CN114528434A - IPTV live channel fusion recommendation method based on self-attention mechanism - Google Patents

IPTV live channel fusion recommendation method based on self-attention mechanism Download PDF

Info

Publication number
CN114528434A
CN114528434A CN202210059426.9A CN202210059426A CN114528434A CN 114528434 A CN114528434 A CN 114528434A CN 202210059426 A CN202210059426 A CN 202210059426A CN 114528434 A CN114528434 A CN 114528434A
Authority
CN
China
Prior art keywords
channel
time
representing
self
layer
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.)
Granted
Application number
CN202210059426.9A
Other languages
Chinese (zh)
Other versions
CN114528434B (en
Inventor
杨灿
杨中伟
曹文志
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.)
South China University of Technology SCUT
Original Assignee
South China University of Technology SCUT
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 South China University of Technology SCUT filed Critical South China University of Technology SCUT
Priority to CN202210059426.9A priority Critical patent/CN114528434B/en
Publication of CN114528434A publication Critical patent/CN114528434A/en
Application granted granted Critical
Publication of CN114528434B publication Critical patent/CN114528434B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
    • G06F16/73—Querying
    • G06F16/735—Filtering based on additional data, e.g. user or group profiles
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
    • G06F16/78—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/7867—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using information manually generated, e.g. tags, keywords, comments, title and artist information, manually generated time, location and usage information, user ratings
    • 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
    • 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
    • G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
    • 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
    • G06N3/049—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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Abstract

The invention discloses an IPTV live channel fusion recommendation method based on a self-attention mechanism, which comprises the following steps: s1, obtaining historical watching log data of the user as original data; s2, preprocessing the original data, and dividing the data into a training set, a verification set and a test set according to a time window; s3, extracting the characteristics of the preprocessed user watching data; s4, constructing an IPTV live channel fusion recommendation neural network model based on a self-attention mechanism; s5, inputting training set data into the fusion network for training, adjusting network hyper-parameters through a verification set, and storing the trained network; and S6, inputting the test set data into the trained fusion network to obtain a personalized channel recommendation list. The method and the device can effectively improve the recommendation accuracy of the IPTV live broadcast channel.

Description

IPTV live channel fusion recommendation method based on self-attention mechanism
Technical Field
The invention relates to the technical field of television recommendation, in particular to an IPTV live channel fusion recommendation method based on a self-attention mechanism.
Background
With the development of Internet and multimedia technology, Internet Protocol Television (IPTV) provides richer live tv channels, and as the number of live tv channels increases rapidly, users have to spend more time searching program contents interested in themselves during channel switching, and this random channel switching action also causes the reduction of user viewing experience and the waste of network resources. Therefore, helping users to improve the quality of the viewing experience through new technical means has also become an important research direction in the IPTV field.
At present, IPTV service providers mainly use Electronic Program Guide (EPG) to reduce the problem of switching delay of users in live tv service. An EPG is a channel list embedded in an IPTV hierarchical menu, where each entry in the list contains a channel identifier and description information of the tv program content that is currently being broadcast by the current channel at the current time. The user can fast browse the program information played by each channel by calling the EPG when needing to switch the channels, and although the time cost of continuously switching the channels by the user can be reduced to a certain extent by the mode, the current EPG menu is difficult to meet the personalized watching requirement of the user. Firstly, a user can find program contents in which the user is interested only by browsing each item in an EPG menu; secondly, the EPG menu information is provided to all users to be the same, which makes it difficult to satisfy the user's personalized needs. The recommendation system can capture the historical interest of the user so as to actively provide a personalized recommendation list for the user, and has been successful in the fields of videos, e-commerce, advertisements and the like. At first, many recommendation method researches pay more attention to the scene of IPTV video on demand, and pay less attention to the scene of recommending television live channels; secondly, the current IPTV live channel recommendation method is limited in the ability of capturing the user watching mode from the historical behavior record of the user watching channel, so that a better recommendation method is to be developed to deeply mine the user interest, and the performance of the live channel recommendation system is further improved.
Disclosure of Invention
In order to overcome the defects and shortcomings in the prior art, the invention provides the IPTV live channel fusion recommendation method based on the self-attention mechanism, which can improve the accuracy of IPTV live channel recommendation.
In order to achieve the purpose, the invention adopts the following technical scheme:
the invention provides an IPTV live channel fusion recommendation method based on a self-attention mechanism, which comprises the following steps of:
acquiring historical viewing log data of a user in terminal equipment as original data to form a multi-tuple;
preprocessing original data, sequencing the preprocessed original data according to the viewing starting time, and dividing the data into a training set, a verification set and a test set according to a set time window;
performing feature extraction operation on the preprocessed original data to construct deep learning features reflecting the watching behavior patterns of the users;
constructing an IPTV live broadcast channel fusion recommendation neural network model combined with a self-attention mechanism based on the preprocessed original data and the deep learning characteristics;
inputting the training set into an IPTV live channel fusion recommendation neural network model for training, adjusting the hyper-parameters of the IPTV live channel fusion recommendation neural network model through the verification set, and obtaining and storing the trained IPTV live channel fusion recommendation neural network model;
and inputting the test set into a trained IPTV live channel fusion recommendation neural network model to obtain an individualized channel recommendation list.
As a preferred technical solution, the preprocessing of the raw data specifically includes the steps of:
and filtering the viewing records of the missing fields and the abnormal fields in the original data and the data records with the viewing time length less than the set threshold value T.
As a preferred technical solution, the constructing of the deep learning feature reflecting the viewing behavior pattern of the user specifically includes:
and acquiring time context characteristics in the user watching data, and calculating the absolute difference value of the time characteristics related to any two channels in the user watching data to construct a relative time interval characteristic.
As a preferred technical solution, the calculating an absolute difference of time characteristics related to any two channels in the user viewing data to construct a relative time interval characteristic includes:
acquiring a channel sequence which is sequenced by users in a training set according to the ascending sequence of the watching starting time, and converting the channel sequence into a sequence with the fixed length of n, wherein the sequence comprises a sequence channel sequence, a time slot characteristic sequence and a week characteristic sequence;
constructing a relative time slot interval characteristic between any two input channels based on the time slot characteristic sequence;
the relative week interval signature between any two input channels is constructed based on a week signature sequence.
As a preferred technical solution, the method for constructing an IPTV live broadcast channel fusion recommendation neural network model combining a self-attention mechanism based on preprocessed original data and the deep learning features includes the specific steps of:
constructing a time-aware self-attention IPTV live channel recommendation neural network model based on a self-attention mechanism and the deep learning characteristics;
constructing a basic IPTV live broadcast channel recommendation model based on an implicit feedback characteristic statistical strategy based on the preprocessed original data;
the method comprises the steps of constructing an IPTV live broadcast channel fusion recommendation neural network model based on a self-attention IPTV live broadcast channel recommendation neural network model based on time perception and a basic IPTV live broadcast channel recommendation model based on an implicit feedback characteristic statistical strategy.
As a preferred technical solution, the building of a time-aware self-attention IPTV live broadcast channel recommendation neural network model specifically includes the steps of:
the input layer converts the discrete features into feature vectors through an embedded matrix;
constructing a time perception self-attention layer, wherein the constructed time perception attention layer captures a user implicit watching sequence mode through the fusion of a user historical watching channel sequence and the deep learning feature, and the method is specifically represented as follows:
Figure BDA0003477640010000041
wherein, giAn output vector representing a temporal perception self-attention layer,
Figure BDA0003477640010000042
indicating an input channel chjCharacteristic vector of (2), WV∈Rd×dA projection matrix representing the value vector in the self-attention layer,
Figure BDA0003477640010000043
a position vector representing the input channel is generated,
Figure BDA0003477640010000044
indicating an input channel chiAnd chjThe relative time interval between the feature vectors, d denotes the dimension of the feature vector, αijRepresenting the weight coefficient between inputs i and j, calculated by the softmax function as follows:
Figure BDA0003477640010000045
Figure BDA0003477640010000046
wherein the content of the first and second substances,
Figure BDA0003477640010000047
indicating an input channel chiCharacteristic vector of (2), WQ∈Rd×dA projection matrix representing the query vector in the self-attention layer,
Figure BDA0003477640010000048
indicating an input channel chjCharacteristic vector of (2), WK∈Rd×dA projection matrix representing the key vectors in the self-attention layer,
Figure BDA0003477640010000049
indicating an input channel chiAnd chjThe relative time interval between feature vectors, d represents the dimension of the feature vector;
constructing a forward connection network layer, wherein the formula of the layer is expressed as follows:
FFN(oi)＝(ReLU(oiW1+b1))W2+b2
wherein FFN (o)i) Output vector, o, representing the forward link network layeri∈RdInput vector, W, representing the Forward connection network layer1,W2∈Rd×dRepresenting a weight transformation matrix, b1,b2∈RdRepresenting a bias vector, ReLU representing an activation function;
the relative score of the channel c at the time step t is calculated by the prediction layer through a latent semantic model, and the specific formula is as follows:
Figure BDA00034776400100000410
wherein r isc,u,tRepresents the predicted score of user u for channel c at time step t for the predicted layer output,
Figure BDA0003477640010000051
a feature vector representing the channel c is shown,
Figure BDA0003477640010000052
representing the user's feature vector at t time steps obtained after forward connecting the network layers in S4.1.3.
As an optimal technical solution, in the process of constructing a time-aware self-attention IPTV live channel recommendation neural network model, the time-aware self-attention layer and the forward connection network layer further include using a residual connection, dropout regularization, and layer normalization optimization method.
As an optimal technical scheme, the method for constructing a basic IPTV live broadcast channel recommendation model based on an implicit feedback feature statistical strategy based on preprocessed original data includes the specific steps of:
the basic IPTV live broadcast channel recommendation model based on the implicit feedback characteristic statistical strategy comprises a basic IPTV live broadcast channel recommendation model which is combined with the implicit feedback characteristic statistical strategy of watching duration, watching frequency, channel switching frequency and channel switching frequency in a time slot;
acquiring a prediction score of a channel based on the total time length of the channel watched by a user in each specific time slot, wherein the specific formula is as follows:
Figure BDA0003477640010000053
wherein the content of the first and second substances,
Figure BDA0003477640010000054
represents the predicted score of user u for channel c at time t, tstartRepresenting the starting time of the history window, ts (T) representing the time slot in which the time T is located, d (u, c, T) representing the total time of the channel c viewed by the user u within the time period T;
normalizing the prediction scores;
the method comprises the following steps of obtaining a predicted score of a channel based on the total frequency of the channels watched by a user in each specific time slot, wherein the specific formula is as follows:
Figure BDA0003477640010000055
wherein the content of the first and second substances,
Figure BDA0003477640010000056
represents the predicted score of user u for channel c at time T, and f (u, c, T) represents user uThe total frequency of channel c viewed during time period T;
the method comprises the following steps of obtaining a predicted score of a channel based on the frequency of switching from a channel watched at the previous time to other channels in user historical data, wherein the specific formula is as follows:
Figure BDA0003477640010000061
wherein the content of the first and second substances,
Figure BDA0003477640010000062
represents the predicted score of the user u for the channel c at time t, c '→ c represents switching from the channel c' to the channel c, and Su(c′→c,[tstartT)) is represented as [ t ]startT) a set of switching actions from channel c' to channel c during a time period;
the method comprises the following steps of obtaining a predicted score of a channel based on the frequency of switching to other channels in a time slot of a channel watched last time in user historical data, wherein the specific formula is as follows:
Figure BDA0003477640010000063
wherein the content of the first and second substances,
Figure BDA0003477640010000064
denotes the predicted score of user u for channel c at time t, TS (c ') denotes the time slot of the channel c' viewed last, Su(c′→c,[tstartT) # TS (c')) is indicated at [ t [startT) a set of switching actions from channel c 'to channel c during the TS (c') time slot of the time period.
As an optimal technical scheme, the self-attention IPTV live channel recommendation neural network model based on time perception and the basic IPTV live channel recommendation model based on an implicit feedback characteristic statistical strategy construct an IPTV live channel fusion recommendation neural network model, and the specific steps include:
the fusion layer constructs a two-layer fully-connected network for obtaining the self-adaptive weight of the basic recommendation model, and the specific formula is as follows:
h1＝σ(W1v1+b1)
β＝σ(W2h1+b2)
wherein v is1∈R3d×1Representing the input eigenvector, W, resulting from concatenation of the user eigenvector, the time-slot eigenvector, and the week eigenvector1∈Rd×3dTransformation matrix representing a fully connected network of the first layer, b1∈Rd×1Represents a bias vector of the first layer of fully-connected network, sigma (-) represents a sigmoid activation function, h1An output vector representing a first layer of fully connected network; w2∈Rd×3dTransformation matrix representing a fully connected network of the second layer, b2∈Rd×1A bias vector representing the second layer of fully-connected network, β ∈ Rk×1Representing an output vector of the second layer of the fully-connected network, k representing a dimension of the output vector;
the fusion prediction layer calculates the relative score of the channel c at the time step t through a latent semantic model, and the specific formula is as follows:
Figure BDA0003477640010000071
wherein R isc,u,tRepresents the prediction score r of the user u output by the fusion prediction layer on the channel c at the time step tc,u,tRepresents the output, β, of the prediction layer of step S4.1.4iThe adaptive weights of the underlying recommendation model representing the fused layer output,
Figure BDA0003477640010000072
method for representing basic recommendation modeliThe output predicted score of user u for channel c at time t'.
As an optimal technical scheme, the training set is input into an IPTV live channel fusion recommendation neural network model for training, the IPTV live channel fusion recommendation neural network model performs network training based on a constructed point-wise ordering loss function, when a loss value does not decrease any more, the network training is stopped and the network is saved, and a specific loss function formula is as follows:
Figure BDA0003477640010000073
where S represents the total sample set in the training set, otThe method comprises the steps of representing positive samples in a training set, sigma (.) representing a sigmoid activation function, lambda representing a regularization coefficient, and theta representing all parameters in an IPTV live channel fusion recommendation neural network model.
Compared with the prior art, the invention has the following advantages and beneficial effects:
(1) the method is characterized in that an IPTV live channel fusion recommendation neural network model is established based on an attention mechanism aiming at the watching data characteristics of IPTV user live channels, a behavior sequence mode of a user historical watching channel is captured through a time-aware self-attention network module, and the implicit feedback characteristics of the user watching behaviors are further mined through a basic recommendation module based on a statistical strategy.
(2) The training and predicting process of the method provided by the invention only needs historical channel data watched by IPTV users, does not need to maintain program source data information played by any channel, and the self-attention network has parallel computing capability, so that compared with the method provided by the prior art, the method provided by the invention consumes less resources and time.
Drawings
Fig. 1 is a schematic flow chart of an IPTV live broadcast channel fusion recommendation method based on a self-attention mechanism according to the present invention;
FIG. 2 is a schematic flow chart of the training set, the verification set and the test set partitioning according to the present invention;
FIG. 3 is a schematic view of a process of obtaining deep learning features of a user viewing behavior pattern according to the present invention;
fig. 4 is an architecture diagram of a time-aware self-attention IPTV live channel recommendation neural network model according to the present invention;
fig. 5 is an overall architecture diagram of the IPTV live broadcast channel fusion recommendation method based on the self-attention mechanism according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Examples
As shown in fig. 1, this embodiment provides an IPTV live channel fusion recommendation method based on a self-attention mechanism, including the following steps:
s1, acquiring historical viewing log data of a user in the terminal equipment as original data to form a multi-tuple, wherein the data content includes but is not limited to < user number, channel number, start time and end time >;
s2, preprocessing the original data, including filtering the missing field data, abnormal field data and noise data with short watching time in the channel switching process; the method comprises the following steps of sequencing preprocessed user data according to the viewing starting time, and dividing the data into a training set, a verification set and a test set according to a set time window, wherein the specific process is as follows:
filtering viewing records containing missing fields and abnormal fields in original data and data records with viewing duration less than a given threshold value T; the channel sequence of the users in the training set after the users sort in the ascending order of the watching starting time is obtained as follows:
Figure BDA0003477640010000091
wherein any one tuple
Figure BDA0003477640010000092
The channel number indicating that user u watched at the t-th time step is
Figure BDA0003477640010000093
Is and watches channel cht uThe number of the relevant time slot is,
Figure BDA0003477640010000094
is and watches the channel
Figure BDA0003477640010000095
The relevant week number; channel sequence SuConverting into fixed length sequence channel sequence ch ═ ch (ch)1,ch2,...,chn) Time slot signature sequence ts ═ ts1,ts2,...,tsn) And week signature sequence dw ═ dw (dw)1,dw2,...,dwn) To satisfy the input of the sequential neural network model Transformer, where n is the maximum input length of the model, and specifically reserve the channel sequence SuFor the most recent n channel records, sequences greater than n will be truncated, and sequences less than n will be padded up to length n;
as shown in fig. 2, in this embodiment, the training window for viewing the data set of the user live channel with a total duration of 31 days is set to 21 days, the verification set window is set to 5 days, and the test set window is set to 5 days; the training set data is used for training the neural network model, the verification set data is used for adjusting key hyper-parameters in the neural network, and the test set data is used for evaluating the performance of the neural network model.
S3, as shown in fig. 3, performing a feature extraction operation on the preprocessed user viewing data, obtaining a deep learning feature that can reflect a user viewing behavior pattern, obtaining a time context feature in the user viewing data, and constructing a relative time interval feature by calculating an absolute difference of time features related to any two channels in the user viewing data, where the specific process is as follows:
s3.1. sequence ts ═ ts (ts) based on time slot characteristics1,ts2,...,tsn) Constructing a relative time slot interval characteristic between any two input frequency channels, wherein the characteristic matrix is expressed as follows:
Figure BDA0003477640010000101
any one of them
Figure BDA0003477640010000102
Indicating the channel chiAnd chjRelative slot spacing characteristics in between;
s3.2 week-based signature sequence dw ═ S (dw)1,dw2,...,dwn) Constructing a relative week interval signature between any two input channels, the signature matrix being represented as follows:
Figure BDA0003477640010000103
any one of them
Figure BDA0003477640010000104
Indicating the channel chiAnd chjRelative week interval characteristics in between;
s4, constructing an IPTV live broadcast channel fusion recommendation neural network model combining a self-attention mechanism based on the user viewing data and deep learning features after preprocessing operation, wherein the specific process is as follows:
s4.1, constructing a time-aware self-attention IPTV live broadcast channel recommendation neural network model based on a self-attention mechanism and deep learning characteristics;
as shown in fig. 4, the method for constructing a time-aware self-attention IPTV live channel recommendation neural network model architecture includes the following steps:
s4.1.1, input layer: converting the discrete features into feature vectors by embedding the matrix; specifically, for the channel sequence ch ═ (ch)1,ch2,...,chn) The channel characteristic matrix output after mapping is
Figure BDA0003477640010000105
Any one of them
Figure BDA0003477640010000106
d is the dimension of the feature vector,r is a real number domain; by a feature matrix
Figure BDA0003477640010000107
And
Figure BDA0003477640010000108
adding position characteristic information for key and value vectors in a self-attention mechanism respectively, wherein any element is
Figure BDA0003477640010000109
Relative time slot interval characteristic matrix RT in step S3.1uThe feature vector matrix output after mapping is:
Figure BDA0003477640010000111
any one of them
Figure BDA0003477640010000112
The mapping maps the feature matrix RTuThe discretization characteristic value in (a) is encoded into a characteristic vector, and then the characteristic vector is used as an input part of S4.1.2 time perception self-attention layer for extracting the relative time slot interval characteristics of any two channels in the input channel sequence ch;
relative week interval signature matrix RW in step S3.2uThe feature vector matrix output after mapping is:
Figure BDA0003477640010000113
any one of them
Figure BDA0003477640010000114
The mapping maps the feature matrix RWuThe discretized feature values in (a) are encoded into feature vectors, and then used as an input part of S4.1.2 time perception self-attention layers for extracting the relative week interval features of any two channels in the input channel sequence ch;
s4.1.2, temporal perception self-attention tier: the original self-attention layer can capture the watching interest of the user through the input historical watching channel sequence of the user without considering the auxiliary context characteristics related to the input; the constructed temporal awareness attention layer jointly captures a user implicit viewing sequence mode by fusing the user historical viewing channel sequence with the temporal context characteristics of step S3, and the formula of the layer is as follows:
Figure BDA0003477640010000115
wherein, giAn output vector representing a temporal perception self-attention layer,
Figure BDA0003477640010000116
indicating an input channel chjCharacteristic vector of (2), WV∈Rd×dA projection matrix representing the value vector in the self-attention layer,
Figure BDA0003477640010000117
a position vector representing the input channel is generated,
Figure BDA0003477640010000118
indicating an input channel chiAnd chjThe relative time interval between the feature vectors, d denotes the dimension of the feature vector, αijRepresenting the weight coefficient between inputs i and j, calculated by the softmax function as follows:
Figure BDA0003477640010000121
Figure BDA0003477640010000122
wherein the content of the first and second substances,
Figure BDA0003477640010000123
presentation inputChannel chiCharacteristic vector of (2), WQ∈Rd×dA projection matrix representing the query vector in the self-attention layer,
Figure BDA0003477640010000124
indicating an input channel chjCharacteristic vector of (2), WK∈Rd×dA projection matrix representing the key vectors in the self-attention layer,
Figure BDA0003477640010000125
indicating an input channel chiAnd chjThe relative time interval between feature vectors, d represents the dimension of the feature vector;
specifically, the value vector of the original self-attention layer is added with the feature vector of the temporal context feature coded in step S3 through step S4.1.1
Figure BDA0003477640010000126
And
Figure BDA0003477640010000127
adding the feature vector of the temporal context feature coded in the step S4.1.1 and described in the step S3 to the key vector of the original self-attention layer
Figure BDA0003477640010000128
And
Figure BDA0003477640010000129
the layer formula is expressed as follows:
Figure BDA00034776400100001210
wherein g isiAn output vector representing a temporal perception self-attention layer,
Figure BDA00034776400100001211
indicating an input channel chjCharacteristic vector of (2), WV∈Rd×dIs the projection of value vectors in the self-attention layerThe matrix is a matrix of a plurality of matrices,
Figure BDA00034776400100001212
a position vector representing the input channel is generated,
Figure BDA00034776400100001213
indicating an input channel chiAnd chjThe relative slot interval feature vector in between,
Figure BDA00034776400100001214
indicating an input channel chiAnd chjRelative week interval eigenvectors between, alphaijRepresenting the weight coefficient between inputs i and j, calculated by the softmax function as follows:
Figure BDA00034776400100001215
Figure BDA0003477640010000131
wherein the content of the first and second substances,
Figure BDA0003477640010000132
indicating an input channel chiCharacteristic vector of (2), WQ∈Rd×dA projection matrix representing the query vector in the self-attention layer,
Figure BDA0003477640010000133
indicating an input channel chjCharacteristic vector of (2), WK∈Rd×dA projection matrix representing the key vectors in the self-attention layer,
Figure BDA0003477640010000134
indicating an input channel chiAnd chjThe relative slot interval feature vector in between,
Figure BDA0003477640010000135
indicating an input channel chiAnd chjThe relative week interval eigenvectors in between, d is the dimension of the eigenvector;
in the present embodiment, wherein
Figure BDA0003477640010000136
Indicating generalized features. Two features are specified in step S3, so that
Figure BDA0003477640010000137
These symbols are substituted
Figure BDA0003477640010000138
S4.1.3, Forward connection network layer: the nonlinear fitting capability of the model is improved by introducing two layers of forward connection networks, and the layer of formula is specifically expressed as follows:
FFN(oi)＝(ReLU(oiW1+b1))W2+b2
wherein, FFN (o)i) Output vector, o, representing the forward link network layeri∈RdInput vector, W, representing the Forward connection network layer1,W2∈Rd×dRepresenting a weight transformation matrix, b1,b2∈RdRepresenting a bias vector, ReLU representing an activation function;
in the process of constructing the time-aware self-attention IPTV live channel recommendation neural network model, the time-aware self-attention layer and the forward connection network layer further comprise the capacity of improving the model by using a residual connection method, a dropout regularization method and a layer normalization optimization method; specifically, residual connection is used for transmitting characteristic information of a low-level network to a deep network, a drop regularization technology is used for relieving an overfitting problem, layer normalization is used for improving training efficiency and model performance, and a formula is expressed as follows:
ui＝xi+Dropout(g(LayerNorm(xi)))
wherein x isiRepresenting the input feature vector of each module at time step t, LayerNorm (.) representing the layer normalization method,g (.) denotes a time-aware self-attention layer or a forward connection network layer, Dropout (.) denotes a regularization method;
s4.1.4, prediction layer: calculating the relative score of the channel c at the time step t through a latent semantic model, wherein the specific formula is expressed as follows:
Figure BDA0003477640010000141
wherein r isc,u,tRepresents the predicted score of user u for channel c at time step t for the predicted layer output,
Figure BDA0003477640010000142
a feature vector representing the channel c is shown,
Figure BDA0003477640010000143
representing the user characteristic vector of t time step obtained after forward connection of S4.1.3 network layer;
s4.2, based on the user watching data obtained after the preprocessing operation in the S2, four basic IPTV live broadcast channel recommendation models based on the implicit feedback characteristic statistical strategy are constructed, specifically: the method comprises the following steps of constructing a basic IPTV live broadcast channel recommendation model combining an implicit feedback characteristic statistical strategy of watching duration, watching frequency, channel switching frequency and channel switching frequency in a time slot, wherein the basic IPTV live broadcast channel recommendation model comprises the following steps:
s4.2.1, obtaining a predicted score (psd) of the channel based on the total time length of the channel watched by the user in each specific time slot, wherein the specific formula is as follows:
Figure BDA0003477640010000144
wherein the content of the first and second substances,
Figure BDA0003477640010000145
represents the predicted score of user u for channel c at time t, tstartShowing the start time of the history window, TS (t) showing the time at which time t isSlot, d (u, c, T) represents the total time channel c was viewed by user u during time period T;
and (3) normalizing the prediction score in S4.2.1 based on a min-max normalization mode, wherein a specific formula is expressed as follows:
Figure BDA0003477640010000146
wherein the content of the first and second substances,
Figure BDA0003477640010000147
represents a vector of prediction scores of the user u for all channels, x represents a min-max normalized minimum mapping value, y represents a min-max normalized maximum mapping value,
Figure BDA0003477640010000151
representing a vector
Figure BDA0003477640010000152
The value of the smallest element in (a),
Figure BDA0003477640010000153
representing a vector
Figure BDA0003477640010000154
Maximum element value of;
s4.2.2, obtaining a predicted score (psf) of the channel based on the total Frequency of the channel watched by the user in each specific time slot, wherein the specific formula is as follows:
Figure BDA0003477640010000155
wherein the content of the first and second substances,
Figure BDA0003477640010000156
representing the predicted score of user u for channel c at time T, and f (u, c, T) representing the total frequency of channel c viewed by user u during time period T;
s4.2.3, obtaining a predicted score (pct) of a Channel based on the frequency of switching from a Channel watched last time to other channels in the user history data, and the specific formula is expressed as follows:
Figure BDA0003477640010000157
wherein the content of the first and second substances,
Figure BDA0003477640010000158
represents the predicted score of the user u for the channel c at time t, c '→ c represents switching from the channel c' to the channel c, and Su(c′→c,[tstartT)) is represented as [ t ]startT) a set of switching actions from channel c' to channel c within a time period;
s4.2.4, obtaining a predicted score (psct) of the Channel based on the frequency of switching to other channels in the time slot of the Channel watched last time in the user history data, wherein the specific formula is as follows:
Figure BDA0003477640010000159
wherein
Figure BDA00034776400100001510
Denotes the predicted score of user u for channel c at time t, TS (c ') denotes the time slot of the channel c' viewed last, Su(c′→c,[tstartT) # TS (c')) is indicated at [ t [startT) a set of switching actions from channel c 'to channel c during the TS (c') time slot of the time segment;
s4.3, constructing an IPTV live broadcast channel fusion recommendation neural network model based on the time-aware self-attention IPTV live broadcast channel recommendation neural network model obtained in S4.1 and four basic IPTV live broadcast channel recommendation models based on implicit feedback characteristic statistical strategies obtained in S4.2; the self-adaptive weight coefficients of the four basic recommendation models are obtained by network joint training;
as shown in fig. 5, the construction of the fusion recommendation neural network model for the IPTV live broadcast channels in this embodiment includes the following steps:
s4.3.1, fusion layer: constructing a two-layer fully-connected network for obtaining the self-adaptive weights of the four basic recommendation models in S4.2, wherein a specific formula is expressed as follows:
h1＝σ(W1v1+b1)
β＝σ(W2h1+b2)
wherein v is1∈R3d×1Representing the input eigenvector, W, resulting from concatenation of the user eigenvector, the time-slot eigenvector, and the week eigenvector1∈Rd×3dTransformation matrix representing a fully connected network of the first layer, b1∈Rd×1Represents a bias vector of the first layer of fully-connected network, sigma (-) represents a sigmoid activation function, h1An output vector representing a first layer of fully connected network; w2∈Rd×3dTransformation matrix representing a fully connected network of the second layer, b2∈Rd×1Offset vector representing the second layer fully-connected network, β ═ β (β)1,β2,β3,β4)TAn output vector representing a second layer of fully connected network;
s4.3.2, fusion prediction layer: calculating the relative score of the channel c at the time step t through a latent semantic model, wherein the specific formula is expressed as follows:
Figure BDA0003477640010000161
wherein R isc,u,tRepresents the prediction score r of the user u output by the fusion prediction layer on the channel c at the time step tc,u,tRepresents the output, β, of the prediction layer of step S4.1.4iRepresents the adaptive weights of the underlying recommendation model of the fused layer output of said step S4.3.1,
Figure BDA0003477640010000162
representing the basic recommendation model Method described in step S4.2iThe output predicted score of the user u on the channel c at the time t';
further, in step S5, the IPTV live broadcast channel fusion recommendation neural network model performs network training based on the constructed point-wise ranking loss function, and when the loss value does not decrease any more, the network training is stopped and the network is saved, where the specific loss function formula is expressed as follows:
Figure BDA0003477640010000171
wherein S represents the total sample set in the training set in step S2, otRepresenting positive samples in the training set in the step S2, sigma (·) representing a sigmoid activation function, lambda representing a regularization coefficient, and theta representing all parameters in the IPTV live channel fusion recommendation neural network model in the step S4;
the specific formula is expressed as follows:
Figure BDA0003477640010000172
wherein R isc,u,tRepresents the prediction score r of the user u output by the fusion prediction layer on the channel c at the time step tc,u,tRepresents the output, β, of the predicted layer of step S4.1.41,β2,β3,β4The self-adaptive weights of the four recommended models psd, psf, pct and psct in S4.2 are respectively set;
in the present embodiment, the formula is used
Figure BDA0003477640010000173
MethodiUsed to represent the underlying recommendation method of generalization. While describing the specific four basic recommendation methods psd, psf, pct, psct, a Method to generalizeiA specific basic recommendation method is substituted.
S5, inputting the training set divided in the step S2 into an IPTV live channel fusion recommendation neural network model for training, stopping network training when the loss value does not decrease any more, adjusting hyper-parameters of the IPTV live channel fusion recommendation neural network model through the verification set divided in the step S2, and obtaining and storing the trained neural network model;
the IPTV live broadcast channel in this embodiment fuses the recommendation neural network model to perform network training based on the constructed point-wise ranking loss function, and the specific formula is expressed as follows:
Figure BDA0003477640010000174
wherein S represents the total sample set in the training set divided in step S2, otPositive samples in the training set divided in step S2 are represented, σ (·) represents a sigmoid activation function, λ represents a regularization coefficient, and Θ represents all parameters in the IPTV live broadcast channel fusion recommendation neural network model in S4.
S6, inputting the test set sample into the trained IPTV live broadcast channel fusion recommendation neural network model to obtain an individualized channel recommendation list;
performing deep learning feature extraction on the test set sample divided in the step S2 through S3, inputting channel data and feature data into the IPTV live broadcast channel fusion recommendation neural network trained in S5, and specifically obtaining adaptive weights β of the four basic recommendation models in S4.2 according to the fusion layer in the step S4.3.11,β2,β3,β4And after the values are obtained, the prediction scores of all the channels are output according to a fusion prediction layer formula in S4.3.2, and finally, the prediction scores of all the channels are arranged in a descending order to obtain a personalized channel recommendation list.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims (10)

1. An IPTV live channel fusion recommendation method based on a self-attention mechanism is characterized by comprising the following steps:
acquiring historical viewing log data of a user in terminal equipment as original data to form a multi-tuple;
preprocessing original data, sequencing the preprocessed original data according to the viewing starting time, and dividing the data into a training set, a verification set and a test set according to a set time window;
performing feature extraction operation on the preprocessed original data to construct deep learning features reflecting the watching behavior patterns of the users;
constructing an IPTV live broadcast channel fusion recommendation neural network model combined with a self-attention mechanism based on the preprocessed original data and the deep learning characteristics;
inputting the training set into an IPTV live channel fusion recommendation neural network model for training, adjusting the hyper-parameters of the IPTV live channel fusion recommendation neural network model through the verification set, and obtaining and storing the trained IPTV live channel fusion recommendation neural network model;
and inputting the test set into a trained IPTV live channel fusion recommendation neural network model to obtain an individualized channel recommendation list.
2. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 1, wherein the pre-processing of the original data comprises the following specific steps:
and filtering the viewing records of the missing fields and the abnormal fields in the original data and the data records with the viewing time length less than the set threshold value T.
3. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 1, wherein said constructing deep learning features reflecting the watching behavior pattern of the user comprises the following steps:
and acquiring time context characteristics in the user watching data, and calculating the absolute difference value of the time characteristics related to any two channels in the user watching data to construct a relative time interval characteristic.
4. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 3, wherein said calculating an absolute difference of time characteristics related to any two channels in the user viewing data to construct a relative time interval characteristic comprises the following specific steps:
acquiring a channel sequence which is sequenced by users in a training set according to the ascending sequence of the watching starting time, and converting the channel sequence into a sequence with the fixed length of n, wherein the sequence comprises a sequence channel sequence, a time slot characteristic sequence and a week characteristic sequence;
constructing a relative time slot interval characteristic between any two input channels based on the time slot characteristic sequence;
the relative week interval signature between any two input channels is constructed based on a week signature sequence.
5. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 1, wherein the IPTV live channel fusion recommendation neural network model combined with the self-attention mechanism is constructed based on the preprocessed original data and the deep learning features, and the specific steps include:
constructing a time-aware self-attention IPTV live channel recommendation neural network model based on a self-attention mechanism and the deep learning characteristics;
constructing a basic IPTV live broadcast channel recommendation model based on an implicit feedback characteristic statistical strategy based on the preprocessed original data;
the method comprises the steps of constructing an IPTV live broadcast channel fusion recommendation neural network model based on a self-attention IPTV live broadcast channel recommendation neural network model based on time perception and a basic IPTV live broadcast channel recommendation model based on an implicit feedback characteristic statistical strategy.
6. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 5, wherein the building of the time-aware self-attention IPTV live channel recommendation neural network model comprises the following specific steps:
the input layer converts the discrete features into feature vectors through an embedded matrix;
constructing a time perception self-attention layer, wherein the constructed time perception attention layer captures a user implicit watching sequence mode through the fusion of a user historical watching channel sequence and the deep learning feature, and the method is specifically represented as follows:
Figure FDA0003477638000000031
wherein, giAn output vector representing a temporal perception self-attention layer,
Figure FDA0003477638000000032
indicating an input channel chjCharacteristic vector of (2), WV∈Rd×dA projection matrix representing the value vector in the self-attention layer,
Figure FDA0003477638000000033
a position vector representing the input channel is generated,
Figure FDA0003477638000000034
indicating an input channel chiAnd chjThe relative time interval between the feature vectors, d denotes the dimension of the feature vector, αijRepresenting the weight coefficient between inputs i and j, calculated by the softmax function as follows:
Figure FDA0003477638000000035
Figure FDA0003477638000000036
wherein the content of the first and second substances,
Figure FDA0003477638000000037
representing input frequencyRoad chiCharacteristic vector of (2), WQ∈Rd×dA projection matrix representing the query vector in the self-attention layer,
Figure FDA0003477638000000038
indicating an input channel chjCharacteristic vector of (2), WK∈Rd×dA projection matrix representing the key vectors in the self-attention layer,
Figure FDA0003477638000000039
indicating an input channel chiAnd chjThe relative time interval between feature vectors, d represents the dimension of the feature vector;
constructing a forward connection network layer, wherein the formula of the layer is expressed as follows:
FFN(oi)＝(ReLU(oiW1+b1))W2+b2
wherein, FFN (o)i) Output vector, o, representing the forward link network layeri∈RdInput vector, W, representing the Forward connection network layer1,W2∈Rd×dRepresenting a weight transformation matrix, b1,b2∈RdRepresenting a bias vector, ReLU representing an activation function;
the prediction layer calculates the relative score of the channel c at the time step t through a latent semantic model, and the specific formula is as follows:
Figure FDA00034776380000000310
wherein r isc,u,tRepresents the predicted score of user u for channel c at time step t for the predicted layer output,
Figure FDA0003477638000000041
a feature vector representing the channel c is shown,
Figure FDA0003477638000000042
indicating forward link network layer backtrackingUser feature vectors at the time step of arrival t.
7. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 5, wherein in the process of constructing the time-aware self-attention IPTV live channel recommendation neural network model, the time-aware self-attention layer and the forward connection network layer further comprise using a residual connection, dropout regularization and layer normalization optimization method.
8. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 5, wherein the basic IPTV live channel recommendation model based on the implicit feedback feature statistical strategy is constructed based on the preprocessed original data, and the specific steps include:
the basic IPTV live broadcast channel recommendation model based on the implicit feedback characteristic statistical strategy comprises a basic IPTV live broadcast channel recommendation model which is combined with the implicit feedback characteristic statistical strategy of watching duration, watching frequency, channel switching frequency and channel switching frequency in a time slot;
acquiring a prediction score of a channel based on the total time length of the channel watched by a user in each specific time slot, wherein the specific formula is as follows:
Figure FDA0003477638000000043
wherein the content of the first and second substances,
Figure FDA0003477638000000044
represents the predicted score of user u for channel c at time t, tstartRepresenting the starting time of the history window, ts (T) representing the time slot in which the time T is located, d (u, c, T) representing the total time of the channel c viewed by the user u within the time period T;
normalizing the prediction scores;
the method comprises the following steps of obtaining a predicted score of a channel based on the total frequency of the channels watched by a user in each specific time slot, wherein the specific formula is as follows:
Figure FDA0003477638000000045
wherein the content of the first and second substances,
Figure FDA0003477638000000046
representing the predicted score of user u for channel c at time T, and f (u, c, T) representing the total frequency of channel c viewed by user u during time period T;
the method comprises the following steps of obtaining a predicted score of a channel based on the frequency of switching from a channel watched at the previous time to other channels in user historical data, wherein the specific formula is as follows:
Figure FDA0003477638000000051
wherein the content of the first and second substances,
Figure FDA0003477638000000052
represents the predicted score of the user u for the channel c at time t, c '→ c represents switching from the channel c' to the channel c, and Su(c′→c,[tstartT)) is represented as [ t ]startT) a set of switching actions from channel c' to channel c during a time period;
the method comprises the following steps of obtaining a predicted score of a channel based on the frequency of switching to other channels in a time slot of a channel watched last time in user historical data, wherein the specific formula is as follows:
Figure FDA0003477638000000053
wherein the content of the first and second substances,
Figure FDA0003477638000000054
denotes the predicted score of user u for channel c at time t, TS (c ') denotes the time slot of the channel c' viewed last, Su(c′→c,[tstartT) # TS (c')) is indicated at [ t [startT) a set of switching actions from channel c 'to channel c during the TS (c') time slot of the time period.
9. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 5, wherein the IPTV live channel fusion recommendation neural network model is constructed by the self-attention IPTV live channel recommendation neural network model based on time perception and the basic IPTV live channel recommendation model based on the implicit feedback feature statistical strategy, and the specific steps include:
the fusion layer constructs a two-layer fully-connected network for obtaining the self-adaptive weight of the basic recommendation model, and the specific formula is as follows:
h1＝σ(W1v1+b1)
β＝σ(W2h1+b2)
wherein v is1∈R3d×1Representing the input eigenvector, W, resulting from concatenation of the user eigenvector, the time-slot eigenvector, and the week eigenvector1∈Rd×3dTransformation matrix representing a fully connected network of the first layer, b1∈Rd×1Represents a bias vector of the first layer of fully-connected network, sigma (-) represents a sigmoid activation function, h1An output vector representing a first layer of fully connected network; w2∈Rd×3dTransformation matrix representing a fully connected network of the second layer, b2∈Rd×1A bias vector representing the second layer of fully-connected network, β ∈ Rk×1Representing an output vector of the second layer of the fully-connected network, k representing a dimension of the output vector;
the fusion prediction layer calculates the relative score of the channel c at the time step t through a latent semantic model, and the specific formula is as follows:
Figure FDA0003477638000000061
wherein R isc,u,tRepresents the prediction score r of the user u output by the fusion prediction layer on the channel c at the time step tc,u,tRepresents the output, β, of the prediction layer of step S4.1.4iThe adaptive weights of the underlying recommended model representing the fused layer output,
Figure FDA0003477638000000062
method for representing basic recommendation modeliThe output predicted score of user u for channel c at time t'.
10. The IPTV live channel fusion recommendation method based on the self-attention mechanism as claimed in claim 1, wherein the training set is input into an IPTV live channel fusion recommendation neural network model for training, the IPTV live channel fusion recommendation neural network model performs network training based on a constructed point-wise ordering loss function, when a loss value does not decrease any more, the network training is stopped and the network is saved, and a specific loss function formula is as follows:
Figure FDA0003477638000000063
where S represents the total sample set in the training set, otThe method comprises the steps of representing positive samples in a training set, sigma (.) representing a sigmoid activation function, lambda representing a regularization coefficient, and theta representing all parameters in an IPTV live channel fusion recommendation neural network model.
CN202210059426.9A 2022-01-19 2022-01-19 IPTV live channel fusion recommendation method based on self-attention mechanism Active CN114528434B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210059426.9A CN114528434B (en) 2022-01-19 2022-01-19 IPTV live channel fusion recommendation method based on self-attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210059426.9A CN114528434B (en) 2022-01-19 2022-01-19 IPTV live channel fusion recommendation method based on self-attention mechanism

Publications (2)

Publication Number Publication Date
CN114528434A true CN114528434A (en) 2022-05-24
CN114528434B CN114528434B (en) 2023-07-21

Family

ID=81620981

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210059426.9A Active CN114528434B (en) 2022-01-19 2022-01-19 IPTV live channel fusion recommendation method based on self-attention mechanism

Country Status (1)

Country Link
CN (1) CN114528434B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116320511A (en) * 2023-02-03 2023-06-23 华南理工大学 Cross-domain fusion recommendation method based on graph convolution network
CN117333950A (en) * 2023-11-30 2024-01-02 苏州元脑智能科技有限公司 Action generation method, device, computer equipment and storage medium

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108347652A (en) * 2018-02-24 2018-07-31 华南理工大学 A kind of method and system for recommending IPTV direct broadcast bands using artificial neural network
CN111461175A (en) * 2020-03-06 2020-07-28 西北大学 Label recommendation model construction method and device of self-attention and cooperative attention mechanism
CN112579914A (en) * 2020-12-31 2021-03-30 安徽大学 Time perception service recommendation system and method based on self-attention factorization machine
WO2021169367A1 (en) * 2020-02-27 2021-09-02 深圳大学 Multi-layer attention based recommendation method
CN113822742A (en) * 2021-09-18 2021-12-21 电子科技大学 Recommendation method based on self-attention mechanism

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108347652A (en) * 2018-02-24 2018-07-31 华南理工大学 A kind of method and system for recommending IPTV direct broadcast bands using artificial neural network
WO2021169367A1 (en) * 2020-02-27 2021-09-02 深圳大学 Multi-layer attention based recommendation method
CN111461175A (en) * 2020-03-06 2020-07-28 西北大学 Label recommendation model construction method and device of self-attention and cooperative attention mechanism
CN112579914A (en) * 2020-12-31 2021-03-30 安徽大学 Time perception service recommendation system and method based on self-attention factorization machine
CN113822742A (en) * 2021-09-18 2021-12-21 电子科技大学 Recommendation method based on self-attention mechanism

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
GUANGYU LI等: "IPTV Channel Zapping Recommendation With Attention Mechanism", 《IEEE TRANSACTIONS ON MULTIMEDIA》 *
GUANGYU LI等: "IPTV Channel Zapping Recommendation With Attention Mechanism", 《IEEE TRANSACTIONS ON MULTIMEDIA》, vol. 23, 23 February 2021 (2021-02-23), pages 538 - 549, XP011835277, DOI: 10.1109/TMM.2020.2984094 *
郭景峰;朱晓松;李爽;: "直播电视推荐系统的评分预测算法研究", 电子学报, no. 09, pages 73 - 78 *
陈细玉等: "一种融合注意力和记忆网络的序列推荐算法", 《信息技术与网络安全》 *
陈细玉等: "一种融合注意力和记忆网络的序列推荐算法", 《信息技术与网络安全》, no. 05, 10 May 2020 (2020-05-10), pages 41 - 45 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116320511A (en) * 2023-02-03 2023-06-23 华南理工大学 Cross-domain fusion recommendation method based on graph convolution network
CN117333950A (en) * 2023-11-30 2024-01-02 苏州元脑智能科技有限公司 Action generation method, device, computer equipment and storage medium
CN117333950B (en) * 2023-11-30 2024-03-12 苏州元脑智能科技有限公司 Action generation method, device, computer equipment and storage medium

Also Published As

Publication number Publication date
CN114528434B (en) 2023-07-21

Similar Documents

Publication Publication Date Title
CN109874053B (en) Short video recommendation method based on video content understanding and user dynamic interest
US20070288965A1 (en) Recommended program information providing method and apparatus
CN111565316B (en) Video processing method, video processing device, computer equipment and storage medium
CN112732936B (en) Radio and television program recommendation method based on knowledge graph and user microscopic behaviors
EP2817970B1 (en) Automatically recommending content
CN114528434B (en) IPTV live channel fusion recommendation method based on self-attention mechanism
CN112597395A (en) Object recommendation method, device, equipment and storage medium
CN111339433A (en) Information recommendation method and device based on artificial intelligence and electronic equipment
CN113207010B (en) Model training method, live broadcast recommendation method, device and storage medium
CN102118639A (en) Information processing apparatus, information processing method, and program
CN112989209B (en) Content recommendation method, device and storage medium
CN112256980A (en) Dynamic graph attention network-based multi-relation collaborative filtering recommendation
CN112100504B (en) Content recommendation method and device, electronic equipment and storage medium
Yang et al. Personalized channel recommendation deep learning from a switch sequence
CN111046298A (en) Method and device for pushing application program, computer equipment and storage medium
CN112434184A (en) Deep interest network sequencing method based on historical movie posters
CN114595383A (en) Marine environment data recommendation method and system based on session sequence
CN107368584B (en) Personalized video recommendation method and system
CN114357201B (en) Audio-visual recommendation method and system based on information perception
CN117197722B (en) User perception and analysis system based on mobile internet video
CN113836406A (en) Information flow recommendation method and device
CN115640449A (en) Media object recommendation method and device, computer equipment and storage medium
CN116662656A (en) Movie recommendation method based on collaborative enhancement and graph annotation intention neural network
CN117494762A (en) Training method of student model, material processing method, device and electronic equipment
CN114357301A (en) Data processing method, device and readable storage medium

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