CN110309359A - Video correlation prediction technique, device, equipment and storage medium - Google Patents
Video correlation prediction technique, device, equipment and storage medium Download PDFInfo
- Publication number
- CN110309359A CN110309359A CN201910420026.4A CN201910420026A CN110309359A CN 110309359 A CN110309359 A CN 110309359A CN 201910420026 A CN201910420026 A CN 201910420026A CN 110309359 A CN110309359 A CN 110309359A
- Authority
- CN
- China
- Prior art keywords
- sample
- neural network
- video
- uncorrelated
- video sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—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
- H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
- H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
- H04N21/4668—Learning process for intelligent management, e.g. learning user preferences for recommending movies for recommending content, e.g. movies
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Signal Processing (AREA)
- Image Analysis (AREA)
Abstract
This application discloses a kind of video correlation prediction techniques, device, equipment and storage medium, relationship between constructing and training half twin neural network model study related or uncorrelated video, it will be in half twin neural network model described in target source video and at least one video input to be predicted, the correlation results between the video to be predicted and the target source video are exported after described half twin neural network model analysis, relative list of videos can be predicted to the video for lacking user behavior information to realize, the correlation given between the correlation size and each associated video of each associated video is symmetrical, effectively increase the accuracy of video recommendations.
Description
Technical field
This application involves video data process field, in particular to a kind of video correlation prediction technique, device, equipment and
Storage medium.
Background technique
With the development of information technology and Internet technology, slowly developed to based on the media data of audio-video from scarcity
Information overload.In the data of magnanimity, the interested information of user how is found, and interested according to some that user finds
Information goes excavation to find that more information that are interested, being associated become to be increasingly difficult to.
Recommender system screens the data of magnanimity, mistake as a kind of information filtering platform by specific algorithm
Filter, for predicting user to " preference " of information.In general, effective solution party of the search engine as information overload, needs
It could be that it realizes accurate commending contents that user, which provides clear demand, this is a very passively process.And recommender system one
As for, do not need user and one clear demand be provided, by the previous habit of analysis user, it is modeled, it is real
Referring now to the personalized recommendation of Different Individual.And video correlation prediction be in online streaming media service most important task it
One and recommender system algorithm research important component, according to user watch or search for video record, recommender system
Personalized recommendation is capable of providing to help user to find more interested video contents, and this will inevitably bring
" cold start-up " problem, i.e. system are difficult to provide associated recommendation because that can not obtain the user behavior record of new video.
Currently, for there are two ways to solving the problems, such as above-mentioned " cold start-up ": one is the collaborative filterings based on user;Separately
One is content-based recommendations.Wherein, the collaborative filtering based on user is the content phase recommended and liked before it to user
As content, but this mode is not from content angle, but likes A and user likes B based on user, then A and B phase
Seemingly.This mode (leads some associated metadatas of new video content because lacking when solving the problems, such as " to be cold-started "
Drill, performer), so recommendation effect is bad.Content-based recommendation is similar with the content that it is being liked to user's recommendation
Other contents, therefore do not need the problem of historical behavior data of user are provided, can solve " cold start-up ".
But current recommender system be based on new video, to user recommend may interested video when, often neglect
Depending on the correlation size relation between video, all associated videos are made no exception, and recommend the correlation between video
And it is asymmetric.Therefore, the technical issues of accuracy for how improving video recommendations is this field urgent need to resolve.
Summary of the invention
The application's is designed to provide a kind of video correlation prediction technique, device, equipment and storage medium, to improve
The accuracy of video recommendations.
In a first aspect, the embodiment of the present application provides a kind of video correlation prediction technique, comprising:
Obtain source video sample and its associated video list and uncorrelated list of videos;Include in the associated video list
Multiple associated video samples sorted from large to small according to correlation include multiple uncorrelated views in the uncorrelated list of videos
Frequency sample;
The primary features of the source video sample, the associated video sample and the uncorrelated video sample are extracted respectively
Vector;
To the primary features vector of the source video sample, the associated video sample and the uncorrelated video sample into
Row resampling constructs sample data pair;
To the sample data to symmetry division is carried out, symmetric data set and asymmetric data set are obtained;
Initial neural network is constructed, the initial neural network is trained according to the symmetric data set, is obtained
Initial neural network model;
Half twin neural network is constructed, the half twin neural network includes two initial neural network models, root
Stand-alone training is carried out to described half twin neural network according to the asymmetric data set, obtains half twin neural network model;
By in half twin neural network model described in target source video and at least one video input to be predicted, through described half
The correlation results between the video to be predicted and the target source video are exported after twin neural network model analysis.
Second aspect, the embodiment of the present application provide a kind of video correlation prediction meanss, comprising:
Neural metwork training module, is used for:
Obtain source video sample and its associated video list and uncorrelated list of videos;Include in the associated video list
Multiple associated video samples sorted from large to small according to correlation include multiple uncorrelated views in the uncorrelated list of videos
Frequency sample;
The primary features of the source video sample, the associated video sample and the uncorrelated video sample are extracted respectively
Vector;
To the primary features vector of the source video sample, the associated video sample and the uncorrelated video sample into
Row resampling constructs sample data pair;
To the sample data to symmetry division is carried out, symmetric data set and asymmetric data set are obtained;
Initial neural network is constructed, the initial neural network is trained according to the symmetric data set, is obtained
Initial neural network model;
Half twin neural network is constructed, the half twin neural network includes two initial neural network models, root
Stand-alone training is carried out to described half twin neural network according to the asymmetric data set, obtains half twin neural network model;
Dependency prediction module, is used for:
By in half twin neural network model described in target source video and at least one video input to be predicted, through described half
The correlation results between the video to be predicted and the target source video are exported after twin neural network model analysis.
The third aspect, the embodiment of the present application provide a kind of electronic equipment, comprising: memory and processor;
The memory, for storing computer program;
Wherein, the processor executes the computer program in the memory, to realize described in above-mentioned first aspect
Method.
Fourth aspect, the embodiment of the present application provide a kind of computer readable storage medium, the computer-readable storage
Computer program is stored in medium, for realizing described in above-mentioned first aspect when the computer program is executed by processor
Method.
Compared with prior art, video correlation prediction technique, device, equipment and storage medium provided by the present application lead to
The relationship constructed and between related/uncorrelated video of half twin neural network model study of training is crossed, is realized to lacking user behavior
The video of information can predict relative list of videos, give the correlation size of each associated video and each related view
Correlation between frequency is symmetrical, effectively increases the accuracy of video recommendations.
Detailed description of the invention
Fig. 1 is the flow diagram for the video correlation prediction technique that the embodiment of the present application one provides;
Fig. 2 is the structural schematic diagram for the video correlation prediction meanss that the embodiment of the present application two provides;
Fig. 3 is the structural schematic diagram for the electronic equipment that the embodiment of the present application three provides.
Specific embodiment
With reference to the accompanying drawing, the specific embodiment of the application is described in detail, it is to be understood that the guarantor of the application
Shield range is not limited by the specific implementation.
Unless otherwise explicitly stated, otherwise in entire disclosure and claims, term " includes " or its change
Changing such as "comprising" or " including " etc. will be understood to comprise stated element or component, and not exclude other members
Part or other component parts.
Fig. 1 is the flow diagram for the video correlation prediction technique that the embodiment of the present application one provides, as shown in Figure 1, should
Method includes:
S101, source video sample and its associated video list and uncorrelated list of videos are obtained;The associated video list
In include multiple associated video samples sorted from large to small according to correlation, include in the uncorrelated list of videos it is multiple not
Associated video sample;
In practical application, the video correlation information data obtained using actual acquisition in daily life, according to its correlation
Property size sequence, obtain associated video list, following format: (source video A: associated video 1, associated video 2 ..., associated video
N).It is far longer than associated video quantity with other incoherent number of videos of source video.
Specifically, for obtained video correlation information data, for each source video vi, obtain relative
List of videosWhereinWherein j represents the significance level of the video, MiIndicate related
The total length of list of videos and the total number of associated video.That is, having its corresponding phase for each source video
List of videos is closed, the video property associated therewith before list is big, and the subsequent video of list property associated therewith is smaller.
S102, the first of the source video sample, the associated video sample and the uncorrelated video sample is extracted respectively
Grade feature vector;
In practical application, S102 specific implementation are as follows: the source is extracted using the convolutional neural networks model of pre-training respectively
The primary features vector of video sample, the associated video sample and the uncorrelated video sample.
Specifically, the convolutional neural networks model can instruct on super large data set (such as ImageNet image data set)
The C3D convolutional neural networks model got, for each specific source video vi, obtain its corresponding primary features
Characteristic dimension is 512.For associated video listObtain its corresponding primary features list relatedWhereinTo put it more simply, willWrite Fi, willWriting
S103, to the primary features of the source video sample, the associated video sample and the uncorrelated video sample
Vector carries out resampling, constructs sample data pair;
Specifically, the sample data to include positive sample data to and negative sample data pair.S103 specific implementation are as follows: really
Total sampling number N of the fixed primary features vector to the source video sample, N are the integer greater than 1;Using harmonious sampling policy
Resampling is carried out to the primary features vector of the associated video sample, constructs positive sample data pair, quantity N/2;It is described
Positive sample data are to the primary features vector F including source video samplea, associated video sample primary features vector FbAnd correlation
Label is denoted as (Fa,Fb,1);Resampling is carried out to the primary features vector of the uncorrelated video sample, constructs negative sample data
It is right, quantity N/2;The negative sample data are to the primary features vector F including source video samplea, uncorrelated video sample
Primary features vector FbWith uncorrelated label, it is denoted as (Fa,Fb,0)。FaIndicate first sample of sample data pair, FbIt indicates
Second sample of sample data pair.
In practical application, for each specific video ViCorresponding primary features Fi, its available corresponding phase
Close list of videos VrWith feature list relatedPrimary features F corresponding to the videoiCarry out construction of the sample data to P, structure
At the sample data pair of following form: (Fi,Fj,y);Wherein, y is correlation label, if video primary features FiAt the beginning of video
Grade feature FjCorrelation, then correlation label y is 1, otherwise is 0.
Therefore, for feature list relatedEach of primary featuresPositive sample data are to can indicate are as follows:
Because with FiIncoherent feature quantity is far longer than relative quantity Mi, in order to balance the number of positive negative sample
Amount, to FiThe positive sample data of construction N/2 are to the negative sample data of (label 1) and N/2 to (label 0).
Further, resampling is carried out using primary features vector of the harmonious sampling policy to the associated video sample,
Include:
The frequency for needing to be sampled using the primary features vector that frequency calculation formula calculates the associated video sample
Then the primary features vector of the associated video sample needs the number that is sampled to be
The frequency calculation formula are as follows:
Wherein,Wherein, i represents source video sample, and j represents associated video
Significance level, the smaller importance of j is stronger, MiIndicate the total number of associated video, P is default hyper parameter.
That is, for any one primary featuresThe number for needing to be resampled isNeed structure
It buildsIt is aPositive sample data pair, then from FiAppoint in incoherent all video features and takes N/2 composition
Negative sample data pair.
Resampling is carried out to each source video according to the above harmonious sampling policy, constitutes the sample data with correlation label
To P.
S104, to the sample data to carry out symmetry division, obtain symmetric data set and asymmetric data set;
In practical application, S104 specific implementation are as follows: according to the whether symmetrical principle of correlation, to the sample data into
Row symmetry divides, and obtains symmetric data set and asymmetric data set.
The whether symmetrical principle of correlation are as follows: for any one data to (Fa,Fb, y), if FbIn FaCorrelation
While in list, FaAlso in FbDependency list in, i.e. ya,b=yb,a, then it is assumed that (Fa,Fb,ya,b) and (Fb,Fa,yb,a) be
Symmetrically, whereas if FbIn FaDependency list in but FaNot in FbDependency list, then it is assumed that (Fa,Fb,ya,b) and
(Fb,Fa,yb,a) it is asymmetric.
According to the above principle, symmetry division is carried out to P to sample data, obtains symmetric data to set P1 and asymmetric
Data are to set P2.
S105, the initial neural network of building, are trained the initial neural network according to the symmetric data set,
Obtain initial neural network model;
S106, half twin neural network of building, the half twin neural network include two initial neural network moulds
Type carries out stand-alone training to described half twin neural network according to the asymmetric data set, obtains half twin neural network
Model;
In practical application, the full Connection Neural Network comprising 5 hidden layers can be designed, the dimension of input layer meets initially
The dimension of feature F, for example, 512.First hidden layer includes 2048 nodes, second and third hidden layer includes 512 nodes,
4th hidden layer includes 1100 nodes, and the 5th hidden layer includes 800 nodes, and ReLU is used after the 5th hidden layer
Activation primitive activates feature, and output layer is 512 nodes, that is, exports and be characterized in 512 dimensions, carries out by tanh function
Activation.
Initial characteristics F can be mapped to a new feature space by the full Connection Neural Network of design, obtain new spy
Levy S.
Further, since P1 is the symmetric data pair divided by screening, that is, (F is existed simultaneouslya,Fb, 1) and (Fb,Fa,
1) either (Fa,Fb, 0) and (Fb,Fa, 0), sum is N1.Using twin neural network to sample datas all in P1 to progress
Supervised training, specific training method are as follows:
(1) feature extraction is carried out to N1 data respectively using initial neural network M, for Fa,Fb, it is obtained new
Feature Mapping in feature space(512 dimension);
(2) Euclidean distance between new feature after calculating each sample data to feature extraction is passed through, forCalculation is as follows:
(3) using losing as follows with feature penalty term purpose comparison loss function calculating neural network, wherein m indicates pre-
The significant interval of definition,For the predefined strength of punishment to feature:
(4) neural network parameter is updated and is learnt using gradient back-propagation algorithm, obtained after the completion of training just
Beginning neural network model M_O.
Further, two models completely the same with initial neural network model M_O are constructed, M_F and M_L are denoted as.
Since P2 is the asymmetric data pair divided by screening, that is, exist simultaneously (Fa,Fb, 1) and (Fb,Fa, 0) or
It is (Fa,Fb, 1) and (Fb,Fa, 0), sum is N2.Using half twin neural network to data all in P2 to the instruction that exercises supervision
Practice, specific training method is as follows:
(1) for first sample of sample data pair each in P2, feature extraction is carried out to it using M_F.For
Two samples carry out feature extraction to it using M_L.For example, for Fa,Fb, obtain its feature in new feature space and reflect
It penetrates(512 dimension).
(2) each sample data is calculated to by the Euclidean distance between the new feature after feature extraction.ForCalculation is as follows:
(3) it is lost using following with feature penalty term purpose comparison loss function calculating neural network:
(4) loss being calculated is passed back M_F and M_L respectively, using gradient back-propagation algorithm to two models point
Not, it independently optimizes, M_F and M_L training are completed, and half twin neural network model is obtained.
S107, by half twin neural network model described in target source video and at least one video input to be predicted, warp
The correlation knot between the video to be predicted and the target source video is exported after the half twin neural network model analysis
Fruit.
In practical application, for the primary features of any number of videos, it is used for M_F and M_L progress feature respectively and mentions
It takes, available two characteristic set RFAnd RL.When needing to obtain a certain video viAssociated video list when, Ke Yifen
It does not calculateWith remove viExcept all videos RLDistance between feature, and according to characteristic distance size to it corresponding to source
Video is ranked up, and obtains the associated video list of correlation from big to small.
Therefore, by constructing and between related/uncorrelated video of the twin neural network model study of training half in this implementation
Relationship, realization can predict relative list of videos to the video for lacking user behavior information, give each associated video
Correlation size and each associated video between correlation it is symmetrical, effectively increase the accuracy of video recommendations.
Following is the application Installation practice, can be used for executing the application embodiment of the method.It is real for the application device
Undisclosed details in example is applied, the application embodiment of the method is please referred to.
Fig. 2 is the structural schematic diagram for the video correlation prediction meanss that the embodiment of the present application two provides, as shown in Fig. 2, should
Device includes:
Neural metwork training module 210, is used for:
Obtain source video sample and its associated video list and uncorrelated list of videos;Include in the associated video list
Multiple associated video samples sorted from large to small according to correlation include multiple uncorrelated views in the uncorrelated list of videos
Frequency sample;
The primary features of the source video sample, the associated video sample and the uncorrelated video sample are extracted respectively
Vector;
To the primary features vector of the source video sample, the associated video sample and the uncorrelated video sample into
Row resampling constructs sample data pair;
To the sample data to symmetry division is carried out, symmetric data set and asymmetric data set are obtained;
Initial neural network is constructed, the initial neural network is trained according to the symmetric data set, is obtained
Initial neural network model;
Half twin neural network is constructed, the half twin neural network includes two initial neural network models, root
Stand-alone training is carried out to described half twin neural network according to the asymmetric data set, obtains half twin neural network model;
Dependency prediction module 220, is used for:
By in half twin neural network model described in target source video and at least one video input to be predicted, through described half
The correlation results between the video to be predicted and the target source video are exported after twin neural network model analysis.
Video correlation prediction meanss provided in this embodiment, by constructing and training half twin neural network model to learn
Relationship between related/uncorrelated video, realizing can predict that relative video arranges to the video for lacking user behavior information
Table, the correlation given between the correlation size and each associated video of each associated video is symmetrical, effectively increases video
The accuracy of recommendation.
Optionally, the neural metwork training module, is specifically used for:
Using the convolutional neural networks model of pre-training extract respectively the source video sample, the associated video sample and
The primary features vector of the uncorrelated video sample.
Optionally, the sample data to include positive sample data to and negative sample data pair;
Optionally, the neural metwork training module, is specifically used for:
Determine that total sampling number N to the primary features vector of the source video sample, N are the integer greater than 1;
Resampling is carried out using primary features vector of the harmonious sampling policy to the associated video sample, constructs positive sample
Data pair, quantity N/2;The positive sample data are to the primary features vector F including source video samplea, associated video sample
This primary features vector FbAnd respective labels, it is denoted as (Fa,Fb, 1);
Resampling is carried out to the primary features vector of the uncorrelated video sample, constructs negative sample data pair, quantity
For N/2;The negative sample data are to the primary features vector F including source video samplea, uncorrelated video sample primary features
Vector FbWith uncorrelated label, it is denoted as (Fa,Fb,0)。
Optionally, the neural metwork training module, is specifically used for:
The frequency for needing to be sampled using the primary features vector that frequency calculation formula calculates the associated video sample
Then the primary features vector of the associated video sample needs the number that is sampled to be
The frequency calculation formula are as follows:
Wherein,Wherein, i represents source video sample, and j represents associated video
Significance level, the smaller importance of j is stronger, MiIndicate the total number of associated video, P is default hyper parameter.
Optionally, the neural metwork training module, is specifically used for:
Symmetric data is obtained to the sample data to symmetry division is carried out according to the whether symmetrical principle of correlation
Set and asymmetric data set.
Optionally, the neural metwork training module, is specifically used for:
Initial neural network M is constructed, using the initial neural network M to sample data pair in the symmetric data set
Two samples carry out feature extraction, obtain the corresponding new feature of two samples;
Calculate the first Euclidean distance between the corresponding new feature of described two samples;
According to first Euclidean distance, neural network loss is calculated using comparison loss function, is reversely passed using gradient
It broadcasts algorithm neural network parameter is updated and is learnt, initial neural network model M_O is obtained after the completion of training.
Optionally, the neural metwork training module, is specifically used for:
Half twin neural network is constructed, which includes two initial neural network model M_O, is remembered respectively
For M_F and M_L;
For first sample of each sample data pair in asymmetric data set, feature is carried out to it using M_F and is mentioned
It takes, obtains corresponding new feature;For second sample of each sample data pair in asymmetric data set, M_L pairs is used
It carries out feature extraction, obtains corresponding new feature;
Calculate the second Euclidean between first sample of each sample data pair and the corresponding new feature of second sample
Distance;
According to second Euclidean distance, neural network loss is calculated using comparison loss function, the loss is distinguished
It passes M_F and M_L back, is separately optimized using parameter of the gradient back-propagation algorithm to M_F and M_L, it is twin to obtain half after the completion of training
Raw neural network model.
Fig. 3 is the structural schematic diagram for the electronic equipment that the embodiment of the present application three provides, as shown in figure 3, the equipment includes: to deposit
Reservoir 301 and processor 302;
Memory 301, for storing computer program;
Wherein, processor 302 executes the computer program in memory 301, to realize each method embodiment as described above
Provided method.
In embodiment, example is carried out to video correlation prediction meanss provided by the present application with an electronic equipment.Processing
Device can be the place of central processing unit (CPU) or the other forms with data-handling capacity and/or instruction execution capability
Unit is managed, and can control the other assemblies in electronic equipment to execute desired function.
Memory may include one or more computer program products, and computer program product may include various forms
Computer readable storage medium, such as volatile memory and/or nonvolatile memory.Volatile memory for example can be with
Including random access memory (RAM) and/or cache memory (cache) etc..Nonvolatile memory for example can wrap
Include read-only memory (ROM), hard disk, flash memory etc..It can store one or more computers on computer readable storage medium
Program instruction, processor can run program instruction, method in each embodiment to realize the application above and/or
Other desired functions of person.Such as input signal, signal component, noise point can also be stored in a computer-readable storage medium
The various contents such as amount.
The embodiment of the present application four provides a kind of computer readable storage medium, stores in the computer readable storage medium
There is computer program, for realizing side provided by each method embodiment as described above when which is executed by processor
Method.
In practical application, the computer program in the present embodiment can be with any group of one or more programming languages
It closes to write the program code for executing the embodiment of the present application operation, programming language includes object-oriented programming
Language, Java, C++, python etc. further include conventional procedural programming language, such as " C " language or similar
Programming language.Program code can be executed fully on the user computing device, partly execute, make on a user device
It is executed for an independent software package, part partially executes on a remote computing on the user computing device or complete
It is executed in remote computing device or server.
In practical application, computer readable storage medium can be using any combination of one or more readable mediums.It can
Reading medium can be readable signal medium or readable storage medium storing program for executing.Readable storage medium storing program for executing for example can include but is not limited to electricity,
Magnetic, optical, electromagnetic, infrared ray or semiconductor system, device or device, or any above combination.Readable storage medium storing program for executing
More specific example (non exhaustive list) includes: electrical connection with one or more conducting wires, portable disc, hard disk, random
It accesses memory (RAM), read-only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, portable
Formula compact disk read-only memory (CD-ROM), light storage device, magnetic memory device or above-mentioned any appropriate combination.
The description of the aforementioned specific exemplary embodiment to the application is in order to illustrate and illustration purpose.These descriptions
It is not wishing to for the application to be limited to disclosed precise forms, and it will be apparent that according to the above instruction, can much be changed
And variation.The purpose of selecting and describing the exemplary embodiment is that explaining the specific principle of the application and its actually answering
With so that those skilled in the art can be realized and utilize the application a variety of different exemplary implementation schemes and
Various chooses and changes.Scope of the present application is intended to be limited by claims and its equivalents.
Claims (10)
1. a kind of video correlation prediction technique characterized by comprising
Obtain source video sample and its associated video list and uncorrelated list of videos;It include multiple in the associated video list
The associated video sample sorted from large to small according to correlation includes multiple uncorrelated video samples in the uncorrelated list of videos
This;
Extract respectively the primary features of the source video sample, the associated video sample and the uncorrelated video sample to
Amount;
Weight is carried out to the primary features vector of the source video sample, the associated video sample and the uncorrelated video sample
Sampling constructs sample data pair;
To the sample data to symmetry division is carried out, symmetric data set and asymmetric data set are obtained;
Initial neural network is constructed, the initial neural network is trained according to the symmetric data set, is obtained initial
Neural network model;
Half twin neural network is constructed, the half twin neural network includes two initial neural network models, according to institute
It states asymmetric data set and stand-alone training is carried out to described half twin neural network, obtain half twin neural network model;
It is twin through described half by half twin neural network model described in target source video and at least one video input to be predicted
The correlation results between the video to be predicted and the target source video are exported after neural network model analysis.
2. the method according to claim 1, wherein described extract the source video sample, the correlation respectively
The primary features vector of video sample and the uncorrelated video sample, comprising:
The source video sample, the associated video sample and described are extracted using the convolutional neural networks model of pre-training respectively
The primary features vector of uncorrelated video sample.
3. the method according to claim 1, wherein the sample data to include positive sample data to negative sample
Notebook data pair;
The primary features vector to the source video sample, the associated video sample and the uncorrelated video sample into
Row resampling constructs sample data pair, comprising:
Determine that total sampling number N to the primary features vector of the source video sample, N are the integer greater than 1;
Resampling is carried out using primary features vector of the harmonious sampling policy to the associated video sample, constructs positive sample data
It is right, quantity N/2;The positive sample data are to the primary features vector F including source video samplea, associated video sample
Primary features vector FbAnd respective labels, it is denoted as (Fa,Fb,1);
Resampling is carried out to the primary features vector of the uncorrelated video sample, constructs negative sample data pair, quantity N/
2;The negative sample data are to the primary features vector F including source video samplea, uncorrelated video sample primary features vector
FbWith uncorrelated label, it is denoted as (Fa,Fb,0)。
4. according to the method described in claim 3, it is characterized in that, described use harmonious sampling policy to the associated video sample
This primary features vector carries out resampling, comprising:
The frequency for needing to be sampled using the primary features vector that frequency calculation formula calculates the associated video sampleThen institute
The primary features vector for stating associated video sample needs the number that is sampled to be
The frequency calculation formula are as follows:
Wherein,Wherein, i represents source video sample, and j represents the important journey of associated video
Degree, the smaller importance of j is stronger, MiIndicate the total number of associated video, P is default hyper parameter.
5. according to the method described in claim 4, it is characterized in that, it is described to the sample data to carry out symmetry division,
Obtain symmetric data set and asymmetric data set, comprising:
Symmetric data set is obtained to the sample data to symmetry division is carried out according to the whether symmetrical principle of correlation
With asymmetric data set.
6. according to the method described in claim 5, it is characterized in that, described construct initial neural network, according to the symmetry number
The initial neural network is trained according to set, obtains initial neural network model, comprising:
Initial neural network M is constructed, using the initial neural network M to two of sample data pair in the symmetric data set
Sample carries out feature extraction, obtains the corresponding new feature of two samples;
Calculate the first Euclidean distance between the corresponding new feature of described two samples;
According to first Euclidean distance, neural network loss is calculated using comparison loss function, is calculated using gradient backpropagation
Method is updated and learns to neural network parameter, and initial neural network model M_O is obtained after the completion of training.
7. according to the method described in claim 6, it is characterized in that, half twin neural network of the building, the half twin mind
It include two initial neural network models through network, according to the asymmetric data set to described half twin neural network
It is trained, obtains half twin neural network model, comprising:
Half twin neural network is constructed, which includes two initial neural network model M_O, is denoted as M_ respectively
F and M_L;
For first sample of each sample data pair in asymmetric data set, feature extraction is carried out to it using M_F, is obtained
To corresponding new feature;For second sample of each sample data pair in asymmetric data set, it is carried out using M_L
Feature extraction obtains corresponding new feature;
Calculate the second Euclidean distance between first sample of each sample data pair and the corresponding new feature of second sample;
According to second Euclidean distance, neural network loss is calculated using comparison loss function, the loss is passed back respectively
M_F and M_L is separately optimized using parameter of the gradient back-propagation algorithm to M_F and M_L, and half twin mind is obtained after the completion of training
Through network model.
8. a kind of video correlation prediction meanss characterized by comprising
Neural metwork training module, is used for:
Obtain source video sample and its associated video list and uncorrelated list of videos;It include multiple in the associated video list
The associated video sample sorted from large to small according to correlation includes multiple uncorrelated video samples in the uncorrelated list of videos
This;
Extract respectively the primary features of the source video sample, the associated video sample and the uncorrelated video sample to
Amount;
Weight is carried out to the primary features vector of the source video sample, the associated video sample and the uncorrelated video sample
Sampling constructs sample data pair;
To the sample data to symmetry division is carried out, symmetric data set and asymmetric data set are obtained;
Initial neural network is constructed, the initial neural network is trained according to the symmetric data set, is obtained initial
Neural network model;
Half twin neural network is constructed, the half twin neural network includes two initial neural network models, according to institute
It states asymmetric data set and stand-alone training is carried out to described half twin neural network, obtain half twin neural network model;
Dependency prediction module, is used for:
It is twin through described half by half twin neural network model described in target source video and at least one video input to be predicted
The correlation results between the video to be predicted and the target source video are exported after neural network model analysis.
9. a kind of electronic equipment, comprising: memory and processor;
The memory, for storing computer program;
Wherein, the processor executes the computer program in the memory, to realize such as any one of claim 1-7 institute
The method stated.
10. a kind of computer readable storage medium, which is characterized in that be stored with computer in the computer readable storage medium
Program, for realizing method such as of any of claims 1-7 when the computer program is executed by processor.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910420026.4A CN110309359B (en) | 2019-05-20 | 2019-05-20 | Video correlation prediction method, device, equipment and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910420026.4A CN110309359B (en) | 2019-05-20 | 2019-05-20 | Video correlation prediction method, device, equipment and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110309359A true CN110309359A (en) | 2019-10-08 |
CN110309359B CN110309359B (en) | 2021-06-15 |
Family
ID=68075484
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910420026.4A Active CN110309359B (en) | 2019-05-20 | 2019-05-20 | Video correlation prediction method, device, equipment and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110309359B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111143604A (en) * | 2019-12-25 | 2020-05-12 | 腾讯音乐娱乐科技(深圳)有限公司 | Audio similarity matching method and device and storage medium |
WO2022171067A1 (en) * | 2021-02-09 | 2022-08-18 | 北京有竹居网络技术有限公司 | Video processing method and apparatus, and storage medium and device |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109117744A (en) * | 2018-07-20 | 2019-01-01 | 杭州电子科技大学 | A kind of twin neural network training method for face verification |
CN109243467A (en) * | 2018-11-14 | 2019-01-18 | 龙马智声(珠海)科技有限公司 | Sound-groove model construction method, method for recognizing sound-groove and system |
US20190138896A1 (en) * | 2017-11-03 | 2019-05-09 | Samsung Electronics Co., Ltd. | Method for Optimizing Neural Networks |
-
2019
- 2019-05-20 CN CN201910420026.4A patent/CN110309359B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190138896A1 (en) * | 2017-11-03 | 2019-05-09 | Samsung Electronics Co., Ltd. | Method for Optimizing Neural Networks |
CN109117744A (en) * | 2018-07-20 | 2019-01-01 | 杭州电子科技大学 | A kind of twin neural network training method for face verification |
CN109243467A (en) * | 2018-11-14 | 2019-01-18 | 龙马智声(珠海)科技有限公司 | Sound-groove model construction method, method for recognizing sound-groove and system |
Non-Patent Citations (1)
Title |
---|
刘琼等: "基于深度信念卷积神经网络的图像识别与分类", 《清华大学学报(自然科学版)》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111143604A (en) * | 2019-12-25 | 2020-05-12 | 腾讯音乐娱乐科技(深圳)有限公司 | Audio similarity matching method and device and storage medium |
CN111143604B (en) * | 2019-12-25 | 2024-02-02 | 腾讯音乐娱乐科技(深圳)有限公司 | Similarity matching method and device for audio frequency and storage medium |
WO2022171067A1 (en) * | 2021-02-09 | 2022-08-18 | 北京有竹居网络技术有限公司 | Video processing method and apparatus, and storage medium and device |
Also Published As
Publication number | Publication date |
---|---|
CN110309359B (en) | 2021-06-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110532479A (en) | A kind of information recommendation method, device and equipment | |
CN110781321B (en) | Multimedia content recommendation method and device | |
US9110983B2 (en) | Traversing data utilizing data relationships | |
CN110377913B (en) | Emotion analysis method and device, electronic equipment and storage medium | |
CN113688313A (en) | Training method of prediction model, information pushing method and device | |
CN110033382A (en) | A kind of processing method of insurance business, device and equipment | |
CN110309359A (en) | Video correlation prediction technique, device, equipment and storage medium | |
Wang | A hybrid recommendation for music based on reinforcement learning | |
CN113934851A (en) | Data enhancement method and device for text classification and electronic equipment | |
CN114417174B (en) | Content recommendation method, device, equipment and computer storage medium | |
US20170286522A1 (en) | Data file grouping analysis | |
CN117009650A (en) | Recommendation method and device | |
Li et al. | From edge data to recommendation: A double attention-based deformable convolutional network | |
CN110267102B (en) | Video data processing method and device, medium and electronic equipment | |
CN116821513B (en) | Parameter searching method, device, equipment and medium under recommended scene | |
CN113590976A (en) | Recommendation method of space self-adaptive graph convolution network | |
US20220147547A1 (en) | Analogy based recognition | |
CN112784157A (en) | Training method of behavior prediction model, behavior prediction method, device and equipment | |
Alluhaidan | Recommender system using collaborative filtering algorithm | |
CN111401641B (en) | Service data processing method and device and electronic equipment | |
CN116340645A (en) | Content recommendation based on graph-enhanced collaborative filtering | |
CN110020118A (en) | A kind of method and device calculating similarity between user | |
Quadrana | Algorithms for sequence-aware recommender systems | |
Xu et al. | How Does the System Perceive Me?—A Transparent and Tunable Recommender System | |
US12002129B1 (en) | Systems and methods of generating a display data structure from an input signal |
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 |