CN114025205A - Intelligent recommendation method for home TV video - Google Patents

Intelligent recommendation method for home TV video Download PDF

Info

Publication number
CN114025205A
CN114025205A CN202111287347.5A CN202111287347A CN114025205A CN 114025205 A CN114025205 A CN 114025205A CN 202111287347 A CN202111287347 A CN 202111287347A CN 114025205 A CN114025205 A CN 114025205A
Authority
CN
China
Prior art keywords
user
content
representing
interest
current
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111287347.5A
Other languages
Chinese (zh)
Inventor
赵心炜
侯永宏
刘传玉
王辉
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.)
Tianjin University
Original Assignee
Tianjin University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tianjin University filed Critical Tianjin University
Priority to CN202111287347.5A priority Critical patent/CN114025205A/en
Publication of CN114025205A publication Critical patent/CN114025205A/en
Pending legal-status Critical Current

Links

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
    • H04N21/252—Processing of multiple end-users' preferences to derive collaborative data
    • 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
    • G06F16/9536—Search customisation based on social or collaborative filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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]
    • H04N21/20—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
    • H04N21/258—Client or end-user data management, e.g. managing client capabilities, user preferences or demographics, processing of multiple end-users preferences to derive collaborative data
    • H04N21/25866—Management of end-user data
    • H04N21/25891—Management of end-user data being end-user preferences
    • 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
    • H04N21/4508—Management of client data or end-user data
    • H04N21/4532—Management of client data or end-user data involving end-user characteristics, e.g. viewer profile, preferences
    • 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
    • H04N21/4661—Deriving a combined profile for a plurality of end-users of the same client, e.g. for family members within a home
    • 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
    • H04N21/4662—Learning process for intelligent management, e.g. learning user preferences for recommending movies characterized by learning algorithms
    • H04N21/4666—Learning process for intelligent management, e.g. learning user preferences for recommending movies characterized by learning algorithms using neural networks, e.g. processing the feedback provided by the user
    • 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

Abstract

The invention discloses an intelligent recommendation method for family TV videos, which comprises a long-term interest video recommendation list generation step for all family members, wherein the long-term interest video recommendation list generation step is used for generating a recommendation list of all family user interest contents according to long-term film watching records and through a collaborative filtering algorithm based on user clustering; the method comprises the steps of generating a short-term interest video recommendation list of a current TV user, wherein the short-term interest video recommendation list is used for acquiring interest preference of the current user through a recurrent neural network combined with an Attention mechanism according to a short-term viewing record and generating an interest video recommendation list of the current user; the method comprises an optimal user recommendation list generation step, wherein the content recommendation lists in the steps (1) and (2) are optimized and combined to generate a final user recommendation list. The method and the device solve the problem that the recommendation accuracy is poor when the current recommendation technology is oriented to a plurality of users with large interest difference, better solve the problem of TV end information overload in the Internet television era, and finally improve the use experience of TV end users.

Description

Intelligent recommendation method for home TV video
Technical Field
The invention belongs to the field of automatic recommendation of internet video contents, and particularly relates to video content recommendation facing a family TV user group by combining a clustering-based collaborative filtering algorithm and a recurrent neural network algorithm of a kneading Attention mechanism.
Background
The phenomenon of information overload of internet television content is becoming obvious day by day, and the use of an automatic recommendation algorithm is an effective means for solving the problem of information overload; the video content recommendation scenes aimed at by common recommendation algorithms in the market at present are all use scenes with strong privacy, for example, video content recommendation of a user by a mobile app is performed, and the algorithms only need to grasp the watching interest of a single user through data; compared with the above scenario, the scenario of the user at the television end is more complex, and mainly appears as follows:
(1) one television terminal is usually oriented to a plurality of members in one family, and different members in the same family have the same film watching hobbies and different film watching requirements;
(2) a television terminal user can switch the television normally, the watching content and the watching interest recorded by the terminal change greatly in a short time, and the watching data shows the characteristic of rapid and dynamic change along with the time.
The recommendation algorithm with strong privacy is directly applied to the television end, so that the recommendation can be completed, but the recommendation effect is poor; therefore, the invention aims to better solve the problem of information overload of Internet television content and improve the use experience of family users on the television.
Disclosure of Invention
The invention aims to provide an automatic recommendation algorithm more suitable for a home television end to solve the problem of overload of on-demand content information on a home Internet television, so that the television use experience of home users is improved.
The technical scheme of the invention is as follows:
an intelligent recommendation method for home TV videos comprises the following steps:
(1) long-term interest video recommendation list generation for all members of TV-end family
And (4) a recommendation algorithm combining clustering and collaborative filtering according to the historical playing preference content of the user. In the algorithm, on one hand, a clustering algorithm is utilized to mine the long-term preference rule of the family user to the content; on the other hand, a long-term interest content list of all the family members is generated by utilizing a user-based collaborative filtering algorithm. The specific process is as follows:
the first step is as follows: screening and preprocessing behavior data of home users
The main purpose of the process is to remove invalid user behavior data and to retain valid user behavior data, for example, to remove data with too short a viewing time.
The second step is that: building implicit scoring interest model
The main logic of the implicit rating model of the user is to calculate the rating of the user to different contents by the ratio of the viewing time length of the user to the total viewing time length. The main process comprises three steps:
A. extracting main fields in the user viewing content behavior data: the "content viewing time length" and the "content viewing time length ratio".
B. And obtaining the implicit rating of the user to the content according to the characteristic field of the user behavior.
C. And sorting the final calculation result to provide data for a subsequent algorithm.
The third step: classifying different families by using a clustering algorithm;
clustering users by using a K-Means + + algorithm, wherein the process mainly comprises the following three steps:
A. of all data points, the k most distant points are selected as the points for which the cluster center is initialized.
B. And classifying all the points in the data set, and respectively classifying the points to the cluster with the shortest distance from the clustering center.
C. And (4) calculating the centers of all the clusters again, finishing clustering if the positions of the center points are not deviated, and otherwise, repeating the steps until the positions of the center points are not deviated.
The fourth step: generating recommendation lists using collaborative filtering algorithms
The content recommendation based on the collaborative filtering algorithm of the user is completed by utilizing a Pearson similarity calculation method, and the specific steps are as follows:
A. dividing implicit grading data of users according to each user, and calculating intermediate variables;
B. integrating intermediate calculation results with the same content, calculating intermediate variable products among users if the users belong to the same category, and not calculating if the users do not belong to the same category;
C. product summation is carried out on the intermediate variables with the same user type, and the similarity between the users is calculated;
D. according to the similarity between the users, obtaining other users similar to the users;
E. and generating a content recommendation list according to the implicit grading content of the similar users.
(2) Current TV user interest video recommendation list generation
The method mainly utilizes the advantages of a Recurrent Neural Network (RNN) and a variant LSTM thereof on processing time sequences to build a generation model of an interest video recommendation list of a current TV user; meanwhile, considering that the interests of different members of the family may be greatly different and interfere with the result, the part proposes to add an attention (attention) mechanism on the basis of a neural network structure to weaken interference items generated by the viewing of other members and improve the recommendation accuracy. The specific process is as follows:
the first step is as follows: screening and preprocessing behavior data of home users
The main purpose of the process is to remove invalid user behavior data and to retain valid user behavior data, for example, to remove data with too short a viewing time.
The second step is that: converting behavior data into data that can be processed by model
The main purpose of the process is to convert the viewing data into data which can be processed by a model, in the process, one-hot coding is firstly carried out on the content, and as the coded data is a high-dimensional sparse vector, the result is processed by an embedding layer, the high-dimensional sparse vector is converted into a low-dimensional dense vector which is beneficial to deep neural network processing, all the viewing behaviors of a user are mapped into vectors with fixed length, and the subsequent network computing processing is facilitated.
The third step: inputting data to LSTM layer for processing
The main purpose of this process is to obtain the relationship that exists between the user's state at time t and the user's viewing behavior.
The fourth step: inputting data into the Attention layer for processing
Since the watching behaviors of all the family are recorded in the playing history behaviors of the family user, namely, the recorded history behaviors can be regarded as a sequence mixed by a plurality of interest topic sampling points. Therefore, the main purpose of introducing the Attention mechanism in the process is to enhance the content with strong correlation with the planning content in the historical viewing record and weaken the content with weak correlation, thereby further predicting the interest degree of the user in the target content.
The fifth step: inputting data into output layer, generating user short-term interest list
The process mainly comprises the steps of outputting a result feature vector of the attribute to a full link layer, calculating result probability through a sigmoid function, and judging whether a current user is interested in target content through short-term behavior data. And finally, generating a current interest recommendation list of the user.
(3) Optimal user recommendation list generation
In the process, the user interest lists generated in the steps (1) and (2) are combined, and a final user recommendation list with priority is generated according to indexes such as watching frequency, popularity and content scores;
the invention has the beneficial effects that:
the method solves the problem of information overload of the Internet television end, builds the on-demand content recommendation algorithm more suitable for the TV end, and can improve the recommendation accuracy by 10% -15% compared with the traditional recommendation algorithm.
The specific implementation mode is as follows:
the present invention is further illustrated by the following specific examples, which are intended to be illustrative, not limiting and are not intended to limit the scope of the invention.
An intelligent recommendation method for home TV videos comprises the following steps:
generating long-term interest video recommendation list of all members of TV-end family
The first step is as follows: the method comprises the following steps of screening and preprocessing behavior data of a family user, wherein the specific processing process in the process comprises the following three steps:
(1) the process faces the case that the data has missing values: the repairable data are supplemented manually; directly deleting unrepairable data; and directly carrying out deduplication processing on the repeated data.
(2) The process is faced with significantly erroneous data: storing the corrected data after manual correction; and directly deleting the data which cannot be corrected.
(3) Abnormal value discrimination and processing: in the first two steps, only obvious dirty data are processed, and some data which cannot be removed and corrected by a simple method may exist in real data; and in the face of the data, removing unreasonable data by adopting a statistical analysis method: if the data meet normal distribution, a confidence probability and a confidence interval can be set to select the data, the data are not in the corresponding interval to indicate that the related data exceed the minimum range allowed by the error, and the data are removed; if the data does not satisfy the normal distribution, the data can be determined by using a box plot, wherein the box plot includes 6 data points, the upper quartile is recorded as Q1, and the lower quartile is recorded as Q2, the quartile distance IQR is Q1-Q2, and if the data is smaller than the values of Q1-1.5IQR or Q2+1.5IQR, the data is considered abnormal, and the data can be rejected.
The second step is that: building an implicit scoring interest model, wherein the main logic of the model is to calculate scores of different contents of a user according to the ratio of the watching time length of the user to the total watching time length, and the main process comprises three steps:
extracting main fields in the content viewing behavior data of the user: the "content viewing time length" and the "content viewing time length ratio".
And obtaining the implicit rating of the user to the content according to the characteristic field of the user behavior.
And thirdly, sorting the final calculation result to provide data for a subsequent algorithm.
The third step: classifying different families by using a K-Means + + clustering algorithm, wherein the specific calculation method is as follows:
(1) firstly, randomly selecting a point in a set as the center of a cluster;
(2) calculating the distance between each point in the set and the center of the cluster, wherein a commonly used method is an Euclidean distance method, and assuming that two points x and y exist in the space, the Euclidean distance formula between the two points is as follows:
Figure BDA0003333385550000041
(3) selecting a new data point as a new cluster center, wherein the selection idea is as follows: (2) the larger the calculated distance is, the higher the probability value of the point as a new cluster center is;
(4) initializing the center points of all clusters by continuously executing the second step and the third step;
(5) and after the central point of each cluster is obtained, calculating by using a K-Means algorithm respectively. The detailed procedure for K-Means is: randomly selecting k values from a data set N as centers of initialization clustering; calculating the distance from all points in the data set N to the central value of each cluster by using the Euclidean distance method in the step (2); dividing the values in the data set N into the cluster with the central point nearest to the values; after all the data are divided, the central position of the cluster is recalculated according to the data condition of the cluster, and the calculation formula is as follows:
Figure BDA0003333385550000042
in the above formula, x represents the number of clusters in each class of the cluster center kCoordinates of points, CenterkRepresenting the coordinates of the newly calculated cluster center. If the center position of the new cluster is changed, the distances from all the points to the central values of all the clusters are recalculated, and if the center of the new cluster is kept unchanged, the related clustering result is output.
The fourth step: the content recommendation based on the collaborative filtering algorithm of the user is completed by utilizing a Pearson similarity calculation method, and the specific steps are as follows:
(1) dividing implicit grading data of users according to each user, and calculating an intermediate variable QuiThe specific calculation formula is as follows:
Figure BDA0003333385550000051
wherein u and i represent the user and the content of the user with implicit rating, PuiRefers to the user u's score for content i,
Figure BDA0003333385550000052
representing the mean of all content implicit scores by user u.
(2) Integrating intermediate calculation results with the same content, and calculating intermediate variable product Q between users if the users belong to the same categoryui×QviIf not, not calculating;
(3) and (3) performing product summation on intermediate variables with the same user type to calculate the similarity between users, wherein a specific calculation formula is as follows:
Figure BDA0003333385550000053
(4) according to the similarity between the users, obtaining other users similar to the users;
(5) and generating a content recommendation list according to the implicit grading content of the similar users.
Introducing the clustered effect:
TABLE 1 comparison of accuracy rates of two recommendation algorithms
Figure BDA0003333385550000054
TABLE 2 two recommendation algorithm recall comparisons
Figure BDA0003333385550000055
As shown in the above table, 300 to 1500 user data are respectively input in the experiment, and the two tables respectively represent the accuracy and the recall rate, and the higher the value is, the better the effect is. The experimental result shows that compared with the traditional collaborative filtering algorithm based on the user, the collaborative filtering algorithm based on the user clustering has more excellent recommendation result.
2. Current TV user interest video recommendation list generation
The same as 1, 2, in the first step, the behavior data of the family user is screened and preprocessed, and the specific processing process in the process is divided into the following three steps:
(1) the process faces the case that the data has missing values: the repairable data are supplemented manually; directly deleting unrepairable data; and directly carrying out deduplication processing on the repeated data.
(2) The process is faced with significantly erroneous data: storing the corrected data after manual correction; and directly deleting the data which cannot be corrected.
(3) Abnormal value discrimination and processing: in the first two steps, only obvious dirty data are processed, and some data which cannot be removed and corrected by a simple method may exist in real data; and in the face of the data, removing unreasonable data by adopting a statistical analysis method: if the data meet normal distribution, a confidence probability and a confidence interval can be set to select the data, the data are not in the corresponding interval to indicate that the related data exceed the minimum range allowed by the error, and the data are removed; if the data does not satisfy the normal distribution, the data can be determined by using a box plot, wherein the box plot includes 6 data points, the upper quartile is recorded as Q1, and the lower quartile is recorded as Q2, the quartile distance IQR is Q1-Q2, and if the data is smaller than the values of Q1-1.5IQR or Q2+1.5IQR, the data is considered abnormal, and the data can be rejected.
In the second step, the behavior data is converted into data which can be processed by the model
In the process, one-hot coding is firstly carried out on the content, and as the coded data is a high-dimensional sparse vector, the result is processed by an embedding layer, the high-dimensional sparse vector is converted into a low-dimensional dense vector beneficial to deep neural network processing, all the watching behaviors of a user are mapped into a vector with a fixed length, and the subsequent network computing processing is facilitated.
Inputting data into an LSTM layer for processing, wherein the process mainly aims to obtain the relationship between the state of a user at the time t and the watching behavior of the user, and the specific calculation process and formula are as follows:
in the whole calculation process, after input reaches an LSTM structure, information needing to be abandoned is judged through a sigmoid function, and the calculation formula is as follows:
ft＝σ(Wf·[ht-1，xt]+bf)
in the above formula, ftRepresenting forgetting gate, WfIs a weight value ht-1Representing the state of the hidden layer at the previous moment, xtRepresenting the input at the current time, bfRepresents a bias;
then, the storage of the neuron information at the current moment is determined through two parts of operation, the first part determines the updating content through a sigmoid function, and the calculation formula is as follows:
it＝σ(Wi·[ht-1，xt]+bi)
in the above formula, itRepresents an input gate, WiIs a weight value ht-1Representing the state of the hidden layer at the previous moment, xtRepresenting the input at the current time, biRepresents a bias;
the second part creates candidate vectors through tanh function
Figure BDA0003333385550000071
The calculation formula is as follows:
Figure BDA0003333385550000072
in the above formula, the first and second carbon atoms are,
Figure BDA0003333385550000073
is the current candidate vector value, WCIs a weight value ht-1Representing the state of the hidden layer at the previous moment, xtRepresenting the input at the current time, bCRepresents a bias;
combining the previous calculation to obtain the neuron state C of the present stagetThe calculation formula is as follows:
Figure BDA0003333385550000074
in the above formula, CtIs the present stage of neuronal state, ftRepresenting forgetting gate, Ct-1Representing the state of the neuron at the previous moment, itWhich represents the input gate or gates, respectively,
Figure BDA0003333385550000075
the current candidate vector value is obtained;
finally, the current stage neuron state C is determinedtBy the tanh function and otMultiply to calculate htThe calculation formula is as follows:
ot＝σ(Wo·[h-1，xt]+bo)
ht＝ot·tanh(Ct)
in the above formula, otRepresents an output gate, WoRepresents the weight value, ht-1Representing the state of the hidden layer at the previous moment, xtRepresenting the input at the current time, boRepresents a bias; h istBeing the state of the current hidden layer, CtThe present stage of the neuronal state.
And fourthly, inputting data into an Attention layer for processing, introducing an Attention mechanism to enhance the content with stronger correlation with the planned content in the historical viewing record and weaken the content with weak correlation, thereby further predicting the interest degree of the user in the target content, wherein the calculation formula is as follows:
Figure BDA0003333385550000076
in the above formula, αtRepresenting an Attention score reflecting the correlation between the video content of the target and the user's historical interest status at each moment, the greater the correlation, alphatThe higher the value of (A); e.g. of the typeaA vector representing that the name of the content watched by the current user passes through the embedding layer, W is a model parameter,
Figure BDA0003333385550000077
wherein n ishRepresents a hidden state htDimension of, naRepresenting the vector dimension after the currently viewed program embedding. Finally, the vector formula output after the Attention mechanism can be expressed as follows:
h′t＝ht×αt
fifthly, inputting data into an output layer to generate a user short-term interest list
The process mainly comprises the steps of outputting a result feature vector of the attribute to a full link layer, calculating result probability through a sigmoid function, and judging whether a current user is interested in target content through short-term behavior data. And finally, generating a current interest recommendation list of the user.
The relevant experimental effects are as follows:
TABLE 3 LSTM model accuracy data
Figure BDA0003333385550000081
TABLE 4 RNN model accuracy data
Figure BDA0003333385550000082
TABLE 5 LSTM model recall data
Figure BDA0003333385550000083
TABLE 6 RNN model recall data
Figure BDA0003333385550000084
TABLE 7 LSTM model f1 data
Figure BDA0003333385550000085
TABLE 8 RNN model f1 data
Figure BDA0003333385550000091
As can be seen from the data in tables 3-8, the LSTM model has better effect than the RNN model, and the accuracy of the LSTM model is 5-10% higher than that of the RNN model in terms of accuracy. At the same time, it is clear that: after the LSTM model and the RNN model are added with the Attention mechanism, the accuracy, the recall rate and the f1 value of the LSTM model and the RNN model are all higher than the model corresponding to the model without the Attention mechanism, wherein the quasi-removal rate is about 5% higher, the recall rate is about 4% higher, and the f1 value is about 5% higher.
The invention is not the best known technology.
The above embodiments are merely illustrative of the technical ideas and features of the present invention, and the purpose thereof is to enable those skilled in the art to understand the contents of the present invention and implement the present invention, and not to limit the protection scope of the present invention. All equivalent changes and modifications made according to the spirit of the present invention should be covered within the protection scope of the present invention.

Claims (10)

1. An intelligent recommendation method for home TV videos is characterized by comprising the following steps:
the method comprises the steps of generating a long-term interest video recommendation list of all family members: generating a recommendation list of all user interest contents of a family through a collaborative filtering algorithm based on user clustering;
the method comprises the steps of generating a video recommendation list of the current TV user interest: acquiring interest preference of a current user through a recurrent neural network combined with an Attention mechanism, and generating an interest video recommendation list of the current user;
the method comprises the following steps of generating an optimal user recommendation list: and (3) optimally combining the content recommendation lists in the steps (1) and (2) to generate a final user recommendation list.
2. The intelligent home TV video recommendation method according to claim 1, wherein: the collaborative filtering algorithm based on user clustering comprises the following steps:
(1) screening and preprocessing the behavior data of the family user;
(2) building an implicit grading interest model;
(3) classifying different families by using a clustering algorithm;
(4) and generating a long-term interest video recommendation list of all the family members by using a collaborative filtering algorithm.
3. The intelligent home TV video recommendation method according to claim 2, wherein: and (3) the implicit rating model in the step (2) is used for calculating the rating of the user to different contents according to the ratio of the watching time length of the user to the total watching time length.
4. The intelligent home TV video recommendation method according to claim 2, wherein: the clustering algorithm in the step (3) is a K-Means + + algorithm, and comprises the following steps:
step one, selecting k points with the farthest distance from all the data points as initialization points of the cluster centers.
Step (ii) ofIIAnd classifying all the points in the data set, and respectively classifying the points to the cluster with the shortest distance from the clustering center.
And step three, calculating the centers of all the clusters again, if the positions of the center points are not deviated, finishing clustering, and otherwise, repeating the steps until the positions of the center points are not deviated.
5. The intelligent home TV video recommendation method according to claim 2, wherein: the collaborative filtering algorithm in the step (4) is a Pearson similarity calculation method, and specifically comprises the following steps:
dividing implicit rating data of users according to each user, and calculating intermediate variables;
step (ii) ofIIIntegrating intermediate calculation results with the same content, calculating the intermediate variable product between users if the users belong to the same category, and not calculating if the users do not belong to the same category;
step (ii) ofIIIProduct summation is carried out on the intermediate variables with the same user type, and the similarity between the users is calculated;
step (ii) ofFourthlyAccording to the similarity between the users, obtaining other users similar to the users;
and step five, generating a content recommendation list according to the implicit grading content of the similar users.
6. The intelligent home TV video recommendation method according to claim 5, wherein: the formula for calculating the intermediate variable in the first step is as follows:
Figure FDA0003333385540000021
wherein u and i represent the user and the content of the user with implicit rating, QuiIs an intermediate variable, PuiRefers to the user u's score for content i,
Figure FDA0003333385540000022
representing the mean of all content implicit scores by user u.
7. The intelligent home TV video recommendation method according to claim 5, wherein: the third step is as follows:
Figure FDA0003333385540000023
wherein u and v represent two users, i represents the user with implicit rating, and QuiFor user u intermediate variables, QviA user v intermediate variable; puiRefers to the user u's score, P, for content iviScoring content i for user v;
Figure FDA0003333385540000024
represents the mean of all content implicit scores by user u,
Figure FDA0003333385540000025
representing the mean of all content implicit scores by user v.
8. The intelligent recommendation method for TV video according to claim 1, wherein: the recurrent neural network combined with the Attention mechanism comprises the following steps:
(1) screening and preprocessing the behavior data of the family user;
(2) converting the behavior data into data which can be processed by a model;
(3) obtaining the relation between the state of the user at the time t and the watching behavior of the user by using an LSTM network;
(4) the Attention network is utilized to weaken the interference items of the viewing records of other members, and the interest degree of the user on the target content is predicted more accurately;
(5) and calculating the result probability by using a sigmoid function, and generating a current TV user interest video recommendation list.
9. The intelligent recommendation method for TV video according to claim 8, wherein: the specific calculation flow and formula of the step (3) are as follows:
in the whole calculation process, after input reaches an LSTM structure, information needing to be abandoned is judged through a sigmoid function, and the calculation formula is as follows:
ft＝σ(Wf·[ht-1，xt]+bf)
in the above formula, ftRepresenting forgetting gate, WfIs a weight value ht-1Representing the state of the hidden layer at the previous moment, xtRepresenting the input at the current time, bfRepresents a bias;
then, the storage of the neuron information at the current moment is determined through two parts of operation, the first part determines the updating content through a sigmoid function, and the calculation formula is as follows:
it＝σ(Wi·[ht-1，xt]+bi)
in the above formula, itRepresents an input gate, WiIs a weight value ht-1Representing the state of the hidden layer at the previous moment, xtRepresenting the input at the current time, biRepresents a bias;
the second part creates candidate vectors through tanh function
Figure FDA0003333385540000031
The calculation formula is as follows:
Figure FDA0003333385540000032
in the above formula, the first and second carbon atoms are,
Figure FDA0003333385540000033
is the current candidate vector value, WCIs a weight value ht-1Representing the state of the hidden layer at the previous moment, xtRepresenting the input at the current time, bCRepresents a bias;
combining the previous calculation to obtain the neuron state C of the present stagetThe calculation formula is as follows:
Figure FDA0003333385540000034
in the above formula, CtIs the present stage of neuronal state, ftRepresenting forgetting gate, Ct-1Representing the state of the neuron at the previous moment, itWhich represents the input gate or gates, respectively,
Figure FDA0003333385540000035
the current candidate vector value is obtained;
finally, the current stage neuron state C is determinedtBy the tanh function and otMultiply to calculate htThe calculation formula is as follows:
ot＝σ(Wo·[ht-1，xt]+bo)
ht＝ot·tanh(Ct)
in the above formula, otRepresents an output gate, WoRepresents the weight value, ht-1Representing the state of the hidden layer at the previous moment, xtRepresenting the input at the current time, boRepresents a bias; h istBeing the state of the current hidden layer, CtThe present stage of the neuronal state.
10. The intelligent recommendation method for TV video according to claim 8, wherein: the calculation formula of the step (4) is as follows:
Figure FDA0003333385540000036
in the above formula, αtRepresenting an Attention score reflecting the correlation between the video content of the target and the user's historical interest status at each moment, the greater the correlation, alphatThe higher the value of (A); e.g. of the typeaA vector representing that the name of the content watched by the current user passes through the embedding layer, W is a model parameter,
Figure FDA0003333385540000037
wherein n ishRepresents a hidden state htDimension of, naRepresenting the vector dimension after the program embedding is watched currently;
finally, the vector formula output after the Attention mechanism is expressed as follows:
h′t＝ht×αt。
CN202111287347.5A 2021-11-02 2021-11-02 Intelligent recommendation method for home TV video Pending CN114025205A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111287347.5A CN114025205A (en) 2021-11-02 2021-11-02 Intelligent recommendation method for home TV video

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111287347.5A CN114025205A (en) 2021-11-02 2021-11-02 Intelligent recommendation method for home TV video

Publications (1)

Publication Number Publication Date
CN114025205A true CN114025205A (en) 2022-02-08

Family

ID=80059897

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111287347.5A Pending CN114025205A (en) 2021-11-02 2021-11-02 Intelligent recommendation method for home TV video

Country Status (1)

Country Link
CN (1) CN114025205A (en)

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110191311A1 (en) * 2010-02-03 2011-08-04 Gartner, Inc. Bi-model recommendation engine for recommending items and peers
CN106649540A (en) * 2016-10-26 2017-05-10 Tcl集团股份有限公司 Video recommendation method and system
CN106651546A (en) * 2017-01-03 2017-05-10 重庆邮电大学 Intelligent community oriented electronic commerce information recommendation method
CN106909607A (en) * 2017-01-06 2017-06-30 南京邮电大学 A kind of collaborative filtering group recommending method based on random perturbation technology
CN107368519A (en) * 2017-06-05 2017-11-21 桂林电子科技大学 A kind of cooperative processing method and system for agreeing with user interest change
CN109862431A (en) * 2019-01-23 2019-06-07 重庆第二师范学院 A kind of TV programme mixed recommendation method based on MCL-HCF algorithm
CN110162553A (en) * 2019-05-21 2019-08-23 南京邮电大学 Users' Interests Mining method based on attention-RNN
CN110430471A (en) * 2019-07-24 2019-11-08 山东海看新媒体研究院有限公司 It is a kind of based on the television recommendations method and system instantaneously calculated
CN111259195A (en) * 2018-11-30 2020-06-09 清华大学深圳研究生院 Video recommendation method and device, electronic equipment and readable storage medium

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110191311A1 (en) * 2010-02-03 2011-08-04 Gartner, Inc. Bi-model recommendation engine for recommending items and peers
CN106649540A (en) * 2016-10-26 2017-05-10 Tcl集团股份有限公司 Video recommendation method and system
CN106651546A (en) * 2017-01-03 2017-05-10 重庆邮电大学 Intelligent community oriented electronic commerce information recommendation method
CN106909607A (en) * 2017-01-06 2017-06-30 南京邮电大学 A kind of collaborative filtering group recommending method based on random perturbation technology
CN107368519A (en) * 2017-06-05 2017-11-21 桂林电子科技大学 A kind of cooperative processing method and system for agreeing with user interest change
CN111259195A (en) * 2018-11-30 2020-06-09 清华大学深圳研究生院 Video recommendation method and device, electronic equipment and readable storage medium
CN109862431A (en) * 2019-01-23 2019-06-07 重庆第二师范学院 A kind of TV programme mixed recommendation method based on MCL-HCF algorithm
CN110162553A (en) * 2019-05-21 2019-08-23 南京邮电大学 Users' Interests Mining method based on attention-RNN
CN110430471A (en) * 2019-07-24 2019-11-08 山东海看新媒体研究院有限公司 It is a kind of based on the television recommendations method and system instantaneously calculated

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
贾毓臻: "《硕士论文》", 31 October 2019 *

Similar Documents

Publication Publication Date Title
CN108920641B (en) Information fusion personalized recommendation method
CN112765486B (en) Knowledge graph fused attention mechanism movie recommendation method
CN108920720B (en) Large-scale image retrieval method based on depth hash and GPU acceleration
Al-Shamri et al. Fuzzy-genetic approach to recommender systems based on a novel hybrid user model
CN102254006B (en) Method for retrieving Internet video based on contents
CN107292097B (en) Chinese medicine principal symptom selection method based on feature group
CN109241366B (en) Hybrid recommendation system and method based on multitask deep learning
CN107633444A (en) Commending system noise filtering methods based on comentropy and fuzzy C-means clustering
CN107635151A (en) A kind of machine learning TV programme suggesting method based on domain disassembler
Li et al. Content-based filtering recommendation algorithm using HMM
CN112950325B (en) Self-attention sequence recommendation method for social behavior fusion
CN111651678A (en) Knowledge graph-based personalized recommendation method
CN111159473A (en) Deep learning and Markov chain based connection recommendation method
CN112380451A (en) Favorite content recommendation method based on big data
CN113761359A (en) Data packet recommendation method and device, electronic equipment and storage medium
CN109508407A (en) The tv product recommended method of time of fusion and Interest Similarity
CN115618128A (en) Collaborative filtering recommendation system and method based on graph attention neural network
Jang et al. Cities: Contextual inference of tail-item embeddings for sequential recommendation
CN108491477B (en) Neural network recommendation method based on multi-dimensional cloud and user dynamic interest
CN111949894B (en) Collaborative filtering personalized recommendation method based on multi-space interaction
CN114025205A (en) Intelligent recommendation method for home TV video
CN108549729B (en) Personalized user collaborative filtering recommendation method based on coverage reduction
CN116452293A (en) Deep learning recommendation method and system integrating audience characteristics of articles
CN113707293B (en) Feature selection-based Chinese medicine main symptom selection method
CN115840853A (en) Course recommendation system based on knowledge graph and attention network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20220208