CN106971053A - A kind of recommendation method based on mixing collaborative filtering - Google Patents
A kind of recommendation method based on mixing collaborative filtering Download PDFInfo
- Publication number
- CN106971053A CN106971053A CN201610012329.9A CN201610012329A CN106971053A CN 106971053 A CN106971053 A CN 106971053A CN 201610012329 A CN201610012329 A CN 201610012329A CN 106971053 A CN106971053 A CN 106971053A
- Authority
- CN
- China
- Prior art keywords
- user
- project
- matrix
- collaborative filtering
- scoring
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16Z—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
- G16Z99/00—Subject matter not provided for in other main groups of this subclass
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of recommendation method based on mixing collaborative filtering, comprise the following steps:A user items rating matrix is set up, using the method for optimization singular value decomposition by user items matrix decomposition into two recessive factor matrixs on user and on project;Using gradient descent method for two continuous iterative approach optimal objective parameters of recessive factor matrix;Two recessive factor matrixs are multiplied, the user items matrix of a full rank is obtained;On the user items matrix of full rank, introduce user's effort analysis and project effort analysis, then it is combined using two kinds of collaborative filterings based on KNN, project-based collaborative filtering prediction scoring and the collaborative filtering prediction scoring based on user are obtained respectively, two kinds of collaborative filtering prediction scoring weighted sums are obtained into prediction of the user to project to score, most the multiple projects of prediction scoring highest generate recommendation list to user at last.
Description
Technical field
The present invention relates to recommended technology field, and in particular to a kind of recommendation method based on mixing collaborative filtering.
Background technology
In fast changing Internet era, network technology high speed development, at the same time, information constantly increase in geometric progression
Long, each user obtainable information in internet is more and more, and information content is far beyond the screening scope of manpower.
This means although information content is increased, the efficiency that user obtains available information is reduced.Therefore, how in magnanimity information
For user select its may it is interested or may be useful to user information just into an important research topic.For solution
Certainly this problem, commending system arises at the historic moment.
Commending system is used for user's recommended products, and these products can be most popular products or be based on
The result that the demography of user is calculated, certainly, the most frequently used mode are also based on the historical behavior of user to user
Hobby be predicted.So, the work of commending system can regard the personalized process of an electric business as, because it causes electricity
Business website is adapted to the demand of each user, and specific commodity are provided for it.In the past, commending system is only by part electric business
Website is as novel, minority a application, and nowadays, commending system has had changed into a very important business tool,
And in the general layout of change internet business slowly.
In industrial quarters, collaborative filtering recommending mode the most frequently used at present is namely based on KNN collaborative filtering recommending, so
And, collaborative filtering based on KNN adjusts the decline that the problem of recommending maximum is exactly the openness precision that can cause to predict the outcome of matrix,
And in actual application, the rating matrix of user-project is generally all that than sparse, therefore, many recommendation results are not
It is preferable.
The content of the invention
In view of this, the invention provides a kind of recommendation method based on mixing collaborative filtering, it can be commented in user-project
Matrix is made up in the case that sub-matrix is sparse openness and produce the recommendation results of better quality.
In order to achieve the above object, technical scheme comprises the following steps:
Step 1, a user-project rating matrix is set up, using the method for optimization singular value decomposition by user-project square
Battle array resolves into two recessive factor matrixs on user and on project.
Step 2, using gradient descent method for two recessive continuous iterative approach optimal objective parameters of factor matrix;
Step 3, two recessive factor matrixs are multiplied, obtain user-project matrix of a full rank.
Step 4, on user-project matrix of full rank, introduce user's effort analysis and project effort analysis, then use
Two kinds of collaborative filterings based on KNN are combined, and project-based collaborative filtering prediction scoring are obtained respectively and based on user
Collaborative filtering prediction scoring, the prediction scoring weighted sums of two kinds of collaborative filterings are obtained into prediction of the user to project and scored, most
The multiple projects of prediction scoring highest generate recommendation list to user at last.
Further, user-project rating matrix includes m user and n project in step 1, by the user-project
Rating matrix carries out singular value decomposition, obtains the recessive factor matrix U on the user and recessive factor matrix M on project;
Wherein U is F × m rank matrix, and M is F × n rank matrix, and F is singular value number, is the number of the recessive factor.
Further, optimal objective parameter is:
ui=[ui1, ui2..., uif..., uiF], mj=[mj1, mj2..., mjf..., mjF];rijFor actual scoring;λ is pre-
If regularization constraint coefficient of balance, diFor the user i of setting effort analysis, djFor the project j of setting effort analysis;f∈
[1, F], uifThe value arranged for the i-th row f of matrix U, mifThe value arranged for matrix M jth row f.
Further, prediction of the user to project, which is scored, is
Project j prediction is scored for user i, diFor the user i of setting effort analysis, djFor the project j of setting
Effort analysis;F ∈ [1, F], uifThe value arranged for the i-th row f of matrix U, mjfThe value arranged for matrix M jth row f.
Further, project-based collaborative filtering predicts that scoring is:Pass through u pairs of item similar to destination item i of user
Scorings of the user u to project i is predicted in the actual scoring of mesh set;
Collaborative filtering based on user predicts that scoring is:User u is predicted to destination item i scoring by similar users
Scoring to project i.
Further, the parallelization of KNN algorithms is completed by the way of MapReduce+Hadoop.
Beneficial effect:
The present invention provides a kind of recommendation method SVD&KNN Hybrid Collaborative based on mixing collaborative filtering
Filtering (SKHCF), comprising:Based on optimization singular value decomposition (Singular Value Decomposition, SVD)
Matrix fill-in technology and collaborative filtering (the Hybrid KNN-Based Collaborative based on KNN of mixing
Filtering, H-KNN), user-project matrix is resolved into user and project by the method first by optimization singular value decomposition
Two recessive factor matrixs, using the continuous iterative approach optimal objective parameter of gradient descent method, then by two matrix multiples, are obtained
User-project matrix of one full rank, then uses H-KNN algorithms on non-singular matrix, predicts scoring of the user to project, most
The recommendation list to targeted customer is produced eventually.
One aspect of the present invention, using based on mixing collaborative filtering recommendation method, solve in traditional collaborative filtering due to
User-project rating matrix is relatively sparse and recommendation results that cause are inaccurate, recommend the problem of precision is not high;On the other hand,
A set of practicable system schema is formd, for qualified input, preferably recommendation results relatively can be produced.
Embodiment
With reference to embodiment, the present invention will be described in detail.
Embodiment 1, present embodiments provide it is a kind of based on mixing collaborative filtering recommendation method, comprise the following steps:
Step 1, a user-project rating matrix is set up, using the method for optimization singular value decomposition by user-project square
Battle array resolves into two recessive factor matrixs on user and on project;In the present embodiment, wrapped in user-project rating matrix
M user and n project are included, the user-project rating matrix is subjected to singular value decomposition, the recessive factor on user is obtained
Matrix U and the recessive factor matrix M on project;Wherein U is F × m rank matrix, and M is F × n rank matrix, and F is strange
Different value number, is the number of the recessive factor.
Step 2, using gradient descent method for two recessive continuous iterative approach optimal objective parameters of factor matrix;This reality
Apply in example, optimal objective parameter is:
ui=[ui1, ui2..., uif..., uiF], mj=[mj1, mj2..., mjf..., mjF];riiFor actual scoring;λ is pre-
If regularization constraint coefficient of balance, diFor the user i of setting effort analysis, djFor the project j of setting effort analysis;f∈
[1, F], uifThe value arranged for the i-th row f of matrix U, mjfThe value arranged for matrix M jth row f.
Step 3, two recessive factor matrixs are multiplied, obtain user-project matrix of a full rank;
Step 4, on user-project matrix of full rank, introduce user's effort analysis and project effort analysis, then use
Two kinds of collaborative filterings based on KNN are combined, and project-based collaborative filtering prediction scoring are obtained respectively and based on user
Collaborative filtering prediction scoring, the prediction scoring weighted sums of two kinds of collaborative filterings are obtained into prediction of the user to project and scored, most
The multiple projects of prediction scoring highest generate recommendation list to user at last.
Completed by the way of singular value decomposition after matrix fill-in, next will be produced using H-KNN algorithms and recommend knot
Really.The groundwork of H-KNN algorithms has two, is the similar neighbours of inquiry and the scoring of prediction project respectively.Most important of which
Step is exactly to inquire about similar neighborhood, needs to build the two-dimensional matrix of a user-project before inquiry similar neighborhood.
H-KNN algorithms are substantially the combination of two kinds of collaborative filterings based on KNN, i.e., project-based collaborative filtering
With the collaborative filtering based on user.Wherein, project-based collaborative filtering passes through u pairs of project similar to destination item i of user
Scorings of the user u to project i is predicted in the actual scoring of set;And the collaborative filtering based on user by similar users to target
Project i's scores to predict scorings of the user u to project i.
Scorings of the user u of two kinds of different modes predictions to project i can act as final result and directly show, and be terrible
To more accurately result, weighted factor t is introduced herein, summation is weighted to two kinds of prediction scorings, in all prediction score values
Middle selection Top-N generates recommendation list.Its specific formula is as follows.
rU, i=t × ru+(1-t)×ri
Wherein, t is the control parameter that introduces, and span is [0,1], and value is at intervals of 0.1, the t after experiment in theory
The optimal value in specific set of data can be got, if experiment condition allows, can also suitably reduce value interval, more be managed
The t values thought.ruTo use scoring of the targeted customer that the collaborative filtering based on user is obtained to a certain project, riTo use
Scoring of the targeted customer that project-based collaborative filtering is obtained to a certain project.When t values are 0, only consider based on use
Influence of the collaborative filtering at family to result, and it is then completely opposite when t values are 1.
So far, final predictions of the user u to project i is obtained to score.In the present embodiment, prediction of the user to project is scored
For:
Project j prediction is scored for user i, diFor the user i of setting effort analysis, djFor the project j of setting
Effort analysis;F ∈ [1, F], uifThe value arranged for the i-th row f of matrix U, mjfThe value arranged for matrix M jth row f.
In embodiment 2, above-described embodiment 1, in actual system, the expense of KNN algorithms is all generally very big, is thought
As the once customer volume of millions and million grades of number of songs, the expense for calculating nearest-neighbors is very fearful.Especially originally
UB-CF and IB-CF are combined by the H-KNN algorithms that text is proposed, result are obtained by way of weighted sum, this means that
Need to calculate similarity twice.Therefore, it is necessary to take certain optimization means solve KNN algorithms time space complexity it is high,
The problems such as arithmetic speed is undesirable.
In view of the limitation of one-of-a-kind system performance, realize that parallel optimization is undoubtedly a preferably choosing using distributed system
Select.The parallelization of KNN algorithms is completed by the way of MapReduce+Hadoop herein, implementation is as follows.
1st, user-project rating matrix is inputted.The Map stages receive key-value pair < key, the value >, wherein key of input
It is the line number of data set, and value is then the content of current line, that is, scoring of the user to a song.Next
The cutting to this content is completed, Shuffle processes is completed according to key value, produces the output key-value pair in Reduce stages, at this moment
Key become user id, value becomes song id and rating.Then, output of the Reduce stages the Map stages
Synthesize user-project rating matrix.
2nd, the similitude between project between the step 1 of H-KNN algorithmic procedures, calculating user is completed.This process Map ranks
The input of section is user-project rating matrix, and scoring of each user of Shuffle procedure extractions to project forms key-value pair, key
Be worth be project to (user id (a), user id (b)), value values are scoring to (rating (a), rating (b)), by key assignments
To the input as the Reduce stages, Similarity Measure between Reduce stage finished items, and result is preserved exported.Return to step
Rapid 1, calculate the similitude between user with similar method.
3rd, this step Map stages are inputted as the similitude between user and project, the N number of arest neighbors of Shuffle processes completion
The calculating in residence forms key-value pair, and key values are that user id or song id, value value are nearest N number of neighbours, Reduce stages
3,4 steps of H-KNN algorithms are completed, prediction of the targeted customer to project is calculated and scores, and recommendation list is formed with this.
Embodiment 3, this method complete target component using stochastic gradient descent method and optimized.Stochastic gradient descent method is one
Individual optimization algorithm, also commonly referred to as steepest descent method.This method is using negative gradient direction as the direction of search, under stochastic gradient
Drop method is closer to desired value, and step-length is smaller, advances slower.Use comprising the following steps that for stochastic gradient descent method:
1st, certainty factor number F, punishment parameter λ and learning rate η, initialising subscriber stealth factor matrix U and project it is recessive because
Submatrix M;
2nd, for each user-project scoring to (user u, project m belongs to matrix R):
201st, error score is calculated
202nd, user u and project m recessive factor vector u is updatedifAnd mjf:
uif+=η (eui·mjf-λ·uif)
mjf+=η (eui·uif-λ·mjf)
3rd, target component E is calculated, if E value makes η=0.9 × η, then proceeded to than small, renewal learning rate before
Step 2, until E values are in interval concussion by a small margin or reached default iterations.
To sum up, presently preferred embodiments of the present invention is these are only, is not intended to limit the scope of the present invention.It is all
Within the spirit and principles in the present invention, any modification, equivalent substitution and improvements made etc. should be included in the protection of the present invention
Within the scope of.
Claims (6)
1. a kind of recommendation method based on mixing collaborative filtering, it is characterised in that comprise the following steps:
Step 1, a user-project rating matrix is set up, using the method for optimization singular value decomposition by user-project matrix point
Solution is on user and on two of project recessive factor matrixs;
Step 2, using gradient descent method for two recessive continuous iterative approach optimal objective parameters of factor matrix;
Step 3, two recessive factor matrixs are multiplied, obtain user-project matrix of a full rank;
Step 4, on user-project matrix of full rank, user's effort analysis and project effort analysis are introduced, then using two kinds
Collaborative filtering based on KNN is combined, and project-based collaborative filtering prediction scoring and the association based on user are obtained respectively
With filtering prediction scoring, two kinds of collaborative filtering prediction scoring weighted sums are obtained into prediction of the user to project and scored, most at last
The multiple projects of prediction scoring highest generate recommendation list to user.
2. a kind of recommendation method based on mixing collaborative filtering as claimed in claim 1, it is characterised in that described step 1
Described in user-project rating matrix include m user and n project, the user-project rating matrix is subjected to singular value
Decompose, obtain the recessive factor matrix U on the user and recessive factor matrix M on project;Wherein U is F × m rank square
Battle array, M is F × n rank matrix, and F is singular value number, is the number of the recessive factor.
3. a kind of recommendation method based on mixing collaborative filtering as claimed in claim 2, it is characterised in that
The optimal objective parameter is:
ui=[ui1, ui2..., uif..., uiF], mj=[mj1, mj2..., mjf..., mjF];rijFor actual scoring;λ is default
Regularization constraint coefficient of balance, diFor the user i of setting effort analysis, dfFor the project j of setting effort analysis;F ∈ [1,
F], uifThe value arranged for the i-th row f of matrix U, mjfThe value arranged for matrix M jth row f.
4. a kind of recommendation method based on mixing collaborative filtering as claimed in claim 3, it is characterised in that
Prediction of the user to project, which is scored, is
Project j prediction is scored for user i, diFor the user i of setting effort analysis, dfFor the project j of setting scoring
Deviation;F ∈ [1, F], uifThe value arranged for the i-th row f of matrix U, mjfThe value arranged for matrix M jth row f.
5. a kind of recommendation method based on mixing collaborative filtering as claimed in claim 4, it is characterised in that described to be based on project
Collaborative filtering prediction scoring be:User is predicted by the actual scoring of u pairs of project set similar to destination item i of user
Scorings of the u to project i;
The collaborative filtering based on user predicts that scoring is:User u is predicted to destination item i scoring by similar users
Scoring to project i.
6. a kind of recommendation method based on mixing collaborative filtering as claimed in claim 5, it is characterised in that use
MapReduce+Hadoop mode completes the parallelization of KNN algorithms.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610012329.9A CN106971053A (en) | 2016-01-08 | 2016-01-08 | A kind of recommendation method based on mixing collaborative filtering |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610012329.9A CN106971053A (en) | 2016-01-08 | 2016-01-08 | A kind of recommendation method based on mixing collaborative filtering |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106971053A true CN106971053A (en) | 2017-07-21 |
Family
ID=59334162
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610012329.9A Pending CN106971053A (en) | 2016-01-08 | 2016-01-08 | A kind of recommendation method based on mixing collaborative filtering |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106971053A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108509573A (en) * | 2018-03-27 | 2018-09-07 | 陕西科技大学 | Book recommendation method based on matrix decomposition collaborative filtering and system |
CN108960986A (en) * | 2018-06-26 | 2018-12-07 | 西安交通大学 | A kind of supplier's recommended method based on web crawlers |
CN109597899A (en) * | 2018-09-26 | 2019-04-09 | 中国传媒大学 | The optimization method of media personalized recommendation system |
CN109670087A (en) * | 2018-11-28 | 2019-04-23 | 平安科技(深圳)有限公司 | Course intelligent recommendation method, apparatus, computer equipment and storage medium |
CN111492392B (en) * | 2017-12-22 | 2023-11-17 | 华为技术有限公司 | Client, server and client-server system for updating a client-project matrix |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120030159A1 (en) * | 2010-07-30 | 2012-02-02 | Gravity Research & Development Kft. | Recommender Systems and Methods |
WO2012013996A1 (en) * | 2010-07-30 | 2012-02-02 | Gravity Research & Development Kft. | Recommender systems and methods |
CN104503973A (en) * | 2014-11-14 | 2015-04-08 | 浙江大学软件学院(宁波)管理中心(宁波软件教育中心) | Recommendation method based on singular value decomposition and classifier combination |
CN105025091A (en) * | 2015-06-26 | 2015-11-04 | 南京邮电大学 | Shop recommendation method based on position of mobile user |
-
2016
- 2016-01-08 CN CN201610012329.9A patent/CN106971053A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120030159A1 (en) * | 2010-07-30 | 2012-02-02 | Gravity Research & Development Kft. | Recommender Systems and Methods |
WO2012013996A1 (en) * | 2010-07-30 | 2012-02-02 | Gravity Research & Development Kft. | Recommender systems and methods |
CN104503973A (en) * | 2014-11-14 | 2015-04-08 | 浙江大学软件学院(宁波)管理中心(宁波软件教育中心) | Recommendation method based on singular value decomposition and classifier combination |
CN105025091A (en) * | 2015-06-26 | 2015-11-04 | 南京邮电大学 | Shop recommendation method based on position of mobile user |
Non-Patent Citations (3)
Title |
---|
CHUN-XIA YIN等: "A careful assessment of recommendation algorithms related to dimension reduction teckniques", 《KNOWLEDGE-BASED SYSTEMS》 * |
邓世龙: "基于近邻的协同过滤算法的优化与实现", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
陈彦萍等: "基于用户-项目的混合协同过滤算法", 《计算机技术与发展》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111492392B (en) * | 2017-12-22 | 2023-11-17 | 华为技术有限公司 | Client, server and client-server system for updating a client-project matrix |
CN108509573A (en) * | 2018-03-27 | 2018-09-07 | 陕西科技大学 | Book recommendation method based on matrix decomposition collaborative filtering and system |
CN108509573B (en) * | 2018-03-27 | 2021-04-27 | 宁夏三得教育科技有限公司 | Book recommendation method and system based on matrix decomposition collaborative filtering algorithm |
CN108960986A (en) * | 2018-06-26 | 2018-12-07 | 西安交通大学 | A kind of supplier's recommended method based on web crawlers |
CN109597899A (en) * | 2018-09-26 | 2019-04-09 | 中国传媒大学 | The optimization method of media personalized recommendation system |
CN109670087A (en) * | 2018-11-28 | 2019-04-23 | 平安科技(深圳)有限公司 | Course intelligent recommendation method, apparatus, computer equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Ba et al. | Optimal sliced Latin hypercube designs | |
CN106971053A (en) | A kind of recommendation method based on mixing collaborative filtering | |
Nariman-Zadeh et al. | Inverse modelling of multi-objective thermodynamically optimized turbojet engines using GMDH-type neural networks and evolutionary algorithms | |
CN105893609A (en) | Mobile APP recommendation method based on weighted mixing | |
CN103309972A (en) | Recommend method and system based on link prediction | |
CN112613602A (en) | Recommendation method and system based on knowledge-aware hypergraph neural network | |
CN102799671A (en) | Network individual recommendation method based on PageRank algorithm | |
CN111179031B (en) | Training method, device and system for commodity recommendation model | |
CN108038629A (en) | A kind of optimization method based on collaborative filtering | |
CN106326351A (en) | Recommendation system cold start solving method based on user feedback | |
CN103971161A (en) | Hybrid recommendation method based on Cauchy distribution quantum-behaved particle swarm optimization | |
CN105976070A (en) | Key-element-based matrix decomposition and fine tuning method | |
Zou et al. | Reinforcement learning to diversify top-n recommendation | |
CN112115378A (en) | Recommendation prediction system and recommendation prediction method based on graph convolution collaborative filtering | |
CN110837578A (en) | Video clip recommendation method based on graph convolution network | |
CN109410001A (en) | A kind of Method of Commodity Recommendation, system, electronic equipment and storage medium | |
CN106202515A (en) | A kind of Mobile solution based on sequence study recommends method and commending system thereof | |
CN105260460A (en) | Diversity-oriented recommendation method | |
CN106777069A (en) | Personalized recommendation method and commending system based on matrix factorisation | |
Dai et al. | GraphCDA: a hybrid graph representation learning framework based on GCN and GAT for predicting disease-associated circRNAs | |
Barsotti et al. | Gravitational wave surrogates through automated machine learning | |
CN106227767A (en) | A kind of based on the adaptive collaborative filtering method of field dependency | |
CN109918564B (en) | Context automatic coding recommendation method and system for complete cold start | |
Szabó et al. | Collaborative filtering via group-structured dictionary learning | |
CN108763515B (en) | Time-sensitive personalized recommendation method based on probability matrix decomposition |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170721 |