the data value which does not exceed 15% is supplemented, wherein the supplementation mode can be set by a person skilled in the art according to actual requirements.
In the embodiment of the invention, a user behavior sequence based on time difference is constructed; grouping the user behavior data according to the user ID, and sequencing the user behavior data according to the time sequence to form a user behavior sequence; for each of the series of actions, the difference between the timestamp of the next action and the timestamp of the current action is used as a time factor feature.
S3, coding and representing the user behavior sequence, the user portrait data and the advertisement data respectively to obtain embedded vectors of corresponding features;
in the embodiment of the invention, the processed user behavior sequence, advertisement data and user portrait data are subjected to one-hot coding, and then the data characteristics can be respectively normalized. And then, using a feature embedding method to convert the input high-dimensional sparse feature vector into a low-dimensional dense vector to obtain embedded representation of feature data, namely outputting low-dimensional dense user behavior sequence features, advertisement features and user portrait features.
In a preferred embodiment of the present invention, the user behavior sequence features employ time-series modeling, that is, it is assumed that the user behavior sequence features are a user behavior-time sequence feature binary set U (B, Δt), specifically expressed as;
U(B,ΔT)＝{(b 1 ,δt 1 ),(b 2 ,δt 2 ),…,(b n ,δt n )}
the user behavior-time series feature binary set U (B, Δt) is defined as a binary set of B and Δt, expressed as:
B＝{b 1 ,b 2 ,...,b n }
ΔT＝{δt 1 ,δt 2 ,...,δt n |δt 1 ＝0，δt i ＝time(b i )-time(b i-1 )i>1}
wherein B represents a historical behavior sequence feature set of the user, deltaT represents the difference between corresponding times of two adjacent user behaviors in B, and Deltat 1 =0 means that the time difference corresponding to the first sequence set is 0.
S4, inputting user behavior sequence characteristics, and outputting interest expression vectors of users by adopting a Time-GRU (Time-based neural network);
in the embodiment of the invention, a gating cycle unit (Time-Gate Recurrent Unit, abbreviated as Time-GRU) based on a Time factor adopts a mode of combining the Time factor with the GRU; the process of simulating interest expression vector of user in the invention mainly adopts time-gating circulation unit to learn static interest group state set of user according to user behavior sequence, expressed as Intres s ＝Time-GRU(e u )＝{h' 1 ,h' 2 ,...,h' n }。
Wherein, static interest group state set Intres s Is defined as a hidden state set of each moment which is output after time sequence modeling after feature embedding processing is carried out on a user behavior-time sequence feature binary set, wherein e u An embedded representation representing a set of user behavior-time series feature tuples, each hidden state in the set reflecting a user interest extracted from the user behavior sequence at that time; time-GRU (e) u ) Representation pair embedded vector e u A result obtained by a time gating circulation unit is adopted; h's' n Indicating the nth hidden interest state.
FIG. 3 is a schematic illustration of a timing model constructed in accordance with an embodiment of the present inventionThe interest feature diagram, as shown in FIG. 3, represents the Time-GRU core structure in the Time series model by combining the hidden interest states h of the last behavior sequence t-1 With the behavior characteristics i of the input user in the current behavior sequence t And a time factor Deltat, outputting the user interest h represented by the next hidden interest state through the actions of the time gate, the update gate and the reset gate t . The time gating circulation unit strengthens the influence of time factors in the user behavior sequence on the user interests, and each middle hidden interest state of the time gating circulation unit highlights the static interests of the user which embody short-term and high-frequency behaviors of the user at a certain moment.
In a preferred embodiment of the present invention, the step S4 specifically includes the following steps:
s41, calculating time gate weight according to the input time interval and the input characteristics, wherein a specific formula is as follows, T g ＝σ(W t [Δt,i t ]): where Δt is the time factor, i.e. the difference between the timestamp of the current activity and the timestamp of the last activity. Considering that the decay of preferential behaviors such as browsing, clicking and the like of a user along with time accords with long tail distribution in a serialization model, the invention adds logarithmic processing to a time factor delta T, and the improved time gate weight T t The formula of (2) is as follows:
T t ＝σ(W t [log(Δt+ζ),i t ])
when the input time factor Δt is smaller, the time gate weight T g The smaller the log-processed time gate weight T t The smaller the information retained by the current step, the more information is retained by the previous step. Namely, the time interval between two adjacent behaviors of the user is shorter, and the dependency relationship between the two behaviors is higher;
s42, respectively updating the reset gates r according to the input characteristics, the time interval and the state of the last state t Update door z t And intermediate hidden interest statesThe specific formula is as follows:
z t ＝σ(W z i t +U z h t-1 +b z )，
r t ＝σ(W r i t +U r h t-1 +b r )，
wherein σ is a sigmoid activation function, omicron is element-wise multiplied, w z ,w r ,U z ,U r ,U h ∈n H ×n H ，n H Is the size of the hidden layer, n I Is the size of the input layer. i.e t Input vector representing Time-GRU, < - > and +.>A temporary state, z, representing the t-th hidden interest state t Is an update gate (update gate), r t Is a reset gate. z t And r t Under the mapping action of the sigmoid function, the value range is 0 to 1;
s43, adding the time gate weight into an updating strategy of an updating gate, wherein the specific formula is as follows:
in the embodiment of the invention, the time factor is newly added in the gate structure as input, and simultaneously, the time weight is internally introduced as auxiliary participation in the updating strategy of the updating gate, so that the time factor can be used as an important factor to participate in the interest simulation process.
S5, inputting an interest expression vector of the user, adopting a deep neural network of the AT-GRU based on an attention mechanism, simulating an interest updating process, and outputting an interest updating vector of the user;
in the embodiment of the invention, a Attention-based gating-and-circulating unit (Attention-Gate Recurrent Unit, AT-GRU for short) adopts a mode of combining an Attention mechanism with the GRU, and the network is described in detail in the future.
FIG. 4 is a diagram of an interest update model for constructing an attention mechanism according to the present invention, as shown in FIG. 4, the structure is a core structure of an attention mechanism-based gate-control loop unit AT-GRU, and an input part of the structure is a model representing a user static interest h simulated by a time sequence model in a current step t t Attention score alpha associated with this interest and targeted advertisement t The user interest is simulated to update along the process related to the target advertisement by the function of the improved updating gate, and the final unit outputs the simulated final interest state. In an embodiment of the present invention, the step S5 may include the following steps:
s51, calculating an attention score set of each interest state and the target advertisement according to the static interest group state set, wherein the attention score set is expressed as Atns= { alpha i |i＝1,2,...T}；
In an embodiment of the invention, α i Representing the ith attention score, T representing the number of states of interest; alpha i Defined as each interest state h 'in SI' i The similarity measurement with the target advertisement q is calculated by a weight parameter distribution mechanism, and important characteristics in the model can be captured.
Wherein s (h' i Q) represents the interest state h' i Similarity weights calculated by a similarity function of the bilinear model with the target advertisement q,
s52, learning an interest final update state by using a gating circulation unit based on an attention mechanism according to the static interest group state set and the attention score set, wherein the interest final update state is expressed as H=AT-GRU (Intres s ，Atns)＝{h i I=1, 2..t }, T being the size of the hidden layer in the AT-GRU;
considering the dynamics and generalization of user interests, the interest update final state is defined as a static interest group state Intres s Under the attention-based mechanism updating strategy, the final interest expression vector is extracted through the interest updating model.
In a preferred embodiment of the present invention, the step S52 specifically includes the steps of:
s521, according to the formula of step S51, the relevance weight of each interest and the candidate advertisement, namely the attention score a, can be obtained i The method is characterized by comprising the following steps:
wherein e ad Is a concatenation of embedded vectors from different categories of advertisement fields,is a parameter matrix, n H Is the dimension of the hidden interest state vector, n A Is the dimension of the embedded vector of the advertisement. Attention score a i Reflecting the correlation between the targeted advertisement and the input interest state, the more relevant the interest state is to the targeted advertisement, the larger the attention score is.
S522, introducing a attention-mechanism-based gating circulating unit AT-GRU according to the attention score calculated in the steps, wherein the structure can determine the updating strength of the hidden interest state according to the size of the attention score, namely the interest state related to the target advertisement can participate in the updating process of the final interest state with greater strength, and the interest not related to the target advertisement can be less or even not participated in the updating process, and the specific updating strategy is as follows:
r’ t ＝σ(W ri i' t +U ri h' t-1 +b ri )，
wherein h' t 、h' t-1 Andare all hidden states, w, of the AT-GRU ri ，/>U ri ,U hi ∈n Hi ×n Hi ，n Hi Is the size of the AT-GRU hidden layer, n Ii Is the size of the AT-GRU input layer. i' t Input vector representing AT-GRU, i.e. user static interest learned by time-gated loop unit, a t Compared with the original GRU structure, the attention score is used by the AT-GRU structure to replace the original update gate, the problem of interest drift caused by generalization and isomerism of the interest of the user can be effectively avoided by the AT-GRU, the process of simulating interest development update from the interest of the user which is continuously developed is realized, and the final interest is promoted to be updated along the direction related to the target advertisement.
S6, inputting user portrait features and advertisement features, and extracting invisible relation vectors between the user portrait features and the advertisement features by adopting a stack type automatic encoder;
in the embodiment of the invention, the implicit relation between the features is calculated according to the portrait feature data of the user and the target advertisement feature data and is expressed as R implicit ＝SAE(I,P)。
Because of the implicit relationship between the target advertisement and the non-time sequence characteristics such as the portrait of the user, the embodiment of the invention defines R implicit The method is a recessive relation between the features extracted by the stack encoder after feature compression, wherein I is a target advertisement feature embedding vector set, and P is a user portrait embedding vector set.
In the embodiment of the present invention, a stacked automatic encoder structure is designed to obtain the relationship of further other non-timing characteristics, wherein the structure of the single-layer automatic encoder is shown in fig. 5, and the structure is composed of an encoder and a decoder and is divided into three parts, namely an input layer, a hidden layer and an output layer. The encoder layer converts the features of the input layer into features of the hidden layer under the action of the encoder function, and then converts the hidden layer features into the output layer under the action of the decoder function. The stack type automatic encoder initializes parameters of the depth network through pre-training of layer-by-layer non-supervised learning, and can learn high-dimensional nonlinear element interaction relations among non-time sequence features by utilizing training parameters after the pre-training is finished.
In a preferred embodiment of the present invention, the step S6 may specifically include the following steps:
s61, according to the input user portrait features and target advertisement features, calculating a coding layer, wherein the coding layer is responsible for converting input data X of the input layer into a state H of a hidden layer, and the specific formula is as follows:
Z＝sigmoid(W 1 X+b 1 )，
wherein W is 1 Is a weight matrix, b 1 Is a first training bias that is set to be the first training bias,is an activation function;
s62, similarly, calculating a decoding layer, wherein the decoding layer converts the state H of the hidden layer into an output layer Y, and the definition is as follows:
Y＝sigmoid(W 2 Z+b 2 )，
wherein W is 2 Is a weight matrix, b 2 Is a second training bias.
S63, calculating a reconstruction error, so that the error between the output Y and the original X is small enough, wherein a specific formula is that,
wherein W is W 1 And W is 2 λ is a regularization coefficient, and a penalty factor λ may be added to control the magnitude of the weights to prevent overfitting.
S64, repeating the training process of the steps S61-S63, obtaining training parameters of the whole stack type automatic encoder in a layer-by-layer superposition training mode, and then learning a hidden relation between non-time sequence features according to the parameters.
S7, inputting the interest expression vector of the user and the invisible relation vector between the portrait characteristic of the user and the advertisement characteristic into a multi-layer perceptron for joint training, and obtaining a prediction result of the advertisement click rate.
In the embodiment of the invention, the interest updating vector of the user and the invisible relation vector are connected, and the connected vector is subjected to smoothing treatment; and respectively carrying out joint training on the auxiliary loss function of the time gating circulating unit in the interest simulation and update model and the predicted target loss function of the multi-layer perceptron, and obtaining the predicted result of the advertisement click rate after training is completed.
When the prediction model is trained, a joint training mode is adopted to perform joint training on the Time-GRU part and the target loss function of the prediction model, and the global loss function of the model is expressed as follows:
L＝L target +λ*L aux
where λ is a hyper-parameter, simulation of user balance interests and prediction of advertisement click rate, L aux Indicating the loss of assistance of the Time-GRU. In the embodiment of the invention, the target loss function L target When the MLP loss function is improved, the MLP loss function is set as weighted mean square error, and the coefficient of the loss function is set according to the proportion of positive and negative samples in the existing data, and the improved target loss function is expressed as:
wherein L is target To improve the multilayer senseKnowing the target loss function of the machine; n1 represents the number of positive samples; n2 represents the number of negative samples; y is an indicator variable, 1 if the class is the same as the class of the sample, or 0 otherwise; p (y= 0|X) and p (y= 1|X) are respectively different prediction probabilities that the network output belongs to the tag. At the same time, the invention also introduces auxiliary loss L aux The auxiliary loss is expressed as:
wherein,the t-th embedded vector representing a user click, G being the entire set of items; />Representing the embedding of samples outside the item clicked by user i at step t; />Is a sigmoid activation function,/->Indicating the t-th hidden interest state of user i in the Time-GRU. Assistance loss uses the next positive and negative click sample behavior to supervise learning of the current state of interest. The design of auxiliary loss introduces feedback information of the whole network behavior of the user, and meanwhile, click deviation among multiple scenes is not introduced and multi-scene coupling is not caused; from an optimization perspective, the auxiliary loss can reduce the difficulty of gradient back propagation in long-sequence modeling of the GRU. Under the action of auxiliary loss, the Time-GRU can correlate the hidden interest state of the output of each unit with the next click action, so that the interest of the user can be better simulated according to the user action sequence; meanwhile, under the update strategy of the Time gate, the Time-GRU strengthens the influence of the Time interval in the user behavior sequence on the interest simulation, so that the user is interested by the behaviors which are shorter in Time interval and more frequently clickedThe greater the impact of (c).
Fig. 6 is a block diagram of an advertisement click rate prediction apparatus according to an embodiment of the present invention, as shown in fig. 6, the apparatus includes:
201. the acquisition module is used for acquiring user behavior data, user portrait data and advertisement data of the electronic commerce platform;
202. the processing module is used for preprocessing the user behavior data of the e-commerce platform and forming a user behavior sequence;
203. the embedding module is used for respectively coding and representing the user behavior sequence, the user portrait data and the advertisement data to obtain an embedding vector of the corresponding feature;
204. the first feature extraction module is used for inputting user behavior sequence features, adopting a Time-GRU (Time-based group) deep neural network based on a Time factor, and outputting interest expression vectors of users;
205. the second feature extraction module is used for inputting interest expression vectors of users, simulating an interest updating process by adopting an AT-GRU deep neural network based on an attention mechanism, and outputting interest updating vectors of the users;
206. the third feature extraction module is used for inputting the portrait features and the advertisement features of the user and extracting the recessive relation vectors between the portrait features and the advertisement features of the user by adopting a stack type automatic encoder;
207. and the advertisement click rate prediction module is used for inputting the interest expression vector of the user and the implicit relation vector between the portrait characteristic of the user and the advertisement characteristic into the multi-layer perceptron for joint training to obtain a prediction result of the advertisement click rate.
In a preferred embodiment of the present invention, a computer device of the present invention may include a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the advertisement click rate prediction method according to the present invention. The computer device may be a terminal or a server. The computer device includes a processor, a memory, and a network interface connected by a system bus. The memory includes a nonvolatile storage medium and an internal memory. The non-volatile storage medium of the computer device may store an operating system and a computer program. The computer program, when executed, may cause the processor to perform a method of advertisement click rate prediction. The processor of the computer device is used to provide computing and control capabilities, supporting the operation of the entire computer device. The internal memory may have stored therein a computer program which, when executed by the processor, causes the processor to perform a preference prediction method. The network interface of the computer device is used for network communication.
In one embodiment, the advertisement click rate prediction apparatus provided in the present application may be implemented in the form of a computer program, where the computer program may run on the above-mentioned computer device, and a nonvolatile storage medium of the computer device may store each program module that constitutes the advertisement click rate prediction apparatus. For example, the acquisition module, the processing module, the embedding module, the first feature extraction module, the second feature extraction module, and the advertisement click rate prediction module shown in fig. 6. The computer program comprising the program modules is for causing the computer device to execute the steps in the advertisement click rate prediction method of the various embodiments of the present application described in the present specification.
In the description of the present invention, it should be understood that the terms "coaxial," "bottom," "one end," "top," "middle," "another end," "upper," "one side," "top," "inner," "outer," "front," "center," "two ends," etc. indicate or are based on the orientation or positional relationship shown in the drawings, merely to facilitate description of the invention and simplify the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and therefore should not be construed as limiting the invention.
In the present invention, unless explicitly specified and limited otherwise, the terms "mounted," "configured," "connected," "secured," "rotated," and the like are to be construed broadly, and may be, for example, fixedly connected, detachably connected, or integrally formed; can be mechanically or electrically connected; either directly or indirectly through intermediaries, or in communication with each other or in interaction with each other, unless explicitly defined otherwise, the meaning of the terms described above in this application will be understood by those of ordinary skill in the art in view of the specific circumstances.
Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.
Claims (8)
1. The advertisement click rate prediction method is characterized by comprising the following steps of:
acquiring user behavior data of an e-commerce platform, user portrait data and advertisement data;
preprocessing user behavior data of an e-commerce platform and forming a user behavior sequence;
coding and representing the user behavior sequence, the user portrait data and the advertisement data respectively to obtain embedded vectors of corresponding features;
inputting user behavior sequence characteristics, and outputting interest expression vectors of users by adopting a Time-GRU (Time-based neural network);
inputting an interest expression vector of a user, adopting a deep neural network of an AT-GRU based on an attention mechanism, simulating an interest updating process, and outputting an interest updating vector of the user;
inputting user portrait features and advertisement features, and extracting invisible relation vectors between the user portrait features and the advertisement features by adopting a stack type automatic encoder;
the interest updating vector of the user, the invisible relation vector between the portrait characteristic and the advertisement characteristic of the user are respectively input into a multi-layer perceptron for joint training, and the prediction result of the advertisement click rate is obtained;
the interest updating vector of the user and the invisible relation vector between the portrait characteristic of the user and the advertisement characteristic are respectively input into a multi-layer perceptron for joint training, and the prediction result of the advertisement click rate is obtained by connecting the interest updating vector of the user and the invisible relation vector and smoothing the connected vector; respectively carrying out joint training on a local loss function of a Time-GRU deep neural network part and a global loss function of a multi-layer perceptron, and obtaining a predicted result of the advertisement click rate after training is completed; wherein the global loss function is expressed as:
L＝L target +λ*L aux
wherein L represents a global loss function of the multi-layer perceptron; lambda is a superparameter, L aux Representing an auxiliary loss function of the time-gated loop unit;n represents the number of users;the t embedded vector representing the single click of the user i, G is the whole behavior sequence embedded set; />An embedded vector representing samples of user i outside the item clicked at step t; sigma represents the sigmoid activation function, +.>Representing the t-th hidden interest state of the user i in the Time-GRU; l (L) target Representing an objective loss function based on positive and negative sample scale improvement;n1 represents the number of positive samples; n2 represents the number of negative samples; y is an indicating variable, 1 if class Y is the same as that of sample X, or 0 otherwise; p represents the prediction probability of the output of the multi-layer perceptron network belonging to the label;
the predicted result of the advertisement click rate is expressed as follows:
y＝sigmoid(W L (...(W c (R c )+b c )...)+b L )
wherein R is c A connection vector representing the update vector of interest of the user and the invisible relation vector; w (W) L The first weight parameter matrix is a first weight parameter matrix; w (W) c B is a second weight parameter matrix c Representing a first training bias; b L Biased for the second training.
2. The method of claim 1, wherein preprocessing the user behavior data of the e-commerce platform comprises removing invalid user behavior data; counting data according to the number of users, and splicing original user behavior data according to the ID, the browse advertisement and the time stamp information of each user; supplementing the counted user behavior data by adopting a multiple interpolation method; constructing a user behavior sequence based on time difference; grouping the user behavior data according to the user ID, and sequencing the user behavior data according to the time sequence to form a user behavior sequence; for each of the series of actions, the difference between the timestamp of the next action and the timestamp of the current action is used as time.
3. The method of claim 1, wherein outputting the user interest representation vector comprises learning a set of static interest group states of the user according to the user behavior sequence using a Time-based Time-GRU and Time-gating loop; namely, according to the input user behavior sequence characteristics, calculating the time gate weight; adding the time gate weight to a first updating strategy of an updating gate; the static interest group state set is selected by a reset gate and an update gate in the time-gated loop unit.
4. The method of claim 3, wherein the step ofThe first update policy is expressed as
Wherein h is t Representing the t-th hidden interest state in the Time-GRU; z t Is an update gate in the Time-GRU;is multiplication element by element, T t Represents the time gate weight, h t-1 Representing the t-1 hidden interest state in the Time-GRU; t (T) t ＝σ(W t [log(Δt+ζ),i t ]) The method comprises the steps of carrying out a first treatment on the surface of the Sigma is a sigmoid activation function, W t A t-th hidden state matrix representing a time gate; Δt represents a time factor, i.e., the difference between the timestamp of the current behavior and the timestamp of the last behavior; ζ represents the time gate bias; i.e t An input vector representing Time-GRU; />Representing the temporary state of the t-th hidden interest state in the Time-GRU.
5. The method of claim 3 or 4, wherein outputting the user's interest update vector comprises calculating an attention score for each interest state and the targeted advertisement based on the static interest group state set; according to the static interest group state set and the attention score, an attention mechanism-based AT-GRU (attention mechanism-based gating circulation unit) is adopted to calculate an interest final update state, namely the attention score is taken as an update gate, the magnitude of the attention score is taken as a second update strategy of the update gate, and the selected interest final update state is selected through the update gate and a reset gate in the time gating circulation unit.
6. The method of claim 5, wherein the second update strategy is expressed as
Wherein h' t Representing a t-th hidden interest state of the AT-GRU; a, a t Represents an attention score;is multiplication by element, h' t-1 Representing the t-1 hidden interest state of the AT-GRU; />Representing the temporary state of the t-th hidden interest state in the AT-GRU.
7. An advertisement click rate prediction apparatus for implementing the advertisement click rate prediction method according to any one of claims 1 to 6, characterized in that the apparatus comprises:
the acquisition module is used for acquiring user behavior data, user portrait data and advertisement data of the electronic commerce platform;
the processing module is used for preprocessing the user behavior data of the e-commerce platform and forming a user behavior sequence;
the embedding module is used for respectively coding and representing the user behavior sequence, the user portrait data and the advertisement data to obtain an embedding vector of the corresponding feature;
the first feature extraction module is used for inputting user behavior sequence features, adopting a Time-GRU (Time-based group) deep neural network based on a Time factor, and outputting interest expression vectors of users;
the second feature extraction module is used for inputting interest expression vectors of users, simulating an interest updating process by adopting an AT-GRU deep neural network based on an attention mechanism, and outputting interest updating vectors of the users;
the third feature extraction module is used for inputting the portrait features and the advertisement features of the user and extracting the recessive relation vectors between the portrait features and the advertisement features of the user by adopting a stack type automatic encoder;
and the advertisement click rate prediction module is used for inputting the interest expression vector of the user and the invisible relation vector between the portrait characteristic of the user and the advertisement characteristic into the multi-layer perceptron for joint training to obtain a prediction result of the advertisement click rate.
8. A computer device comprising a memory and a processor, the memory having stored therein a computer program which, when executed by the processor, causes the processor to perform the steps of the method of any of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111243596.4A CN113888238B (en) | 2021-10-25 | 2021-10-25 | Advertisement click rate prediction method and device and computer equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111243596.4A CN113888238B (en) | 2021-10-25 | 2021-10-25 | Advertisement click rate prediction method and device and computer equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113888238A CN113888238A (en) | 2022-01-04 |
| CN113888238B true CN113888238B (en) | 2024-04-09 |
Family
ID=79014053
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111243596.4A Active CN113888238B (en) | 2021-10-25 | 2021-10-25 | Advertisement click rate prediction method and device and computer equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113888238B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109960759A (en) * | 2019-03-22 | 2019-07-02 | 中山大学 | Recommender system clicking rate prediction technique based on deep neural network |
| CN112307257A (en) * | 2020-11-25 | 2021-02-02 | 中国计量大学 | Short video click rate prediction method based on multi-information node graph network |
| CN112381581A (en) * | 2020-11-17 | 2021-02-19 | 东华理工大学 | Advertisement click rate estimation method based on improved Transformer |
| CN112395505A (en) * | 2020-12-01 | 2021-02-23 | 中国计量大学 | Short video click rate prediction method based on cooperative attention mechanism |
| CN112700274A (en) * | 2020-12-29 | 2021-04-23 | 华南理工大学 | Advertisement click rate estimation method based on user preference |
-
2021
- 2021-10-25 CN CN202111243596.4A patent/CN113888238B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109960759A (en) * | 2019-03-22 | 2019-07-02 | 中山大学 | Recommender system clicking rate prediction technique based on deep neural network |
| CN112381581A (en) * | 2020-11-17 | 2021-02-19 | 东华理工大学 | Advertisement click rate estimation method based on improved Transformer |
| CN112307257A (en) * | 2020-11-25 | 2021-02-02 | 中国计量大学 | Short video click rate prediction method based on multi-information node graph network |
| CN112395505A (en) * | 2020-12-01 | 2021-02-23 | 中国计量大学 | Short video click rate prediction method based on cooperative attention mechanism |
| CN112700274A (en) * | 2020-12-29 | 2021-04-23 | 华南理工大学 | Advertisement click rate estimation method based on user preference |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113888238A (en) | 2022-01-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111177575B (en) | Content recommendation method and device, electronic equipment and storage medium | |
| CN113626719B (en) | Information recommendation method, device, equipment, storage medium and computer program product | |
| CN111339415B (en) | Click rate prediction method and device based on multi-interactive attention network | |
| CN115082147B (en) | Sequence recommendation method and device based on hypergraph neural network | |
| CN110717098B (en) | Meta-path-based context-aware user modeling method and sequence recommendation method | |
| CN111080400B (en) | Commodity recommendation method and system based on gate control graph convolution network and storage medium | |
| CN110955826B (en) | Recommendation system based on improved cyclic neural network unit | |
| CN111241394B (en) | Data processing method, data processing device, computer readable storage medium and electronic equipment | |
| Chiang et al. | Hybrid time-series framework for daily-based PM 2.5 forecasting | |
| Wang et al. | A hierarchical attention model for CTR prediction based on user interest | |
| CN109189922B (en) | Comment evaluation model training method and device | |
| Ibrahim et al. | An intelligent hybrid neural collaborative filtering approach for true recommendations | |
| CN115438787A (en) | Training method and device of behavior prediction system | |
| Khan et al. | Comparative analysis on Facebook post interaction using DNN, ELM and LSTM | |
| CN114529077A (en) | Click rate prediction method based on intra-session heterogeneous behaviors | |
| CN112270570B (en) | Click conversion rate prediction method based on feature combination and representation learning | |
| CN117392686A (en) | Improved dynamic graph neural network-based unrealistic information detection method | |
| CN113888238B (en) | Advertisement click rate prediction method and device and computer equipment | |
| CN116910357A (en) | Data processing method and related device | |
| CN116204723A (en) | Social recommendation method based on dynamic hypergraph representation learning | |
| CN115293812A (en) | E-commerce platform session perception recommendation prediction method based on long-term and short-term interests | |
| CN116089708A (en) | Agricultural knowledge recommendation method and device | |
| CN114519600A (en) | Graph neural network CTR estimation algorithm fusing adjacent node variances | |
| CN114329193B (en) | Click rate prediction method based on time perception interest evolution | |
| TOUHAMI et al. | Session-based Recommendation Systems with Graph ATtention Networks |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |