CN111294619B - Long-short term interest modeling method for IPTV field - Google Patents

Long-short term interest modeling method for IPTV field Download PDF

Info

Publication number
CN111294619B
CN111294619B CN202010129277.XA CN202010129277A CN111294619B CN 111294619 B CN111294619 B CN 111294619B CN 202010129277 A CN202010129277 A CN 202010129277A CN 111294619 B CN111294619 B CN 111294619B
Authority
CN
China
Prior art keywords
term
long
short
click sequence
attention
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.)
Expired - Fee Related
Application number
CN202010129277.XA
Other languages
Chinese (zh)
Other versions
CN111294619A (en
Inventor
李恒
雷航
杨茂林
曾敬鸿
朱迪
付守伟
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.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
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 University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN202010129277.XA priority Critical patent/CN111294619B/en
Publication of CN111294619A publication Critical patent/CN111294619A/en
Application granted granted Critical
Publication of CN111294619B publication Critical patent/CN111294619B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
    • H04N21/25—Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
    • H04N21/251—Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • 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
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
    • H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • H04N21/4667—Processing of monitored end-user data, e.g. trend analysis based on the log file of viewer selections
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
    • H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
    • H04N21/4668—Learning process for intelligent management, e.g. learning user preferences for recommending movies for recommending content, e.g. movies

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Two-Way Televisions, Distribution Of Moving Picture Or The Like (AREA)

Abstract

The invention provides a long-term and short-term interest modeling method for the IPTV field, which distinguishes a long-term click sequence Lt‑1And short-term click sequence StLong term click sequence Lt‑1And short-term click sequence StEmbedding mapping to obtain long-term click sequence embedded representation
Figure DDA0002395348820000011
And short-term click sequence embedded representation
Figure DDA0002395348820000012
Further calculating the long-term preference of the family and the short-term interest U of the userSThen according to family long-term preference and user short-term interest USObtaining user mix preferences UhybirdAnd according to the user mixing preference UhybirdCalculating click rate Ru. According to the invention, more accurate interest and hobby prediction and push are realized through the operation.

Description

Long-short term interest modeling method for IPTV field
Technical Field
The invention belongs to the field of deep learning modeling, and particularly relates to a long-short term interest modeling method for the field of IPTV.
Background
In the era of information overload, recommendation systems have become very important for internet services and are also widely used in different fields, such as: e-commerce websites, video websites, and the like. The television, one of the most commonly used household appliances in daily life, has gradually developed towards the internet. Currently, a large amount of video content has been embedded in IPTV. Therefore, the IPTV field also needs to introduce a recommendation system to solve the problem of how to filter the content meeting the preference of the user. In the prior art, a method and system for modeling the interests of a television viewer is disclosed, for example, in patent application No. 201610485614.2.
However, the IPTV application and the video website in the internet mainly have the following two problems:
1) there is a lot of implicit feedback in IPTV applications, but the explicit feedback is very poor. This makes it impossible to tell whether the user dislikes or does not notice his or her non-interactive items;
2) more particularly, the user of IPTV is typically the entire home, not an individual. The preferences of each person in the family may be different, which greatly increases the difficulty of recommending tasks.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a long-term and short-term interest modeling method for the IPTV field, which constructs the long-term preference of a family and the short-term interest of a current user according to the historical click records of the family group and the user by using a sequence modeling mode, thereby realizing more accurate recommendation.
The specific implementation content of the invention is as follows:
a long-short term interest modeling method for IPTV field divides the history click record into long-term click sequence Lt-1And short-term click sequence StProcessing the data and obtaining the long-term preference U of the family through self-attention calculationLUser short-term interest US。
In order to better implement the invention, the method further specifically comprises the following steps:
s1, obtaining a long-term click sequence L according to historical click recordst-1Short term click sequence St(ii) a Will long term click sequence Lt-1Short term click sequence StInputting the code into an embedded code layer to be coded, wherein the code is represented in a dense matrix form; the long-term click sequence obtained after coding is then embedded into the representation
Figure BDA0002395348800000011
Short term click sequence embedded representation
Figure BDA0002395348800000012
Outputting to a long-term and short-term modeling layer;
s2, embedding and representing long-term click sequences in long-term and short-term modeling layers
Figure BDA0002395348800000013
Short term click sequence embedded representation
Figure BDA0002395348800000014
Self-attention calculation is carried out, and family long-term preference U is obtained through mapping respectivelyLUser short-term interest USAnd make the family prefer U for a long timeLUser short-term interest USInputting the long-short term preference fusion layer;
s3, long-term preference U of the family in the long-term and short-term preference fusion layerLUser short-term interest USPerforming self-attention decoding and mapping to obtain user mixed preference Uhybird；
S4, mixing preference U to usershybirdLinear mapping is carried out to obtain the click rate RuAnd (6) predicting.
In order to better implement the present invention, further, in the step s2, the short-term interest U of the user is obtainedSThe method comprises the following specific steps:
step SA. calculation of user short-term interests U in a self-attention mechanismSThe three inputs of Query, Key and Value are completely consistent and are short-term click sequencesColumn embedded representation
Figure BDA0002395348800000021
Embedding three-input short-term click sequences into a representation
Figure BDA0002395348800000022
Mapping into different spaces respectively through linear transformation to obtain embedded representation in corresponding space
Figure BDA0002395348800000023
Wherein Wq,Wk,WvHas the dimension of
Figure BDA0002395348800000024
Therefore, it is
Figure BDA0002395348800000025
Has the dimension of
Figure BDA0002395348800000026
C is a fixed value;
then will be
Figure BDA0002395348800000027
Is split into nhHead, obtained in the dimension of
Figure BDA0002395348800000028
Tensor of
Figure BDA0002395348800000029
Wherein d ═ C/nhThen transposing each tensor, the tensor
Figure BDA00023953488000000210
From dimension of
Figure BDA00023953488000000211
Is transformed into
Figure BDA00023953488000000212
Step SB. divides each header after splitting
Figure BDA00023953488000000213
As an input of the dot product attention scaled, the following operations are performed: will be provided with
Figure BDA00023953488000000214
Performing matrix multiplication to obtain dimension of
Figure BDA00023953488000000215
The dot product attention of (1); then reducing the dot product attention by square root times of dimensionality through scale transformation; then, performing softmax normalization processing on the accumulated attention; finally, the dot product is added with attention
Figure BDA00023953488000000216
Matrix multiplication outputs scaled dot product attention;
step SC. splices the scaled dot product attentiveness of each output, and then performs linear transformation to output short-term multi-head attentiveness
Figure BDA00023953488000000217
Short term multi-headed attention
Figure BDA00023953488000000218
Has the dimension of
Figure BDA00023953488000000219
Step SD. directs short term multiple attention
Figure BDA00023953488000000220
Obtaining user short-term interest U by input point type feedforward networkS。
To better implement the present invention, in step SD., before outputting the multi-headed attention to the point-type feedforward network, residual concatenation and layer normalization are performed, and then the short-term multi-headed attention is focused
Figure BDA00023953488000000221
Inputting a point type feedforward network, and performing residual error connection and layer normalization on the output of the point type feedforward network to obtain the user short-term interest US。
In order to better implement the present invention, further, the step s2 obtains the long-term preference U of the familyLThe method comprises the following specific steps:
step Sa. for Long term click sequence Lt-1Encoded long-term click sequence embedded representation for each time step
Figure BDA0002395348800000031
The long-term click sequence embedded representation
Figure BDA0002395348800000032
Has the dimension of
Figure BDA0002395348800000033
Embedding a representation for each long-term click sequence
Figure BDA0002395348800000034
Short-term multi-head attention is calculated as described above
Figure BDA0002395348800000035
Is processed to obtain the long-term preference U for calculating the familyLLong-term multi-head attention of
Figure BDA0002395348800000036
The long-term multi-head attention
Figure BDA0002395348800000037
Has the dimension of
Figure BDA0002395348800000038
Step Sb. combines the long-term click sequence Lt-1Long-term multi-head attention per step in
Figure BDA0002395348800000039
Is compressed and then the average value of each column is calculated to obtain the dimension of
Figure BDA00023953488000000310
A row vector of (a); then the long-term click sequence Lt-1All the line vectors of the time step are spliced to finally obtain the dimension of
Figure BDA00023953488000000311
Family long-term preference UL。
In order to better implement the present invention, further, the step s3. specifically includes: preference of family for long term ULUser short-term interest USCalculating multi-head attention through a common attention mechanism, performing residual connection and layer normalization on the processed output, inputting the result into a point type feedforward network for further processing, performing residual connection and layer normalization on the further processed result, and finally obtaining user mixed preference U after linear change and Sigmoid function processinghybirdThe user mix preferences UhybirdDimension of
Figure BDA00023953488000000312
In order to better implement the present invention, further, the specific operation of step s4 is: first mix the user preferences U by global average poolinghybirdReduce the dimension to
Figure BDA00023953488000000313
The click rate R is then obtained by linear mapping, i.e. matrix multiplicationuThe click rate RuHas the dimension of
Figure BDA00023953488000000314
To better implement the invention, further, the long-term click sequence Lt-1And short-term click sequence StDividing according to t time step: the t time step is the last time segment in the acquired click sequence to be processed, and all time steps before the t time step in the acquired click sequence to be processed are defined as a long-term click sequence Lt-1The long-term click sequence Lt-1＝S1∪S2∪...∪St-1Defining t time step as a short-term click sequence StThe short-term click sequence St＝{i1,i2,...,im}。
To better implement the invention, further, the encoded long-term click sequence Lt-1The specific operation is as follows: will long term click sequence Lt-1Each time step S iniEmbedding mapping to obtain long-term click sequence embedded representation
Figure BDA00023953488000000315
The long-term click sequence embedded representation
Figure BDA0002395348800000041
Dimension of
Figure BDA0002395348800000042
Wherein | SiI represents the number of media assets in the processed time step, and K represents the size of the embedding layer;
the encoded short-term click sequence StThe specific operation is as follows: short-term click sequence StEmbedding mapping to obtain short-term click sequence embedded representation
Figure BDA0002395348800000043
The short term click sequence embedded representation
Figure BDA0002395348800000044
Has the dimension of
Figure BDA0002395348800000045
Wherein | StAnd | represents the number of assets in the t time step, and K represents the size of the embedding layer.
In order to better implement the invention, further, the point feed-forward network is composed of two fully-coupled layers, and a ReLU activation function is set between the two fully-coupled layers.
Compared with the prior art, the invention has the following advantages and beneficial effects:
1) meanwhile, the hobbies of family groups and the interests of users in a short period are considered, so that the interest recommendation result is more accurate;
2) through the click sequence, the user can be better distinguished whether the user dislikes or does not notice the non-recommended videos;
3) self-attention computation is used, which refines the representation by matching individual sequences to themselves; unlike general attention, the mechanism of self-attention reduces reliance on external information and is better at capturing internal correlations of sequences or features.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a model structure and operation diagram of the present invention;
FIG. 3 shows the long term preference U of a familyLUser short-term interest USCalculating a flow chart;
FIG. 4 is family long term preference ULUser short-term interest USA processing flow chart of multi-head attention in the calculation process;
FIG. 5 is a flow chart of the process of scaling dot product attention by ratio during a multi-head attention process;
FIG. 6 is a calculation of user mix preferences U for the decoding processhybirdIs described.
Detailed Description
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all embodiments, and therefore should not be considered as a limitation to the scope of protection. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, are within the scope of the present invention.
Example 1:
a method for modeling long-short term interest in IPTV field, as shown in FIG. 1, comprises the following steps:
s1, obtaining a long-term click sequence L according to historical click recordst-1Short term click sequence St(ii) a The long-term click sequence Lt-1And short-term click sequence StDividing according to t time step: the t time step is the last time segment in the acquired click sequence to be processed, and in the acquired click sequence to be processed, the time step before the t time step is defined as a long-term click sequence Lt-1The long-term click sequence Lt-1＝S1∪S2∪...∪St-1Defining t time step as a short-term click sequence StThe short-term click sequence St＝{i1,i2,...,im}; for long-term click sequence Lt-1Encoding is carried out, and the specific operation is as follows: will long term click sequence Lt-1Each time step S iniEmbedding mapping to obtain long-term click sequence embedded representation
Figure BDA0002395348800000051
The long-term click sequence embedded representation
Figure BDA0002395348800000052
Dimension of
Figure BDA0002395348800000053
Wherein | SiI represents the number of media assets in the processed time step, and K represents the size of the embedding layer; for short-term click sequence StEncoding is carried out, and the specific operation is as follows: short-term click sequence StEmbedding mapping to obtain short-term click sequence embedded representation
Figure BDA0002395348800000054
The short term click sequence embedded representation
Figure BDA0002395348800000055
Has the dimension of
Figure BDA0002395348800000056
Wherein | StI represents the number of media assets in the time step of t, and K represents the size of the embedding layer; the long-term click sequence obtained after coding is then embedded into the representation
Figure BDA0002395348800000057
Short term click sequence embedded representation
Figure BDA0002395348800000058
Outputting to a long-term and short-term modeling layer;
s2, embedding and representing long-term click sequences in long-term and short-term modeling layers
Figure BDA0002395348800000059
Short term click sequence embedded representation
Figure BDA00023953488000000510
Self-attention calculation is carried out, and family long-term preference U is obtained through mapping respectivelyLUser short-term interest USAnd make the family prefer U for a long timeLUser short-term interest USInputting the long-short term preference fusion layer;
s3, long-term preference U of the family in the long-term and short-term preference fusion layerLUser short-term interest USPerforming self-attention decoding and mapping to obtain user mixed preference Uhybird；
S4, mixing preference U to usershybirdLinear mapping is carried out to obtain the click rate RuAnd (6) predicting.
The working principle is as follows: through the division of the t time step, the preference of the family user before the t time step can be divided, the preference interest of the user using the television at present is the t time step, the preference interest and the preference interest are processed after the t time step is divided, and the final result can be more accurate. The specific division of the time step needs to be determined according to the scene, assuming that it is first time-dependentThe inter-step is divided in hours, so if the user clicks, the behavior of nearly one hour is considered as a short-term sequence and the earlier behavior is considered as a long-term sequence. Then coding the long-term and short-term sequence to obtain the family long-term preference ULAnd user short-term interests USThen using the family long-term preference ULDecoding user short-term interests USObtaining user mix preferences Uhybird. Finally, the user is mixed with the preference UhybirdLinear mapping is carried out to finally obtain the click rate R of the useruAnd (6) predicting. And in actual operation, in our recommendation system, there would be two phases, namely a recall phase and a sort phase, and the model would be used in the sort phase. First, in the recall phase, the recommendation system generates a longer candidate list, and then filters each candidate in the candidate list through the ranking model. Let us assume that the user clicks each candidate item and clicks, and calculate the click rate R according to the above-mentioned calculation procedureu. Finally the click rate RuAnd sequencing to obtain a final recommendation list.
Example 2:
the present invention is based on the above embodiment 1, as shown in FIG. 2, and further, as shown in FIG. 2, the present invention is to arrange the long-term click sequence Lt-1And short-term click sequence StAre respectively input into an embedded coding layer which maps the long-term click sequence L through embeddingt-1And short-term click sequence StEncoded as a long-term click sequence Lt-1Vector and short-term click sequence embedded representation
Figure BDA0002395348800000061
The long-term click sequence is then embedded into the representation
Figure BDA0002395348800000062
And short-term click sequence embedded representation
Figure BDA0002395348800000063
Sending the data into a long-short term modeling layer, and carrying out long-term click sequence on the long-short term modeling layer through a Transformer encoderColumn embedded representation
Figure BDA0002395348800000064
And short-term click sequence embedded representation
Figure BDA0002395348800000065
Calculating the long-term preference U of the family by self attentionLAnd user short-term interests USNote that the representation is embedded for short-term click sequences
Figure BDA0002395348800000066
Directly obtaining the short-term interest U of the user after being processed by a Transformer encoderSAnd for long-term click sequences the representation is embedded
Figure BDA0002395348800000067
After being processed by a Transformer encoder, the long-term preference U of the family can be obtained through global average poolingL(ii) a Because of the long-term click sequence Lt-1And short-term click sequence StDivided by t time steps, short-term click sequences StOne time step, and a long-term click sequence Lt-1For short-term click sequences StA plurality of time steps before the time step, so the long-term preference U of the family can be obtained after one global average poolingL. Family long-term preference U to be generatedLAnd user short-term interests USSending the information into the long and short term preference fusion layer, and carrying out attention calculation in a Transformer decoder of the fusion layer of the long and short term preference to obtain the user mixed preference UhybirdFinally, mix the user preferences UhybirdDimension reduction by global average pooling to
Figure BDA0002395348800000068
The click rate R is then obtained by linear mapping, i.e. matrix multiplicationuThe click rate RuHas the dimension of
Figure BDA0002395348800000069
Other parts of this embodiment are the same as those of embodiment 1, and thus are not described again.
Example 3:
in this embodiment, based on any one of the above embodiments 1-2, and as shown in fig. 3, 4, and 5, further, the ordinary attention mechanism has three inputs, which are Query (request), Key (primary Key), and Value (Value), and the calculation method is as follows:
Figure BDA00023953488000000610
and in the self-attention mechanism, for calculating the short-term interest U of the userSAre identical, all short term click sequence embedded representations
Figure BDA0002395348800000071
Embedding three-input short-term click sequences into a representation
Figure BDA0002395348800000072
Respectively by linear transformation Wq,Wk,WvMapping into different spaces
Figure BDA0002395348800000073
Wherein Wq,Wk,WvThe dimensions of (a) are respectively:
Figure BDA0002395348800000074
therefore, it is
Figure BDA0002395348800000075
The dimensions of (a) are respectively:
Figure BDA0002395348800000076
then will be
Figure BDA0002395348800000077
Is split into nhHead, obtainingDimension of
Figure BDA0002395348800000078
Tensor of
Figure BDA0002395348800000079
Wherein d ═ C/nhThen transposing each tensor, the tensor
Figure BDA00023953488000000710
From m nhX d is converted into nh×m×d；
As shown in FIG. 3, the user's short-term interests U are calculated using self-attention in the long-and-short-term modeling layerSThe process of (2) is as follows: first embedding the short-term click sequence of the input into the representation
Figure BDA00023953488000000711
Treatment as short-term multi-head attention
Figure BDA00023953488000000712
Then the obtained short-term attention of the head
Figure BDA00023953488000000713
Input into a point feed-forward network, where attention is paid to the point feed-forward network and the short term
Figure BDA00023953488000000714
The output ends of the two-dimensional model are all provided with a residual error connection and are subjected to layer normalization processing, and the short-term interest U of the user can be obtained after the point type feedforward network, the residual error connection and the layer normalization processingSIt is used.
Short-term Bull attention with respect to FIG. 3
Figure BDA00023953488000000715
The calculation method of (2) is, as shown in fig. 4, to make linear change to the inputted multi-head data, then to process it into scaled dot product attention, then to splice the multi-head scaled dot product attention, and finally to make linear transformationOutputting short-term multi-head attention
Figure BDA00023953488000000716
On the other hand, as for the processing method of dot product attention by scaling in fig. 4, as shown in fig. 5, the tensor divided into multiple tensors
Figure BDA00023953488000000717
As an input of the dot product attention scaled, the following operations are performed: will be provided with
Figure BDA00023953488000000718
Performing matrix multiplication to obtain a dot product attention with dimension of m multiplied by m; then reducing the dot product attention by square root times of dimensionality through scale transformation; then, performing softmax normalization processing on the accumulated attention; finally, the dot product is added with attention
Figure BDA00023953488000000719
Matrix multiplication outputs scaled dot product attention;
likewise, the family long term preference U is calculated for the use of self-attention in the long and short term modeling layerLAnd the above-mentioned process for calculating the short-term interest U of the userSThe operation is approximately the same, and the long-term click sequence needs to be embedded with a representation
Figure BDA00023953488000000720
Wherein the short-term click sequence embedding representation E is performed for each time stepStOperation of the treatment to obtain long-term attention
Figure BDA00023953488000000721
The difference is that global average pooling treatment is required, and the specific steps are as follows: long-term multi-head attention per time step will be obtained
Figure BDA00023953488000000722
Compressing the dimensionality, and then calculating the average value of each column to obtain a row vector; then click long termSequence Lt-1All the line vectors of the time step are spliced to finally obtain the long-term preference U of the familyL(ii) a The formula for the global average pooling is as follows:
Figure BDA0002395348800000081
wherein GAP represents Global Average Pooling (Global Average Pooling), ULPreference of family for long term UL，
Figure BDA0002395348800000082
The output of attention for each step.
Other parts of this embodiment are the same as any of embodiments 1-2 described above, and thus are not described again.
Example 4:
this embodiment is different from the transform encoder in that the transform decoder uses a normal attention mechanism rather than a self-attention mechanism, based on any of the embodiments 1 to 3 described above, as shown in fig. 6. The user's short-term interest U is processed through similar calculation steps as a transform encoder, including a multi-head attention mechanism, residual connection and regularization and a point feed-forward networkSAnd inputting the family long-term interest preference to obtain the attention output of the transform decoder, wherein the specific user mixed preference UhybirdThe calculation formula of (2) is as follows:
Uhybird＝Attention(UL,US,US)
wherein, UhybirdMix preferences for users with dimensions of
Figure BDA0002395348800000083
After obtaining the user mixing preference UhybirdThen, the users are mixed with the preference U through global average poolinghybirdReducing dimension, and then obtaining click rate R through linear mapping, namely matrix multiplicationuThe resulting click rate RuHas the dimension of
Figure BDA0002395348800000084
The specific calculation formula is as follows:
Ru＝Linear(GAP(Uhybird))。
other parts of this embodiment are the same as any of embodiments 1 to 3, and thus are not described again.
The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, and all simple modifications and equivalent variations of the above embodiments according to the technical spirit of the present invention are included in the scope of the present invention.

Claims (4)

1. A long-term and short-term interest modeling method for the IPTV field is characterized in that a long-term click sequence L is extracted from a historical click recordt-1And short-term click sequence StProcessing the data and obtaining the long-term preference U of the family through self-attention calculationLUser short-term interest US(ii) a The method specifically comprises the following steps:
s1, obtaining a long-term click sequence L according to historical click recordst-1Short term click sequence St(ii) a Will long term click sequence Lt-1Short term click sequence StInputting the code into an embedded code layer to be coded, wherein the code is represented in a dense matrix form; the long-term click sequence obtained after coding is then embedded into the representation
Figure FDA0003057713560000011
Short term click sequence embedded representation
Figure FDA0003057713560000012
Outputting to a long-term and short-term modeling layer;
s2, embedding and representing long-term click sequences in long-term and short-term modeling layers
Figure FDA0003057713560000013
Short term click sequence embedded representation
Figure FDA0003057713560000014
Self-attention calculation is carried out, and family long-term preference U is obtained through mapping respectivelyLUser short-term interest USAnd make the family prefer U for a long timeLUser short-term interest USInputting the long-short term preference fusion layer;
s3, long-term preference U of the family in the long-term and short-term preference fusion layerLUser short-term interest USPerforming self-attention decoding and mapping to obtain user mixed preference Uhybird；
S4, mixing preference U to usershybirdLinear mapping is carried out to obtain click rate R for predictionu；
In the step S2, the short-term interest U of the user is obtainedSThe method comprises the following specific steps:
step SA. calculation of user short-term interests U in a self-attention mechanismSThe three inputs of Query, Key and Value are completely consistent and are embedded expressions of short-term click sequences
Figure FDA0003057713560000015
Embedding three-input short-term click sequences into a representation
Figure FDA0003057713560000016
Mapping into different spaces respectively through linear transformation to obtain embedded representation in corresponding space
Figure FDA0003057713560000017
Wherein the weight matrix W of the linear transformationq,Wk,WvHas the dimension of
Figure FDA0003057713560000018
Therefore, it is
Figure FDA0003057713560000019
Has the dimension of
Figure FDA00030577135600000110
C is a fixed value;
then will be
Figure FDA00030577135600000111
Is split into nhHead, obtained in the dimension of
Figure FDA00030577135600000112
Tensor of
Figure FDA00030577135600000113
Wherein d ═ C/nhThen transposing each tensor, the tensor
Figure FDA00030577135600000114
From dimension of
Figure FDA00030577135600000115
Is transformed into
Figure FDA00030577135600000116
Step SB. divides each header after splitting
Figure FDA00030577135600000117
The matrix is used as an input of the dot product attention of scaling, and the following operations are carried out: will be provided with
Figure FDA00030577135600000118
Performing matrix multiplication to obtain dimension of
Figure FDA00030577135600000119
The dot product attention of (1); then reducing the dot product attention by square root times of dimensionality through scale transformation; then, performing softmax normalization processing on the accumulated attention; finally, the dot product is added with attention
Figure FDA0003057713560000021
Matrix multiplication outputs scaled dot product attention;
step SC. splices the scaled dot product attentiveness of each output, and then performs linear transformation to output short-term multi-head attentiveness
Figure FDA0003057713560000022
The short-term multi-headed attention
Figure FDA0003057713560000023
Has the dimension of
Figure FDA0003057713560000024
Step SD. directs short term multiple attention
Figure FDA0003057713560000025
Obtaining user short-term interest U by input point type feedforward networkS；
In said step SD., attention is paid to the short-term bull
Figure FDA0003057713560000026
Before inputting to the point type feedforward network, residual error connection and layer normalization are firstly carried out, and the residual error connection and layer normalization are also carried out on the output of the point type feedforward network, and finally the short-term interest U of the user is obtainedS；
The long-term preference U of the family is obtained in the step S2LThe method comprises the following specific steps:
step Sa. for Long term click sequence Lt-1Encoded long-term click sequence embedded representation for each time step
Figure FDA0003057713560000027
The long-term click sequence embedded representation
Figure FDA0003057713560000028
Has the dimension of
Figure FDA0003057713560000029
Embedding a representation for each long-term click sequence
Figure FDA00030577135600000210
Short-term multi-head attention is calculated as described above
Figure FDA00030577135600000211
Is processed to obtain the long-term preference U for calculating the familyLLong-term multi-head attention of
Figure FDA00030577135600000212
The long-term multi-head attention
Figure FDA00030577135600000213
Has the dimension of
Figure FDA00030577135600000214
Step Sb. combines the long-term click sequence Lt-1Long-term multi-head attention per step in
Figure FDA00030577135600000215
Is compressed and then the average value of each column is calculated to obtain the dimension of
Figure FDA00030577135600000216
A row vector of (a); then the long-term click sequence Lt-1All the line vectors of the time step are spliced to finally obtain the dimension of
Figure FDA00030577135600000217
Family long-term preference UL；
The step S3 specifically comprises the following operations: preference of family for long term ULUser short-term interest USCalculating multi-head attention through a common attention mechanism, performing residual connection and layer normalization on the processed output, inputting the result into a point type feedforward network for further processing, performing residual connection and layer normalization on the further processed result, and finally obtaining user mixed preference U after linear change and Sigmoid function processinghybirdThe user mix preferences UhybirdDimension of
Figure FDA00030577135600000218
The specific operation of the step S4 is as follows: first mix the user preferences U by global average poolinghybirdReduce the dimension to
Figure FDA00030577135600000219
The click rate R is then obtained by linear mapping, i.e. matrix multiplicationuThe click rate RuHas the dimension of
Figure FDA00030577135600000220
2. The method of claim 1, wherein the long-term click sequence L is a long-term short-term interest modeling method in IPTV fieldt-1And short-term click sequence StDividing according to t time step: the t time step is the last time segment in the acquired click sequence to be processed, and all time steps before the t time step in the acquired click sequence to be processed are defined as a long-term click sequence Lt-1The long-term click sequence Lt-1＝S1∪S2∪...∪St-1Defining t time step as a short-term click sequence StThe short-term click sequence St＝{i1,i2,...,im}。
3. The method as claimed in claim 2, wherein the method for modeling the long-short term interest in IPTV field is characterized in thatThen, the code long-term click sequence Lt-1The specific operation is as follows: will long term click sequence Lt-1Each time step S iniEmbedding mapping to obtain long-term click sequence embedded representation
Figure FDA0003057713560000031
The long-term click sequence embedded representation
Figure FDA0003057713560000032
Dimension of
Figure FDA0003057713560000033
Wherein | SiI represents the number of media assets in the processed time step, and K represents the size of the embedding layer;
the encoded short-term click sequence StThe specific operation is as follows: short-term click sequence StEmbedding mapping to obtain short-term click sequence embedded representation
Figure FDA0003057713560000034
The short term click sequence embedded representation
Figure FDA0003057713560000035
Has the dimension of
Figure FDA0003057713560000036
Wherein | StAnd | represents the number of assets in the t time step, and K represents the size of the embedding layer.
4. The method as claimed in claim 1, wherein the point feed-forward network comprises two fully-coupled layers, and a ReLU activation function is set between the two fully-coupled layers.
CN202010129277.XA 2020-02-28 2020-02-28 Long-short term interest modeling method for IPTV field Expired - Fee Related CN111294619B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010129277.XA CN111294619B (en) 2020-02-28 2020-02-28 Long-short term interest modeling method for IPTV field

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010129277.XA CN111294619B (en) 2020-02-28 2020-02-28 Long-short term interest modeling method for IPTV field

Publications (2)

Publication Number Publication Date
CN111294619A CN111294619A (en) 2020-06-16
CN111294619B true CN111294619B (en) 2021-09-10

Family

ID=71029259

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010129277.XA Expired - Fee Related CN111294619B (en) 2020-02-28 2020-02-28 Long-short term interest modeling method for IPTV field

Country Status (1)

Country Link
CN (1) CN111294619B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114422859B (en) * 2020-10-28 2024-01-30 贵州省广播电视信息网络股份有限公司 Deep learning-based ordering recommendation system and method for cable television operators
CN113344662A (en) * 2021-05-31 2021-09-03 联想(北京)有限公司 Product recommendation method, device and equipment

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110008409A (en) * 2019-04-12 2019-07-12 苏州市职业大学 Based on the sequence of recommendation method, device and equipment from attention mechanism
CN110489639A (en) * 2019-07-15 2019-11-22 北京奇艺世纪科技有限公司 A kind of content recommendation method and device
CN110599280A (en) * 2018-06-12 2019-12-20 阿里巴巴集团控股有限公司 Commodity information preference model training and predicting method and device and electronic equipment
CN110610168A (en) * 2019-09-20 2019-12-24 合肥工业大学 Electroencephalogram emotion recognition method based on attention mechanism
CN110807156A (en) * 2019-10-23 2020-02-18 山东师范大学 Interest recommendation method and system based on user sequence click behaviors

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107995535B (en) * 2017-11-28 2019-11-26 百度在线网络技术（北京）有限公司 A kind of method, apparatus, equipment and computer storage medium showing video
CN110162553A (en) * 2019-05-21 2019-08-23 南京邮电大学 Users' Interests Mining method based on attention-RNN
CN110446112A (en) * 2019-07-01 2019-11-12 南京邮电大学 IPTV user experience prediction technique based on two-way LSTM-Attention
CN110796313B (en) * 2019-11-01 2022-05-31 北京理工大学 Session recommendation method based on weighted graph volume and item attraction model

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110599280A (en) * 2018-06-12 2019-12-20 阿里巴巴集团控股有限公司 Commodity information preference model training and predicting method and device and electronic equipment
CN110008409A (en) * 2019-04-12 2019-07-12 苏州市职业大学 Based on the sequence of recommendation method, device and equipment from attention mechanism
CN110489639A (en) * 2019-07-15 2019-11-22 北京奇艺世纪科技有限公司 A kind of content recommendation method and device
CN110610168A (en) * 2019-09-20 2019-12-24 合肥工业大学 Electroencephalogram emotion recognition method based on attention mechanism
CN110807156A (en) * 2019-10-23 2020-02-18 山东师范大学 Interest recommendation method and system based on user sequence click behaviors

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Adaptive User Modeling with Long and Short-Term Preferences for Personalized Recommendation;Zeping Yu;《Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence(IJCAI-19)》;20191231;全文 *
基于神经网络的IPTV用户体验预测系统设计与实现;毛佳丽;《中国优秀硕士学位论文全文数据库 信息科技辑》;20200215;全文 *

Also Published As

Publication number Publication date
CN111294619A (en) 2020-06-16

Similar Documents

Publication Publication Date Title
JP6523498B1 (en) Learning device, learning method and learning program
Pfaff et al. Neural network based intra prediction for video coding
CN111294619B (en) Long-short term interest modeling method for IPTV field
CN111310045A (en) Network-embedded movie recommendation method based on meta-path
CN107562787B (en) POI (point of interest) encoding method and device, POI recommendation method and electronic equipment
Chang et al. On the design fundamentals of diffusion models: A survey
CN113094587B (en) Implicit recommendation method based on knowledge graph path
CN113255908A (en) Method, neural network model and device for service prediction based on event sequence
CN114386513A (en) Interactive grading prediction method and system integrating comment and grading
CN107506479B (en) A kind of object recommendation method and apparatus
CN113033090A (en) Push model training method, data push device and storage medium
CN114357201A (en) Audio-visual recommendation method and system based on information perception
Sharifi et al. A new algorithm for solving data sparsity problem based-on Non negative matrix factorization in recommender systems
He et al. A time-context-based collaborative filtering algorithm
CN113918764A (en) Film recommendation system based on cross modal fusion
CN117171440A (en) News recommendation method and system based on news event and news style joint modeling
CN116680456A (en) User preference prediction method based on graph neural network session recommendation system
CN117251622A (en) Method, device, computer equipment and storage medium for recommending objects
CN111046280A (en) Cross-domain recommendation method for application FM
CN114782995A (en) Human interaction behavior detection method based on self-attention mechanism
CN111222722B (en) Method, neural network model and device for business prediction for business object
CN111460302B (en) Data processing method, device, electronic equipment and computer readable storage medium
CN112132345A (en) Method and device for predicting user information of electric power company, electronic equipment and storage medium
Kumari et al. Movie Recommendation System for Cold-Start Problem Using User's Demographic Data
Ahirwadkar et al. Deepautoencf: A Denoising Autoencoder For Recommender Systems

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20210910