CN110188283A - Information recommendation method and system based on joint neural network collaborative filtering - Google Patents
Information recommendation method and system based on joint neural network collaborative filtering Download PDFInfo
- Publication number
- CN110188283A CN110188283A CN201910484886.4A CN201910484886A CN110188283A CN 110188283 A CN110188283 A CN 110188283A CN 201910484886 A CN201910484886 A CN 201910484886A CN 110188283 A CN110188283 A CN 110188283A
- Authority
- CN
- China
- Prior art keywords
- user
- article
- collaborative filtering
- information
- network
- 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
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—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
- G06N3/045—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
Abstract
The invention provides an information recommendation method and system based on joint neural network collaborative filtering, which comprises the following steps: extracting user characteristic information and article characteristic information based on the user-article scoring matrix and by combining a deep neural network A; modeling the interaction relation between the user and the article by taking the user characteristic information and the article characteristic information as input and combining a deep neural network B; and outputting a predicted value of the interaction behavior between the user and the article according to the model obtained by modeling to provide data support for information recommendation. The invention provides a recommendation method based on joint neural network collaborative filtering, which adopts a joint neural network to tightly combine a depth feature extraction process and a depth interactive behavior modeling process, and can quickly and accurately obtain a predicted value.
Description
Technical field
The invention belongs to information resources to recommend field, be specifically related to a kind of based on united NNs collaborative filtering
Information recommendation method and its system.
Background technique
In face of complicated and a large amount of information resources, recommender system can effectively help user to obtain the letter that they want
Breath.Collaborative filtering is one of the recommended method being widely used at present.Traditional collaborative filtering is all based on what matrix decomposition obtained,
Such as present factor model LFM (latent factor model, hidden semantic model), user and article are expressed as potential
Vector, then with the correlativity between an interior product representation user and article, but the mode based on this inner product is all
A kind of linear correlativity.
The method of deep learning shows good effect in recommender system, overcomes many conventional methods before and exists
The problem of, such as the modeling etc. to complicated user-article relationship modeling, to user's dynamic hobby.But at present
The method of the deep learning of most of recommender system is all the excavation carried out to an auxiliary information, such as: text information, phonotape and videotape
Information etc..The characteristic information of article is enriched by these.But for user-article interbehavior, most side
The method that method still uses traditional matrix decomposition.
RBM ((Restricted Boltzmann Machine, limited Boltzmann machine) is first with neural network
Method carrys out the interbehavior between analog subscriber-article, and effect will be good than traditional method, but it only has two layers of network
The method for being not construed as a deep learning.CDAE (Collaborative Denoising Auto-Encoders, collaboration
Noise reduction autocoder) it is also a method neural network based, but it is mainly used for the prediction scored user.NCF
(Neural Collaborative filter, neural collaborative filtering) uses a deep neural network between user-object
Interactive information modeled, but not excavated to the characteristic information of user and article, the method for CDAE and NCF
The display feedback information of user is not all used.DMF's (Deep Matrix Factorization, matrix of depths are decomposed)
Method uses neural network and models to user-article marking, and user characteristics and article characteristics effectively mention
It takes, but about user-article interbehavior, still using the linear method as LFM.
Summary of the invention
The technical problem to be solved in the present invention is to provide a kind of information recommendation sides based on united NNs collaborative filtering
Method and its system solve the technical problems existing in the prior art.
The contents of the present invention include:
Propose a kind of information recommendation method based on united NNs collaborative filtering, comprising the following steps:
User's characteristic information and article characteristics are extracted in conjunction with deep-neural-network A based on user-article scoring matrix
Information；
Using user's characteristic information and article characteristics information as input, in conjunction with deep neural network B, to user-article it
Between interactive relation modeled；
The predicted value of interbehavior for information recommendation between the model output user obtained according to modeling and article to provide
Data are supported.
Preferably, deep-neural-network A includes two parallel network N etuserAnd Netitem, the marking of user and article
Information indicates are as follows: v respectively as the input of the two networksu=< yu1..., yuN>, vi=< y1i..., yMi>, wherein
Preferably, using multi-layer perception (MLP) by NetuserThe DUAL PROBLEMS OF VECTOR MAPPING of this higher-dimension is obtained to the vector space of low-dimensional:
…
WhereinThe weight vectors in xth layer perceptron, bias vector and activation primitive are respectively indicated, X is indicated
The network number of plies, z possessed by deep-neural-network AuIt is a depth representing to user characteristics；Z can similarly be obtainediIt is right for one
The depth representing of article characteristics.
Preferably, using ReLU as activation primitive.
Preferably, the interactive relation between user-article is modeled by the way of linear and nonlinear combination.
Preferably, the interactive relation between user and article indicates are as follows:
Preferably, using multi-layer perception (MLP) to auiIt is handled to obtain:
…
WhereinRespectively indicate the weight vectors in y layers of perceptron, bias vector and activation primitive, Y table
Show the network number of plies possessed by DF network, using ReLU as activation primitive.
Preferably, the predicted value to interbehavior between user and article is exported by sigmoid function:
Wherein by export-restriction between (0,1), vector h is used to control z sigmoid functionuiDifferent dimensions position in vector
The weight set, this parameter can be obtained by training.
Preferably, it is also necessary to predicted value accuracy be promoted by loss function, loss function is assembling loss letter
Number:
α is the weight coefficient for controlling two kinds of losses,Max(Ru) indicate all scorings that user u is provided
Middle highest must score, and Ω (θ) indicates regularization term,WithRespectively indicate the marking of the prediction to positive sample and negative sample, NS
Indicate the sampling set of negative sample.
The above method is relied on, the invention also provides a kind of information recommendation systems based on united NNs collaborative filtering
System including memory, processor and stores the computer program that can be run on a memory and on a processor, and processor is held
The step of realizing any of the above-described method when row computer program.
The beneficial effect comprise that
1, the present invention proposes a recommended method based on united NNs collaborative filtering, uses an association nerve
Depth characteristic extraction process and depth interbehavior modeling process are combined closely, can fast and accurately be predicted by network
Value.
2, the present invention proposes a kind of new loss function, it is contemplated that display and implicit feedback, single-point loss and peering loss,
To having significant raising in the promotion of predicted value accuracy.
Detailed description of the invention
Attached drawing 1 is a kind of information recommendation method flow chart based on united NNs collaborative filtering of the invention；
Attached drawing 2 is the structural schematic diagram of J-NCF model of the invention；
Attached drawing 3 is that the different data of the preferred embodiment of the present invention concentrates interactive relation schematic diagram；
Attached drawing 4 is the different model performance comparison schematic diagrams of the preferred embodiment of the present invention；
Attached drawing 5 is J-NCF in the preferred embodiment of the present invention to the sensibility schematic diagram of the sparse degree of different data.
Specific embodiment
Embodiment 1:
Traditional recommended method for top n recommend task, most of recommender systems be based on collaborative filtering, wherein recommend according to
Rely in behavior over (grading) but regardless of domain knowledge how.The task that top-N recommends is the history according to user to article
Marking, recommends the list of an article to user, and the article for allowing user to like as far as possible comes the front end of list.We will use
Family is expressed as U={ user1..., userM, article is expressed as I={ item1..., itemN, R ∈ iM·NIt is user-article
Scoring matrix, marking are a kind of display feedback informations, it is demonstrated by user for the preference of article, and those are not given a mark
Item, then it is assumed that be that user does not know or do not like these articles, therefore, it is considered that this is implicit feedback information, both are anti-
Feedforward information is all important information in recommendation
There are two kinds of main CF (collaborative filtering, Collaborative Filtering) methods for traditional recommendation: being based on
Neighborhood is based on potential factor.Neighborhood method is similar to the project model of user individual, this with the present invention is based on user items
The method of model is different.Therefore, the present invention is absorbed in potential factor method.Most of researchs to potential factor modeling are to be based on
Factorization is carried out to user items rating matrix, this matrix is referred to as SVD (Singular Value
Decomposition, singular value decomposition).User items rating matrix is decomposed into the product of two lower grade matrixes by SVD,
One includes " user's factor ", another includes " project factor ".Then, for internal product and deviation, user is to project
Preference can be generated.Model of the another kind based on SVD is SVD++, and the model of SVD++ includes explicit and implicit feedback, and
Show the improved performance on many MF (Matrix Factorization, matrix decomposition) model.This exists with the present invention
It is combined in J-NCF (recommended method based on united NNs collaborative filtering) explicit consistent with the motivation of implicit feedback.However,
It is highly difficult for carrying out sparse evaluation matrix using traditional MF method.Many traditional recommender systems by linear kernel and user and
The inner product of project vector is applied to model user items interaction.Linear function is possibly can not accurate description user (project)
With the feature of user items interaction, the non-linear performance for improving recommender system has potential advantage.
The recommender system that recommended method based on deep learning is based on DL (Deep learning, deep learning) can divide
For two classes, i.e., single neural network model and depth Integrated Models, it is to rely solely on depth learning technology that this, which depends on them,
Or conventional recommendation model is combined with deep learning.
For the first kind, RBM is the neural recommender system of early stage.It simulates list data, example using two layers of non-directed graph
Clearly scoring such as user to film.Result's management system is directed to top n suggestion, loss function for grading prediction
Only consider the grading observed.The training that RBM is included in negative sampling is technically challenging, this is for N first
It is recommended that being required.Score in predicting is carried out using Auto-Encoder and only considers the grading observed in loss function, this is not
It can guarantee the good behaviour recommended preceding N.In order to avoid autocoder learns identity recognition function and fails to be summarised as seeing
The data loseed, DAE (Denoising Autoencoder denoises autocoder) have been applied to from the defeated of wilful damage
Enter middle study.The most of publications listed so far all lay particular emphasis on specific feedback, therefore use can not be understood from implicit feedback
The preference at family.The DAEs of extension, input are the partially observable implicit feedbacks of user.Be different from the present invention, DAEs and
CDAE uses project model to carry out personalisation process, represents user and uses its scoring item.For depth Integrated Models, CDL
(Collaborative deep learn, cooperate deep learning) is the Bayesian model of a stacking, it is whole by the DAE of stacking
It closes in traditional probability MF.It is different at two aspects with the present invention: (1) it extracts project from content information
Depth characteristic indicates；(2) it still come using linear kernel the relationship between analog subscriber and project user and project to
Amount.Another Integrated Models directly related with the present invention is DMF.It will be used using the depth MF model with neural network
Family and article are mapped to common lower dimensional space.But it follows LFM using internal product to calculate between user and project
Interaction.Different from DMF, the combination of user and item feature vector is used as input to simulate by present invention application multilayer perceptron
User items interaction.It is not only does this facilitate model and has more expressive force than linear product in terms of modeling user items interaction, and
Also contribute to the accuracy of user and item characteristic extraction.
In view of the above-mentioned problems, the invention proposes a kind of information recommendations based on united NNs collaborative filtering
Method, referring to Fig. 1, comprising the following steps:
User's characteristic information and article characteristics are extracted in conjunction with deep-neural-network A based on user-article scoring matrix
Information.
It can be with structuring user's-article scoring matrix according to marking record of each user of system record to article.
Using user's characteristic information and article characteristics information as input, in conjunction with deep neural network B, to user-article it
Between interactive relation modeled.
Deep-neural-network A and deep neural network B is multi-layer perception (MLP) network, because the network can be very good
User and article characteristics are portrayed, and the non-linear relation between user and article is modeled.Deep layer nerve net
Network A includes two parallel network N etuserAnd Netitem, the scoring information of user and article is respectively as the defeated of the two networks
Enter.Output is the character representation of user and article.The input of deep-neural-network B is the combination of user and article characteristics vector,
Output is the predicted value to interbehavior between user and article.
It is input, output article information preference arrangement with user's characteristic information according to the model that modeling obtains.According to modeling
The predicted value of interbehavior is supported between obtained model output user and article with providing data for information recommendation.
Referring to fig. 2, Fig. 2 illustrates the structure of J-NCF model, and J-NCF model includes two main neural networks, DF net
Network is used to excavate the characteristic information of user and article, and DI network is used to model the interaction between user-article.
Input of the output of DF network as DI network.
DF network mainly respectively excavates the characteristic information of user and article with two parallel networks, finally by this
The output of two networks combines, therefore we claim the two networks to be respectively NetuserAnd Netitem.We are by user and article
Scoring information indicates are as follows: v respectively as the input of the two networksu=< yu1..., yuN>, vi=< y1i..., yMi>, wherein
Then with multi-layer perception (MLP) by the vector space of the DUAL PROBLEMS OF VECTOR MAPPING of this higher-dimension to low-dimensional.Because of NetuserWith
NetitemIt is only upper different in input, therefore the present invention is with NetuserFor, NetitemAnd so on.Using Multilayer Perception
Machine is available:
…
WhereinRespectively indicate the weight vectors in xth layer perceptron, bias vector and activation primitive.Here
We use ReLU as activation primitive, because it has better information ability to express and biological similarities, while can also be very
Good solves the problems, such as that gradient disappears.X indicates the network number of plies possessed by DF network.It exports zuIt is one to user characteristics
Depth representing can similarly obtain ziIt is a depth representing to article characteristics.ReLU (line rectification function, Rectified
Linear Unit) it is a kind of activation primitive.
For the modeling of user-article interbehavior, traditional method mostly use by user vector and article vector into
The mode of row dot product measures the relationship between user and article.But this is a kind of linear modeling pattern, due to nonlinear
It is more preferable than linear modeling pattern.The feature vector of user and article is combined by we using following two form:
First way is directly to splice two vectors, this is a kind of nonlinear combination, is for second
Two vectors are carried out corresponding element to be multiplied to obtain new vector, this is a kind of linear combination.We are based on both
Mode proposes the variant of two different J-NCF models.
Generate auiOnly to user-article interactive modeling first step, this be also not enough to user-article interbehavior into
Row is accurately portrayed, therefore we continue to handle it using multi-layer perception (MLP):
…
WhereinRespectively indicate the weight vectors in y layers of perceptron, bias vector and activation primitive.Here
We equally use ReLU as activation primitive.Y indicates the network number of plies possessed by DF network.Finally we pass through one
Sigmoid function exports the predicted value to interbehavior between user and article:
Wherein sigmoid function can be by export-restriction between (0,1), and vector h is used to control zuiDifferent dimensional in vector
The weight of position is spent, this parameter can be obtained by training.
Predicted value also needs to be handled by loss function.The loss function of general training has single-point loss and peering damage
Lose function.Single-point loss function is mainly used for improving the accuracy to the prediction of some marking value, is more suitable for pushing away for score in predicting
It recommends in task, peering loss function focuses on user to difference preference's degree of different articles, is more applicable for the recommendation of top-N
In task.
A loss function is indicated with l (g), indicates regularization term with Ω (θ), for single-point loss function, usually
Calculation it is as follows:
Specifically, difference of two squares loss function is more suitable for display feedback information:
For implicit feedback information, cross entropy loss function is more applicable:
Peering loss function considers the opposite of user's preference and the two articles different for two articles
Sequence, therefore be more applicable for top-N and recommend task.
Recommend task to also proposed a kind of peering loss function TOP1 top-N, calculate as shown below:
WhereinWithRespectively indicate the marking of the prediction to positive sample and negative sample, NSIndicate the sampling set of negative sample.
Be mostly based on the recommended method of deep learning all using single-point loss function, using peering loss function as
Later research work.Whether single-point loss function or peering loss function all have certain Pros and Cons.For
For single-point loss function, it has ignored the correlativity between article scoring, and for peering loss function, he only considers
Correlativity does not account for user to the fancy grade of some special article.Therefore the present invention mutually ties two kinds of loss functions
It closes, obtains:
L=α lpoint-wise+(1-α)lpoint-wise
Wherein α is the weight coefficient for controlling two kinds of losses.
We further implicitly and explicitly feedback information will take into account, and obtain:
WhereinMax(Ru) indicate that highest must score in all scorings for providing of user u.It is different in this way
Score value different influences can be generated to loss function.Loss function proposed in this paper is known as assembling loss by us
Function.
J-NCF model is shown by algorithm below.In the initialization that 1-4 step is parameter, in step 9 to step 10
It is to be extracted to the characteristic information of user and article, 11 and 12 steps are to combine DI neural network forecast user-article interbehavior.
Finally optimize the parameter in network in 13 and 14 step loss functions and backpropagation.
Input:Epoches:training iterations；
R:the oriainal rating matrix；
U:user set；
I:item set；
Output:Weight matrix of Netuser；
Bias of Netuser；
Weight matrix of Netitem；
Bias of Netitem；
Weight matrix of Di network；
Bias matrix of Di network；
The above method is relied on, the invention also provides a kind of information recommendation systems based on united NNs collaborative filtering
System including memory, processor and stores the computer program that can be run on a memory and on a processor, and processor is held
The step of realizing any of the above-described method when row computer program.
J-NCF model proposed in this paper uses a united NNs, and depth characteristic extraction process and depth are handed over
Mutual behavior modeling process is combined closely.Wherein, depth characteristic extraction process is based on user-article scoring matrix, in conjunction with deep layer
Neural network extracts the characteristic information of user and article.Depth interbehavior modeling process is with the above-mentioned user extracted and object
The feature of product, in conjunction with deep neural network, models the non-linear interactive relation between user-article as input.J-
NCF combines the two processes, and the two processes is allowed constantly mutually to be optimized by training, to mention
The effect that height is recommended.
Experimental result shows that method of the invention, compared to existing method, all has been improved on three data sets,
In, with regard to 10 index of HR@, in MovieLens100K, ML1M has been respectively increased 8.24%, 10.81% on AMovies data set,
10.21%, with regard to 10 index of NDCG@, it has been respectively increased 12.42%, 14.24%, 15.06%.Meanwhile experimental result also indicates that J-
NCF model is also got well than existing method effect on sparse data set and some sluggish users.
Embodiment 2:
The present embodiment mainly uses two datasets: (1) MoviesLens: wherein more comprising the website MovieLens
A score data collection.Data set in different times collect by section.MovieLens100K (ML100K) includes 1,682 films
100,000 of 943 users gradings and 3 are more than 1,000,000 comprising 6040 users in 706 films and grade
MovienLens 1M(ML1M).(2) Amazon film (AMovies), wherein including the film 4,607,047 from Amazon
Click score information, it is bigger more sparse than MovieLens data set, and assessment is widely used in recommender system.Specifically
Treated, and data set information is as shown in table 1 below:
Table 1
The present embodiment depicts the user distribution .X in all three data sets as shown in Figure 3 with distinct interaction number
Axis indicates grading quantity, and Y-axis indicates number of users corresponding with grading.The most users that three data are concentrated are only few
Number grading, it is " inactive users " that these, which are considered these users, and the grading of few " any active ues " is higher.In ML100K number
According to concentration, 61.72% user's scoring is less than 100, and 32.66% scoring is between 100 and 300, only 5.6% use
Family is scored above 300.
When on the data set with different characteristic in use, the present embodiment consider model can obtain different scores,
That is number of users and item number.Therefore, the performance for assessment models on the data set with different sparse degree, keeps user
It is identical with the quantity of project.That is, each concentrated for three data, creates the version of three different degree of rarefications
This.For each data set, the subset of a user and project are randomly choosed first from master data set.The data set is with ' -1'
Postfix notation.Same group of user and project are kept, have ' first sparse version of the data set of -2' suffix be by from the
Random erasure entry creates in the user items matrix of one data set.Second of data set with ' -3' suffix is sparse
Version is created similarly by entry is removed at random from the user items matrix of the second data set.Table 2 summarizes all numbers
According to the feature of collection.
Table 2
It can be seen that DMF is more preferable than the performance of other baselines in terms of HR@10 and NDCG@10 from table 3.Therefore, only make
Use DMF as the optimal baseline compared in later experiments.BPR (Bayesian Personalized Ranking, Bayes
Propertyization grading) it clearly illustrates, the improvement in terms of 10 aspect of NDCG@is than Item-pop baseline is higher than changing for 10 aspect of HR@
Kind, this shows to lose in pairs very strong performance in terms of ranking prediction.NCF and DMF model CF model all more traditional than two
Better performance is shown, this shows that DL technology is improving the practicability for recommending aspect of performance.
Table 3
Baseline is compared with J-NCF model.NCF and DMF HR@10 and NDCG@10 aspect with J-NCF model
Compared to being declined.The united NNs structure of this show to combine closely depth characteristic study and deep layer interactive modeling facilitates
It improves and recommends performance.For J-NCF model, the selection independently of user and project vector is combined, and the performance that J-NCF is realized is better than
DMF baseline causes the improvement threshold of HR10 in ML100K data set from 5.04% to 8.24%, ML1M data set be 5.62% to
10.81%, AMovies data set are 7.21% to 10.21%.The improvement threshold of NDCG@10 is from ML100K data set
The 10.44% of 6.25% to the 14.24% and AMovies data set of 7.22% to 12.42%, ML1M data set arrives
15.06%.
By comparing J-NCFcAnd J-NCFm, it is seen that J-NCFcReach optimum performance, is respectively increased in terms of HR10
3.05%, 3.51% and 2.81%, and 4.85%, 7.51% and 4.18% has been respectively increased in performance on three data sets
NDCG@10 respectively in J-NCFmOn.This is because complex relationship between user and project can with nonlinear kernel come
Description, rather than linear kernel.
When the size of top n recommendation list is between 1 to 10, as the aggregate performance of HR and NDCG increases, because
It includes User Preferences purpose probability in recommendation list that the value of big N, which is increased,.J-NCFhybridIt consistently realizes pair
The improvement of DMF and two kinds of models have individual loss function between the different positions, this demonstrate that our loss functions
Practicability.Based on ML100K database, J-NCFhybridCompare J-NCF respectively at 10 aspect of HR@pointAnd J-NCFpairIt improves
2.68% and 7.61%；To J-NCFpointAnd J-NCFpairThe improvement of NDCG@10 be 3.99% and 2.36% respectively.Referring to figure
4, compare J-NCFpointAnd J-NCFpair, it has been found that J-NCFpointIt is better than J-NCF in terms of HRpair, and J-NCFpair?
It is shown in terms of NDCG and compares J-NCFpointHigher competitive performance.Peering loss function has very strong performance to ranking prediction.Cause
This combines the point-by-point loss in losses by mixture function and in pairs loss.Obviously, it is based on J-NCFcModel, i.e. J-
NCFpoint, J-NCFpairAnd J-NCFhybridIt shows than DMF better performance, it is reasonable that this, which also demonstrates association nerve structure,
's.Fig. 4 (a)-Fig. 4 (f) be respectively HR@N of ML00K, HR@N of ML1M, HR@N of AMovies, ML00K NDCG@N,
The NDCG@N of NDCG@N of ML1M, AMovies.J-NCFpointJ-NCF model when for using single-point loss function, J-NCFpair
J-NCF model when for using peering loss function, J-NCFhybridJ-NCF model when for using losses by mixture function.
In J-NCFcIn, the present embodiment not only learns user and article by the DF neural network with multiple hidden layers
Feature, but also in DI network analog subscriber item and multilayer perceptron interaction.Therefore, whether DL facilitates our mould
Type is vital.The present embodiment checks respectively for the J-NCF of the various numbers of plies in DF and DI network by testingcPerformance.
As a result it is shown in Table 4.The i in DF-i and DI-i in table 4 respectively indicates J-NCFcDF network and DI network in the number of plies.
Table 4
As shown in table 4, in DF network the number of plies increase to 5, DI network from 1 and increase to 4 from 1, recommend performance to improve, this
Show DL technology to the validity of recommender system.In particular, the deep layer of 2 layers of addition or more appears not to use compared with DMF
Place, J-NCFcMore layers or both is stacked in DI or DF network all realizes further improvement.In addition, recommendatory enhancing
It is more helpful to stack more layers ratio DI network in DF network for energy aspect.For example, ML100K data set is based on, relative to (DF-
2, DI-2), the improvement of (DF-3, DI-2) is respectively 2.82% and 4.31% in terms of HR@10 and NDCG@10, (DF-2, DI-
3) improvement rate of (DF-2, DI-2) is respectively 1.05% and 2.62%.When we stack 4 layers or more in DI network (DI-5)
When, J-NCFcPerformance not will increase.However, more layers is stacked in DF network (DF-5) seems still helpful, every number
J-NCF is all based on according to the optimum that collection generatesc(DF-5, DI-4) configuration.
As shown in table 5, J-NCFcIt is better than optimal baseline model DMF in all activity levels, that is, constitutes most of activities
" inactive " user and relatively small number of " very active " user it is higher to the scoring of all data sets.In addition, J-
NCFcModel consistently achieves optimum performance in terms of HR@10 and NDCG@10.
Table 5
Specifically, the user of J-NCF ratio " very active " shows bigger change to the DMF model of " inactive " user
Into.For example, when being included in has more interactive users, i.e., from 50% to 90%, according to HR10 respectively from 11.08% to
7.85%, it is calculated with the NDCG@10 of ML100K, 9.57% to 7.32% data set.This is because the user of " very active " with
The less project that scores has many interactions, and collaborative filtering lacks the information for being based only upon rating matrix recommended project, this shows can be with
J-NCF is extended, in conjunction with more auxiliary informations, such as content information, to explore the relationship between more accurate project.
In order to study J-NCF to the sensibility of the sparse degree of different data, the present embodiment is investigated in table 3 different sparse
Spend the recommendation performance of data set.Fig. 5 is shown as a result, as shown in figure 5, when being applied to the data set comprising more users and project
When, all models can all obtain better performance.For example, the overall performance of all models is better than it on AMovies data set
His two datasets.Therefore, in order to study the model sensibility in the data set with different sparse degree, it is necessary to by user
Scale identical with data set is maintained at the quantity of project.Fig. 5 (a)-Fig. 5 (f) is respectively HR@10, the ML1M of ML00K
The NDCG@10 of HR@10, the HR@10 of AMovies, the NDCG@10 of ML00K, the NDCG@10 of ML1M, AMovies.
Specifically, for ML100K data set, ML1M data set and AMovies data set, in HR@10 and NDCG@10
Aspect, with different degree of rarefications all Sub Data Sets in, J-NCF model is better than baseline model DMF.In addition, the present embodiment
It was found that best model is J-NCFcHigher improvement is shown on sparse data set.For example, it is based on ML100K data set,
In HR10 and NDCG10 index, in terms of ML100K-1 subset (density=4.413%), J-NCFcImprovement compared to DMF reaches
To 4.91% and 9.12%, and improving in ML100K-3 subset (density=0.630%) HR10 and NDCG10 is respectively 7.77%
With 12.02%.
Above-mentioned J-NCFmIt is that J-NCF uses the multiplication based on element that user and item feature vector group are combined into DI network
Input, there is inside DI linear kernel；J-NCFcBeing J-NCF is combined into DI layers for user and item feature vector group using series connection
Input, this is nonlinear way.
The foregoing is only a preferred embodiment of the present invention, is not intended to restrict the invention, for the skill of this field
For art personnel, the invention may be variously modified and varied.All within the spirits and principles of the present invention, made any to repair
Change, equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.
Claims (10)
1. a kind of information recommendation method based on united NNs collaborative filtering, characterized in that the following steps are included:
User's characteristic information and article characteristics information are extracted in conjunction with deep-neural-network A based on user-article scoring matrix；
Using the user's characteristic information and the article characteristics information as input, in conjunction with deep neural network B, to user-object
Interactive relation between product is modeled；
The predicted value of interbehavior is between the model output user obtained according to modeling and article to provide data for information recommendation
It supports.
2. a kind of information recommendation method based on united NNs collaborative filtering as described in claim 1, characterized in that deep
Layer neural network A includes two parallel network N etuserAnd Netitem, the scoring information of user and article is respectively as the two
The input of network indicates are as follows: vu=< yu1..., yuN>, vi=< y1i..., yMi>, wherein
3. a kind of information recommendation method based on united NNs collaborative filtering as claimed in claim 2, characterized in that adopt
With multi-layer perception (MLP) by NetuserThe DUAL PROBLEMS OF VECTOR MAPPING of this higher-dimension is obtained to the vector space of low-dimensional:
WhereinThe weight vectors in xth layer perceptron, bias vector and activation primitive are respectively indicated, X indicates deep layer
The network number of plies, z possessed by neural network AuIt is a depth representing to user characteristics；Z can similarly be obtainediFor one to article
The depth representing of feature.
4. a kind of information recommendation method based on united NNs collaborative filtering as claimed in claim 3, characterized in that adopt
Use ReLU as activation primitive.
5. a kind of information recommendation method based on united NNs collaborative filtering as described in claim 1, characterized in that adopt
The interactive relation between user-article is modeled with the mode that linear and nonlinear combines.
6. a kind of information recommendation method based on united NNs collaborative filtering as claimed in claim 5, characterized in that use
Interactive relation between family and article indicates are as follows:
7. a kind of information recommendation method based on united NNs collaborative filtering as claimed in claim 6, characterized in that adopt
With multi-layer perception (MLP) to auiIt is handled to obtain:
WhereinThe weight vectors in y layers of perceptron, bias vector and activation primitive are respectively indicated, Y indicates DF net
The network number of plies possessed by network, using ReLU as activation primitive.
8. a kind of information recommendation method based on united NNs collaborative filtering as claimed in claim 7, characterized in that logical
Sigmoid function is crossed to export the predicted value to interbehavior between user and article:
Wherein by export-restriction between (0,1), vector h is used to control z sigmoid functionuiDifferent dimensions position in vector
Weight, this parameter can be obtained by training, and σ indicates sigmoid function, i.e.,
9. a kind of information recommendation method based on united NNs collaborative filtering as described in claim 8 is any, feature
It is, it is also necessary to predicted value accuracy be promoted by loss function, the loss function is assembling loss function:
α is the weight coefficient for controlling two kinds of losses,Max(Ru) indicate in all scorings for providing of user u most
High to score, Ω (θ) indicates regularization term,WithRespectively indicate the marking of the prediction to positive sample and negative sample, NSIt indicates
The sampling set of negative sample,The set of all users is respectively indicated, all article set given a mark of some user u.
10. a kind of information recommendation system based on united NNs collaborative filtering, characterized in that including memory, processor
And store the computer program that can be run on a memory and on a processor, which is characterized in that the processor executes institute
The step of any the method for the claims 1 to 9 is realized when stating computer program.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910484886.4A CN110188283B (en) | 2019-06-05 | 2019-06-05 | Information recommendation method and system based on joint neural network collaborative filtering |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910484886.4A CN110188283B (en) | 2019-06-05 | 2019-06-05 | Information recommendation method and system based on joint neural network collaborative filtering |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110188283A true CN110188283A (en) | 2019-08-30 |
| CN110188283B CN110188283B (en) | 2021-11-23 |
Family
ID=67720399
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910484886.4A Active CN110188283B (en) | 2019-06-05 | 2019-06-05 | Information recommendation method and system based on joint neural network collaborative filtering |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110188283B (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110990624A (en) * | 2019-12-13 | 2020-04-10 | 上海喜马拉雅科技有限公司 | Video recommendation method, device, equipment and storage medium |
| CN110991601A (en) * | 2019-11-29 | 2020-04-10 | 中山大学 | Neural network recommendation method based on multi-user behaviors |
| CN111459781A (en) * | 2020-04-01 | 2020-07-28 | 厦门美图之家科技有限公司 | Behavior relation determination method and device, computer equipment and readable storage medium |
| CN111553743A (en) * | 2020-05-08 | 2020-08-18 | 深圳前海微众银行股份有限公司 | Federal product recommendation method, device, equipment and computer storage medium |
| CN112328908A (en) * | 2020-11-11 | 2021-02-05 | 北京工业大学 | Personalized recommendation method based on collaborative filtering |
| CN112395514A (en) * | 2020-12-08 | 2021-02-23 | 杭州电子科技大学 | Article collaborative filtering recommendation method based on memory network |
| CN112529414A (en) * | 2020-12-11 | 2021-03-19 | 西安电子科技大学 | Article scoring method based on multitask neural collaborative filtering network |
| CN112765474A (en) * | 2021-01-28 | 2021-05-07 | 武汉大学 | Deep collaborative filtering-based recommendation method and system |
| CN112818256A (en) * | 2021-02-05 | 2021-05-18 | 武汉大学 | Recommendation method based on neural collaborative filtering |
| CN112905895A (en) * | 2021-03-29 | 2021-06-04 | 平安国际智慧城市科技股份有限公司 | Similar item recommendation method, device, equipment and medium |
| CN113158024A (en) * | 2021-02-26 | 2021-07-23 | 中国科学技术大学 | Causal reasoning method for correcting popularity deviation of recommendation system |
| WO2021184640A1 (en) * | 2020-03-16 | 2021-09-23 | 平安国际智慧城市科技股份有限公司 | Sparse matrix-based product pushing method and apparatus, computer device, and medium |
| CN113486257A (en) * | 2021-07-01 | 2021-10-08 | 湖北工业大学 | Convergence matrix decomposition-based coordinated filtering convolutional neural network recommendation system and method |
| CN113590964A (en) * | 2021-08-04 | 2021-11-02 | 燕山大学 | Deep neural network Top-N recommendation algorithm based on heterogeneous modeling |
| WO2022028140A1 (en) * | 2020-08-04 | 2022-02-10 | 深圳前海微众银行股份有限公司 | Item recommendation method and device, and computer readable storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108090093A (en) * | 2016-11-22 | 2018-05-29 | 华为技术有限公司 | The method and apparatus for generating recommendation results |
| WO2018148493A1 (en) * | 2017-02-09 | 2018-08-16 | Painted Dog, Inc. | Methods and apparatus for detecting, filtering, and identifying objects in streaming video |
| CN108809948A (en) * | 2018-05-21 | 2018-11-13 | 中国科学院信息工程研究所 | A kind of abnormal network connecting detection method based on deep learning |
| CN108874914A (en) * | 2018-05-29 | 2018-11-23 | 吉林大学 | A kind of information recommendation method based on the long-pending and neural collaborative filtering of picture scroll |
| US20190102678A1 (en) * | 2017-09-29 | 2019-04-04 | Samsung Electronics Co., Ltd. | Neural network recogntion and training method and apparatus |
| CN109740743A (en) * | 2019-03-21 | 2019-05-10 | 中国人民解放军国防科技大学 | Hierarchical neural network query recommendation method and device |
| CN109785062A (en) * | 2019-01-10 | 2019-05-21 | 电子科技大学 | A kind of hybrid neural networks recommender system based on collaborative filtering model |
-
2019
- 2019-06-05 CN CN201910484886.4A patent/CN110188283B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108090093A (en) * | 2016-11-22 | 2018-05-29 | 华为技术有限公司 | The method and apparatus for generating recommendation results |
| WO2018148493A1 (en) * | 2017-02-09 | 2018-08-16 | Painted Dog, Inc. | Methods and apparatus for detecting, filtering, and identifying objects in streaming video |
| US20190102678A1 (en) * | 2017-09-29 | 2019-04-04 | Samsung Electronics Co., Ltd. | Neural network recogntion and training method and apparatus |
| CN108809948A (en) * | 2018-05-21 | 2018-11-13 | 中国科学院信息工程研究所 | A kind of abnormal network connecting detection method based on deep learning |
| CN108874914A (en) * | 2018-05-29 | 2018-11-23 | 吉林大学 | A kind of information recommendation method based on the long-pending and neural collaborative filtering of picture scroll |
| CN109785062A (en) * | 2019-01-10 | 2019-05-21 | 电子科技大学 | A kind of hybrid neural networks recommender system based on collaborative filtering model |
| CN109740743A (en) * | 2019-03-21 | 2019-05-10 | 中国人民解放军国防科技大学 | Hierarchical neural network query recommendation method and device |
Non-Patent Citations (1)
| Title |
|---|
| ANU TANEJA ET AL: "Modeling user preferences using neural networks and tensor factorization", 《INTERNATIONAL JOURNAL OF INFORMATION MANAGEMENT》 * |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110991601A (en) * | 2019-11-29 | 2020-04-10 | 中山大学 | Neural network recommendation method based on multi-user behaviors |
| CN110991601B (en) * | 2019-11-29 | 2023-04-21 | 中山大学 | Neural network recommendation method based on multi-user behavior |
| CN110990624B (en) * | 2019-12-13 | 2024-03-01 | 上海喜马拉雅科技有限公司 | Video recommendation method, device, equipment and storage medium |
| CN110990624A (en) * | 2019-12-13 | 2020-04-10 | 上海喜马拉雅科技有限公司 | Video recommendation method, device, equipment and storage medium |
| WO2021184640A1 (en) * | 2020-03-16 | 2021-09-23 | 平安国际智慧城市科技股份有限公司 | Sparse matrix-based product pushing method and apparatus, computer device, and medium |
| CN111459781A (en) * | 2020-04-01 | 2020-07-28 | 厦门美图之家科技有限公司 | Behavior relation determination method and device, computer equipment and readable storage medium |
| CN111459781B (en) * | 2020-04-01 | 2022-08-26 | 厦门美图之家科技有限公司 | Behavior relation determination method and device, computer equipment and readable storage medium |
| CN111553743A (en) * | 2020-05-08 | 2020-08-18 | 深圳前海微众银行股份有限公司 | Federal product recommendation method, device, equipment and computer storage medium |
| WO2022028140A1 (en) * | 2020-08-04 | 2022-02-10 | 深圳前海微众银行股份有限公司 | Item recommendation method and device, and computer readable storage medium |
| CN112328908A (en) * | 2020-11-11 | 2021-02-05 | 北京工业大学 | Personalized recommendation method based on collaborative filtering |
| CN112395514A (en) * | 2020-12-08 | 2021-02-23 | 杭州电子科技大学 | Article collaborative filtering recommendation method based on memory network |
| CN112529414B (en) * | 2020-12-11 | 2023-08-01 | 西安电子科技大学 | Article scoring method based on multi-task neural collaborative filtering network |
| CN112529414A (en) * | 2020-12-11 | 2021-03-19 | 西安电子科技大学 | Article scoring method based on multitask neural collaborative filtering network |
| CN112765474A (en) * | 2021-01-28 | 2021-05-07 | 武汉大学 | Deep collaborative filtering-based recommendation method and system |
| CN112818256A (en) * | 2021-02-05 | 2021-05-18 | 武汉大学 | Recommendation method based on neural collaborative filtering |
| CN112818256B (en) * | 2021-02-05 | 2022-06-03 | 武汉大学 | Recommendation method based on neural collaborative filtering |
| CN113158024A (en) * | 2021-02-26 | 2021-07-23 | 中国科学技术大学 | Causal reasoning method for correcting popularity deviation of recommendation system |
| CN113158024B (en) * | 2021-02-26 | 2022-07-15 | 中国科学技术大学 | Causal reasoning method for correcting popularity deviation of recommendation system |
| CN112905895B (en) * | 2021-03-29 | 2022-08-26 | 平安国际智慧城市科技股份有限公司 | Similar item recommendation method, device, equipment and medium |
| CN112905895A (en) * | 2021-03-29 | 2021-06-04 | 平安国际智慧城市科技股份有限公司 | Similar item recommendation method, device, equipment and medium |
| CN113486257B (en) * | 2021-07-01 | 2023-07-11 | 湖北工业大学 | Coordinated filtering convolutional neural network recommendation system and method based on countermeasure matrix decomposition |
| CN113486257A (en) * | 2021-07-01 | 2021-10-08 | 湖北工业大学 | Convergence matrix decomposition-based coordinated filtering convolutional neural network recommendation system and method |
| CN113590964A (en) * | 2021-08-04 | 2021-11-02 | 燕山大学 | Deep neural network Top-N recommendation algorithm based on heterogeneous modeling |
| CN113590964B (en) * | 2021-08-04 | 2023-05-23 | 燕山大学 | Deep neural network Top-N recommendation method based on heterogeneous modeling |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110188283B (en) | 2021-11-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110188283A (en) | Information recommendation method and system based on joint neural network collaborative filtering | |
| Khan et al. | CNN with depthwise separable convolutions and combined kernels for rating prediction | |
| CN109213863B (en) | Learning style-based adaptive recommendation method and system | |
| Zuo et al. | Tag-aware recommender systems based on deep neural networks | |
| CN111259133B (en) | Personalized recommendation method integrating multiple information | |
| CN109785062B (en) | Hybrid neural network recommendation system based on collaborative filtering model | |
| CN103324657B (en) | Extended using the multi-dimensional query of semantic and service condition statistics | |
| CN108959603A (en) | Personalized recommendation system and method based on deep neural network | |
| Rad et al. | Clustering and ranking university majors using data mining and AHP algorithms: A case study in Iran | |
| Salehi et al. | Hybrid recommendation approach for learning material based on sequential pattern of the accessed material and the learner’s preference tree | |
| CN109145112A (en) | A kind of comment on commodity classification method based on global information attention mechanism | |
| Sharma et al. | A document-level sentiment analysis approach using artificial neural network and sentiment lexicons | |
| CN107577682A (en) | Users' Interests Mining and user based on social picture recommend method and system | |
| CN107526754A (en) | A kind of user's portrait platform method for building up based on big data | |
| Luaces et al. | A factorization approach to evaluate open-response assignments in MOOCs using preference learning on peer assessments | |
| CN111259263A (en) | Article recommendation method and device, computer equipment and storage medium | |
| CN109871485A (en) | A kind of personalized recommendation method and device | |
| CN112016002A (en) | Mixed recommendation method integrating comment text level attention and time factors | |
| Anand et al. | Adaptive user similarity measures for recommender systems: a genetic programming approach | |
| Zhong et al. | Design of a personalized recommendation system for learning resources based on collaborative filtering | |
| Zhang et al. | Micro-blog topic recommendation based on knowledge flow and user selection | |
| Sharma et al. | Comparative Analysis of Data Storage Solutions for Responsive Big Data Applications | |
| CN112328832B (en) | Movie recommendation method integrating labels and knowledge graph | |
| CN110795640B (en) | Self-adaptive group recommendation method for compensating group member difference | |
| Santhosh et al. | Machine learning based ideal job role fit and career recommendation system |
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 |