CN116228368A - Advertisement click rate prediction method based on deep multi-behavior network - Google Patents
Advertisement click rate prediction method based on deep multi-behavior network Download PDFInfo
- Publication number
- CN116228368A CN116228368A CN202310247095.6A CN202310247095A CN116228368A CN 116228368 A CN116228368 A CN 116228368A CN 202310247095 A CN202310247095 A CN 202310247095A CN 116228368 A CN116228368 A CN 116228368A
- Authority
- CN
- China
- Prior art keywords
- behavior
- advertisement
- user
- term
- network
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 45
- 230000004927 fusion Effects 0.000 claims abstract description 50
- 238000012512 characterization method Methods 0.000 claims abstract description 20
- 238000007781 pre-processing Methods 0.000 claims abstract description 13
- 239000013598 vector Substances 0.000 claims description 100
- 230000007774 longterm Effects 0.000 claims description 56
- 230000008569 process Effects 0.000 claims description 21
- 238000012545 processing Methods 0.000 claims description 20
- 238000012549 training Methods 0.000 claims description 15
- 238000010606 normalization Methods 0.000 claims description 12
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 2
- 238000000926 separation method Methods 0.000 claims description 2
- 230000007613 environmental effect Effects 0.000 claims 1
- 230000006399 behavior Effects 0.000 description 134
- 230000006870 function Effects 0.000 description 19
- 239000011159 matrix material Substances 0.000 description 10
- 230000009466 transformation Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 7
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 6
- 238000011156 evaluation Methods 0.000 description 6
- 230000007246 mechanism Effects 0.000 description 6
- 230000004913 activation Effects 0.000 description 5
- 238000013461 design Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000004880 explosion Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 238000013210 evaluation model Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 229940070527 tourmaline Drugs 0.000 description 1
- 229910052613 tourmaline Inorganic materials 0.000 description 1
- 239000011032 tourmaline Substances 0.000 description 1
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
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
-
- 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
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Strategic Management (AREA)
- Physics & Mathematics (AREA)
- Development Economics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Marketing (AREA)
- Economics (AREA)
- General Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (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)
- Information Transfer Between Computers (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to the field of advertisement click rate prediction, and particularly relates to an advertisement click rate prediction method based on a deep multi-behavior network, which comprises the following steps: acquiring corresponding data information of a user and an advertisement, wherein the corresponding data information comprises basic information data of the user and the advertisement, advertisement exposure data and user behavior log data; preprocessing corresponding data information; extracting characteristics of the preprocessed data information, wherein the characteristics comprise sequence characteristics, user characteristics, advertisement characteristics and context characteristics; inputting the data information characteristics into a trained advertisement click rate prediction model based on a depth multi-behavior network to obtain an advertisement click rate prediction result; the invention provides an interest fusion module based on dynamic Dropout, which can capture the difference of user behavior distribution, effectively fuse the user interests and avoid the overfitting of a model to a certain behavior characterization.
Description
Technical Field
The invention belongs to the field of advertisement click rate prediction, and particularly relates to an advertisement click rate prediction method based on a deep multi-behavior network.
Background
At present, society is in an information explosion age, and the commercial products with the full view of the tourmaline often make it difficult for users to choose, and especially when selecting the commercial products on a mobile phone, the commercial products of the heart instrument are selected from a large number of commercial products to be important. Therefore, recommending a commodity to a user through click through rate (Click through rate, abbreviated as CTR) estimation is an important technology.
The existing sequence model obtains better prediction accuracy than the traditional recommendation algorithm after introducing the user behavior into CTR estimation, but the type of the user behavior sequence which is usually introduced by the existing technology is single, only the historical click behavior sequence of the user is usually introduced, and a certain improvement space is provided for the modeling method of the user behavior.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides an advertisement click rate prediction method based on a deep multi-behavior network, which comprises the following steps: acquiring corresponding data information of a user and an advertisement, wherein the corresponding data information comprises basic information data of the user and the advertisement, advertisement exposure data and user behavior log data; preprocessing corresponding data information; extracting characteristics of the preprocessed data information, wherein the characteristics comprise sequence characteristics, user characteristics, advertisement characteristics and context characteristics; inputting the data information characteristics into a trained advertisement click rate prediction model based on a depth multi-behavior network to obtain an advertisement click rate prediction result;
the process of training the advertisement click rate prediction model based on the depth multi-behavior network comprises the following steps:
s1: acquiring historical advertisement click data of a user, and preprocessing the data; wherein the user history advertisement click data comprises user behavior sequence characteristics, advertisement characteristics, user characteristics and environment characteristics;
s2: inputting the user behavior sequence feature, the advertisement feature, the user feature and the environment feature into a feature embedding layer, and generating a user behavior sequence feature vector representation, an advertisement feature vector representation, a user feature vector representation and an environment feature vector representation;
s3: inputting the user behavior sequence feature vector representation into a deep multi-behavior network, and extracting behavior features of a user;
s4: inputting all the user behavior characteristics into a multi-behavior fusion module to obtain user behavior fusion characteristics;
s5: inputting the advertisement feature vector representation, the user feature vector representation and the environment feature vector representation into a deep cross network to obtain advertisement context fusion features;
s6: fusing the user behavior fusion characteristics and the advertisement context fusion characteristics to obtain an advertisement click rate prediction result;
s7: and calculating a loss function of the model according to the advertisement click rate prediction result, optimizing parameters of the model by adopting an Adam optimization algorithm, and completing training of the model when the loss function converges.
Preferably, the user behavior sequence features include a long-term post-click behavior sequence, a long-term click behavior sequence, a short-term click behavior sequence, and a short-term exposure behavior sequence.
Preferably, the deep multi-behavior network comprises a long-term click behavior sequence modeling module, a short-term click behavior sequence modeling module and a short-term exposure news sequence modeling module;
the process of processing the input data by the behavior sequence modeling module after long-term clicking comprises the following steps: after long-term clicking, the behavior sequence and the candidate advertisement feature vector are input into a sparse multi-head attention layer to extract attention features; gold-adding and normalizing the extracted features and the input sequence; inputting the normalized data into a full-connection layer to obtain fusion characteristics; adding and normalizing the fusion features and the input features to obtain a long-term click interest characterization of the user;
the long-term click behavior sequence modeling module processes the input data including: inputting the short-term click behavior sequence into an encoder for encoding processing, and inputting the encoded data combined with the candidate advertisement feature vector into a decoder for decoding to obtain interest characterization after long-term clicking;
the process of processing the input data by the short-term exposure behavior sequence modeling module comprises the following steps: inputting the short-term sequence and the candidate advertisement feature vector into a multi-head attention layer, and carrying out addition normalization processing on the output result of the multi-head attention layer and input data; inputting the normalization processing result into a multi-layer two-dimensional convolution network to obtain short-term exposure interest characterization;
the short-term click behavior sequence modeling module processes input data, including: inputting the behavior sequence after long-term clicking and the candidate advertisement feature vector into a sparse multi-head attention layer for extracting attention features; gold-adding and normalizing the extracted features and the input sequence; inputting the normalized data into a full-connection layer to obtain fusion characteristics; and adding and normalizing the fusion features and the input features to obtain the short-term click interest characterization.
Preferably, the process of fusing the user behavior features by the multi-behavior fusion module includes: the method comprises the steps of respectively inputting a long-term clicking interest representation, a short-term clicking interest representation and a short-term exposure interest representation of a user into four Dropout layers, and fusing type embedded vectors; and inputting the characteristics of the fusion type embedded vector into a full connection layer to generate a final fusion interest vector.
Further, the monotonic function of the Dropout layer is expressed as:
wherein S is the true length of the sequence, θ 1 ,θ 2 To control the monotonicity and slope of the superparameter, p (S) is the resulting Dropout ratio.
The invention has the beneficial effects that:
according to the invention, a plurality of different user behavior sequences are considered to be introduced, an optimized transformation former structure is designed aiming at the different behavior sequences, and a long-term clicking behavior sequence modeling module, a short-term exposure behavior sequence modeling module and a long-term clicking behavior sequence modeling module are respectively designed by introducing different optimized transformation former structures; for the long-term clicking behavior sequence and the long-term clicking behavior sequence, the influence of the sequence length on the model performance is considered, a sparse attention mechanism is introduced, and the performance of the model is improved on the premise of not losing the model effect; for a short-term exposure behavior sequence, taking the fact that the proportion of noise in exposure data is large into consideration, denoising is carried out by introducing a two-dimensional convolution network de-exposure behavior sequence representation, and the effect of sequence modeling is further improved; the invention provides an interest fusion module based on dynamic Dropout, which can capture the difference of user behavior distribution, effectively fuse the user interests and avoid the overfitting of a model to a certain behavior characterization.
Drawings
FIG. 1 is a flow chart of an advertisement click rate prediction method based on a deep multi-behavior network according to the present invention;
FIG. 2 is a block diagram of the overall system framework of the present invention;
FIG. 3 is a block diagram of an input/output module of the present invention;
FIG. 4 is a block diagram of a long-term post-click behavior sequence modeling of the present invention;
FIG. 5 is a block diagram of a long-term click behavior sequence modeling of the present invention;
FIG. 6 is a block diagram of a short-term exposure behavior sequence modeling of the present invention;
FIG. 7 is a block diagram of a short-term click behavior sequence modeling of the present invention;
FIG. 8 is a block diagram of a multi-row fusion module of 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.
An advertisement click rate prediction method based on a deep multi-behavior network, as shown in fig. 1, comprises the following steps: acquiring corresponding data information of a user and an advertisement, wherein the corresponding data information comprises basic information data of the user and the advertisement, advertisement exposure data and user behavior log data; preprocessing corresponding data information; extracting characteristics of the preprocessed data information, wherein the characteristics comprise sequence characteristics, user characteristics, advertisement characteristics and context characteristics; and inputting the data information characteristics into the trained advertisement click rate prediction model based on the depth multi-behavior network to obtain an advertisement click rate prediction result.
The process of training the advertisement click rate prediction model based on the depth multi-behavior network comprises the following steps:
s1: acquiring historical advertisement click data of a user, and preprocessing the data; wherein the user history advertisement click data comprises user behavior sequence characteristics, advertisement characteristics, user characteristics and environment characteristics;
s2: inputting the user behavior sequence feature, the advertisement feature, the user feature and the environment feature into a feature embedding layer, and generating a user behavior sequence feature vector representation, an advertisement feature vector representation, a user feature vector representation and an environment feature vector representation;
s3: inputting the user behavior sequence feature vector representation into a deep multi-behavior network, and extracting behavior features of a user;
s4: inputting all the user behavior characteristics into a multi-behavior fusion module to obtain user behavior fusion characteristics;
s5: inputting the advertisement feature vector representation, the user feature vector representation and the environment feature vector representation into a deep cross network to obtain advertisement context fusion features;
s6: fusing the user behavior fusion characteristics and the advertisement context fusion characteristics to obtain an advertisement click rate prediction result;
s7: and calculating a loss function of the model according to the advertisement click rate prediction result, optimizing parameters of the model by adopting an Adam optimization algorithm, and completing training of the model when the loss function converges.
An embodiment of an advertisement click rate prediction method based on a deep multi-behavior network comprises the following steps:
and step 1, acquiring basic information data of users and advertisements, advertisement exposure data and user behavior log data. Preprocessing the data, and extracting the historical behavior sequence characteristics, the user characteristics, the advertisement characteristics and the context characteristics of the user from the data.
And 2, inputting the characteristics into an advertisement click rate prediction model based on a depth multi-behavior network, and obtaining a final advertisement click rate prediction result through calculation of the model.
Preprocessing data includes: the post click behavior in the data is changed to < browse, buy, like, buy, etc. <1,2,3,4 according to the mapping. The continuous type characteristic in the data is changed into discrete type characteristic through the barrel separation process.
In this embodiment, preprocessing of the data set is mainly divided into two steps: first, user behavior sequence features are spliced. The advertisement exposure click sample and the user behavior record in the user behavior log are single behavior records with time stamps, and the model provided by the invention requires that the user behavior history is input into the model in a sequence form. Therefore, for the advertisement exposure click sample and the data in the user behavior log, the data are aggregated according to the user ID and the behavior type, and for a plurality of behavior records of each behavior type of each user, the behavior records are ordered from small to large according to the time stamp, and are organized into an ordered sequence. Based on the characteristics of the present dataset, the historical behavior within 7 days was considered short-term behavior, and no restrictions other than the total length of the sequence were imposed on long-term behavior, so that limited training data was utilized as fully as possible. For each sample in the advertisement exposure click samples, the text spells up four behavior sequences of long-term clicking, short-term exposure and post-click (including collection, shopping cart joining and purchase) of the corresponding user. Of course, the behavior sequence is truncated according to the timestamp of the current sample during splicing, so that information leakage caused by the fact that future behavior information is introduced into the sample is prevented.
In order to fully utilize the user behavior sequence data, the data set preprocessing link also divides the training set and the testing set, wherein the date is used as the basis for dividing the training set and the testing set, and the first 6 days are divided into the training set and the last 1 day are divided into the testing set in all 7 days of data.
The user behavior sequence features include a long-term post-click behavior sequence, a long-term click behavior sequence, a short-term click behavior sequence, and a short-term exposure behavior sequence.
The four sequence formats were designed as follows: a, a i Representing the advertisement interacted with by the ith user action.
Long-term click behavior sequence and short-term exposure behavior sequence: [ a ] 1 ,a 2 ,...a L ]Where L represents the upper limit of the behavioral sequence length.
Short-term click behavior sequence: [ (a) 11 ,a 12 ,...a 1M ),(a 21 ,a 22 ,...a 2M ),...(a N1 ,a N2 ,...a NM )]Where N represents the upper limit of the number of sessions and M represents the upper limit of the number of behaviors within a single session.
Behavior sequence after long-term clicking: [ < a ] 1 ,t 1 ＞,＜a 2 ,t 2 ＞,...＜a L ,t L ＞]In t i The behavior category (browse, buy, like, buy, etc.) of the advertisement to which the ith user behavior is to interact is represented.
As shown in fig. 2, the advertisement click rate prediction model based on the deep multi-behavior network includes: the system comprises an input/output module, a feature embedding module, a multi-behavior processing module, a multi-behavior fusion module and a deep cross network module.
As shown in fig. 3, the input of the model: firstly, the user behavior sequence features, wherein the user behavior sequence comprises four kinds of behavior sequences after long-term clicking, short-term clicking and short-term exposure in the DMBN. Other features including advertisement features, user features, context features, etc.
The output layer of the model receives the output of the underlying network and converts it into a CTR estimate. The output layer of the DMBN model is a fully connected layer with the output size of 1 by taking a Sigmoid function as an activation function.
In this embodiment, the deep multi-behavior network includes a long-term post-click behavior sequence modeling module, a long-term click behavior sequence modeling module, a short-term click behavior sequence modeling module, and a short-term exposure news individual sequence modeling module.
As shown in fig. 4, the process of the long-term click behavior sequence modeling module for processing input data includes: after long-term clicking, the behavior sequence and the candidate advertisement feature vector are input into a sparse multi-head attention layer to extract attention features; gold-adding and normalizing the extracted features and the input sequence; inputting the normalized data into a full-connection layer to obtain fusion characteristics; and adding and normalizing the fusion features and the input features to obtain the long-term click interest characterization of the user.
Specifically, let the long-term click advertisement vector sequence be
The position embedding vector sequence is
Wherein the maximum length of the L user behavior sequences, the elements in both sequences, i.e. the vector E dimensions, are identical. The two are added by position and the resulting sequence is input into the upper network. Let the user behavior matrix after the above conversion be s= [ x ] 1 ,x 2 ,...x L ]Wherein the commodity to be estimated is t +.>
x i ∈R D ,t∈R D The dimension of the advertisement vector to be estimated is D. />
The sequence S and the advertisement t to be input are input into a sparse multi-head attention layer, then an addition and normalization layer is input, and finally the full connection layer is entered. The expression is as follows:
O A ＝Concat(head 1 ,...head H )W O
in the middle of
The trainable parameter matrix representing random initialization, K is hidden vector dimension, H is attention head number, both are super parameters, head i For the output of the ith attention head, O A ∈R D The result of the transformation is input to the sum and normalization layer, which is as follows:
O AN ＝LayerNorm(S+O A )
then, the full connection layer is reached, the layer is essentially a multi-layer full connection network, the input vector is transformed and then output to another summation and normalization layer, and the operation can be formally described as follows:
O LC ＝LayerNorm(O AN +FC(O AN ))
where FC represents a fully connected network transformation.
Because the clicking behaviors of the user exist in the long-term clicking behavior sequence of the user for the last half year or even one year, the behavior sequence is long, and the traditional multi-head attention mechanism is not easy to process, although the traditional multi-head attention mechanism is efficient, the length of the sequence is thousands or even tens of thousands. Thus, the multi-headed attention module in the long-term click behavior sequence is changed to a sparse multi-headed attention module herein.
As shown in fig. 5, the long-term click behavior sequence modeling module processes input data including: and (3) inputting the short-term click behavior sequence into an encoder for encoding processing, and inputting the encoded data combined with the candidate advertisement feature vector into a decoder for decoding to obtain the interest characterization after long-term clicking.
Specifically, the short-term clicking behavior is mainly characterized by sparse data, low noise, high matching degree of the reflected interests and the current mind of the user, and great influence on final prediction. In other words, the modeling of the model has a high lower limit, and a good effect can be obtained in a simple way such as pooling. On the other hand, the excessive dependence on the recent clicking behavior of the user is often an important factor causing a feedback circulation phenomenon, and merely recommending results with high similarity with the recent clicking advertisements of the user easily causes the user to generate the feeling of' what is pushed by what is more, so that the user experience is influenced, the evaluation and interest of the user on the recommended results are reduced, on the other hand, the advertisement side Martai effect is continuously increased, a large number of advertisements cannot be displayed, and the recommendation system deviates from the original purpose of relieving information overload.
Similar to the long-term click behavior modeling module, the short-term click advertisement vector sequence is first added with a position embedding vector before entering the upper-level network. For the decoder section, the Masked Multi-Head Attention (mask) section in the original converter is removed here, and the design shown in the above figure is modified.
As shown in fig. 6, the process of the short-term exposure behavior sequence modeling module processing the input data includes: inputting the short-term sequence and the candidate advertisement feature vector into a multi-head attention layer, and carrying out addition normalization processing on the output result of the multi-head attention layer and input data; and inputting the normalization processing result into a multi-layer two-dimensional convolution network to obtain the short-term exposure interest characterization.
Specifically, the short-term exposure behavior sequence modeling module adopts a scheme of combining a multi-head attention mechanism with a convolutional neural network: the multi-head attention mechanism is responsible for extracting effective information from sequence data and is expressed as a plurality of vectors, and the convolutional neural network is good at extracting multi-level information from a matrix formed by the plurality of vectors.
Although the underlying structure of the short-term exposure behavior modeling module appears similar to the long-term click behavior modeling module, it is slightly different in implementation due to the differences in the overlying structure. Let the short-term exposure advertisement vector sequence be
The position embedding vector sequence is +.>
Where L is the maximum length of the sequence of user actions, the elements in both sequences, i.e. vector E i The dimensions are consistent. The two are added by position and the resulting sequence is input into the upper network. Let the user behavior matrix after the above conversion be s= [ x ] 1 ,x 2 ,...x L ]Wherein->
The advertisement to be estimated is t, and the dimension of the vector of the advertisement to be estimated is D. The sequence S and the advertisement t to be estimated are input into a multi-head attention layer, wherein the specifically performed transformation can be formally described as follows:
O A ＝Concat(head 1 ,...head H )W O
wherein,,
the training parameter matrix representing random initialization, K is hidden vector dimension, H is attention head number, both are super parameters, diag is a function of creating corresponding diagonal matrix according to vector, head i For the output of the ith attention head, O A ∈R (L×D) The result of the transformation is input to the sum and normalization layer. Note that here unlike long-term click behavior sequences, the resulting transformation results are matrices rather than vectors, which facilitate processing using a two-dimensional convolutional neural network. The result obtained after transformation is a matrix, the multi-layer two-dimensional convolutional neural network is transformed, the convolution kernel size of the two-dimensional convolutional neural network and the layer number of the convolutional neural network are both super-parameters, and the method is determined through experimental trial.
As shown in fig. 7, the process of the short-term click behavior sequence modeling module processing the input data includes: inputting the behavior sequence after long-term clicking and the candidate advertisement feature vector into a sparse multi-head attention layer for extracting attention features; gold-adding and normalizing the extracted features and the input sequence; inputting the normalized data into a full-connection layer to obtain fusion characteristics; and adding and normalizing the fusion features and the input features to obtain the short-term click interest characterization.
Specifically, the main characteristics of the behavior after clicking are sparse data, long behavior occurrence time interval and multiple types of behaviors. Therefore, in the modeling mode design, the different behavior types are distinguished mainly by splicing the behaviors of multiple types into a unified sequence and adding a behavior type vector.
Similar to the long-term click behavior modeling module, information related to the advertisement vector to be estimated is extracted from the sequence through a multi-head sparse attention mechanism. The difference is that the vector input to the sparse multi-head attention module by the long-term post-click behavior sequence is the post-click advertisement vector sequence, the post-click advertisement type vector sequence and the position embedding vector are added according to the positions, and the obtained sequence is input to the upper network. Other parts are completely consistent with the long-term click behavior modeling module, and are not described in detail herein.
As shown in fig. 8, the process of fusing the user behavior features by the multi-behavior fusion module includes: the method comprises the steps of respectively inputting a long-term clicking interest representation, a short-term clicking interest representation and a short-term exposure interest representation of a user into four Dropout layers, and fusing type embedded vectors; and inputting the characteristics of the fusion type embedded vector into a full connection layer to generate a final fusion interest vector.
Specifically, after modeling of the behavior sequence, the long-term clicking interest characterization, the short-term clicking interest characterization and the short-term exposure interest characterization of the user can be obtained. Because of the dispersion of user behaviors, the short-term clicking sequence of the user and the short-term exposure sequence of the user can be sparse, and the sequence characterization after the long-term clicking and the long-term clicking is dominant, so that the short-term interest expression of the user is submerged. Based on the observation, the invention provides an interest fusion layer based on dynamic Dropout, and the module can capture the difference of user behavior distribution, effectively fuse the user interests and avoid the overfitting of a model to a certain behavior characterization.
The core idea of Dropout is to randomly discard a proportion of neurons during model training, thereby reducing the risk of model overfitting. It is believed herein that the large variance in user behavior distribution may lead to model overfitting to some type of behavior characterization, and thus an interest fusion layer based on dynamic Dropout was designed. The model inputs the real four lengths as a monotonic function to obtain four Dropout ratios, and the monotonic function is realized as follows:
wherein S is the true length of the sequence, θ 1 ,θ 2 To control the monotonicity and slope of the superparameter, p (S) is the resulting Dropout ratio.
The obtained four Dropout ratios are used as the probability of the real Dropout of the four behavior sequences, and the probability is used for carrying out Dropout on interest vectors of the four behavior sequences, so that the weight of short-term behavior sequence expression and long-term behavior sequence expression is adaptively controlled, the adaptive fusion of the four different types of behavior sequences is realized in a training stage, and the influence of over-fitting on a model is reduced.
In addition, considering that the four modules are relatively independent, the generated user interest vectors may not be in the same semantic space, and the user interest vectors need to be converted first, which can be accomplished by a fully connected layer.
Next, in order to explicitly add the type information of the four interest vectors to the model, referring to the position embedded vectors in the transducer, the interest vectors also need to be added with the type embedded vectors before accessing the full connection layer.
And adding type embedded vectors after connecting different interest vectors, inputting the type embedded vectors into a full-connection layer, generating a final fusion interest vector, connecting the interest vector and the output of a deep cross network into another full-connection layer, and finally calculating to obtain a CTR estimated value.
The depth and crossover network of the present invention generally follow the structure of the master DCN and are optimized in part detail. First, in the original DCN network, the activation functions of the depth network are all ReLU functions, and the depth network part of the invention replaces the Leaky ReLU activation functions. Second, the Batch Normalization and Dropout techniques are not applied in the original DCN, but the present invention adds a BN layer after each layer in the depth network and a Dropout layer after the last layer. In addition, in terms of input features, the invention also selects the features of the depth and cross networks that are accessed by the depth network and the cross network respectively. According to the design thought and application experience of DCN, the depth network is mainly responsible for the generalization capability of the model, the cross network is mainly responsible for the memory capability of the model, meanwhile, due to the characteristics of the cross layer, the output width is equal to the input width, and the triangle structure gradually decreasing from the input layer to the output layer cannot be designed like a full connection layer, so that the characteristics of the depth network and the cross network access should be selected to balance the prediction effect and the resource consumption of the model.
The processing of the advertisement feature vector representation, the user feature vector representation, and the environment feature vector representation using the deep crossover network comprises: the model input also includes other features such as advertisement features, user features, contextual features, and the like. The method mainly comprises three types of numerical value characteristics, single-value category characteristics and multi-value category characteristics. Firstly, numerical characteristics, such as commodity selling price, are typical, the characteristics are real numbers, normalization is needed when data preprocessing is carried out, or the numerical characteristics are divided into barrels according to the numerical values, and are converted into category characteristics, so that the robustness is improved, and certain nonlinear factors are introduced. Second, in the CTR estimation problem, category features take a very important role. Class features refer to features whose values are class codes, the magnitude of which often does not have a numerical meaning, but rather merely represents that they belong to a class. Class features can be further divided into single-valued and multi-valued, as the name implies, single-valued features contain only one value, while multi-valued features may contain multiple values. The most typical single value category characteristics are the user and ad ID characteristics, while the most typical multi-value category characteristics are the user's historical behavior. The multi-value class features are similar in form to the sequence features described above, even though in some data formats both are stored in the same way, the multi-value class features often do not distinguish between the sequence of values within the feature, which is strictly sequential. The CTR pre-estimated scene has a large number of category features, namely user and advertisement ID features, the feature value space is large, single values are less frequently generated in data, if the single values are coded in a simple One-hot mode, feature dimension explosion is easy to cause, and meanwhile, due to the data sparseness problem, the model cannot be fully trained, so that the model effect is influenced.
Therefore, after the original input is processed, all the characteristics are converted into category vectors, and the high-dimensional sparse category characteristics are generally converted into low-dimensional dense real vectors by adopting a characteristic embedding layer for the category characteristics, so that an upper network can be processed efficiently. The implementation principle of the feature embedding layer can be regarded as a table look-up process. For example, for the user ID feature, a matrix is established with a number of rows equal to the total number of user ID values and a number of columns equal to the dimension of the embedded vector, so that each user ID can be mapped to a row in the matrix, i.e., each user ID can be mapped to a corresponding embedded vector. During model training, the matrix is updated through a back propagation algorithm, so that the updating of the Embedding vector along with the model training process is realized.
A LeakyReLU activation function is used in the deep network and the multi-behavior fusion module, and a Swish activation function is adopted in each behavior modeling module.
The loss function expression of the model is:
L total ＝L CE +L AUC
wherein L is CE Is a classical cross entropy loss function, L AUC Is the hinge ordering loss function.
In the present embodiment, the model evaluation criteria selection includes: the CTR estimated model has inconsistency among an online evaluation index, an offline evaluation index and a model loss function. In the invention, the merits of the evaluation model are that an off-line experiment method is adopted, so the section mainly discusses the selection problem of the model evaluation standard. The most commonly used model evaluation criteria in CTR estimation is AUC, the range of AUC is [0,1], and the physical meaning is that any positive sample and one negative sample are taken from the data set, and the probability that the predicted value of the positive sample is larger than that of the negative sample, namely the probability that the ordering is correct. That is, AUC is an indicator of model ordering ability. For the advertisement CTR estimation model, not only the order of the ordering result is required to be correct, but also the absolute value of the CTR estimation is required to be accurate, the ordering capability of the model can only be judged by examining AUC indexes, and if the absolute value of the CTR estimation is to be judged to be accurate, other indexes are also required to be examined.
For a plurality of CTR estimated samples, whether the CTR estimated value is accurate or not is calculated, and Logloss is selected as an index for measuring the accuracy of the CTR estimated absolute value. The calculation formula is as follows:
wherein n is the number of samples, y i The label for the i-th sample,
is the predicted value of the i-th sample.
While the foregoing is directed to embodiments, aspects and advantages of the present invention, other and further details of the invention may be had by the foregoing description, it will be understood that the foregoing embodiments are merely exemplary of the invention, and that any changes, substitutions, alterations, etc. which may be made herein without departing from the spirit and principles of the invention.
Claims (8)
1. The advertisement click rate prediction method based on the deep multi-behavior network is characterized by comprising the following steps of: acquiring corresponding data information of a user and an advertisement, wherein the corresponding data information comprises basic information data of the user and the advertisement, advertisement exposure data and user behavior log data; preprocessing corresponding data information; extracting characteristics of the preprocessed data information, wherein the characteristics comprise sequence characteristics, user characteristics, advertisement characteristics and context characteristics; inputting the data information characteristics into a trained advertisement click rate prediction model based on a depth multi-behavior network to obtain an advertisement click rate prediction result;
the process of training the advertisement click rate prediction model based on the depth multi-behavior network comprises the following steps:
s1: acquiring historical advertisement click data of a user, and preprocessing the data; wherein the user history advertisement click data comprises user behavior sequence characteristics, advertisement characteristics, user characteristics and environment characteristics;
s2: inputting the user behavior sequence feature, the advertisement feature, the user feature and the environment feature into a feature embedding layer, and generating a user behavior sequence feature vector representation, an advertisement feature vector representation, a user feature vector representation and an environment feature vector representation;
s3: inputting the user behavior sequence feature vector representation into a deep multi-behavior network, and extracting behavior features of a user;
s4: inputting various user behavior characteristics of the output of the depth multi-behavior network into a multi-behavior fusion module to obtain user behavior fusion characteristics;
s5: inputting the advertisement feature vector representation, the user feature vector representation and the environment feature vector representation into a deep cross network to obtain advertisement context fusion features;
s6: fusing the user behavior fusion characteristics and the advertisement context fusion characteristics, and inputting the fused user behavior fusion characteristics and the fused advertisement context fusion characteristics into a full-connection layer to obtain an advertisement click rate prediction result;
s7: and calculating a loss function of the model according to the advertisement click rate prediction result, optimizing parameters of the model by adopting an Adam optimization algorithm, and completing training of the model when the loss function converges.
2. The method of claim 1, wherein the user behavior sequence features include a long-term post-click behavior sequence, a long-term click behavior sequence, a short-term click behavior sequence, and a short-term exposure behavior sequence.
3. The method for predicting advertisement click rate based on deep multi-behavior network of claim 1, wherein preprocessing the data comprises: changing post click behavior in the data to < browse, buy, like, buy, … > <1,2,3,4 according to the mapping; the continuous type characteristic in the data is changed into discrete type characteristic through the barrel separation process.
4. The method for predicting advertisement click rate based on deep multi-behavior network according to claim 1, wherein the deep multi-behavior network comprises a long-term post-click behavior sequence modeling module, a long-term click behavior sequence modeling module, a short-term click behavior sequence modeling module and a short-term exposure news sequence modeling module;
the process of processing the input data by the behavior sequence modeling module after long-term clicking comprises the following steps: after long-term clicking, the behavior sequence and the candidate advertisement feature vector are input into a sparse multi-head attention layer to extract attention features; gold-adding and normalizing the extracted features and the input sequence; inputting the normalized data into a full-connection layer to obtain fusion characteristics; adding and normalizing the fusion features and the input features to obtain a long-term click interest characterization of the user;
the long-term click behavior sequence modeling module processes the input data including: inputting the short-term click behavior sequence into an encoder for encoding processing, and inputting the encoded data combined with the candidate advertisement feature vector into a decoder for decoding to obtain interest characterization after long-term clicking;
the process of processing the input data by the short-term exposure behavior sequence modeling module comprises the following steps: inputting the short-term sequence and the candidate advertisement feature vector into a multi-head attention layer, and carrying out addition normalization processing on the output result of the multi-head attention layer and input data; inputting the normalization processing result into a multi-layer two-dimensional convolution network to obtain short-term exposure interest characterization;
the short-term click behavior sequence modeling module processes input data, including: inputting the behavior sequence after long-term clicking and the candidate advertisement feature vector into a sparse multi-head attention layer for extracting attention features; gold-adding and normalizing the extracted features and the input sequence; inputting the normalized data into a full-connection layer to obtain fusion characteristics; and adding and normalizing the fusion features and the input features to obtain the short-term click interest characterization.
5. The method for predicting advertisement click rate based on deep multi-behavior network of claim 1, wherein the process of fusing the user behavior features by the multi-behavior fusion module comprises: the method comprises the steps of respectively inputting a long-term clicking interest representation, a short-term clicking interest representation and a short-term exposure interest representation of a user into four Dropout layers, and fusing type embedded vectors; and inputting the characteristics of the fusion type embedded vector into a full connection layer to generate a final fusion interest vector.
6. The method for predicting advertisement click-through rate based on deep multi-behavior network of claim 5, wherein the monotonic function of Dropout layer is expressed as:
wherein S is the true length of the sequence, θ 1 ,θ 2 To control the monotonicity and slope of the superparameter, p (S) is the resulting Dropout ratio.
7. The method of claim 1, wherein processing the advertisement feature vector representation, the user feature vector representation, and the environmental feature vector representation using the deep cross network comprises: the deep crossover network includes a deep network and a crossover network; wherein the depth network consists of a plurality of full-connection layers, and the cross network consists of a plurality of cross layers; the full connection layer is used for acquiring deep characteristic information of advertisement characteristic vector representation, user characteristic vector representation and environment characteristic vector representation, and the cross layer is used for acquiring cross characteristic information of advertisement characteristic vector representation, user characteristic vector representation and environment characteristic vector representation; and splicing the feature information extracted by the depth network, the cross feature information extracted by the cross network and the fusion interest vector output by the multi-type interest vector fusion module, and inputting the spliced feature information and the fusion interest vector into the full-connection layer to obtain a final result.
8. The method for predicting advertisement click rate based on deep multiple behavior network of claim 1, wherein the model has a loss function expression of:
wherein y is a true value,
n is the number of samples, which is the predicted probability value. />
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310247095.6A CN116228368A (en) | 2023-03-15 | 2023-03-15 | Advertisement click rate prediction method based on deep multi-behavior network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310247095.6A CN116228368A (en) | 2023-03-15 | 2023-03-15 | Advertisement click rate prediction method based on deep multi-behavior network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116228368A true CN116228368A (en) | 2023-06-06 |
Family
ID=86571223
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310247095.6A Pending CN116228368A (en) | 2023-03-15 | 2023-03-15 | Advertisement click rate prediction method based on deep multi-behavior network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116228368A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116757748A (en) * | 2023-08-14 | 2023-09-15 | 广州钛动科技股份有限公司 | Advertisement click prediction method based on random gradient attack |
| CN117522479A (en) * | 2023-11-07 | 2024-02-06 | 北京创信合科技有限公司 | Accurate Internet advertisement delivery method and system |
-
2023
- 2023-03-15 CN CN202310247095.6A patent/CN116228368A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116757748A (en) * | 2023-08-14 | 2023-09-15 | 广州钛动科技股份有限公司 | Advertisement click prediction method based on random gradient attack |
| CN116757748B (en) * | 2023-08-14 | 2023-12-19 | 广州钛动科技股份有限公司 | Advertisement click prediction method based on random gradient attack |
| CN117522479A (en) * | 2023-11-07 | 2024-02-06 | 北京创信合科技有限公司 | Accurate Internet advertisement delivery method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110196946B (en) | Personalized recommendation method based on deep learning | |
| CN112214685B (en) | Knowledge graph-based personalized recommendation method | |
| CN111127142B (en) | Article recommendation method based on generalized nerve attention | |
| CN110781409B (en) | Article recommendation method based on collaborative filtering | |
| CN111737578B (en) | Recommendation method and system | |
| CN116228368A (en) | Advertisement click rate prediction method based on deep multi-behavior network | |
| Wen et al. | Neural attention model for recommendation based on factorization machines | |
| CN109785062A (en) | A kind of hybrid neural networks recommender system based on collaborative filtering model | |
| CN114693397B (en) | Attention neural network-based multi-view multi-mode commodity recommendation method | |
| CN111127146A (en) | Information recommendation method and system based on convolutional neural network and noise reduction self-encoder | |
| CN112800344B (en) | Deep neural network-based movie recommendation method | |
| CN113505307B (en) | Social network user region identification method based on weak supervision enhancement | |
| CN118132856A (en) | Intelligent analysis method and system based on big data | |
| Gong | Deep Belief Network‐Based Multifeature Fusion Music Classification Algorithm and Simulation | |
| CN115374288A (en) | Recommendation method based on knowledge graph and attention mechanism | |
| CN111428181A (en) | Bank financing product recommendation method based on generalized additive model and matrix decomposition | |
| CN114781503A (en) | Click rate estimation method based on depth feature fusion | |
| CN117194771B (en) | Dynamic knowledge graph service recommendation method for graph model characterization learning | |
| CN118246939A (en) | Method and system for processing DeepFM model browsing data based on attention mechanism | |
| CN114996566A (en) | Intelligent recommendation system and method for industrial internet platform | |
| Qi et al. | Deep Field‐Aware Interaction Machine for Click‐Through Rate Prediction | |
| CN114519600A (en) | Graph neural network CTR estimation algorithm fusing adjacent node variances | |
| Sun et al. | DeepPRFM: Pairwise Ranking Factorization Machine Based on Deep Neural Network Enhancement | |
| CN118400423B (en) | Intelligent recommendation system and method based on knowledge graph | |
| CN117575744B (en) | Article recommendation method and system based on user association relation |
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 |