[go: up one dir, main page]
More Web Proxy on the site http://driver.im/

CN115905617B - Video scoring prediction method based on deep neural network and double regularization - Google Patents

Video scoring prediction method based on deep neural network and double regularization Download PDF

Info

Publication number
CN115905617B
CN115905617B CN202310187456.2A CN202310187456A CN115905617B CN 115905617 B CN115905617 B CN 115905617B CN 202310187456 A CN202310187456 A CN 202310187456A CN 115905617 B CN115905617 B CN 115905617B
Authority
CN
China
Prior art keywords
user
video
matrix
representing
users
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.)
Active
Application number
CN202310187456.2A
Other languages
Chinese (zh)
Other versions
CN115905617A (en
Inventor
赵学健
张晶晶
孙知信
孙哲
曹亚东
宫婧
汪胡青
胡冰
徐玉华
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.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
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 Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN202310187456.2A priority Critical patent/CN115905617B/en
Publication of CN115905617A publication Critical patent/CN115905617A/en
Application granted granted Critical
Publication of CN115905617B publication Critical patent/CN115905617B/en
Priority to PCT/CN2023/135369 priority patent/WO2024179089A1/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
    • 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
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0499—Feedforward 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • H04N21/43—Processing of content or additional data, e.g. demultiplexing additional data from a digital video stream; Elementary client operations, e.g. monitoring of home network or synchronising decoder's clock; Client middleware
    • H04N21/44—Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs
    • 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
    • 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
    • H04N21/47—End-user applications
    • H04N21/475—End-user interface for inputting end-user data, e.g. personal identification number [PIN], preference data
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Signal Processing (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Human Computer Interaction (AREA)
  • Library & Information Science (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a video scoring prediction method based on a deep neural network and double regularization, which comprises the steps of reconstructing a user-video scoring matrix, introducing a video association regular term fused with user activity and a reliable nearest neighbor regular term, constructing a matrix decomposition recommendation model fused with the video association regular term fused with user activity and the reliable nearest neighbor regular term, inputting potential features into the deep neural network to obtain a result of the deep neural network model, combining the result of the deep neural network model with a matrix decomposition structure to obtain a final prediction score, and improving the accuracy of the prediction score; and the LDA model is utilized to mine relevant information in the user video comments, a user type potential feature matrix and a video type potential feature matrix are generated, the user type potential feature matrix and the video type potential feature matrix are combined to obtain a hidden information matrix, and then the hidden information matrix is combined with the original user video scoring matrix to generate a new user-video scoring matrix, so that the problems of cold start and data sparsity are relieved.

Description

Video scoring prediction method based on deep neural network and double regularization
Technical Field
The invention relates to a video scoring prediction method based on a deep neural network and double regularization, and belongs to the field of scoring prediction.
Background
Along with the rapid development of internet technology, video resources in each network platform are more and more, so that abundant video resources are provided for users, and when more choices are provided for users, trouble and trouble are brought to the users, and huge video resources not only increase the difficulty of finding favorite videos of the users, but also make the process of finding videos quite time-consuming. In order to solve the information overload problem, the personalized recommendation system becomes an effective tool for solving the problem. Score prediction is an important component of the recommendation algorithm. The existing recommendation algorithm mainly comprises three main categories: collaborative filtering-based recommendation algorithms, content-based recommendation algorithms, and hybrid recommendation algorithms. The most used collaborative filtering-based recommendation algorithm at present, and the most used collaborative filtering recommendation algorithm is the model-based collaborative filtering recommendation algorithm, and several more common algorithms in the model-based collaborative filtering recommendation algorithm include: matrix decomposition model, singular value decomposition, cluster analysis, etc. However, the existing collaborative filtering recommendation algorithm has the problems of sparse data, cold start and the like, so that the recommendation video resource scoring prediction is inaccurate, the personalized recommendation result is affected, the accuracy of the video resource scoring prediction is improved, and the recommendation precision is further improved to be one of hot spots of the current research.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a video scoring prediction method based on a depth neural network and double regularization, a user-video scoring matrix is reconstructed, a video association regular term and a reliable nearest neighbor regular term which are integrated with the activity of a user are introduced to restrain the learning of a potential feature matrix during matrix decomposition, the depth neural network is introduced, the restriction of linear dot products in the matrix decomposition process is relieved by utilizing the nonlinear features of the depth neural network, the result of a depth neural network model is combined with the result of the double regularization matrix decomposition, and the accuracy of video scoring prediction is improved.
The technical scheme adopted by the invention is as follows: a video scoring prediction method based on a deep neural network and double regularization is used for improving the accuracy of scoring prediction of recommended videos, and specifically comprises the following steps:
step S1: processing the video comments, mining hidden information, combining the hidden information matrix with the original user-video scoring matrix to generate a new user-video scoring matrix, and entering into step S2;
step S2: the method comprises the steps that a dual regularization term is added to learn a potential feature matrix when a user-video scoring matrix is decomposed, each user can make a certain contribution to video similarity, the user contributions are not identical, the user can be divided into active users and inactive users from the consideration of the liveness of the users, the active users refer to users with a large number of scoring records on the video, the inactive users refer to users with scoring records on a small number of videos, so that the contributions of the active users and the inactive users are distinguished when the video similarity is calculated, and the liveness of the users is defined as follows:
Figure SMS_1
equation 1
In the case of the formula 1 of the present invention,
Figure SMS_2
the total score of the user u is represented, so that the video similarity calculation method obtained by combining the activity coefficient of the user with the corrected cosine similarity is as follows: />
Figure SMS_3
Equation 2
In the case of the formula 2 of the present invention,
Figure SMS_4
representing the score of user u for video i, +.>
Figure SMS_5
Representing the score of user u for video j, +.>
Figure SMS_6
Score representing user u, ++>
Figure SMS_7
Representing a set of users who have scored video i and j simultaneously; the learning of the potential feature matrix of the video association regularization term constraint project which is integrated with the user activity is introduced during matrix decomposition, and the video association regularization constraint function formula which is integrated with the user activity at the moment is as follows:
Figure SMS_8
equation 3
In equation 3, V represents the video feature matrix, V j Is the potential feature vector of video j, V i Is a potential feature vector of the video i, and proceeds to step S3;
step S3: taking the potential feature vectors decomposed by the matrix as the input of the multi-layer perceptron, processing the potential feature vectors by the multi-layer perceptron to obtain a multi-layer perceptron model prediction result, and entering into step S4;
s4, a step of S4; and combining the result of the multi-layer perceptron model prediction with the result of matrix decomposition at the merging layer, and optimizing the model by using a normalized cross entropy method to obtain a final prediction score.
As a preferred technical scheme of the invention: in the step S1, firstly, the LDA model is utilized to mine relevant type hidden information in the user video evaluation, generate a user type latent feature matrix LU and a video type latent feature matrix LV, and combine the user type latent feature matrix and the video type latent feature matrix to obtain
To the hidden information matrix L, the calculation formula is:
Figure SMS_9
equation 4
And combining the hidden information matrix L with the original user-video scoring matrix R to generate a new user-video scoring matrix
Figure SMS_10
The calculation formula is as follows:
Figure SMS_11
equation 5
As a preferred technical scheme of the invention: in the step S2, users with the same interests may affect each other, and the user similarity may be calculated using the weighted pearson correlation coefficient:
Figure SMS_12
equation 6
In the case of the formula 6 of the present invention,
Figure SMS_13
and->
Figure SMS_14
Represents the average score of users u and v, respectively, < >>
Figure SMS_15
Representing the score of user u for video i, +.>
Figure SMS_16
Representing the score of user v for video i, +.>
Figure SMS_17
Representing a video set reviewed by user u, +.>
Figure SMS_18
Representing a video set reviewed by user v, +.>
Figure SMS_19
The Jaccard correlation coefficient, which is a term capable of affecting the user similarity calculation, is calculated as follows:
Figure SMS_20
equation 7
Wherein,,
Figure SMS_21
representing a video set reviewed by user u, +.>
Figure SMS_22
Representing a video set reviewed by the user v;
the scoring of the item by the user depends on the influence of the adjacent user, and the influence of the adjacent user by the adjacent user can be also possible, but the adjacent user after a certain distance cannot be influenced by the user, namely the adjacent user becomes unreliable, so that a reliable value is introduced, the scoring of the item by the adjacent user with the reliable value being larger than a certain value can be influenced, and the reliable value is calculated in the following way:
Figure SMS_23
equation 8
In the case of the formula 8 of the present invention,
Figure SMS_26
representing the score of video i with u, +.>
Figure SMS_28
Representing the score of user v for video i, +.>
Figure SMS_30
Representing a video set reviewed by user u, +.>
Figure SMS_25
Representing a video set reviewed by user v, +.>
Figure SMS_27
Represents the maximum value of the score,/->
Figure SMS_29
Indicating trust distance, i.e. presence between user u and user vIs->
Figure SMS_31
Represents the maximum distance allowed between two users, +.>
Figure SMS_24
Is a correction parameter, is a number greater than 0 and less than 1, and is a reliable nearest neighbor user
Figure SMS_32
Equation 9
The reliable nearest-neighbor regular term is introduced to restrain the learning of the potential feature matrix of the user when the matrix is decomposed, and the reliable nearest-neighbor regular term restraining function is as follows:
Figure SMS_33
. Equation 10
Wherein,,
Figure SMS_34
for the potential vector of user u +.>
Figure SMS_35
Is a potential feature vector for user v.
As a preferred technical scheme of the invention: in the step S3, the user potential feature vector and the video potential feature vector are used as input of a multi-layer perceptron, wherein the deep neural network is composed of the multi-layer perceptron and a single-layer perceptron, the multi-layer perceptron comprises an input layer, a plurality of hidden layers allowing nonlinearity of the neural structure and an output layer, and the result of the multi-layer perceptron model is obtained through processing of the multi-layer perceptron by utilizing the nonlinearity characteristics of the hidden layers.
As a preferred technical scheme of the invention: in the step S4, in the deep neural network structure, the single-layer perceptron is a merging layer, and in the merging layer, the prediction result of the multi-layer perceptron model is combined with the result of the dual regularization matrix decomposition model, and the calculation formula is as follows:
Figure SMS_36
equation 11
In equation 11
Figure SMS_37
To activate the function +.>
Figure SMS_38
For the matrix weight set between the output layer and the merge layer,/for the matrix weight set between the output layer and the merge layer>
Figure SMS_39
For outputting the result of the layer->
Figure SMS_40
For user potential vector, ++>
Figure SMS_41
For video potential vector, ++>
Figure SMS_42
And optimizing the model by using a normalized cross entropy method for merging the bias terms of the layers, and finally obtaining the prediction scores.
The beneficial effects are that:
1. according to the method, the video association regular term and the reliable nearest neighbor regular term which are integrated with the user activity are introduced to restrain the learning of the potential feature matrix, the nonlinear structure of the deep neural network is utilized to relieve the limit of linear dot products in the matrix decomposition process, and the result of the deep neural network model is combined with the result of the double regularization matrix decomposition, so that the accuracy of video scoring prediction is improved.
2. The invention utilizes the LDA model to mine the relevant information in the user video comment, generates the user type potential feature matrix and the video type potential feature matrix, combines the user type potential feature matrix and the video type potential feature matrix to obtain the hidden information matrix, combines the hidden information matrix and the original user video scoring matrix to generate the new user-video scoring matrix, and alleviates the problems of cold start and data sparsity.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a diagram of a multi-layer perceptron of the present invention;
fig. 3 is a block diagram of a deep neural network of the present invention.
Detailed Description
The following describes the embodiments of the present invention in further detail with reference to the drawings.
The invention reconstructs a user-video scoring matrix based on a traditional matrix decomposition model, and introduces video association regular terms and reliable nearest neighbor regular terms which are integrated with the activity of the user to limit the learning of potential feature matrices. The nonlinear structure of the deep neural network is utilized, the limit of linear dot products in the matrix decomposition process is relieved, potential feature vectors decomposed by the matrix are used as input of the deep neural network, the result of the MLP model is obtained through multi-layer perceptron processing, the result of the MLP model is combined with the result of the double regularization matrix decomposition model in a single-layer perceptron layer, namely a merging layer, and the model is optimized through a normalized cross entropy method, so that the accuracy of scoring prediction is improved.
As shown in fig. 1.
The invention designs a video scoring prediction method based on a deep neural network and double regularization, which is used for improving the precision of scoring prediction of recommended videos and comprises the following steps of:
step S1: processing the video comments, mining hidden information, combining the hidden information matrix with the original user-video scoring matrix to generate a new user-video scoring matrix, and entering into step S2;
step S2: adding a double-regular term for restraining the learning of the potential feature matrix when decomposing the user-video scoring matrix, and entering into step S3;
step S3: taking the potential feature vectors decomposed by the matrix as the input of the multi-layer perceptron, processing the potential feature vectors by the multi-layer perceptron to obtain a multi-layer perceptron model prediction result, and entering into step S4;
step S4: and combining the result of the multi-layer perceptron model prediction with the result of matrix decomposition at the merging layer, and optimizing the model by using a normalized cross entropy method to obtain a final prediction score.
The method comprises the following specific steps:
the step S1 comprises the following steps: firstly, utilizing an LDA model to mine relevant type hidden information in a user video evaluation, generating a user type latent feature matrix LU and a video type latent feature matrix LV, combining the user type latent feature matrix and the video type latent feature matrix to obtain a hidden information matrix L, wherein the calculation formula is as follows
Figure SMS_43
And combining the hidden information matrix L with the original user-video scoring matrix R to reconstruct the user-video scoring matrix R, wherein the calculation formula is as follows:
Figure SMS_44
。
The step S2 comprises the following steps: performing matrix decomposition on the user-video scoring matrix, and decomposing the high-dimensional user-video scoring matrix into a low-dimensional user feature matrix and a video feature matrix, wherein the formula is as follows:
Figure SMS_45
wherein U represents a user feature matrix, U i Representing the potential feature vector of user i, V represents the video feature matrix, V j Is a potential feature vector for video j. The low-dimensional matrix decomposition method approximately calculates a scoring matrix R by the product of d rank factors. The predictive score of user i for video j is expressed as
Figure SMS_46
The square of the error between the predicted score and the original score is taken as a loss function, and the loss function is minimized to approach the scoring matrix R. The loss function is:
Figure SMS_47
，
in the above-mentioned formula(s),
Figure SMS_48
is an indication function that indicates that 1 is equal if user i scores item j, and 0 is otherwise equal.
Figure SMS_49
And->
Figure SMS_50
Overfitting is prevented for two regular terms. Since each user can make a certain contribution to the video similarity, but each user contribution is not exactly the same, the user can be classified into active users and inactive users from the viewpoint of the activity of the user, the active users refer to users with a large number of scoring records on the video, and the inactive users refer to users with only a small number of scoring records on the video, so that the contribution of the active users and the inactive users should be distinguished when the video similarity is calculated, and the activity of the user can be defined as:
Figure SMS_51
in the above-mentioned formula(s),
Figure SMS_52
representing the total amount of scoring for user u. Therefore, the video similarity calculation method obtained by combining the activity coefficient of the user with the modified cosine similarity comprises the following steps:
Figure SMS_53
in the above-mentioned formula(s),
Figure SMS_54
representing the score of user u for video i, +.>
Figure SMS_55
Representing the score of user u for video j, +.>
Figure SMS_56
A score representing user u. Introducing a video association regularization term which is integrated with the user activity during matrix decomposition to restrict the learning of the project potential feature matrix, wherein a video association regularization constraint function formula which is integrated with the user activity at the moment is as follows:
Figure SMS_57
wherein V represents a video feature matrix, V j Is the potential feature vector of video j, V i Is a potential feature vector for video i. Users of the same interest will interact and user similarity can be calculated using weighted pearson correlation coefficients:
Figure SMS_58
in the above-mentioned formula(s),
Figure SMS_59
and->
Figure SMS_60
Represents the average score of users u and v, respectively, < >>
Figure SMS_61
The Jaccard correlation coefficient, which is a term capable of affecting the user similarity calculation, is calculated as follows:
Figure SMS_62
wherein,,
Figure SMS_63
representing a video set reviewed by user u, +.>
Figure SMS_64
Representing the video collection reviewed by user v.
The scoring of the item by the user depends on the influence of the adjacent user, and the influence of the adjacent user by the adjacent user can be also possible, but the adjacent user after a certain distance cannot be influenced by the user, namely the adjacent user becomes unreliable, so that a reliable value is introduced, the scoring of the item by the adjacent user with the reliable value being larger than a certain value can be influenced, and the reliable value is calculated in the following way:
Figure SMS_65
，
in the above-mentioned formula(s),
Figure SMS_67
representing the score of video i with u, +.>
Figure SMS_69
Representing the score of user v for video i, +.>
Figure SMS_72
Representing a video set reviewed by user u, +.>
Figure SMS_68
Representing a video set reviewed by user v, +.>
Figure SMS_70
Represents the maximum value of the score,/->
Figure SMS_71
Representing the trust distance, i.e. the number of users present between user u and user v>
Figure SMS_73
Represents the maximum distance allowed between two users, +.>
Figure SMS_66
Is a correction parameter, which is a number greater than 0 and less than 1. Reliable nearest neighbor user is
Figure SMS_74
，
The reliable nearest-neighbor regular term is introduced to restrain the learning of the potential feature matrix of the user when the matrix is decomposed, and the reliable nearest-neighbor regular term restraining function is as follows:
Figure SMS_75
wherein,,
Figure SMS_76
for the potential vector of user u +.>
Figure SMS_77
Is a potential feature vector for user v.
Adding a video association regularization term fused with user liveness and a reliable nearest neighbor regularization term, wherein the final optimization loss function is as follows:
Figure SMS_78
and searching an optimal solution by adopting a random gradient descent method, and finding out an optimal potential feature matrix.
As shown in fig. 2.
The step S3 comprises the following steps: latent feature vector U of user u And video potential feature vector V i As input to the multi-layer perceptron, as shown in FIG. 2, comprising an input layer L in Several hidden layers and output layers L allowing for neural structure nonlinearity out Input layer L in The output vector of (2) is:
Figure SMS_79
the vector is output after the first hidden layer processing:
Figure SMS_80
wherein the method comprises the steps of
Figure SMS_81
Is the set of weights contained in the matrix between the input layer and the first layer hidden layer L1,/and>
Figure SMS_82
is the deviation of the L1 layer, < >>
Figure SMS_83
Is an activation function of
Figure SMS_84
So conceal layer L k The output vector of (2) is:
Figure SMS_85
wherein the method comprises the steps of
Figure SMS_86
For the activation function of neurons, +.>
Figure SMS_87
Is a weight matrix>
Figure SMS_88
Is the deviation. Output layer L of multi-layer perceptron out The output vector is:
Figure SMS_89
as shown in fig. 3.
The step S4 includes: the single-layer perceptron is a merging layer in the deep neural network structure, wherein the deep neural network structure is shown in fig. 3, the multi-layer perceptron model prediction result is combined with the result of the double regularization matrix decomposition model in the merging layer, and the calculation formula is as follows:
Figure SMS_90
wherein,,
Figure SMS_91
to activate the function +.>
Figure SMS_92
For the matrix weight set between the output layer and the merge layer,/for the matrix weight set between the output layer and the merge layer>
Figure SMS_93
For outputting the result of the layer->
Figure SMS_94
For user potential vector, ++>
Figure SMS_95
For video potential vector, ++>
Figure SMS_96
Bias term for merging layers
The proposed model is continuously optimized by adopting a normalized cross entropy method through the following cost functions:
Figure SMS_97
in the above-mentioned formula(s),
Figure SMS_98
for merging layer neuron number, +.>
Figure SMS_99
For the predicted score, ++>
Figure SMS_100
For training example true score, +.>
Figure SMS_101
Representing the maximum value of the score. And continuously optimizing the model by using a gradient descent method to obtain a final prediction score by the cost function.
The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims (2)

1. The video scoring prediction method based on the deep neural network and the double regularization is characterized by comprising the following steps of:
step S1: processing the video comments, mining out hidden information, combining the hidden information matrix with the original user-video scoring matrix to generate a new user-video scoring matrix, and entering into step S2, wherein the step S2 comprises the following steps:
firstly, mining relevant type hidden information in a user video comment by using an LDA model, generating a user type potential feature matrix LU and a video type potential feature matrix LV, and combining the user type potential feature matrix and the video type potential feature matrix to obtain a hidden information matrix L, wherein the calculation formula is as follows:
Figure QLYQS_1
and combining the hidden information matrix L with the original user-video scoring matrix R to generate a new user-video scoring matrix
Figure QLYQS_2
The calculation formula is as follows:
Figure QLYQS_3
step S2: the method comprises the steps that a dual regularization term is added to learn a potential feature matrix when a user-video scoring matrix is decomposed, each user can make a certain contribution to video similarity, the user contributions are not identical, the user can be divided into active users and inactive users from the consideration of the liveness of the users, the active users refer to users with a large number of scoring records on the video, the inactive users refer to users with scoring records on a small number of videos, so that the contributions of the active users and the inactive users are distinguished when the video similarity is calculated, and the liveness of the users is defined as follows:
Figure QLYQS_4
in formula 1, |m (u) | represents the total score of user u, so the video similarity calculation method obtained by combining the activity coefficient of the user with the modified cosine similarity is as follows:
Figure QLYQS_5
in formula 2, R ui Representing the score of user u on video i, R uj Representing the score of user u for video j,
Figure QLYQS_6
representing the average score for user U, which represents the set of users that have both video i and j scored; the learning of the potential feature matrix of the video association regularization term constraint project which is integrated with the user activity is introduced during matrix decomposition, and the video association regularization constraint function formula which is integrated with the user activity at the moment is as follows:
Figure QLYQS_7
in equation 3, V represents the video feature matrix, V j Is the potential feature vector of video j, V i Is a potential feature vector of video i, and proceeds to step S3, which includes:
users of the same interest will interact and user similarity is calculated using weighted pearson correlation coefficients:
Figure QLYQS_8
in the case of the formula 6 of the present invention,
Figure QLYQS_9
and->
Figure QLYQS_10
Representing the average score, r, of users u and v, respectively ui Representing the score of user u on video i, r vi Representing the score of user v on video I u Representing a video set reviewed by user u, I v Video collection representing user v commented, JCC uv The Jaccard correlation coefficient, which is a term capable of affecting the user similarity calculation, is calculated as follows:
Figure QLYQS_11
wherein I is u Representing a video set reviewed by user u, I v Representing a video set reviewed by the user v;
the scoring of the item by the user depends on the influence of the adjacent user and can be influenced by the adjacent user of the adjacent user, but the adjacent user after a certain distance cannot be influenced by the user, namely becomes unreliable, so that a reliable value is introduced, the scoring of the item by the adjacent user with the reliable value larger than a certain value can be influenced, and the reliable value is calculated in the following way:
Figure QLYQS_12
in formula 8, R ui Representing the score of video i with u, R vi Representing the score of user v on video I u Representing a video set reviewed by user u, I v Representing video collections reviewed by user v, R max Represents the maximum value of the score, d uv Representing the trust distance, i.e. the number of users present between user u and user v, d max Representing the maximum distance allowed between two users, ε being a correction parameter, a number greater than 0 and less than 1, may beThe nearest neighbor user is
Near＝{v∈U|sim uv Not less than gamma and C uv ∈delta } formula 9
The reliable nearest-neighbor regular term is introduced to restrain the learning of the potential feature matrix of the user when the matrix is decomposed, and the reliable nearest-neighbor regular term restraining function is as follows:
Figure QLYQS_13
wherein U is u As a potential vector for user U, U v Potential feature vectors for user v;
step S3: taking the potential feature vectors decomposed by the matrix as the input of the multi-layer perceptron, processing the potential feature vectors by the multi-layer perceptron to obtain a multi-layer perceptron model prediction result, and entering into step S4;
step S4: combining the result of the multi-layer perceptron model prediction with the result of matrix decomposition at a merging layer, optimizing the model by using a normalized cross entropy method to obtain a final prediction score, wherein the method comprises the following steps:
in the deep neural network structure, a single-layer perceptron is a merging layer, a prediction result of a multi-layer perceptron model is combined with a result of a double regularization matrix decomposition model at the merging layer, and a calculation formula is as follows:
Figure QLYQS_14
ρ in equation 11 m To activate the function, w m For the matrix weight set between the output layer and the merging layer, y out To output the result of the layer, U u For the user potential vector, V i A is a video potential vector m And optimizing the model by using a normalized cross entropy method for merging the bias terms of the layers, and finally obtaining the prediction scores.
2. The video scoring prediction method based on deep neural network and double regularization according to claim 1, wherein the method is characterized in that: in the step S3, the user potential feature vector and the video potential feature vector are used as input of a multi-layer perceptron, wherein the deep neural network is composed of the multi-layer perceptron and a single-layer perceptron, the multi-layer perceptron comprises an input layer, a plurality of hidden layers allowing nonlinearity of the neural structure and an output layer, and the result of the multi-layer perceptron model is obtained through processing of the multi-layer perceptron by utilizing the nonlinearity characteristics of the hidden layers.
CN202310187456.2A 2023-03-02 2023-03-02 Video scoring prediction method based on deep neural network and double regularization Active CN115905617B (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202310187456.2A CN115905617B (en) 2023-03-02 2023-03-02 Video scoring prediction method based on deep neural network and double regularization
PCT/CN2023/135369 WO2024179089A1 (en) 2023-03-02 2023-11-30 Video score prediction method based on deep neural network and double regularization

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310187456.2A CN115905617B (en) 2023-03-02 2023-03-02 Video scoring prediction method based on deep neural network and double regularization

Publications (2)

Publication Number Publication Date
CN115905617A CN115905617A (en) 2023-04-04
CN115905617B true CN115905617B (en) 2023-05-30

Family

ID=85744775

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310187456.2A Active CN115905617B (en) 2023-03-02 2023-03-02 Video scoring prediction method based on deep neural network and double regularization

Country Status (2)

Country Link
CN (1) CN115905617B (en)
WO (1) WO2024179089A1 (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115905617B (en) * 2023-03-02 2023-05-30 南京邮电大学 Video scoring prediction method based on deep neural network and double regularization

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112668374A (en) * 2019-10-16 2021-04-16 北京灵汐科技有限公司 Image processing method and device, re-recognition network training method and electronic equipment
CN111581333B (en) * 2020-05-07 2023-05-26 重庆大学 Text-CNN-based video and audio play list pushing method and video and audio play list pushing system
CN114117191A (en) * 2020-08-26 2022-03-01 付昳漫 Context-aware recommendation method based on deep neural network
CN112328900A (en) * 2020-11-27 2021-02-05 北京工业大学 Deep learning recommendation method integrating scoring matrix and comment text
CN114625980A (en) * 2020-12-13 2022-06-14 国网吉林省电力有限公司培训中心 Dual-regularization collaborative filtering recommendation method
CN113157957A (en) * 2021-03-05 2021-07-23 北京工业大学 Attribute graph document clustering method based on graph convolution neural network
CN113807422B (en) * 2021-09-07 2024-05-31 南京邮电大学 Weighted graph convolutional neural network scoring prediction model integrating multi-feature information
CN115905617B (en) * 2023-03-02 2023-05-30 南京邮电大学 Video scoring prediction method based on deep neural network and double regularization

Also Published As

Publication number Publication date
WO2024179089A1 (en) 2024-09-06
CN115905617A (en) 2023-04-04

Similar Documents

Publication Publication Date Title
Pan et al. A correlative denoising autoencoder to model social influence for top-N recommender system
Bansal et al. Ask the gru: Multi-task learning for deep text recommendations
Ni et al. An effective recommendation model based on deep representation learning
CN111079409B (en) Emotion classification method utilizing context and aspect memory information
CN110309195B (en) FWDL (full Width Domain analysis) model based content recommendation method
CN110866145B (en) Co-preference-assisted deep single-class collaborative filtering recommendation method
CN111461157A (en) Self-learning-based cross-modal Hash retrieval method
CN107507073A (en) Based on the service recommendation method for trusting extension and the sequence study of list level
CN112800344B (en) Deep neural network-based movie recommendation method
CN111104601B (en) Antagonistic multi-feedback-level paired personalized ranking method
Chen et al. A survey on heterogeneous one-class collaborative filtering
CN111241425A (en) POI recommendation method based on hierarchical attention mechanism
CN116071128A (en) Multitask recommendation method based on multi-behavioral feature extraction and self-supervision learning
CN115905617B (en) Video scoring prediction method based on deep neural network and double regularization
Li et al. Coltr: Semi-supervised learning to rank with co-training and over-parameterization for web search
Li et al. Capsule neural tensor networks with multi-aspect information for Few-shot Knowledge Graph Completion
CN116662564A (en) Service recommendation method based on depth matrix decomposition and knowledge graph
Lin et al. Transfer learning for collaborative recommendation with biased and unbiased data
CN110889505A (en) Cross-media comprehensive reasoning method and system for matching image-text sequences
Lian et al. Financial product recommendation system based on transformer
Rauf et al. BCE4ZSR: Bi-encoder empowered by teacher cross-encoder for zero-shot cold-start news recommendation
CN114841778B (en) Commodity recommendation method based on dynamic graph neural network
CN117056609A (en) Session recommendation method based on multi-layer aggregation enhanced contrast learning
CN117539999A (en) Cross-modal joint coding-based multi-modal emotion analysis method
Kim et al. COVID-19 outbreak prediction using Seq2Seq+ Attention and Word2Vec keyword time series data

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