CN109785062A - A kind of hybrid neural networks recommender system based on collaborative filtering model - Google Patents
A kind of hybrid neural networks recommender system based on collaborative filtering model Download PDFInfo
- Publication number
- CN109785062A CN109785062A CN201910023725.5A CN201910023725A CN109785062A CN 109785062 A CN109785062 A CN 109785062A CN 201910023725 A CN201910023725 A CN 201910023725A CN 109785062 A CN109785062 A CN 109785062A
- Authority
- CN
- China
- Prior art keywords
- user
- commodity
- layer
- score
- module
- 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
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The hybrid neural networks recommender system based on collaborative filtering model that the invention discloses a kind of, including data preprocessing module, composite nerve collaborative filtering module, score in predicting module, recommending module and database, wherein, data preprocessing module obtains initial data from database, and initial data is converted to the form of matrix;Composite nerve collaborative filtering module obtains the probability value to score user by way of to multi-model simultaneously training;Score in predicting module exports final score in predicting value using composite nerve collaborative filtering module and descriptor matrix decomposition and combination, finally stores score in predicting value as score in predicting result in the database;The score in predicting that recommending module is calculated using score in predicting module by the sequence to evaluation metrics to user as a result, recommend.The present invention utilizes different structure hybrid neural networks, has used more neural net layers, forecasting accuracy is obviously improved and the information obtained is also more diversified.
Description
Technical field
It is that one kind is passed through based on hybrid neural networks the invention belongs to the recommender system research field in machine learning field
The recommender system that collaborative filtering excavates user interest preference to be recommended.
Background technique
With the fast development of internet with it is universal, people be faced with dazzling online content (such as film, books and
Music), cause user to search out oneself desired information needs and takes considerable time, that is, the information overload phenomenon mentioned.In order to
This phenomenon is solved, recommender system is come into being.Currently, recommender system be widely used in as Taobao, Netflix,
The domestic and international electric business platform such as Amazon and YouTube.It is reported that 80% user watches film using the recommended engine of Netflix,
The homepage recommendation of YouTube is more than 60% video click.Recommender system is while more easily services client, also substantially
Improve to degree the profit of company.
One of most important challenge is how to improve the accuracy of algorithm in recommender system.Solve the key of this problem
It is to need to build an accurate model to indicate the relationship between user and commodity.Such as matrix decomposition (matrix
Factorization, MF) it is each in latent space respectively to user and commodity Joint Mapping into a latent space
Vector represents a user or commodity, predicts a certain user to the emerging of a certain commodity by latent factor (the hidden factor) inner product operation
Interesting degree.Inner product operation method is actually a linear model, and linear method carries out the relationship between user and commodity
Modeling is incomplete, because there is also non-linear relations.Neural collaborative filtering (Neural is proposed as a result,
Collaborative Filtering, NCF), it is learned using multilayer perceptron (Multi-Layer Perceptron, MLP)
Commonly use the non-linear relation before family and commodity.
NCF uses the multilayer perceptron of tower, wherein the number of nodes of each nervous layer is twice of next nervous layer, when
When the depth of neural network increases, algorithm exponentially increases, this number of plies for having resulted in neural network cannot be too big.And node
The identical multilayer perceptron of number can solve this problem substantially, wherein the number of plies of each neural network has identical node.
But for this structure, it is possible that degenerate problem, the i.e. accuracy of algorithm will not be with the depth of neural network
Increase and improve, if this allows for a neural network number of plies deepens that the accuracy of algorithm may be jeopardized.
Summary of the invention
Goal of the invention of the invention is: in view of the above problems, providing a kind of by global neural block and part mind
The recommender system of composite nerve collaborative filtering through block composition, to promote recommendation performance.
For achieving the above object, the present invention uses composite nerve Collaborative Filtering Recommendation System, and system includes following several
A module:
Data preprocessing module: user is extracted from system database to the score information of commodity, and is filtered out to commodity
Scoring number be less than the user of preset threshold (such as 20), i.e., user is divided by two classes by a scoring commodity number threshold value: work
Jump user and inactive users, is based only upon any active ues building user-commodity rating matrix table;
And identifier of the user to commodity with the presence or absence of scoring is set, it scores if it exists, then the identifier is arranged 1;It is no
Then it is set as 0;The scoring of aobvious feedback is converted into hidden feedback in the module, if user scores to commodity, phase
The entry answered will be marked as 1, conversely, entry is then marked as 0.
User-commodity rating matrix table is constructed based on identifier of the user to commodity again and is stored in database, it is each
That row represents is the information of a user, i.e. user characteristics vector (vu);Each column represent the information of a commodity, i.e. commodity
Feature vector (oi), numerical value represents corresponding user whether there is or not scorings to the commodity in table.
Composite nerve collaborative filtering module, including embeding layer, global neural blockWith local nerve blockIt constitutes
Hybrid neural networks;
The global neural block uses tower multilayer neural network, and each layer of neural network is one layer of global nerve net
Network layers;The network layer number of nodes of global neural net layer from front to back successively decreases, and the number of nodes of preceding layer overall situation neural net layer is
2 times of later layer overall situation neural net layer;I.e. global neural block is by tower multilayer perceptron to the friendship between user-commodity
Mutually (i.e. scoring yui, wherein subscript u indicates that user, i indicate commodity) and it is modeled;
Meanwhile the neural net layer of a certain number of stackings is inserted into two neighboring global neural net layer, constitute two
Local nerve block between a adjacent global neural net layer;
Composite nerve collaborative filtering module reads user characteristics vector sum product features vector from database, and inputs embedding
Enter layer;
Family feature vector and product features DUAL PROBLEMS OF VECTOR MAPPING are respectively thick based on the hidden vector being arranged thereon by the embeding layer
After close vector, then export to first layer overall situation neural net layer;
By the training to hybrid neural networks, the interlayer weight information of hybrid neural networks is obtained;And it is based on training
Interlayer weight information obtain user characteristics vector sum product features vector currently entered in the output of hybrid neural networks
Information;The hidden vector for initializing embeding layer, is become sparse user characteristics vector sum product features vector by connecting entirely
Dense vector, then it is passed through into global neural block and local nerve block to user-commodity weight wuiIt is configured, passes through later
The last layer of training output composite nerve collaborative filtering module.
Preferably, local nerve block uses depth residual error neural network.The problem of depth residual error network can solve degeneration,
It is wherein potential to be mapped asIt goes to be fitted some stack layers, stack layer can be with approximate representation
Score in predicting module: composite nerve collaborative filtering module and descriptor matrix are decomposed into (Generalized Matrix
Factorization, GMF) it combines, it is connected entirely with their hidden layers of respective the last layer, exports user couple
The prediction of commodity is scoredWherein subscript u indicates that user, i indicate commodity.And the prediction scoring based on outputBuilding is used
Family-commodity score in predicting table is simultaneously deposited into database.Since the data of input are hidden feedbacks, so the prediction score value of output
Then represent the probability whether user has selected the commodity, i.e., 1, which represents user, has selected the commodity, on the contrary then do not select the quotient
Product.What wherein every a line represented in user-commodity score in predicting table is score in predicting information of the user to all commodity, often
One column represent the information of a commodity, and the size of numerical value between zero and one, indicates prediction scoring of the user to commodity in table
Probability represents user closer to 1 and selects the chance of the commodity larger, on the contrary then small.
Recommending module: user-commodity score in predicting table, the commodity bought except user are read from database
Outside, a certain number of commodity (such as 100 commodity) are randomly selected, pass through evaluation sequence index in user's non-purchased goods
Prediction scoring is ranked up according to sequence from high to low, for each user, selected and sorted is most preceding(example
As 10) part commodity as the recommending data table of active user and are recommended;The current recommending data table of user is stored simultaneously
In the database.Wherein every a line in recommending data table indicates that the recommendation information to a user, each column indicate commodity quilt
The case where recommendation.
Goal of the invention of the invention is achieved in that
Present invention hybrid neurosystem filtered recommendation system, by data preprocessing module, composite nerve collaborative filtering mould
Block, score in predicting module, recommending module composition.Data preprocessing module obtains initial data from database, and by original number
According to the form for being converted to matrix;Composite nerve collaborative filtering module obtains by way of to multi-model simultaneously training to user
The probability value of scoring;Score in predicting module is exported finally using composite nerve collaborative filtering module and descriptor matrix decomposition and combination
Score in predicting value finally stores score in predicting value as score in predicting result in the database;Recommending module is pre- using scoring
The score in predicting that survey module is calculated by the sequence to evaluation metrics to user as a result, recommend.
In conclusion by adopting the above-described technical solution, the beneficial effects of the present invention are:
(1) in the case that the present invention is in view of non-linear relation between user and commodity, without additional information (when as scored
Between, customer relationship etc.) so that recommending more personalized.
(2) depth residual error network mind is added in local nerve block to increase the depth of global tower multi-layer perception (MLP) in the present invention
Degree, the complexity for alleviating degenerate problem and algorithm to a certain extent also do not significantly improve.
(3) algorithm that uses of the present invention, which is scored in user in less sparse matrix, can also very good recommendation effect
Fruit, the problem of Deta sparseness can be efficiently solved.
(4) present invention utilizes different structure hybrid neural networks, has used more neural net layers, forecasting accuracy
It is obviously improved and the information obtained is also more diversified.
(5) the recommendations necks such as the method for the present invention is extensive with field, such as can be used for books recommendation, and music is recommended, and film is recommended
Domain.
Detailed description of the invention
Fig. 1 is that the present invention is based on each module collaboration process figures of the recommender system of composite nerve collaborative filtering;
Fig. 2 is architecture diagram of the present invention, the schematic diagram by taking depth residual error neural network as an example;
Fig. 3 is the schematic diagram of the local nerve block of the identical multi-layer perception (MLP) building of number of nodes;
Fig. 4 is the local nerve block schematic diagram of the multi-layer perception (MLP) building of tower;
Fig. 5 is the schematic diagram of the local nerve block of the multi-layer perception (MLP) building of depth residual error network;
Fig. 6 is the schematic diagram of hybrid neural networks and descriptor matrix decomposition and combination.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, below with reference to embodiment and attached drawing, to this hair
It is bright to be described in further detail.
The hybrid neural networks recommender system based on collaborative filtering model that the invention discloses a kind of, including data prediction
Module, composite nerve collaborative filtering module, score in predicting module, recommending module and database, wherein data preprocessing module from
Initial data is obtained in database, and initial data is converted to the form of matrix;Composite nerve collaborative filtering module by pair
The mode of multi-model training simultaneously, obtains the probability value to score user;Score in predicting module utilizes composite nerve collaborative filtering
Module and descriptor matrix decomposition and combination export final score in predicting value, finally deposit score in predicting value as score in predicting result
Storage is in the database;The score in predicting that recommending module is calculated using score in predicting module is as a result, by evaluation metrics
It sorts and recommends to user.
Fig. 1 is that the present invention is based on each module collaboration process figures of composite nerve Collaborative Filtering Recommendation System.The present invention includes 4 altogether
A module, data preprocessing module, composite nerve collaborative filtering module, score in predicting module and recommending module.
It will specifically introduce that the present invention is based on the workflows of composite nerve Collaborative Filtering Recommendation System mainly in combination with attached drawing below
Journey.
1. data preprocessing module: data preprocessing module reads user by row from database and believes the scoring of commodity
Breath, as soon as every reading information, will score and be put into user-corresponding position of commodity rating matrix, after reading information,
1 value of excessive element will be commented to fill again, the element not scored is filled with 0 value, and finally matrix is stored in database.
2. composite nerve collaborative filtering module: composite nerve collaborative filtering module is by global neural block and local nerve block group
At composed net definitions are hybrid neural networks.Global nerve block is the multilayer perceptron (Global by tower
Layer1~Global LayerX) composition, wherein being inserted into the neural net layer of stacking in adjacent global neural net layer
(Local Layer 1~Local Layer Z), i.e. local nerve block.As shown in Fig. 2, wherein bottom input layer (Input
Layer) by user characteristics vector vuWith product features vector oiComposition passes through the embeding layer above input layer later
Sparse vector is become that dense (the cross arrows line in figure indicates mapping, i.e., sparse to dense by (Embedding Layer)
Mapping), and then embeding layer is input to composite nerve collaborative filtering layer (Hybrid Neural CF Layer), final output
Layer (Output Layer) output prediction score value
Again willWith user to the true scoring y of commodityuiBetween carry out minimum training.The main method of this system
Calculation formula is as follows:
Wherein, P ∈ RM×K, Q ∈ RN×KThe hidden factor matrix of user and the hidden factor matrix of commodity are respectively represented, wherein M table
Show that number of users, N indicate that commodity number, K indicate hidden because of subnumber, i.e., hidden vector number, i.e., by carrying out square to user-commodity rating matrix
Battle array decomposes available P and Q,For global neural block,For local nerve block, function f can be indicated are as follows:
Wherein, φoutIt is the mapping function of output layer,Indicate the mapping function of xth overall situation neural net layer,Table
Show a series of mapping functions of xth Local neural network layer, can be indicated with following formula:
Wherein,The mapping function of x-th of z layers of local nerve block is represented, local nerve block is Z layers shared.
The minimum value of objective function, objective function are sought using Adaptive Moment Estimation (Adam) method
It is as follows:
Wherein,Indicate the set for having scoring,Indicate negative example collection, i.e. missing value set, wuiIt is neural network
Super ginseng, indicate user's commodity weight.
Local nerve block (Local Layer 1~Local Layer Z) of the invention can be by node (neuron)
Three kinds of structure compositions of multilayer perceptron or depth residual error neural network of the identical multilayer perceptron of number, tower.
Referring to Fig. 3, constructed in local nerve block with the identical multi-layer perception (MLP) of number of nodes, each nervous layer (Local
Layer 1~Local Layer Z) all with the node of identical quantity, node in upper one layer is carried out with next layer of node
Full connection.Although identical, the adjacent local nerve of the number of nodes of each Local neural network layer in each local nerve block
Block is still in tower: the Local neural network i.e. between xth layer overall situation neural net layer and its later layer overall situation neural net layer
Layer, with x compared with the Local neural network layer between its preceding layer overall situation neural net layer, number of nodes gradually tails off.And this tool
In body embodiment, the node of each layer of the local nerve block after xth layer overall situation neural net layer and x layers of global neural net layer
Number of nodes it is identical
In the local nerve block being made of the identical multilayer perceptron of number of nodes, architecture be may be defined as:
......,
......,
Wherein, pu、qiRespectively indicate the user characteristics vector sum commodity user characteristics vector after embeding layer, i.e., it is dense
The dense commodity user characteristics vector of user characteristics vector sum;Wx、bx、Respectively indicate xth layer overall situation neural network
Weight matrix, bias vector, activation primitive and the output of layer, wherein x=1,2 ..., X;Wx′,z、b′x,z、WithRespectively
The weight matrix of z layers of Local neural network layer, bias vector, activation primitive after indicating xth layer overall situation neural net layer and
It exports, wherein z=1,2 ..., Z;h,aout, b respectively indicate the weight vectors, activation primitive and bias vector of output layer, symbol
()TRepresenting matrix transposition.
Referring to fig. 4, local nerve block of the invention is also possible to the multi-layer perception (MLP) of tower, i.e., for every adjacent global mind
For each local nerve block in network layer, the inside of local nerve block is in tower, i.e., the node of each Local neural network layer
Number gradually successively decreases from front to back;And adjacent local nerve block is still in tower.
In the structure of the multilayer perceptron of tower, there is identical neural network structure: (n-1)th layer of mind with global neural block
Number of nodes through layer is twice of n-th layer, i.e., the number of nodes of the 1st layer of Local neural network layer after global neural net layer is it
Twice of nearest global neural net layer, architecture may be defined as:
......,
......,
Wherein, Wx、bx、And zxRespectively indicate the weight matrix of xth layer overall situation neural net layer, bias vector,
Activation primitive and output, wherein x=1,2 ..., X;W′x,z、b′x,z、WithRespectively indicate xth layer overall situation neural net layer
Weight matrix, bias vector, the activation primitive of z layers of Local neural network layer afterwards export, wherein z=1,2 ..., Z;h,
aout, b respectively indicate the weight vectors, activation primitive and bias vector of output layer.
In addition, local nerve block of the invention can also be made of depth residual error neural network, as shown in Figure 5.Fig. 5 is deep
Degree residual error neural network and global neural block are combined.In depth residual error neural network structure, useIndicate a bottom
Mapping, it can be matched by the layer of several stackings, and x indicates the input to first layer.These intermediate layers stacked are similar to
Another residual error functionWhereinIt is fully-connected network mapping before summing,It is after being input to summation
Fully-connected network mapping.By depth residual error network, the structure can be in the present invention is defined as:
......,
Wherein, Wx、bx、zxWithIt respectively indicates the weight matrix of xth layer overall situation neural net layer, be biased towards
The final output and x-th of local nerve block the inside that amount, activation primitive, output, xth layer global network are combined with local nerve block
The set of all W, functionIndicate fully-connected network mapping before summing, wherein x=1,2 ..., X;Wx′,z、b′x,z、WithWeight matrix, the bias vector, activation of z layers of Local neural network layer after respectively indicating xth layer overall situation neural net layer
Function and output, wherein z=1,2 ..., Z;h,aout, b respectively indicates the weight vectors of output layer, activation primitive and is biased towards
Amount.
3. score in predicting module:
If Fig. 6 illustrates the combination that hybrid neural networks and descriptor matrix decompose, their the last one hidden layer is connected
It picks up and, finally export user to the prediction score value of commodityAnd user-commodity score in predicting matrix is deposited into database
In.Wherein, the MF User Verctor in Fig. 6 indicates user characteristics vector (matrix decomposition), and MLP User Verctor is indicated
User characteristics vector (multilayer perceptron), MF Item Verctor indicate product features vector (matrix decomposition), MLP Item
Verctor indicates product features vector (multilayer perceptron), and GMF Layer indicates descriptor matrix decomposition network layer, Hybrid
NeuMF Layer indicates hybrid matrix neural network.
Composite nerve collaborative filtering in the present invention and descriptor matrix decomposition and combination can be with is defined as:
......,
......,
Wherein,Respectively indicate user characteristics vector vuWith product features vector oiAfter descriptor matrix decomposes
Feature vector,Respectively correspond the dense commodity user characteristics vector of dense user characteristics vector sum.For
Saving space, in local nerve block, the present invention distinguishes multi-layer perception (MLP) and depth residual error nerve net with a matrix W
Network.WithFor, as W=0, which indicates multi-layer perception (MLP), and as W=I, which indicates residual error
Neural network.
4. recommending module: recommending module reads prediction rating matrix from database, for user-commodity projection scoring square
Every a line of battle array after the commodity bought except user, is calculated by prediction scoring using sequence index in remaining commodity
Being ranked up from high to low to commodity out, selects to score highest preceding 10 commodity storages in recommending column, and will recommend column
Table stores in the database.
Embodiment
When this system carries out recommended work, including the following steps:
Firstly, (being stored with user in database to comment to the score information of commodity from user is read by row in system database
Divide table, every a line indicates a user, and each column indicate different commodity, and content is user's score value in table), reading is completed
Afterwards, the user that user is less than 20 to the scoring of commodity is filtered out.This is because such user is inactive, it cannot accurately be it
Recommend.Then scoring identifier is arranged: if user scores to commodity, the grade form of corresponding user's commodity will be marked
It is denoted as 1;Conversely, will then be marked as 0.To convert user-commodity rating matrix table for score information, numerical value is represented in table
Whether corresponding user scores to the commodity, as shown in the table:
Grade form of 1 user of table to commodity
It can be seen that user User1 is 1 to the scoring of commodity Item1 in table 1, indicate that the user comments the commodity
Valence;It is 0 to Item2, indicates that the user does not score to the user.
User-commodity rating matrix table is deposited into database by second step.Wherein user-commodity rating matrix table
In every a line represent be a user information, be mapped as user characteristics vector vu;Each column represent the letter of a commodity
Breath, that is, be mapped as product features vector oi, the two vectors are decomposed respectively as descriptor matrix and the input of hybrid neural networks
Feature vector.
Third step, what the end value and the last one hidden layer of hybrid neural networks for calculating separately descriptor matrix decomposition exported
Value, the final value of the two is connected entirely.Wherein, it is exactly that general matrix disassembling method is promoted that descriptor matrix, which decomposes,
It is calculated using two feature vectors as eigenmatrix.And this system has been divided into two parts in hybrid neural networks, one
It is divided into what global neural block was built by the multi-layer perception (MLP) of tower, joined the neural net layer of tower in global neural block,
Here it is local nerve block, nervous layer can be randomly provided, in this way it can be found that user and some of commodity interactive mode dive
In feature.For example, in the neural block of the overall situation connected entirely at 15 layers, two layers of local nerve layer connected entirely of each layer of addition, and
The network struction of depth residual error is carried out to local nerve block, relatively good in this way solves degenerate problem.This system uses
Adaptive Moment Estimation (Adam) carrys out the parameter in learning model, calculates the adaptive learning of various parameters
Rate, the final predicted value of final output.
Divide finally, last calculated prediction value information is filled into user-commodity Matrix prediction table by this system
Analysis, for the every a line of user-commodity projection rating matrix, after the commodity bought except user, in remaining commodity
By predicted value size, using sorting, index is calculated to be from high to low ranked up commodity, which is between 0 to 1
A range, closer to 1 indicate the user it is higher to the favorable rating of the commodity, sort it is more forward, final choice prediction
It is worth highest preceding 10 commodity storages in the list of recommendation, and in the database by recommendation list storage
The above description is merely a specific embodiment, any feature disclosed in this specification, except non-specifically
Narration, can be replaced by other alternative features that are equivalent or have similar purpose;Disclosed all features or all sides
Method or in the process the step of, other than mutually exclusive feature and/or step, can be combined in any way.
Claims (4)
1. a kind of hybrid neural networks recommender system based on collaborative filtering model, which is characterized in that including data prediction mould
Block, composite nerve collaborative filtering module, score in predicting module, recommending module and database;
Wherein, data preprocessing module for extracting user from database to the score information of commodity, and is filtered out to commodity
Scoring number be less than preset threshold user;
And user is set to the scoring identifier of commodity, and it scores if it exists, then it will scoring identifier setting 1;It is otherwise provided as 0;
User is then based on to construct user-commodity rating matrix to the scoring identifier of commodity and be stored in database, user-quotient
Judge every a line representative of sub-matrix is the information of a user, i.e. user characteristics vector;Each column represent a commodity
Information, i.e. product features vector, numerical value represents scoring identifier in table;
Composite nerve collaborative filtering module, the hybrid neural networks constituted including embeding layer, global neural block and local nerve block;
The global neural block uses tower multilayer neural network, and each layer of neural network is one layer of global neural network
Layer;The network layer number of nodes of global neural net layer from front to back successively decreases, after the number of nodes of preceding layer overall situation neural net layer is
2 times of one layer of global neural net layer;
Meanwhile the neural net layer of a certain number of stackings is inserted into two neighboring global neural net layer, constitute two phases
Local nerve block between adjacent global neural net layer;
Composite nerve collaborative filtering module reads user characteristics vector sum product features vector from database, and inputs insertion
Layer;
The embeding layer based on the hidden vector being arranged thereon, respectively by family feature vector and product features DUAL PROBLEMS OF VECTOR MAPPING be it is dense to
After amount, then export to first layer overall situation neural net layer;
By the training to hybrid neural networks, the interlayer weight information of hybrid neural networks is obtained;And it is based on trained layer
Between weight information obtain user characteristics vector sum product features vector currently entered in the output information of hybrid neural networks;
Score in predicting module, the nerve that hybrid neural networks and realization descriptor matrix to composite nerve collaborative filtering module decompose
The hidden layer of the last layer of network is connected entirely, and output user scores to the prediction of commodityWherein subscript u indicates to use
Family, i indicate commodity;And it is scored based on the predictionBuilding user-commodity score in predicting table is simultaneously deposited into database;
Recommending module: reading user-commodity score in predicting table from database, rejects outside the commodity that user had bought, with
Machine extracts a certain number of commodity, to active user's non-purchased goods based on the pre- assessment in user-commodity score in predicting table
PointDescending sort is carried out, and selected and sorted is most precedingPart commodity as active user recommending data table and pushed away
It recommends;Simultaneously in the database by the current recommending data table storage of user;Every a line in the recommending data table is indicated to one
The recommendation information of position user, each column indicate the recommended situation of commodity.
2. the system as claimed in claim 1, which is characterized in that the local nerve block is the identical Multilayer Perception of number of nodes
Device, the multilayer perceptron of tower or depth residual error neural network structure composition.
3. the system as claimed in claim 1, which is characterized in that data preprocessing module constructs user-commodity rating matrix
Mode are as follows: user is read to the score information of commodity by row from database, as soon as every reading information, will score and be put into use
In family-corresponding position of commodity rating matrix, after reading information, then the element of scoring will be present and filled with 1 value, do not exist
The element of scoring is filled with 0 value.
4. the system as claimed in claim 1, which is characterized in that in data preprocessing module, filter out scoring number and be less than 20
After user, then construct user-commodity rating matrix.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910023725.5A CN109785062B (en) | 2019-01-10 | 2019-01-10 | Hybrid neural network recommendation system based on collaborative filtering model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910023725.5A CN109785062B (en) | 2019-01-10 | 2019-01-10 | Hybrid neural network recommendation system based on collaborative filtering model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109785062A true CN109785062A (en) | 2019-05-21 |
CN109785062B CN109785062B (en) | 2021-09-24 |
Family
ID=66499275
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910023725.5A Active CN109785062B (en) | 2019-01-10 | 2019-01-10 | Hybrid neural network recommendation system based on collaborative filtering model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109785062B (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110188283A (en) * | 2019-06-05 | 2019-08-30 | 中国人民解放军国防科技大学 | Information recommendation method and system based on joint neural network collaborative filtering |
CN110489223A (en) * | 2019-08-26 | 2019-11-22 | 北京邮电大学 | Method for scheduling task, device and electronic equipment in a kind of isomeric group |
CN111127142A (en) * | 2019-12-16 | 2020-05-08 | 东北大学秦皇岛分校 | Article recommendation method based on generalized neural attention |
CN111241408A (en) * | 2020-01-21 | 2020-06-05 | 武汉轻工大学 | Recommendation model construction system and method |
CN111311324A (en) * | 2020-02-18 | 2020-06-19 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
CN111368995A (en) * | 2020-02-14 | 2020-07-03 | 中国科学院深圳先进技术研究院 | General network compression framework and compression method based on sequence recommendation system |
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 |
CN112214687A (en) * | 2020-09-29 | 2021-01-12 | 华南师范大学 | Paper recommendation method, system and medium for temporal perception academic information |
CN112256965A (en) * | 2020-10-23 | 2021-01-22 | 海南大学 | Neural collaborative filtering model recommendation method based on lambdamat |
CN112287222A (en) * | 2020-10-29 | 2021-01-29 | 深圳大学 | Content recommendation method based on heterogeneous feature depth residual error network |
CN112765204A (en) * | 2021-02-04 | 2021-05-07 | 西安电子科技大学 | Method, device and equipment for predicting RLOC (radio link performance) based on neural network and computer storage medium |
CN112883264A (en) * | 2021-02-09 | 2021-06-01 | 联想(北京)有限公司 | Recommendation method and device |
CN113553000A (en) * | 2018-07-18 | 2021-10-26 | 成都忆芯科技有限公司 | Method for reducing power consumption of integrated circuit and control circuit thereof |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090281923A1 (en) * | 2008-05-06 | 2009-11-12 | David Selinger | System and process for improving product recommendations for use in providing personalized advertisements to retail customers |
CN106157156A (en) * | 2016-07-29 | 2016-11-23 | 电子科技大学 | A kind of cooperation recommending system based on communities of users |
CN107038609A (en) * | 2017-04-24 | 2017-08-11 | 广州华企联信息科技有限公司 | A kind of Method of Commodity Recommendation and system based on deep learning |
CN107563841A (en) * | 2017-08-03 | 2018-01-09 | 电子科技大学 | A kind of commending system decomposed that scored based on user |
CN108090229A (en) * | 2018-01-10 | 2018-05-29 | 广东工业大学 | A kind of method and apparatus that rating matrix is determined based on convolutional neural networks |
CN108520450A (en) * | 2018-03-21 | 2018-09-11 | 电子科技大学 | Local low-rank matrix based on implicit feedback information approximately recommends method and system |
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 |
CN108959603A (en) * | 2018-07-13 | 2018-12-07 | 北京印刷学院 | Personalized recommendation system and method based on deep neural network |
US20180374138A1 (en) * | 2017-06-23 | 2018-12-27 | Vufind Inc. | Leveraging delayed and partial reward in deep reinforcement learning artificial intelligence systems to provide purchase recommendations |
-
2019
- 2019-01-10 CN CN201910023725.5A patent/CN109785062B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090281923A1 (en) * | 2008-05-06 | 2009-11-12 | David Selinger | System and process for improving product recommendations for use in providing personalized advertisements to retail customers |
CN106157156A (en) * | 2016-07-29 | 2016-11-23 | 电子科技大学 | A kind of cooperation recommending system based on communities of users |
CN107038609A (en) * | 2017-04-24 | 2017-08-11 | 广州华企联信息科技有限公司 | A kind of Method of Commodity Recommendation and system based on deep learning |
US20180374138A1 (en) * | 2017-06-23 | 2018-12-27 | Vufind Inc. | Leveraging delayed and partial reward in deep reinforcement learning artificial intelligence systems to provide purchase recommendations |
CN107563841A (en) * | 2017-08-03 | 2018-01-09 | 电子科技大学 | A kind of commending system decomposed that scored based on user |
CN108090229A (en) * | 2018-01-10 | 2018-05-29 | 广东工业大学 | A kind of method and apparatus that rating matrix is determined based on convolutional neural networks |
CN108520450A (en) * | 2018-03-21 | 2018-09-11 | 电子科技大学 | Local low-rank matrix based on implicit feedback information approximately recommends method and system |
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 |
CN108959603A (en) * | 2018-07-13 | 2018-12-07 | 北京印刷学院 | Personalized recommendation system and method based on deep neural network |
Non-Patent Citations (4)
Title |
---|
RUIBIN XIONG等: "Deep hybrid collaborative filtering for Web service recommendation", 《EXPERT SYSTEMS WITH APPLICATIONS》 * |
刘晓琳: "基于协同过滤及关联规则的混合推荐算法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
曾伟: "基于复杂网络的个性化推荐算法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
曾伟: "推荐算法与推荐网络研究", 《中国优秀博士学位论文全文数据库信息科技辑》 * |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113553000A (en) * | 2018-07-18 | 2021-10-26 | 成都忆芯科技有限公司 | Method for reducing power consumption of integrated circuit and control circuit thereof |
CN110188283A (en) * | 2019-06-05 | 2019-08-30 | 中国人民解放军国防科技大学 | Information recommendation method and system based on joint neural network collaborative filtering |
CN110188283B (en) * | 2019-06-05 | 2021-11-23 | 中国人民解放军国防科技大学 | Information recommendation method and system based on joint neural network collaborative filtering |
CN110489223A (en) * | 2019-08-26 | 2019-11-22 | 北京邮电大学 | Method for scheduling task, device and electronic equipment in a kind of isomeric group |
CN110489223B (en) * | 2019-08-26 | 2022-03-29 | 北京邮电大学 | Task scheduling method and device in heterogeneous cluster and electronic equipment |
CN111127142A (en) * | 2019-12-16 | 2020-05-08 | 东北大学秦皇岛分校 | Article recommendation method based on generalized neural attention |
CN111127142B (en) * | 2019-12-16 | 2023-09-08 | 东北大学秦皇岛分校 | Article recommendation method based on generalized nerve attention |
CN111241408A (en) * | 2020-01-21 | 2020-06-05 | 武汉轻工大学 | Recommendation model construction system and method |
CN111368995A (en) * | 2020-02-14 | 2020-07-03 | 中国科学院深圳先进技术研究院 | General network compression framework and compression method based on sequence recommendation system |
CN111368995B (en) * | 2020-02-14 | 2023-05-02 | 中国科学院深圳先进技术研究院 | Universal network compression framework and method based on sequence recommendation system |
CN111311324B (en) * | 2020-02-18 | 2022-05-20 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
CN111311324A (en) * | 2020-02-18 | 2020-06-19 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
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 |
CN112214687A (en) * | 2020-09-29 | 2021-01-12 | 华南师范大学 | Paper recommendation method, system and medium for temporal perception academic information |
CN112214687B (en) * | 2020-09-29 | 2023-06-06 | 华南师范大学 | Paper recommendation method, system and medium for tense perception academic information |
CN112256965A (en) * | 2020-10-23 | 2021-01-22 | 海南大学 | Neural collaborative filtering model recommendation method based on lambdamat |
CN112287222A (en) * | 2020-10-29 | 2021-01-29 | 深圳大学 | Content recommendation method based on heterogeneous feature depth residual error network |
CN112287222B (en) * | 2020-10-29 | 2023-12-15 | 深圳大学 | Content recommendation method based on heterogeneous characteristic depth residual error network |
CN112765204A (en) * | 2021-02-04 | 2021-05-07 | 西安电子科技大学 | Method, device and equipment for predicting RLOC (radio link performance) based on neural network and computer storage medium |
CN112765204B (en) * | 2021-02-04 | 2023-04-07 | 西安电子科技大学 | Method, device and equipment for predicting RLOC (radio link performance) based on neural network and computer storage medium |
CN112883264A (en) * | 2021-02-09 | 2021-06-01 | 联想(北京)有限公司 | Recommendation method and device |
Also Published As
Publication number | Publication date |
---|---|
CN109785062B (en) | 2021-09-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109785062A (en) | A kind of hybrid neural networks recommender system based on collaborative filtering model | |
CN109299396B (en) | Convolutional neural network collaborative filtering recommendation method and system fusing attention model | |
CN111310063B (en) | Neural network-based article recommendation method for memory perception gated factorization machine | |
Castro et al. | A consensus‐driven group recommender system | |
CN108959603A (en) | Personalized recommendation system and method based on deep neural network | |
CN108874914A (en) | A kind of information recommendation method based on the long-pending and neural collaborative filtering of picture scroll | |
CN111932336A (en) | Commodity list recommendation method based on long-term and short-term interest preference | |
CN107220365A (en) | Accurate commending system and method based on collaborative filtering and correlation rule parallel processing | |
Chen et al. | Deep transfer tensor decomposition with orthogonal constraint for recommender systems | |
CN113918832B (en) | Graph convolution collaborative filtering recommendation system based on social relationship | |
CN113918833B (en) | Product recommendation method realized through graph convolution collaborative filtering of social network relationship | |
Boratto et al. | ART: group recommendation approaches for automatically detected groups | |
CN113918834B (en) | Graph convolution collaborative filtering recommendation method fusing social relations | |
CN113379494A (en) | Commodity recommendation method and device based on heterogeneous social relationship and electronic equipment | |
CN112699310A (en) | Cold start cross-domain hybrid recommendation method and system based on deep neural network | |
Ahamed et al. | A recommender system based on deep neural network and matrix factorization for collaborative filtering | |
CN115329215A (en) | Recommendation method and system based on self-adaptive dynamic knowledge graph in heterogeneous network | |
Leng et al. | Dynamically aggregating individuals’ social influence and interest evolution for group recommendations | |
Qiao et al. | TAG: Joint triple-hierarchical attention and GCN for review-based social recommender system | |
Fu et al. | A unified framework for multi-domain ctr prediction via large language models | |
Hassan et al. | Evaluating the performance of a neural network-based multi-criteria recommender system | |
Wang et al. | Online course recommendation algorithm based on multilevel fusion of user features and item features | |
CN108681570A (en) | A kind of individualized webpage recommending method based on multi-objective Evolutionary Algorithm | |
Sarker et al. | A hybrid collaborative recommendation system based on matrix factorization and deep neural network | |
CN115525819A (en) | Cross-domain recommendation method for information cocoon room |
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 |