CN111369278A - Click rate prediction method based on long-term interest modeling of user - Google Patents

Click rate prediction method based on long-term interest modeling of user Download PDF

Info

Publication number
CN111369278A
CN111369278A CN202010103305.0A CN202010103305A CN111369278A CN 111369278 A CN111369278 A CN 111369278A CN 202010103305 A CN202010103305 A CN 202010103305A CN 111369278 A CN111369278 A CN 111369278A
Authority
CN
China
Prior art keywords
user
interest
term
short
click
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202010103305.0A
Other languages
Chinese (zh)
Inventor
汤景凡
张秀杰
张旻
姜明
黄涛
吴鑫强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Dianzi University
Original Assignee
Hangzhou Dianzi University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Dianzi University filed Critical Hangzhou Dianzi University
Priority to CN202010103305.0A priority Critical patent/CN111369278A/en
Publication of CN111369278A publication Critical patent/CN111369278A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0202—Market predictions or forecasting for commercial activities
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions

Abstract

The invention discloses a click rate prediction method based on long-term interest modeling of a user. The invention comprises the following steps: step (1) dividing a user behavior sequence into different sessions, and then extracting short-term interest and hobbies of a user by using a self-attention mechanism module; step (2) extracting the long-term interest and interest deviation of the user through a multi-channel memory network; step (3), the interest activation unit associates the short-term interest and hobby of the user, the long-term interest and hobby deviation characteristics of the user with the target object; step (4) linking the three characteristics processed in the step (3) and inputting the three characteristics into a multiple sensing machine to predict the click probability; and (5) calculating a negative log-likelihood function for the output probability, wherein the smaller the negative log-likelihood function value is, the better the corresponding effect of the method is. The method solves the problems of low interpretability and low accuracy in click rate prediction, and can obtain better effect and interpretability.

Description

Click rate prediction method based on long-term interest modeling of user
Technical Field
The invention relates to the field of click rate prediction in a design recommendation system, in particular to a method for realizing click rate prediction by a multi-channel memory network and a self-attention mechanism
Background
Predicting the click rate of the user for purchasing the commodity, and the key for solving the problems is how to effectively process the serialized information. Currently, there are three types of methods for processing sequence information
First, the serialized information is processed by a simple LSTM model or a CNN model, and we know that the LSTM finally outputs a hidden vector, so that the interpretability is not strong. LSTM, on the other hand, does not distinguish the role of each feature well when processing a sequence of features. This is also why we use a memory network to accurately process serialized information. As for the CNN model, we can find that CNN cannot handle the long interval association problem of features in the sequence well. This may result in failure to extract the variation feature of the user's taste.
The second method is to extract the characteristic information in the sequence through a memory network. The method can well measure the action of each clicked item in the sequence and can provide good interpretability, but the problem of session of the user clicking the characteristic sequence is not well considered in the memory network. It may not be possible to effectively predict the click probability of other items in the same session.
The third method extracts the deviations of the long-term and interest interests of the user in the serialized information by using a multi-channel memory network. The memory network has good interpretability, and can measure different functions of each article through the Attention. The migration change characteristics of the interest and the hobbies of the user on the same type of goods are mined through multiple channels, and each channel represents the migration change of the interest of the user on a certain type of goods. The user historical behavior sequence is divided into different sessions according to time intervals, and then the user short-term interest is represented by the latest session sequence. Due to the consistency of the user behavior sequence, the users click on similar items in the same time interval. Therefore, the method of selecting the latest k behavior sequences as the latest interest expression of the user can cause the items in the sequences to belong to a plurality of different sessions, so that the short-term interest and hobbies of the user cannot be effectively expressed.
Disclosure of Invention
The invention aims to provide a method for predicting long-term and short-term click rate of a user by combining a self-attention mechanism and a multi-channel memory network aiming at the defects of the prior art so as to solve the problems of poor interpretability and low accuracy in click rate prediction.
The technical scheme adopted by the invention for solving the technical problem comprises the following steps:
step (1) dividing a user behavior sequence into different sessions, and then extracting short-term interest and hobbies of a user by using a self-attention mechanism module;
step (2) extracting the long-term interest and interest deviation of the user through a multi-channel memory network;
step (3), the interest activation unit associates the short-term interest and hobby of the user, the long-term interest and hobby deviation characteristics of the user with the target object;
step (4) linking the three characteristics processed in the step (3) and inputting the three characteristics into a multiple sensing machine to predict the click probability;
and (5) calculating a negative log-likelihood function for the output probability, wherein the smaller the negative log-likelihood function value is, the better the corresponding effect of the method is.
The session division in the step (1) is specifically realized by the following steps:
the 1-1 user click sequence has the characteristics of timeliness and continuity, and the user click sequences are divided into different sessions according to time intervals, such as clothes or food. Dividing a user click sequence into different sessions according to a time interval, specifically, if the click time interval of two adjacent articles in the sequence exceeds a threshold (30 minutes is set in an experiment), dividing the two adjacent articles into different session intervals, otherwise, dividing the two adjacent articles into the same session; the short-term interest of the user has the characteristic of timeliness, so that the short-term interest of the user can be represented only by using the click sequence in the last session interval; compared with methods for selecting the latest k point hit item sequences as short-term interests of the user, the session can more reasonably consider the characteristics of timeliness and continuity of the user hit sequences.
1-2, embedding and converting sparse characteristics of user, commodity and user behavior sequence into low-dimensional dense characteristic representation, and representing all user information into matrix
Figure BDA0002387593690000021
Wherein N isuRepresenting the number of users; all article information can be expressed as
Figure BDA0002387593690000031
Wherein N isiIndicating the number of the articles; the user behavior sequence is expressed as
Figure BDA0002387593690000032
N represents the number of commodities clicked by the user's historical behavior, biThe representation represents the ith behavior embedding vector in the corresponding session.
In step 1, a self-attention module is constructed to extract short-term interests and hobbies of the user in the latest session, and the method is specifically realized as follows:
dividing a user behavior sequence S into different sessions R according to time intervals, wherein the latest session sequence of a user u at the time t is expressed as
Figure BDA0002387593690000033
Figure BDA0002387593690000034
Figure BDA0002387593690000035
Representing the number of times of the user u click behavior in the latest session sequence at the time t, biRepresenting the i-th behavior embedding vector in the corresponding session, dmodelRepresenting item embedding dimensions in per click behavior；
First, a query is generated using non-linear variation of a shared parameter
Figure BDA0002387593690000036
Figure BDA0002387593690000037
Keys (keys) and
Figure BDA0002387593690000038
the value (value) maps to the same space;
Figure BDA0002387593690000039
Figure BDA00023875936900000310
Figure BDA00023875936900000311
wherein the content of the first and second substances,
Figure BDA00023875936900000312
is the corresponding mapping weight matrix; relu is used as an activation function that learns nonlinear interactions in features;
Figure BDA00023875936900000313
respectively representing a query (query), a key (key) and a value (value) matrix after mapping;
the correlation matrix is then calculated as follows:
Figure BDA00023875936900000314
Figure BDA00023875936900000315
wherein one is output
Figure BDA00023875936900000316
Attention moment diagram of
Figure BDA00023875936900000317
It represents
Figure BDA00023875936900000318
Similarity between individual items; wherein
Figure BDA00023875936900000319
For the attention value after the zoom point multiplication,
Figure BDA00023875936900000320
is a matrix representation of the user's short-term preferences.
And (3) extracting the long-term interest and hobby deviation conditions of the user by the multi-channel memory network in the step (2). We first describe the memory network used in the model and then the processing of the multiple channels.
Extracting the long-term interest and the interest deviation of the user through the multi-channel memory network in the step (2), which is specifically realized as follows:
define the memory matrix of user u as
Figure BDA0002387593690000041
Wherein
Figure BDA0002387593690000042
Expressing preference expression vectors of corresponding characteristics of the historical item records by the user u; the memory matrix is used for storing long-term preference of users, and the user memory matrix needs to be read and updated every time new data is added.
The operation of reading and updating the user memory matrix is as follows:
2-1. read operation in memory network
Figure BDA0002387593690000043
Figure BDA0002387593690000044
Wherein α is an enhancement parameter, | | | | represents a modulo operation, btThe representation represents the t-th behavior embedding vector in the corresponding session,
Figure BDA0002387593690000045
representing the weight values without regularization,
Figure BDA0002387593690000046
representing weight values after regularization;
then the weight sum of the vectors in the memory matrix is used as the final output
Figure BDA0002387593690000047
Figure BDA0002387593690000048
2-2. write operation of memory network
eraset＝δ(ETbt+be)
Figure BDA0002387593690000049
Where δ (.) denotes sigmoid activation function, ⊙ denotes dot-by-element multiplication operation, E and beIs the erasure parameter to learn;
Figure BDA00023875936900000410
an erasure vector representing the characteristic information; btRepresenting the t-th behavior embedding vector in the corresponding session;
Figure BDA00023875936900000411
wherein A and baIs a learnable parameter, tanh represents an activation function; such erasing, adding, or further erasingThe new strategy allows for forgetting and enhancing user interest during learning, and the model can automatically determine signals to be attenuated and enhanced using update and erase operations;
2-3, adding multiple channels to enhance the mining of the memory network on the characteristics of the bias of the user interests and hobbies; using additional interest offset matrices
Figure BDA0002387593690000051
To store the bias change of interest of the user u in the same type of item in the sequence t, firstly
Figure BDA0002387593690000052
The method comprises the following steps that m unit slots are included, wherein each unit slot is a channel which represents a characteristic channel for the migration of interest and hobbies of a user on a certain type of articles; at the time t of the sequence of user u, from the set
Figure BDA0002387593690000053
Wherein k feature indexes are selected, wherein
Figure BDA0002387593690000054
Is a weight vector in the read operation of the memory network, and updates the corresponding interest and hobby offset channel for the index number i selected arbitrarily as described in the following formula
Figure BDA0002387593690000055
Figure BDA0002387593690000056
Wherein
Figure BDA0002387593690000057
Is the ith cell slot of the user memory matrix, btA behavior embedding vector; multiple channel unit utilizes the behavior sequence information and the interest shift information of the same channel at the previous time
Figure BDA0002387593690000058
And memory momentsArray corresponding cell slot information
Figure BDA0002387593690000059
As input, then using GRU to further mine and update the change situation of user interest and hobbies under the corresponding channel; since the parameters of the GRU are shared for multiple channels, many training parameters are not added.
The interest activation unit in step (3) associates the short-term interest and hobby of the user, the long-term interest and hobby deviation characteristics of the user with the target item, and the specific implementation is as follows:
3-1, when the click rate is predicted, if the target object accords with the short-term interest of the user, the short-term interest of the user has great influence on the click event; otherwise, the short-term interest and hobbies of the user do not interfere with the click event of the user; therefore, it is stated that the short-term interest needs to be associated with the corresponding target object to effectively extract the final short-term interest expression vector of the user
Figure BDA00023875936900000510
Dynamically distributing the weight of each unit vector in the short-term preference matrix by using an interest activation unit; the specific formula for calculating the weight by using the interest activation unit is as follows:
Figure BDA00023875936900000511
wherein the content of the first and second substances,
Figure BDA00023875936900000512
a learnable interaction matrix representing the association target item and the user's short-term interest, d represents the final vector dimension of the item after passing through the embedding layer,
Figure BDA00023875936900000513
an embedded vector representation representing the item v,
Figure BDA0002387593690000061
indicating the short-term interest of user u at time tInterest and hobby matrix
Figure BDA0002387593690000062
The ith row vector of (1); presentation of results after user long-term preference associated with target item
Figure BDA0002387593690000063
The calculation is as follows:
Figure BDA0002387593690000064
wherein the content of the first and second substances,
Figure BDA0002387593690000065
a learnable interaction matrix representing long-term interests and interests of the associated user and the target item,
Figure BDA0002387593690000066
representing the long-term interest preference matrix of user u at time t
Figure BDA0002387593690000067
The ith row vector of (1); representation vector after user interest migration change feature is associated with target item
Figure BDA0002387593690000068
As described by the following equation:
Figure BDA0002387593690000069
wherein the content of the first and second substances,
Figure BDA00023875936900000610
a learnable parameter matrix representing associated target items and user interest migration changes,
Figure BDA00023875936900000611
an interest migration matrix representing user u at time t
Figure BDA00023875936900000612
The ith row vector of (1); user information, article information, Us、UlAnd UdThe embedded vectors of (a) are concatenated, flattened and then input to a multi-layered perceptron to implement the score prediction.
Step (4) connecting the long-term hobby, short-term hobby and hobby deviation characteristics of the user, and then sending the connected characteristic vector into DNN for training to obtain a final click probability p (x); the corresponding loss function is as follows:
Figure BDA00023875936900000613
where x represents all the feature vector connections, D represents the training set, N represents the number of training sets, y ∈ {0,1} represents whether the user will click on the item, and p (.) is the final predicted result value of the network model, which represents the probability that the user will click on the item.
The invention has the following beneficial effects:
first, the user's short-term interests are time-efficient, while the user's click sequences are consistent. This means that the user clicks on items belonging to the same category, such as clothing or cake, in a sequence over a period of time. The traditional method directly selects the latest K serial items as the user's short-term interest and well expresses, and neglects the timeliness and coherence of the user's short-term interest.
Secondly, the inherent interests and hobbies of the user are relatively fixed, and meanwhile, the interests and hobbies of the user on the same type of articles can also be migrated and changed along with the change of time, and the long-term interests and hobbies deviation of the user are mined by utilizing a multi-channel memory network. The traditional long-term interest hobbies neglect the migration change characteristics of the user interests, so that the traditional long-term interest hobbies do not have good effect.
And finally, when the click rate is predicted, if the target object accords with the short-term hobbies of the user, the short-term hobbies of the user have great influence on the click event. Otherwise, the user's short-term hobbies have no effect on the user's click event. This means that the user interests are dynamic and related to the target item to be predicted. The activation unit is used for associating the interest and the hobbies of the user with the target commodity so as to generate a dynamic representation of the hobbies of the user.
The invention utilizes the self-attention mechanism to extract the short-term interest and preference of the user, and simultaneously uses the multi-channel memory network to extract the long-term interest and the interest deviation condition of the user.
Drawings
FIG. 1 is a diagram of a model of the present invention;
FIG. 2 is a schematic diagram of the short term preference extraction from attention module according to the present invention.
Detailed Description
The attached drawings disclose a flow chart of the preferred embodiment of the invention in a non-limiting way; the technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
As shown in fig. 1, a click-through rate prediction method based on long-term interest modeling of a user includes the following basic steps:
step (1) dividing a user behavior sequence into different sessions, and then extracting short-term interest and hobbies of a user by using a self-attention mechanism module;
step (2) extracting the long-term interest and interest deviation of the user through a multi-channel memory network;
step (3), the interest activation unit associates the short-term interest and hobby of the user, the long-term interest and hobby deviation characteristics of the user with the target object;
step (4) linking the three characteristics processed in the step (3) and inputting the three characteristics into a multiple sensing machine to predict the click probability; short-term interest and hobbies excavated in the self-attention mechanism module, long-term interest and hobbies excavated in the multi-channel memory network are connected and offset characteristics are input into DNN to realize click rate prediction;
and (5) calculating a negative log-likelihood function for the output probability, wherein the smaller the negative log-likelihood function value is, the better the corresponding effect of the method is.
The session division in the step (1) is specifically realized by the following steps:
the 1-1 user click sequence has the characteristics of timeliness and continuity, and the user click sequences are divided into different sessions according to time intervals, such as clothes or food. Dividing a user click sequence into different sessions according to a time interval, specifically, if the click time interval of two adjacent articles in the sequence exceeds a threshold (30 minutes is set in an experiment), dividing the two adjacent articles into different session intervals, otherwise, dividing the two adjacent articles into the same session; the short-term interest of the user has the characteristic of timeliness, so that the short-term interest of the user can be represented only by using the click sequence in the last session interval; compared with methods for selecting the latest k point hit item sequences as short-term interests of the user, the session can more reasonably consider the characteristics of timeliness and continuity of the user hit sequences.
1-2, embedding and converting sparse characteristics of user, commodity and user behavior sequence into low-dimensional dense characteristic representation, and representing all user information into matrix
Figure BDA0002387593690000081
Wherein N isuRepresenting the number of users; all article information can be expressed as
Figure BDA0002387593690000082
Wherein N isiIndicating the number of the articles; the user behavior sequence is expressed as
Figure BDA0002387593690000083
N represents the number of commodities clicked by the user's historical behavior, biThe representation represents the ith behavior embedding vector in the corresponding session.
In step 1, a self-attention module is constructed to extract short-term interests and hobbies of the user in the latest session, and the method is specifically realized as follows:
dividing the user behavior sequence S into different according to time intervalssession R, where the last session sequence of user u at time t is denoted as
Figure BDA0002387593690000084
Figure BDA0002387593690000091
Figure BDA0002387593690000092
Representing the number of times of the user u click behavior in the latest session sequence at the time t, biRepresenting the i-th behavior embedding vector in the corresponding session, dmodelRepresenting the item embedding dimension in each click behavior;
as shown in FIG. 2, first, a query is queried using a non-linear change in a shared parameter
Figure BDA0002387593690000093
Figure BDA0002387593690000094
Figure BDA0002387593690000095
Keys (keys) and
Figure BDA0002387593690000096
the value (value) maps to the same space;
Figure BDA0002387593690000097
Figure BDA0002387593690000098
Figure BDA0002387593690000099
wherein the content of the first and second substances,
Figure BDA00023875936900000910
is the corresponding mapping weight momentArraying; relu is used as an activation function that learns nonlinear interactions in features;
Figure BDA00023875936900000911
respectively representing a query (query), a key (key) and a value (value) matrix after mapping;
the correlation matrix is then calculated as follows:
Figure BDA00023875936900000912
Figure BDA00023875936900000913
wherein one is output
Figure BDA00023875936900000914
Attention moment diagram of
Figure BDA00023875936900000915
It represents
Figure BDA00023875936900000916
Similarity between individual items; wherein
Figure BDA00023875936900000917
For the attention value after the zoom point multiplication,
Figure BDA00023875936900000918
is a matrix representation of the user's short-term preferences.
And (3) extracting the long-term interest and hobby deviation conditions of the user by the multi-channel memory network in the step (2). We first describe the memory network used in the model and then the processing of the multiple channels.
Extracting the long-term interest and the interest deviation of the user through the multi-channel memory network in the step (2), which is specifically realized as follows:
define the memory matrix of user u as
Figure BDA00023875936900000919
Wherein
Figure BDA00023875936900000920
Expressing preference expression vectors of corresponding characteristics of the historical item records by the user u; the memory matrix is used for storing long-term preference of users, and the user memory matrix needs to be read and updated every time new data is added.
The operation of reading and updating the user memory matrix is as follows:
2-1. read operation in memory network
Figure BDA0002387593690000101
Figure BDA0002387593690000102
Wherein α is an enhancement parameter, | | | | represents a modulo operation, btThe representation represents the t-th behavior embedding vector in the corresponding session,
Figure BDA0002387593690000103
representing the weight values without regularization,
Figure BDA0002387593690000104
representing weight values after regularization;
then the weight sum of the vectors in the memory matrix is used as the final output
Figure BDA0002387593690000105
Figure BDA0002387593690000106
2-2. write operation of memory network
eraset＝δ(ETbt+be)
Figure BDA0002387593690000107
Where δ (.) denotes sigmoid activation function, ⊙ denotes dot-by-element multiplication operation, E and beIs the erasure parameter to learn;
Figure BDA0002387593690000108
an erasure vector representing the characteristic information; btRepresenting the t-th behavior embedding vector in the corresponding session;
Figure BDA0002387593690000109
wherein A and baIs a learnable parameter, tanh represents an activation function; this erase, add, update strategy allows for forgetting and enhancing user interest during learning, and the model can automatically determine signals to be attenuated and enhanced using update and erase operations;
2-3, adding multiple channels to enhance the mining of the memory network on the characteristics of the bias of the user interests and hobbies; using additional interest offset matrices
Figure BDA00023875936900001010
To store the bias change of interest of the user u in the same type of item in the sequence t, firstly
Figure BDA00023875936900001011
The method comprises the following steps that m unit slots are included, wherein each unit slot is a channel which represents a characteristic channel for the migration of interest and hobbies of a user on a certain type of articles; at the time t of the sequence of user u, from the set
Figure BDA00023875936900001012
Wherein k feature indexes are selected, wherein
Figure BDA00023875936900001013
Is a weight vector in the read operation of the memory network, and updates the index number i selected arbitrarily as described in the following formulaCorresponding interest bias channel
Figure BDA0002387593690000111
Figure BDA0002387593690000112
Wherein
Figure BDA0002387593690000113
Is the ith cell slot of the user memory matrix, btA behavior embedding vector; multiple channel unit utilizes the behavior sequence information and the interest shift information of the same channel at the previous time
Figure BDA0002387593690000114
Cell slot information corresponding to memory matrix
Figure BDA0002387593690000115
As input, then using GRU to further mine and update the change situation of user interest and hobbies under the corresponding channel; since the parameters of the GRU are shared for multiple channels, many training parameters are not added.
The interest activation unit in step (3) associates the short-term interest and hobby of the user, the long-term interest and hobby deviation characteristics of the user with the target item, and the specific implementation is as follows:
3-1, when the click rate is predicted, if the target object accords with the short-term interest of the user, the short-term interest of the user has great influence on the click event; otherwise, the short-term interest and hobbies of the user do not interfere with the click event of the user; therefore, it is stated that the short-term interest needs to be associated with the corresponding target object to effectively extract the final short-term interest expression vector of the user
Figure BDA0002387593690000116
Dynamically distributing the weight of each unit vector in the short-term preference matrix by using an interest activation unit; the specific formula for calculating the weight by using the interest activation unit is as follows:
Figure BDA0002387593690000117
wherein the content of the first and second substances,
Figure BDA0002387593690000118
a learnable interaction matrix representing the association target item and the user's short-term interest, d represents the final vector dimension of the item after passing through the embedding layer,
Figure BDA0002387593690000119
an embedded vector representation representing the item v,
Figure BDA00023875936900001110
representing a short-term interest preference matrix for user u at time t
Figure BDA00023875936900001111
The ith row vector of (1); presentation of results after user long-term preference associated with target item
Figure BDA00023875936900001112
The calculation is as follows:
Figure BDA00023875936900001113
wherein the content of the first and second substances,
Figure BDA00023875936900001114
a learnable interaction matrix representing long-term interests and interests of the associated user and the target item,
Figure BDA00023875936900001115
representing the long-term interest preference matrix of user u at time t
Figure BDA00023875936900001116
The ith row vector of (1); representation vector after user interest migration change feature is associated with target item
Figure BDA0002387593690000121
As described by the following equation:
Figure BDA0002387593690000122
wherein the content of the first and second substances,
Figure BDA0002387593690000123
a learnable parameter matrix representing associated target items and user interest migration changes,
Figure BDA0002387593690000124
an interest migration matrix representing user u at time t
Figure BDA0002387593690000125
The ith row vector of (1); user information, article information, Us、UlAnd UdThe embedded vectors of (a) are concatenated, flattened and then input to a multi-layered perceptron to implement the score prediction.
Step (4) connecting the long-term hobby, short-term hobby and hobby deviation characteristics of the user, and then sending the connected characteristic vector into DNN for training to obtain a final click probability p (x); the corresponding loss function is as follows:
Figure BDA0002387593690000126
where x represents all the feature vector connections, D represents the training set, N represents the number of training sets, y ∈ {0,1} represents whether the user will click on the item, and p (.) is the final predicted result value of the network model, which represents the probability that the user will click on the item.

Claims (7)

1. A click rate prediction method based on long-term interest modeling of a user is characterized by comprising the following steps:
step (1) dividing a user behavior sequence into different sessions, and then extracting short-term interest and hobbies of a user by using a self-attention mechanism module;
step (2) extracting the long-term interest and interest deviation of the user through a multi-channel memory network;
step (3), the interest activation unit associates the short-term interest and hobby of the user, the long-term interest and hobby deviation characteristics of the user with the target object;
step (4) linking the three characteristics processed in the step (3) and inputting the three characteristics into a multiple sensing machine to predict the click probability;
and (5) calculating a negative log-likelihood function for the output probability, wherein the smaller the negative log-likelihood function value is, the better the corresponding effect of the method is.
2. The method for predicting click rate based on long and short term interests of users according to claim 1, wherein the session division in the step (1) is implemented as follows:
1-1 user click sequences have the characteristics of timeliness and continuity, and are divided into different sessions according to time intervals, specifically: if the click time interval of two adjacent articles in the user click sequence exceeds a set threshold, dividing the two adjacent articles into different session intervals, otherwise, dividing the two adjacent articles into the same session; the short-term interest of the user has the characteristic of timeliness, so that the short-term interest of the user can be represented only by using the click sequence in the last session interval;
1-2, embedding and converting sparse characteristics of user, commodity and user behavior sequence into low-dimensional dense characteristic representation, and representing all user information into matrix
Figure FDA0002387593680000011
Wherein N isuRepresenting the number of users; all article information can be expressed as
Figure FDA0002387593680000012
Wherein N isiIndicating the number of the articles; the user behavior sequence is expressed as
Figure FDA0002387593680000013
N represents the number of commodities clicked by the user's historical behavior, biThe representation represents the ith behavior embedding vector in the corresponding session.
3. The click rate prediction method based on the long-term interest modeling of the user as claimed in claim 1, wherein a self-attention module is constructed in step 1 to extract the short-term interest and hobbies of the user in the latest session, and the specific implementation is as follows:
dividing a user behavior sequence S into different sessions R according to time intervals, wherein the latest session sequence of a user u at the time t is expressed as
Figure FDA0002387593680000021
Figure FDA0002387593680000022
Figure FDA0002387593680000023
Representing the number of times of the user u click behavior in the latest session sequence at the time t, biRepresenting the i-th behavior embedding vector in the corresponding session, dmtoelRepresenting the item embedding dimension in each click behavior;
first, a query is generated using non-linear variation of a shared parameter
Figure FDA0002387593680000024
Keys (keys) and
Figure FDA0002387593680000025
the value (value) maps to the same space;
Figure FDA0002387593680000026
Figure FDA0002387593680000027
Figure FDA0002387593680000028
wherein the content of the first and second substances,
Figure FDA0002387593680000029
is the corresponding mapping weight matrix; relu is used as an activation function that learns nonlinear interactions in features;
Figure FDA00023875936800000210
respectively representing a query (query), a key (key) and a value (value) matrix after mapping;
the correlation matrix is then calculated as follows:
Figure FDA00023875936800000211
Figure FDA00023875936800000212
wherein one is output
Figure FDA00023875936800000213
Attention moment diagram of
Figure FDA00023875936800000214
It represents
Figure FDA00023875936800000215
Similarity between individual items; wherein
Figure FDA00023875936800000216
For the attention value after the zoom point multiplication,
Figure FDA00023875936800000217
is a matrix representation of the user's short-term preferences.
4. The click rate prediction method based on user long-term interest modeling according to claim 3, wherein the step (2) of extracting the user long-term interest and interest bias through the multi-channel memory network is implemented as follows:
define the memory matrix of user u as
Figure FDA00023875936800000218
Wherein
Figure FDA00023875936800000219
Expressing preference expression vectors of corresponding characteristics of the historical item records by the user u; the memory matrix is used for storing long-term preference of users, and the user memory matrix needs to be read and updated every time new data is added.
5. The click rate prediction method based on user long-term interest modeling according to claim 4, wherein the operations of reading and updating the user memory matrix are as follows:
2-1. read operation in memory network
Figure FDA0002387593680000031
Figure FDA0002387593680000032
Wherein α is an enhancement parameter, | | | | represents a modulo operation, btThe representation represents the t-th behavior embedding vector in the corresponding session,
Figure FDA0002387593680000033
representing the weight values without regularization,
Figure FDA0002387593680000034
representing weight values after regularization;
then the weight sum of the vectors in the memory matrix is used as the final output
Figure FDA0002387593680000035
Figure FDA0002387593680000036
2-2. write operation of memory network
eraset=δ(ETbt+be)
Figure FDA0002387593680000037
Where δ (.) denotes sigmoid activation function, ⊙ denotes dot-by-element multiplication operation, E and beIs the erasure parameter to learn;
Figure FDA0002387593680000038
an erasure vector representing the characteristic information; btRepresenting the t-th behavior embedding vector in the corresponding session;
Figure FDA0002387593680000039
wherein A and baIs a learnable parameter, tanh represents an activation function; this erase, add, update strategy allows for forgetting and enhancing user interest during learning, and the model can automatically determine signals to be attenuated and enhanced using update and erase operations;
2-3, adding multiple channels to enhance the mining of the memory network on the characteristics of the bias of the user interests and hobbies; using additional interest offset matrices
Figure FDA00023875936800000310
For storageThe user u is interested in the bias change situation of the same type of articles in the sequence t, firstly
Figure FDA00023875936800000311
The method comprises the following steps that m unit slots are included, wherein each unit slot is a channel which represents a characteristic channel for the migration of interest and hobbies of a user on a certain type of articles; at the time t of the sequence of user u, from the set
Figure FDA00023875936800000312
Wherein k feature indexes are selected, wherein
Figure FDA00023875936800000313
Is a weight vector in the read operation of the memory network, and updates the corresponding interest and hobby offset channel for the index number i selected arbitrarily as described in the following formula
Figure FDA00023875936800000314
Figure FDA0002387593680000041
Wherein
Figure FDA0002387593680000042
Is the ith cell slot of the user memory matrix, btA behavior embedding vector; multiple channel unit utilizes the behavior sequence information and the interest shift information of the same channel at the previous time
Figure FDA0002387593680000043
Cell slot information corresponding to memory matrix
Figure FDA0002387593680000044
As input, then using GRU to further mine and update the change situation of user interest and hobbies under the corresponding channel; since the parameters of the GRU are shared for multiple channels, many training parameters are not added.
6. The click rate prediction method based on user long-term interest modeling according to claim 1 or 5, wherein the interest activation unit in step (3) associates the characteristics of the user short-term interest, the user long-term interest and the interest bias with the target item, and the method is implemented as follows:
3-1, when the click rate is predicted, if the target object accords with the short-term interest of the user, the short-term interest of the user has great influence on the click event; otherwise, the short-term interest and hobbies of the user do not interfere with the click event of the user; therefore, it is stated that the short-term interest needs to be associated with the corresponding target object to effectively extract the final short-term interest expression vector of the user
Figure FDA0002387593680000045
Dynamically distributing the weight of each unit vector in the short-term preference matrix by using an interest activation unit; the specific formula for calculating the weight by using the interest activation unit is as follows:
Figure FDA0002387593680000046
wherein the content of the first and second substances,
Figure FDA0002387593680000047
a learnable interaction matrix representing the association target item and the user's short-term interest, d represents the final vector dimension of the item after passing through the embedding layer,
Figure FDA0002387593680000048
an embedded vector representation representing the item v,
Figure FDA0002387593680000049
representing a short-term interest preference matrix for user u at time t
Figure FDA00023875936800000410
The ith row vector of (1); presentation of results after user long-term preference associated with target item
Figure FDA00023875936800000411
The calculation is as follows:
Figure FDA00023875936800000412
wherein the content of the first and second substances,
Figure FDA00023875936800000413
a learnable interaction matrix representing long-term interests and interests of the associated user and the target item,
Figure FDA00023875936800000414
representing the long-term interest preference matrix of user u at time t
Figure FDA00023875936800000415
The ith row vector of (1); representation vector after user interest migration change feature is associated with target item
Figure FDA00023875936800000416
As described by the following equation:
Figure FDA0002387593680000051
wherein the content of the first and second substances,
Figure FDA0002387593680000052
a learnable parameter matrix representing associated target items and user interest migration changes,
Figure FDA0002387593680000053
an interest migration matrix representing user u at time t
Figure FDA0002387593680000054
The ith row vector of (1); user information, article information, us、ulAnd udThe embedded vectors of (a) are concatenated, flattened and then input to a multi-layered perceptron to implement the score prediction.
7. The method of claim 6, wherein the step (4) of linking the long-term interest, short-term interest and interest bias characteristics of the user, and then inputting the linked feature vectors into DNN to train and obtain a final click probability p (x); the corresponding loss function is as follows
Figure FDA0002387593680000055
Where x represents all the feature vector connections, D represents the training set, N represents the number of training sets, y ∈ {0,1} represents whether the user will click on the item, and p (.) is the final predicted result value of the network model, which represents the probability that the user will click on the item.
CN202010103305.0A 2020-02-19 2020-02-19 Click rate prediction method based on long-term interest modeling of user Pending CN111369278A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010103305.0A CN111369278A (en) 2020-02-19 2020-02-19 Click rate prediction method based on long-term interest modeling of user

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010103305.0A CN111369278A (en) 2020-02-19 2020-02-19 Click rate prediction method based on long-term interest modeling of user

Publications (1)

Publication Number Publication Date
CN111369278A true CN111369278A (en) 2020-07-03

Family

ID=71210975

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010103305.0A Pending CN111369278A (en) 2020-02-19 2020-02-19 Click rate prediction method based on long-term interest modeling of user

Country Status (1)

Country Link
CN (1) CN111369278A (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814051A (en) * 2020-07-08 2020-10-23 北京奇艺世纪科技有限公司 Resource type determination method and device
CN112084450A (en) * 2020-09-09 2020-12-15 长沙理工大学 Click rate prediction method and system based on convolutional attention network deep session sequence
CN112182387A (en) * 2020-09-29 2021-01-05 中国人民大学 Personalized search algorithm for enhancing time information
CN112381615A (en) * 2020-11-27 2021-02-19 华中科技大学 Short sequence recommendation method based on user repeated behavior pattern mining
CN112492396A (en) * 2020-12-08 2021-03-12 中国计量大学 Short video click rate prediction method based on fine-grained multi-aspect analysis
CN112700274A (en) * 2020-12-29 2021-04-23 华南理工大学 Advertisement click rate estimation method based on user preference
CN112733030A (en) * 2021-01-13 2021-04-30 重庆邮电大学 User interest preference capturing method
CN113468434A (en) * 2021-09-06 2021-10-01 北京搜狐新动力信息技术有限公司 Resource recommendation method, device, readable medium and equipment
CN115048586A (en) * 2022-08-11 2022-09-13 广东工业大学 Multi-feature-fused news recommendation method and system

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108830416A (en) * 2018-06-13 2018-11-16 四川大学 Ad click rate prediction framework and algorithm based on user behavior
CN109960759A (en) * 2019-03-22 2019-07-02 中山大学 Recommender system clicking rate prediction technique based on deep neural network
CN110659742A (en) * 2019-09-10 2020-01-07 阿里巴巴集团控股有限公司 Method and device for acquiring sequence representation vector of user behavior sequence

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108830416A (en) * 2018-06-13 2018-11-16 四川大学 Ad click rate prediction framework and algorithm based on user behavior
CN109960759A (en) * 2019-03-22 2019-07-02 中山大学 Recommender system clicking rate prediction technique based on deep neural network
CN110659742A (en) * 2019-09-10 2020-01-07 阿里巴巴集团控股有限公司 Method and device for acquiring sequence representation vector of user behavior sequence

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
QI PI 等: "Practice on Long Sequential User Behavior Modeling for Click-Through Rate Prediction", 《ARXIV E-PRINTS》 *
SHUAI ZHANG 等: "Next Item Recommendation with Self-Attention", 《ARXIV E-PRINTS》 *
YUFEI FENG 等: "Deep Session Interest Network for Click-Through Rate Prediction", 《ARXIV E-PRINTS》 *

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814051B (en) * 2020-07-08 2023-07-28 北京奇艺世纪科技有限公司 Resource type determining method and device
CN111814051A (en) * 2020-07-08 2020-10-23 北京奇艺世纪科技有限公司 Resource type determination method and device
CN112084450A (en) * 2020-09-09 2020-12-15 长沙理工大学 Click rate prediction method and system based on convolutional attention network deep session sequence
CN112182387A (en) * 2020-09-29 2021-01-05 中国人民大学 Personalized search algorithm for enhancing time information
CN112182387B (en) * 2020-09-29 2023-08-25 中国人民大学 Personalized search method with time information enhancement
CN112381615A (en) * 2020-11-27 2021-02-19 华中科技大学 Short sequence recommendation method based on user repeated behavior pattern mining
CN112492396A (en) * 2020-12-08 2021-03-12 中国计量大学 Short video click rate prediction method based on fine-grained multi-aspect analysis
CN112700274A (en) * 2020-12-29 2021-04-23 华南理工大学 Advertisement click rate estimation method based on user preference
CN112733030A (en) * 2021-01-13 2021-04-30 重庆邮电大学 User interest preference capturing method
CN113468434B (en) * 2021-09-06 2021-12-24 北京搜狐新动力信息技术有限公司 Resource recommendation method, device, readable medium and equipment
CN113468434A (en) * 2021-09-06 2021-10-01 北京搜狐新动力信息技术有限公司 Resource recommendation method, device, readable medium and equipment
CN115048586A (en) * 2022-08-11 2022-09-13 广东工业大学 Multi-feature-fused news recommendation method and system
CN115048586B (en) * 2022-08-11 2023-02-21 广东工业大学 Multi-feature-fused news recommendation method and system

Similar Documents

Publication Publication Date Title
CN111369278A (en) Click rate prediction method based on long-term interest modeling of user
Casalicchio et al. Visualizing the feature importance for black box models
CN108446374B (en) User's Intention Anticipation method, apparatus, electronic equipment, storage medium
CN108647251B (en) Recommendation sorting method based on wide-depth gate cycle combination model
CN109785062B (en) Hybrid neural network recommendation system based on collaborative filtering model
CN111191092B (en) Label determining method and label determining model training method
CN106202519A (en) A kind of combination user comment content and the item recommendation method of scoring
CN111046294A (en) Click rate prediction method, recommendation method, model, device and equipment
US20230153857A1 (en) Recommendation model training method, recommendation method, apparatus, and computer-readable medium
CN109087178A (en) Method of Commodity Recommendation and device
CN104281718A (en) Method for performing data mining intelligent recommendation based on user group behavior
CN113222711B (en) Commodity information recommendation method, system and storage medium
Wang et al. Modelling local and global dependencies for next-item recommendations
Wang et al. Multi-scale graph attention subspace clustering network
CN112749330B (en) Information pushing method, device, computer equipment and storage medium
CN110110372A (en) A kind of user's timing behavior automatic segmentation prediction technique
CN113344648B (en) Advertisement recommendation method and system based on machine learning
CN115880027A (en) Electronic commerce website commodity seasonal prediction model creation method
Medico et al. Learning multivariate shapelets with multi-layer neural networks for interpretable time-series classification
Wang et al. Dmfp: a dynamic multi-faceted fine-grained preference model for recommendation
CN112150206B (en) Method and equipment for predicting user interested article
CN115048530A (en) Graph convolution recommendation system fusing neighbor importance and feature learning
CN106980685A (en) Data processing method and data processing equipment
CN115659063B (en) Relevance information enhancement recommendation method for user interest drift, computer device, storage medium and program product
Li et al. Self-learning tags and hybrid responses for deep knowledge tracing

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20200703

RJ01 Rejection of invention patent application after publication