CN110110214A - Dynamic recommendation and plus method for de-noising based on bidirectional weighting value and user behavior - Google Patents
Dynamic recommendation and plus method for de-noising based on bidirectional weighting value and user behavior Download PDFInfo
- Publication number
- CN110110214A CN110110214A CN201810072226.0A CN201810072226A CN110110214A CN 110110214 A CN110110214 A CN 110110214A CN 201810072226 A CN201810072226 A CN 201810072226A CN 110110214 A CN110110214 A CN 110110214A
- Authority
- CN
- China
- Prior art keywords
- matrix
- user
- weighting value
- user behavior
- uproar
- 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
- 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
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
Abstract
The invention discloses a kind of Dynamic recommendation based on bidirectional weighting value and user behavior and add method for de-noising, by introducing user oriented weighted value recording mode and the weighted value recording mode towards things, so that algorithm, which need to only be run, once can be completed the prediction comprising complicated emotion, without calling repeatedly, time complexity is substantially reduced, is promoted and recommends efficiency.It is proposed through the invention during prediction simultaneously plus mechanism of making an uproar, prevents " trend concentration phenomenon ", guarantees the quality of recommendation results.
Description
Technical field
The present invention relates to computer field, in particular to a kind of Dynamic recommendation based on bidirectional weighting value and user behavior and
Add method for de-noising.It can effectively solve the problem that existing proposed algorithm is existing to iterate without standard measure measuring and calculating user feeling with recommendation results
The problems such as generating " trend concentration phenomenon " afterwards.
Background technique
With the fast development of internet, internet content is in explosive growth, and the content speed of response is considerably beyond user
Receive range.How by reasonable algorithm, receives in range user is limited, recommend the content for being most suitable for user to it
As research hotspot instantly.
Meanwhile in existing proposed algorithm, with iterating to recommendation results, final result tends to be a certain solid
Definite value, this phenomenon are known as " trend concentration phenomenon "." analysis is weary of as written by California, USA university Fischbach
(Tired Analyses) " etc. described in documents, " repeat to be also easy to produce in face of similar things and is weary of psychology ", " trend concentration phenomenon " produces
Raw immobilized substance also easily causes user's dislike.Reasonable plus mechanism of making an uproar how is introduced in proposed algorithm also to be become when business
It is anxious.
Latent factor Generalization bounds are one kind by the proposition when carrying out statistical relational learning research of masschusetts, U.S.A Polytechnics
Efficient Generalization bounds (Statistical Relational Learning, SRL), it be it is a kind of utilize matrix associated entity
With a kind of efficient Generalization bounds of attribute.Rating matrix can be decomposed into user (P), project latent factor square by matrix decomposition
Battle array (Q).P, prediction matrix can be obtained in Q dot product
Traditional latent factor proposed algorithm can only predict whether user may be interested be in a certain things, and cannot describe
To the interest level of the things；And in order to describe user's interest level, then it needs that the algorithm is run multiple times.With attribute with
Things increases, and the time complexity exponentially grade of algorithm entirety increases, and causes the inefficiency of system entirety.
In traditional latent factor proposed algorithm, it can only predict whether user may like a certain things by the algorithm
Vigorously, it cannot describe to like degree to the things.
Summary of the invention
In view of this, technical problem to be solved by the invention is to provide a kind of based on bidirectional weighting value and user behavior
Dynamic recommendation and plus method for de-noising.By introducing user oriented weighted value recording mode and the weighted value record side towards things
Formula, so that algorithm, which need to only be run, once can be completed the prediction comprising complicated emotion, without calling repeatedly, when substantially reducing
Between complexity.It is proposed through the invention during prediction simultaneously plus mechanism of making an uproar, prevents " trend concentration phenomenon ", protects
Demonstrate,prove the quality of recommendation results.
An object of the present invention is to propose a kind of dynamic recommendation method based on bidirectional weighting value and user behavior, will be passed
Proposed algorithm of uniting is expanded from the single dimension that can only describe to like or do not like by introducing bidirectional weighting value as that can describe to like
The various dimensions model of degree；The second object of the present invention is to propose a kind of carrying out on the basis of purpose one plus mechanism of making an uproar, in mesh
One on the basis of, by Gaussian Profile generate random perturbation the degree of liking of different attribute is carried out in the reasonable scope
It rationally plus makes an uproar, prevents " trend concentration phenomenon " while guaranteeing recommendation results precision.
An object of the present invention is achieved through the following technical solutions:
A kind of dynamic recommendation method based on bidirectional weighting value and user behavior provided by the invention, comprising the following steps:
S1: initial method establishes project model；
S2: the matrix R for having recorded user behavior is read from database；
S3: newly-built user-Factors Weighting value matrix P and things-Factors Weighting value matrix Q；
S4: matrix decomposition is carried out to user behavior matrix R, is decomposed using SVD matrix algorithm and generates P, Q matrix respectively, made
Obtain sufficient R=PQT；
S5: user-Factors Weighting value matrix P is carried out plus is made an uproar；
S6: will add the P after making an uproar and Q dot product, obtain the prediction matrix of generation added after making an uproar
S7: recommendation results are returned toMethod terminates.
Further, it is weighted using included in user-Factors Weighting value matrix P and things-Factors Weighting value matrix Q
Value, the restriction range transmitted when being called according to interface are chosen in P, Q matrix and meet the factor of restriction, then be multiplied, estimated
Matrix is calculated, is recommended further according to this to user；
Bidirectional weighting value, that is, user-Factors Weighting value matrix the P and things-Factors Weighting value matrix Q.User-the factor
It weights value matrix P and things-Factors Weighting value matrix Q and acquisition is decomposed by SVD by user behavior matrix R, to weight value matrix
Mode save user and event respectively to the level of interest of certain Graph One factor.
The calling interface works in the following manner:
When external call is generated because of sub-interface, the above method is executed.It obtains saving user's row by accessing database
For matrix R, then execute the above method, obtain user-Factors Weighting value matrix P and things-Factors Weighting value matrix Q；
When external call analysis factor interface, range parameter, the i.e. range of the factor are limited from outside is incoming.From database
User-Factors Weighting value matrix the P and things-Factors Weighting value matrix Q of the factor of the middle taking-up only comprising incoming parameters dictate.
The two, which is multiplied, obtains prediction matrix R ', returns to prediction matrix；
When external call generates prediction result interface, it is passed to parameter and is prediction object user, it is expected to generate prediction result
Number and prediction matrix R '.If predicting object user not in prediction matrix R ' or the number of expectation generation prediction result being greater than
The affairs number of the middle record of prediction matrix R ', then prediction of failure, is returned as sky；Otherwise, by prediction matrix R ', prediction object is used
Things corresponding to family is ranked up from high to low on the basis of the numerical value of its weighted value；And it is returned with sequence from high to low
It is expected that generating things required by the number of prediction result.
Its mathematical derivation is as follows:
Remember that u row i column element is r in user behavior matrixui, u row i column element is r ' in prediction matrix R 'ui, then remember
Prediction error is eui, calculation is as follows:
eui=rui-r′ui
It is apparent from according to residual sum of squares (RSS) definition:
SSE is acquired in P using gradient descent methodukThe gradient at place:
The derivation of above formula both sides is obtained again:
By error euiDefinition known to:
Thus it easily demonstrate,proves:
Objective function SSE is also deployable are as follows:
If assuming pukUpdate step-length on objective function SSE is η, by pukOptimize along negative gradient direction, then pukMore
It is new-type are as follows:
puk:=puk-η(-euiqki) :=puk+ηeuiqki
Same easily card qkiNewer are as follows:
qki:=qki-η(-euipuk) :=qki+ηeuipuk
The second object of the present invention is to what is be realized by the following method:
After user-Factors Weighting value matrix P and things-Factors Weighting value matrix Q are generated, by every to P, Q matrix
Partial data is disturbed in row, to prevent " trend concentration phenomenon ".It is high by introducing before matrix carries out SVD Orthogonal Decomposition
The random perturbation that this distribution generates, to realize that reasonable noise introduces.
In purpose one, P, Q matrix that SVD is obtained after decomposing are all orthogonal matrix, ∑kThere is k singular value for maximum
Diagonal matrix:
Due to only recording user to the level of interest of part objects in incoming user data vector t, therefore there may be
Vacancy needs to carry out completion to vector t before being decomposed, a perfect matrix is made, to carry out SVD decomposition.
To guarantee not influencing original matrix value, the average value arranged by calculating matrixIt is inserted into, it is new after obtaining completion
VectorWith new user behavior matrixMean value calculation method is as follows:
Again willWith original subscriber's behavioural matrixSpliced, obtain new matrix r ':
By rightThe random perturbation based on Gaussian Profile is introduced, the new vector after obtaining plus making an uproar
Singular value decomposition is carried out to it on this basis, according to QR rule, is then had:
Wherein InFor n rank unit matrix.Thus it easily demonstrate,provesAnd qtFor orthogonal matrix, stFor upper triangular matrix.
Therefore it knows:
Wherein unusual value part can abbreviation in the case where losing part precision:
What i.e. acquisition was final adds result of making an uproar
One of beneficial effects of the present invention are: the present invention converts 01 matrix in traditional latent factor proposed algorithm to
Value matrix is weighted, so that not needing repeatedly this method to be called to can be obtained subject to more in the complicated user-things model of analysis
True user-Factors Weighting value matrix and user-Factors Weighting value matrix.Traditional latent factor algorithm, can only predict user couple
Whether the things is liked, its degree liked can not directly be described；In order to realize that changing demand then needs repeatedly to call；Root of the present invention
Factually border service condition carries out quantitative calculating to user behavior by introducing weighted value, is achieved the preservation to favorable rating,
It is realized simultaneously by SVD method and matrix decomposition is carried out to weighting value matrix.Use method provided by the invention, it is only necessary to calculate
It once can be obtained user-Factors Weighting value matrix and user-Factors Weighting value matrix, greatly reduce time complexity and sky
Between complexity, improve the execution efficiency of system.
The two of beneficial effects of the present invention are: the Gauss adds model of making an uproar successfully to evade " trend concentration phenomenon ".?
Guarantee recommendation results quality while, by introduce Gaussian Profile generate random perturbation, ensure that noise distribution it is uniform with
Controllably, recommendation results distribution can be quantitatively controlled, user experience is improved.
Detailed description of the invention
A kind of Dynamic recommendation and plus method for de-noising embodiment process based on bidirectional weighting value and user behavior of Fig. 1 present invention
Figure.
Specific embodiment
A specific embodiment of the invention and working principle are further detailed with reference to the accompanying drawing.
Embodiment 1
Fig. 1 is flow chart provided in an embodiment of the present invention, a kind of based on bidirectional weighting value and user behavior as shown in the figure
Dynamic recommendation and add method for de-noising, it is characterised in that: the corresponding weighting of user behavior is had recorded in the matrix R of the user behavior
Value, all behaviors of user will store after being converted into corresponding weighted value into matrix R.The generation is because sub-interface is outer
When portion is called, it is performed by described plus method for de-noising with the matrix disassembling method, the matrix disassembling method passes through aforementioned SVD
It decomposes, user behavior matrix R is decomposed into user-Factors Weighting value matrix P and things-Factors Weighting value matrix Q, then pass through institute
It states and adds method for de-noising, completion is carried out to the vector that in P, Q matrix, outside is newly inputted and is simultaneously added using the random number that Gaussian Profile generates
It makes an uproar, the prediction matrix after finally the two dot product is obtained plus made an uproar
Accordingly, the present embodiment the following steps are included:
S1: initial method, i.e., according to the user property and goods attribute defined in advance, system is suitably deposited for its distribution
Space is stored up, to store user behavior matrix R；
S2: reading the matrix R for having recorded user behavior from database, if external never call is generated because of sub-interface, user
It constantly records in external operation into matrix R；
S3: newly-built user-Factors Weighting value matrix P and things-Factors Weighting value matrix Q, i.e., according to the use defined in advance
Family attribute and goods attribute, system distribute suitable memory space for it；
S4: matrix decomposition is carried out to user behavior matrix R, is decomposed using above-mentioned SVD matrix algorithm and generates P, Q square respectively
Battle array, so that foot R=PQT；
S5: by described plus method for de-noising, user-Factors Weighting value matrix P is carried out plus is made an uproar；
S6: will add the P after making an uproar and Q dot product, obtain the prediction matrix of generation added after making an uproar
S7: recommendation results are returned toMethod terminates.
It should be pointed out that the above description is not a limitation of the present invention, the present invention is also not limited to the example above,
Variation, modification, addition or the replacement that those skilled in the art are made within the essential scope of the present invention, are also answered
It belongs to the scope of protection of the present invention.
Claims (7)
1. a kind of Dynamic recommendation based on bidirectional weighting value and user behavior and adding method for de-noising, it is characterised in that: user's row
It is memory module (1) by processing step (2), generates the user-Factors Weighting value matrix (3) and things-Factors Weighting value
Matrix (4), the recommendation results after ultimately producing plus making an uproar.
2. the Dynamic recommendation based on bidirectional weighting value and user behavior and add method for de-noising according to claim 1, it is special
Sign is: the user behavior memory module is made of database (1a) and user behavior matrix (1b), and user behavior matrix is
Two-dimensional matrix comprising user and things, the value of storage are weighted value.
3. the Dynamic recommendation based on bidirectional weighting value and user behavior and add method for de-noising according to claim 1, it is special
Sign is: processing step added by SVD singular value decomposition (2a) with Gaussian Profile make an uproar (2b) form.
4. the Dynamic recommendation based on bidirectional weighting value and user behavior and add method for de-noising according to claim 1, it is special
Sign be the following steps are included:
S1: initial method establishes project model (1)；
S2: the matrix R (1b) for having recorded user behavior is read from database (1a)；
S3: newly-built user-Factors Weighting value matrix P (3) and things-Factors Weighting value matrix Q (4)；
S4: matrix decomposition is carried out to user behavior matrix R, is decomposed using SVD matrix algorithm (2a) and generates P, Q matrix respectively, made
Obtain sufficient R=PQT；
S5: user-Factors Weighting value matrix P (3) is carried out plus makes an uproar (2b)；
S6: will add the P after making an uproar and Q dot product, obtain the prediction matrix of generation added after making an uproar
S7: recommendation results are returned toMethod terminates.
5. the Dynamic recommendation based on bidirectional weighting value and user behavior and add method for de-noising according to claim 4, it is special
Sign is: when establishing project model (1), i.e., according to the user property and goods attribute defined in advance, computer closes for its distribution
Suitable memory space, to store user behavior matrix R (1b).
6. the Dynamic recommendation based on bidirectional weighting value and user behavior and add method for de-noising according to claim 4, it is special
Sign is: the matrix R (1b) for having recorded user behavior is read from database (1a), if external never call is generated because of sub-interface,
User constantly records in external operation into matrix R (1b).
7. the Dynamic recommendation based on bidirectional weighting value and user behavior and add method for de-noising according to claim 4, it is special
Sign is: newly-built user-Factors Weighting value matrix P (3) and things-Factors Weighting value matrix Q (4), i.e. basis are defined in advance
User property and goods attribute, computer distribute suitable memory space for it.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810072226.0A CN110110214A (en) | 2018-01-25 | 2018-01-25 | Dynamic recommendation and plus method for de-noising based on bidirectional weighting value and user behavior |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810072226.0A CN110110214A (en) | 2018-01-25 | 2018-01-25 | Dynamic recommendation and plus method for de-noising based on bidirectional weighting value and user behavior |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110110214A true CN110110214A (en) | 2019-08-09 |
Family
ID=67483526
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810072226.0A Pending CN110110214A (en) | 2018-01-25 | 2018-01-25 | Dynamic recommendation and plus method for de-noising based on bidirectional weighting value and user behavior |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110110214A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115357781A (en) * | 2022-07-13 | 2022-11-18 | 辽宁工业大学 | Deep confidence network interest point recommendation algorithm based on bidirectional matrix |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20020107853A1 (en) * | 2000-07-26 | 2002-08-08 | Recommind Inc. | System and method for personalized search, information filtering, and for generating recommendations utilizing statistical latent class models |
| US20090099984A1 (en) * | 2007-10-10 | 2009-04-16 | Nec Laboratories America, Inc. | Systems and methods for generating predictive matrix-variate t models |
| US20090299996A1 (en) * | 2008-06-03 | 2009-12-03 | Nec Laboratories America, Inc. | Recommender system with fast matrix factorization using infinite dimensions |
| WO2012013996A1 (en) * | 2010-07-30 | 2012-02-02 | Gravity Research & Development Kft. | Recommender systems and methods |
| US20120143802A1 (en) * | 2010-12-02 | 2012-06-07 | Balakrishnan Suhrid | Adaptive Pairwise Preferences in Recommenders |
| US20120185172A1 (en) * | 2011-01-18 | 2012-07-19 | Barash Joseph | Method, system and apparatus for data processing |
| WO2015069607A2 (en) * | 2013-11-08 | 2015-05-14 | Microsoft Technology Licensing, Llc | Hierarchical statistical model for behavior prediction and classification |
| WO2015188349A1 (en) * | 2014-06-12 | 2015-12-17 | Hewlett-Packard Development Company, L.P. | Recommending of an item to a user |
| CN106446015A (en) * | 2016-08-29 | 2017-02-22 | 北京工业大学 | Video content access prediction and recommendation method based on user behavior preference |
| CN107545471A (en) * | 2017-09-19 | 2018-01-05 | 北京工业大学 | A kind of big data intelligent recommendation method based on Gaussian Mixture |
-
2018
- 2018-01-25 CN CN201810072226.0A patent/CN110110214A/en active Pending
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20020107853A1 (en) * | 2000-07-26 | 2002-08-08 | Recommind Inc. | System and method for personalized search, information filtering, and for generating recommendations utilizing statistical latent class models |
| US20090099984A1 (en) * | 2007-10-10 | 2009-04-16 | Nec Laboratories America, Inc. | Systems and methods for generating predictive matrix-variate t models |
| US20090299996A1 (en) * | 2008-06-03 | 2009-12-03 | Nec Laboratories America, Inc. | Recommender system with fast matrix factorization using infinite dimensions |
| WO2012013996A1 (en) * | 2010-07-30 | 2012-02-02 | Gravity Research & Development Kft. | Recommender systems and methods |
| US20120143802A1 (en) * | 2010-12-02 | 2012-06-07 | Balakrishnan Suhrid | Adaptive Pairwise Preferences in Recommenders |
| US20120185172A1 (en) * | 2011-01-18 | 2012-07-19 | Barash Joseph | Method, system and apparatus for data processing |
| WO2015069607A2 (en) * | 2013-11-08 | 2015-05-14 | Microsoft Technology Licensing, Llc | Hierarchical statistical model for behavior prediction and classification |
| WO2015188349A1 (en) * | 2014-06-12 | 2015-12-17 | Hewlett-Packard Development Company, L.P. | Recommending of an item to a user |
| CN106446015A (en) * | 2016-08-29 | 2017-02-22 | 北京工业大学 | Video content access prediction and recommendation method based on user behavior preference |
| CN107545471A (en) * | 2017-09-19 | 2018-01-05 | 北京工业大学 | A kind of big data intelligent recommendation method based on Gaussian Mixture |
Non-Patent Citations (1)
| Title |
|---|
| 曹向前: ""基于贝叶斯网络的协同过滤推荐算法"", 《软件导刊》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115357781A (en) * | 2022-07-13 | 2022-11-18 | 辽宁工业大学 | Deep confidence network interest point recommendation algorithm based on bidirectional matrix |
| CN115357781B (en) * | 2022-07-13 | 2024-02-23 | 辽宁工业大学 | Deep confidence network interest point recommendation algorithm based on bidirectional matrix |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Lian et al. | Scaling & shifting your features: A new baseline for efficient model tuning | |
| Kitamura et al. | Nonparametric analysis of random utility models | |
| Pachter et al. | Tropical geometry of statistical models | |
| Mehrkanoon et al. | Approximate solutions to ordinary differential equations using least squares support vector machines | |
| Frank | The common patterns of nature | |
| CN106250464B (en) | Training method and device of ranking model | |
| CN111798273A (en) | Training method of purchase probability prediction model of product and purchase probability prediction method | |
| Mocenni et al. | Identifying the dynamics of complex spatio-temporal systems by spatial recurrence properties | |
| CN113420421B (en) | QoS prediction method based on time sequence regularized tensor decomposition in mobile edge calculation | |
| Jagalur-Mohan et al. | Batch greedy maximization of non-submodular functions: Guarantees and applications to experimental design | |
| Colson et al. | Optimizing partially separable functions without derivatives | |
| Wang et al. | Orthogonal subspace learning for language model continual learning | |
| Fang et al. | LASSO isotone for high-dimensional additive isotonic regression | |
| CN110110214A (en) | Dynamic recommendation and plus method for de-noising based on bidirectional weighting value and user behavior | |
| Li et al. | Switching diffusion logistic models involving singularly perturbed Markov chains: weak convergence and stochastic permanence | |
| Kook et al. | Estimating conditional distributions with neural networks using r package deeptrafo | |
| Bronstein et al. | How to speed up the quantization tree algorithm with an application to swing options | |
| Cropp et al. | Parameterizing plankton functional type models: insights from a dynamical systems perspective | |
| CN116383441A (en) | Community detection method, device, computer equipment and storage medium | |
| Van Dijk | Some advances in Bayesian estimation methods using Monte Carlo integration | |
| US20130325786A1 (en) | Sparse higher-order markov random field | |
| Picchini | Likelihood-free stochastic approximation EM for inference in complex models | |
| Baillie et al. | Inference for impulse response coefficients from multivariate fractionally integrated processes | |
| EP3602353A1 (en) | Statistics-based multidimensional data cloning | |
| CN112381147B (en) | Dynamic picture similarity model establishment and similarity calculation method and device |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190809 |