CN107256241A - The film recommendation method for improving multi-objective genetic algorithm is replaced based on grid and difference - Google Patents
The film recommendation method for improving multi-objective genetic algorithm is replaced based on grid and difference Download PDFInfo
- Publication number
- CN107256241A CN107256241A CN201710381920.6A CN201710381920A CN107256241A CN 107256241 A CN107256241 A CN 107256241A CN 201710381920 A CN201710381920 A CN 201710381920A CN 107256241 A CN107256241 A CN 107256241A
- Authority
- CN
- China
- Prior art keywords
- individual
- film
- population
- grid
- algorithm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
- G06F16/355—Class or cluster creation or modification
-
- 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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Genetics & Genomics (AREA)
- General Health & Medical Sciences (AREA)
- Physiology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The film recommendation method for improving multi-objective genetic algorithm is replaced based on grid and difference, innovatory algorithm GDNSGA II is proposed for not enough of both distributivity present in NSGA II and convergence, can be used for solving multiple target combinatorial optimization problem.Algorithm design multiple target mesh generation mode initializes population, it is to avoid individual uneven and cause distributivity to lack；And Evolution of Population process is safeguarded with Clustering and selection and difference replacement operator, an appropriate number of worst individual Local Search is selected, the convergence and distributivity of population is maintained.With reference to the excavation of user behavior and film native, by the algorithm applied to this practical problem of film personalized recommendation, by carrying out versatility and validity that test comparison illustrates algorithm with existing algorithm, more excellent recommendation results are obtained, improve F mediations rate, diversity and the novelty of recommendation, and combined there is provided the suggested design of more horn of plenty, be conducive to the interest for fully excavating user to provide more structurally sound recommendation service.
Description
Technical field
The invention belongs to personalized recommendation technical field.The many mesh being improved with the deficiency for NSGA-II algorithms
Marking genetic algorithm GDNSGA-II algorithms, (specifically related to NSGA-II algorithms, multiple target mesh generation mode and cluster are trimmed and poor
Different replacement operator) realize the personalized recommendation to film.
Background technology
The fast development of application popularization and hyundai electronicses commercial affairs with Internet technologies, is full of the money in internet
The situation that source quantity is exponentially increased.Substantial amounts of information is presented simultaneously often causes user to feel at loose ends, it is difficult to therefrom seek
Oneself real resource interested is found, so as to occur in that so-called " information explosion " and " information overload " phenomenon.Search engine
Occur with information retrieval technique to alleviate problem of information overload.In information-based today, user is commonly using search
Engine finds the resource required for oneself.But traditional search engine technique does not account for the individual difference of user, by institute
There are user's equivalent processes, return to what the resource of user was just as, simultaneously because feedack amount is also very big so that use
Family is difficult to choose the resource oneself liked.Therefore, how according to the preference characteristics of each user, largely believe from internet
The information for meeting user's request is found in breath, and then recommends user, is had become when previous urgently to be resolved hurrily studies a question.
Personalized recommendation system (personlized recommender systems) be exactly meet the tendency of under this background and
Raw.It obtains the preference characteristics of user by the behavioural characteristic of user in collection system, and then according to these preference characteristics
Excavate that user is potential interested or resources of needs from the bulk information on network, and make corresponding recommendation.Recommend
It is exactly to predict user to non-selected resource (such as music, film, book by analyzing resource that user selected on question essence
Nationality, webpage, restaurant, tourist attractions etc.) fancy grade, and the result of prediction is presented to use in certain effective form
Family, such as by the higher resource recommendation of predicted value to user.
Commending system is widely used in many fields at present, and common website is such as in daily life：Purchase by group net,
Jingdone district store, Taobao, only product meeting, Amazon, Dangdang.com etc. are all typical commending systems.And it is popular at present go where
Can the website of the characteristic theory such as net, ctrip.com, effectively find and draw over to one's side client, allow the recommendation system in client's dependence
Can system, keeps supply of long duration relation, so as to improving sales achievement, namely choose ripe effective commending system relation
To the final and decisive juncture of e-commerce venture.Also, it is recommended to system regions are always by academia as one of temperature research topic, and
It is progressively independent into a special disciplines.
As academia, engineering circles and business circles are to commending system in-depth study, other field is also widely applied to
Commending system and its technology.Nowadays, commending system related algorithm or by applied to data and the network information including library
The information service that retrieval and DTV are watched etc, or even some fairly simple commending systems are mostly applied such as
In " bean cotyledon net ", the network forum of " Baidu's mhkc ".It can be seen that by development for many years, commending system application field also will increasingly
Wide in range, related personnel and scholar are increasing to the research interest in the field.
Commending system performance depends entirely on selected proposed algorithm.The quality of the quality of recommendation has many evaluation marks
Standard, such as the precision recommended, the personalization level of recommendation, the accuracy rate of recommendation, recall rate etc..Generally some or it is a few
When individual standard is optimal, recommend quality relatively good.Many optimization problems can all sum up in scientific research and engineering practice
For multi-objective optimization question (MOP), personalized recommendation is also the multi-objective optimization question for considering many aspects.
Multiple-objection optimization is a new branch of science of the applied mathematics developed rapidly over nearly more than 20 years.It studies vector
Optimization problem of object function when meeting certain constraints under certain meaning.Due to a large amount of optimization problems of real world,
The optimization problem containing multiple targets can be all attributed to, since the seventies, for the research of multiple-objection optimization, at home and state
People greatly concern and attention are all caused on border.Particularly during the nearly last ten years, theory study deepens continuously, application day
Beneficial extensive, studying team grows rapidly, shows vitality.Meanwhile, with medium-and-large-sized multiple to social economy and engineering design
Miscellaneous system research is goed deep into, and the theory and method of multiobjective optimization also constantly by severe challenge and are rapidly developed.
Multiple-objection optimization has it to be widely applied field as an important research direction in optimization field, and research solves it and effectively calculated
Method has great scientific meaning and application value.
The research purpose of multi-objective Evolutionary Algorithm, which is mainly, makes the disaggregation tried to achieve as much as possible close to the Pareto of problem
Preferable forward position, and it is widely distributed and uniform, and this just determines that the performance indications of evaluation algorithms are distributivity and convergence two
Aspect.Distributivity and convergence have great importance for solving multi-objective optimization question, and good distributivity can be to certainly
Plan person provides more rationally effective selection schemes；Good convergence can more accurately solve the answer of practical problem.
In recent years, evolutionary computation field proposes some multi-objective Evolutionary Algorithms in succession.Wherein, it is most representational mainly to have：
The SPEA (Strength Pareto EvolutionaryAlgorithm) that Zitzler and Thiele is proposed, Kim et al. is at it
On the basis of the non-dominated sorted genetic algorithm NSGA (Non-dominated that propose of the SPEA2, Srinivas and the Deb that propose
Sorting GeneticAlgorithm), and the proposition such as the NSGA-II that Deb etc. is proposed on its basis, Corne PESA
(Pareto Envelope-based SelectionAlgorithm) and PESA-II.NSGA-II is that one kind is most widely used
Multi-objective Evolutionary Algorithm, be to determine that individual is suitable according to the Pareto dominance relations between individual and density information the characteristics of the algorithm
It should be worth, but such fitness calculation there is distributivity and convergence safeguards improperly defect.The bases such as Wen Shihua
The mode of distance metric retains some representative individuals and NSGA-II is improved, and this mode, which only considered, gathers around
Influence of the distance to holding Species structure is squeezed, does not consider that feature is close individual and bad comprehensively from the angle of similar individual collections
The problem of convergence and distributivity that the presence of matter individual is caused are lacked.For the defect of NSGA-II these two aspects, set forth herein
A kind of improved convergence and distributivity keep strategy, and algorithm is set before evolution by the way of multiple target mesh generation
Initialization population is put, prevents population to be easily trapped into local convergence because of random initializtion or restrained slowly, and is avoided initial
Population at individual is uneven and causes distributivity to lack；During evolution, the evolution result to every generation is clustered algorithm,
The dynamics selected according to the size dynamic regulation of similarity in class, by the adaptive an appropriate number of feature of selection it is close and
Non-dominated ranking and individual in the poor class of crowding distance and a small amount of individual away from leading surface, and carried out in smaller range
Local Search carries out population maintenance, accelerates to maintain distributivity again while convergence in population.The innovatory algorithm is applied to film
In the particular problem of personalized recommendation, by with NSGA-II and the Collaborative Filtering Recommendation Algorithm based on user, based on article
The conventional recommendation such as Collaborative Filtering Recommendation Algorithm and content-based recommendation algorithm algorithm carries out contrast experiment under the same conditions, tests
The effect of algorithm is demonstrate,proved.
The content of the invention
It is used for film the purpose of the present invention is to propose to a kind of improvement multi-objective genetic algorithm replaced based on grid and difference
The solution of this practical problem of TOP-N in personalized recommendation, it is pre- with the scoring of N portions film user's behavior prediction and N portions film native
Test and appraisal are divided into two targets and optimized, and realize personalized recommendation.
The improvement multi-objective genetic algorithm (GDNSGA-II) replaced based on grid and difference of the present invention, it is characterised in that：
Algorithm is set initialization population by the way of multiple target mesh generation, prevents population because of random initializtion before evolution
It is easily trapped into local convergence or restrained slow, and it is uneven and cause distributivity to lack to avoid initial population individual；Algorithm
During evolution, the evolution result to every generation is clustered, and is selected in class according to the size dynamic regulation of similarity
Dynamics, selects an appropriate number of feature close and non-dominated ranking and individual in the poor class of crowding distance by adaptive
And a small amount of individual away from leading surface, and population maintenance is carried out in smaller range Local Search, while accelerating convergence in population
Distributivity is maintained again.
The film recommendation method for improving multi-objective genetic algorithm is replaced based on grid and difference, is comprised the following steps：
S1 carries out individual UVR exposure, initialization data, and setup parameter
The individual represents the numbering of N number of film, and wherein N represents the film number for needing to recommend；Use diRepresent to need for i-th
The film to be recommended numbering, in order to applied to the TOP-N for solving the problems, such as personalized recommendation field, be numbered and be combined using N number of film,
Individual UVR exposure form is：<d1, d2, di...dN>, using real coding mode, coding range be scope where film is numbered simultaneously
And be integer form；Population Size is initialized as popszie by the initialization data, and it is big that each offspring produces popsize
Small population；The setup parameter includes：Crossover probability Pc is set as 0.9, mutation probability Pm is 0.1, individual lengths are N,
Film in coding is kept to number in order and do not repeat, using a kind of combination as the different films in N portions to be recommended.
S2 multiple target mesh generation modes initialize population
S2.1 randomly generates s times of population scale number n interim population P, i.e. population P individual amount for s × n, this
In s be decimal system positive integer；
S2.2 calculates each individual in interim population P the numerical value fit of each target in problem to be optimized, is expressed as
Numerical value of i-th of individual on t-th of object function；
S2.3 carries out non-branch to each individual in interim population P according to the good and bad relation between each target value fit
Calculating with sequence layering, and the level ci where each individual is marked, it is expressed as the level where i-th of individual；
S2.4 calculates crowding distance di to each individual in population P, is expressed as i-th using population P as complete or collected works space
The numerical value of body crowding distance；
S2.5, using each individual non-dominated ranking level ci and crowding distance di as two targets, is commented in population P
Valency individual between non-dominant relation come select it is optimal before (s × n)/2 individual composition candidate collection P1, wherein, n is population
Scale number.
S2.6 according in candidate collection P1 numerical value a little in each target, find the maximum on each optimization direction
Value and minimum value, and numerical value is normalized.
S2.7 divides the space into K grid according to the normalization maxima and minima in S2.6, and will belong to correspondence
The numbering k of the individual mark grid of grid.
S2.8 first time individual choices：Number M of the statistics containing all grids a little, and calculate the centre bit of these grids
The coordinate put, if M be more than population scale n, from M grid select n grid, then calculate this n grid point and
The distance of center, selects point composition initial population nearest in each grid, terminates the operating process；Planted if M is equal to
Group scale n, then calculate the point of each grid to the distance of center, select nearest point just to constitute initial population, terminate
The operating process；If M is less than population scale n, the point of each grid is calculated to the distance of center, nearest point is selected
As first time selection set, and continue executing with S2.9 second of individual choice of progress.
S2.9 carries out second according to the difference u between the individual amount of population scale number and first time individual choice
Body is selected, if u<W, then randomly choose u grid in all w grids for containing a little, randomly choosed from each grid
One point merges with the point set selected for the first time before, initial population is generated, if u>W (potentially possible rare occasion),
Then selection random point is gathered as second of selection in w grid is randomly choosed, and continues the step and carries out third time selection, directly
Untill reaching population scale.
S3 selection operations
Individual choice is carried out in the way of championship, i.e., randomly chooses k, k from this popszie individual<Popszie
Individual, it is n/2 (rounding) that k is taken here, and an optimum individual is chosen from this k individual.
Selection standard is with number of plies S where non-dominated rankingiWith crowding distance numerical value diAs two targets, according to non-branch
With winning relation, each individual is compared, relatively winning individual is more excellent individual.
S4 crossover operations
Crossover operation uses SBX crossover operators, it is assumed that when former generation (the t generations in evolutionary process) two individuals to be intersected
For XA t、XB t, α is to intersect the parameter (span is 0~1) being related to, then XA t+1、XB t+1Two individuals produced for the next generation.
Form is as follows：
S5 mutation operations
To the gen in evolutionary process for population PgenIn any individual pi=(pi1, pi2…piN), i ∈ { 1,2 ...
N }, mutation operation is participated in probability P m：Produce a decimal r ∈ [0,1], and a random integers j ∈ [1, N]；Make pi, j=
Lj+r* (uj-lj), to colony PgenEnter row variation and produce new population PgenNew。
S6 Clustering and selections replace Local Search with difference
A S6.1 generation, which is evolved in the population Pnew after terminating, n individual, each individual coordinate in each dimension
Numerical value is xim, represents the value of the individual in m dimensions, calculates the numerical value fit of each target in problem to be optimized, table
Numerical value of the individual on t-th of object function is shown as, wherein, i is each individual sequentially numbering；
S6.2 carries out non-dominant to each individual in population Pnew according to the good and bad relation between each target value fit
The calculating of sequence layering, the level ci where each individual of mark, and the crowding distance di of each individual of calculating；
Coordinate value xims and non-dominant layering ranking ci and crowding distance of the S6.3 in each dimension according to where individual
Di numerical value, is clustered to the individual in the population using adaptive minimum spanning tree clustering algorithm, obtains k classification, made
For the close individual collections of k feature { U1, U2 ..., Uk }；
Individual in each close to feature S6.4 set { U1, U2 ..., Uk }, each dimension according to where individual
On coordinate value xim and each target numerical value fit calculate two-by-two individual between similarity pij, wherein, i, j is each
The sequentially numbering of individual；
S6.5 calculate class in two-by-two individual between similarity pij, and calculate each set in average similarity Pk；
S6.6 is calculated each average similarity Pk in gathering, and calculates the individual amount for needing to retain in each class
nk；
The individual amount nk that S6.7 retains according to each class of calculating, selection non-dominant layering ranking and crowding distance numerical value
Poor (n (Uk)-nk) individual and the individual away from preferable leading surface, wherein, n (Uk)
For the individual amount of k-th of set；
The solution numerical value in each individual dimension that S6.8 is selected to these random replacement in smaller range, formed it is multigroup can
To replace the alternative set for being chosen individual, and by the individual in these alternative set according to non-dominated ranking and crowding distance
Selection mode selection optimum individual is compared with original individual, is replaced if better than original individual, otherwise, retains former
There is individual.
The defect individual that S6.9 passes through these in each set of replacement operation merges with the individual collections retained, again
The new population Pnew of composition, continues executing with follow-on evolutionary process.
S7 end conditions judge
If algebraically as defined in reaching obtains satisfied result, terminate and output result, otherwise turn S3 steps.
S8 enters next heredity circulation
By population PgenNewThe initial population evolved as the next generation, proceeds S3 steps.
It is the main process flow steps for the improvement multi-objective genetic algorithm replaced based on grid and difference above, uses the algorithm
Flow is applied to the recommendation list that film personalized recommendation problem obtains film.
Compared with prior art, the present invention has the advantages that.
The improvement multi-objective genetic algorithm replaced based on grid and difference, algorithm is before evolution, using multiple target grid
The mode of division sets initialization population, prevents population to be easily trapped into local convergence because of random initializtion or restrained slowly, and
And it is uneven and cause distributivity to lack to avoid initial population individual；Algorithm during evolution, to the evolution knot of every generation
Fruit is clustered, the dynamics selected in class according to the size dynamic regulation of similarity, passes through adaptive selection proper number
Feature is close and non-dominated ranking and individual in the poor class of crowding distance and a small amount of individual away from leading surface, and
Local Search carries out population maintenance in smaller range, accelerates to maintain distributivity again while convergence in population.By the way that this is improved
Algorithm is applied in the particular problem of film personalized recommendation, with NSGA-II and the Collaborative Filtering Recommendation Algorithm based on user
Both traditional proposed algorithms carry out contrast experiment under the same conditions with content-based recommendation algorithm, demonstrate algorithm
Practical function.
Brief description of the drawings
Fig. 1 is the flow chart for the improvement multi-objective genetic algorithm replaced based on grid and difference.
Fig. 2 is that multiple target mesh generation mode initializes population broad flow diagram.
Fig. 3 is that Clustering and selection replaces operator broad flow diagram with difference.
Fig. 4 is General Implementing method flow diagram of the present invention.
Fig. 5 is the Collaborative Filtering Recommendation Algorithm broad flow diagram based on user.
Fig. 6 is the Collaborative Filtering Recommendation Algorithm broad flow diagram based on article.
Fig. 7 is content-based recommendation algorithm broad flow diagram.
Embodiment
The present invention will be further described with specific embodiment below in conjunction with the accompanying drawings.
The data set that the present invention is recommended from MovieLens as film, the data set, which is related to 100,000 film, to be commented
Point, 943 users and 1682 films.By the collaborative filtering (UserCF) based on user, the collaboration based on article
Filter algorithm (ItemCF), content-based recommendation algorithm (Content), tetra- algorithms progress pair of NSGA-II and GDNSGA-II
Than experiment.
In two kinds of multi-objective optimization algorithms of NSGA-II and GDNSGA-II, numbering is the film sequence number of N portions film
(N values are 3,6,9,12,15,18,21,24,27,30 here), numbering is limited to not repeat in order, both operation algebraically
Gen is 200, and population scale popsize is 50, sets crossover probability Pc as 0.9, and mutation probability Pm is 0.1, to use user's phase
Like K user most close in degree matrix calculates the corresponding prediction scoring sum of N portions film numbering, scored similarity using article
K article most close calculates the corresponding prediction scoring sum of N portions film numbering and uses film native similarity in matrix
The corresponding prediction scoring sum of N number of film numbering that matrix computations similar movies are obtained is used as three optimization object functions.For
The single portion's film prediction scoring calculated using above-mentioned three kinds of modes is recorded, and calculates same film numbering next time
Corresponding film scoring need not then be calculated, and directly read score value summation.Adjusted with conventional recommendation evaluation criterion F
Carry out the effect of verification algorithm as index with the abundant degree situation of rate, diversity, novelty and suggested design.Its calculating side
Method is as follows：
If on the basis of data acquisition system I, | | I | | represent all things numbers；| | U | | represent all number of users；r
(u) refer to calculating obtained recommendation things combination；T (u) refers to user u things combinations interested；||r(u)∩t(u)|
| refer to that the things recommended combines the same number that the things interested with user u is combined.Based on given variable, to above-mentioned
The definition of several crucial criterions is described as follows with calculation.
(1) accuracy rate is a kind of calculation for measuring recommendation effect levels of precision.
For accuracy rate calculation we can provide form described below：
From this calculation it will be seen that accuracy rate represent be user recommend hit things number of sets with
The ratio size of the number of sets of all recommendation things.
(2) recall rate be judge recommendation results accurately on the basis of whether comprehensively a kind of calculation.
Comparatively, based on the above-mentioned established condition provided, the calculating for recall rate can use form described below：
By above-mentioned calculation, it has been found that the angle for stressing user is compared in the calculating of recall rate, the canonical representation is recommended
The things number of sets of hit and the ratio size of all things number of sets interested of user.
For the staff of platform, obtained recommendation combination is calculated using the recommendation method of this platform to try one's best
All it is user's things interested, so it is considered that accuracy rate is higher with recall rate, the good reality of recommendation method can be embodied by getting over
Border performance.
(3) F mediations rate is a kind of calculation for the above-mentioned accuracy rate of overall measure and the aspect standard of recall rate two.Number
Learn calculation formula as follows：
Because both accuracy rate and recall rate are often restricted each other, raising on the one hand means decrease on the other hand, F
Mediation rate can be with the levels of precision of relatively reasonable evaluation recommendation effect.
(4) diversity is a kind of measurement mode for measuring difference degree between things in the result recommended.
For multifarious calculation, it is as follows that we can provide mathematical formulae for user u：
Wherein, sim (i, j) represents the similarity between two articles i, j.
And then obtain for all users and diversity quality and the proportional representation of evaluation size
It is as follows：
For unique user, diversity represents things similarity sum and all two two-by-two in the things list recommended
The ratio of the possible number of combinations of two things；And for all users, be expressed as 1 and calculate the obtained numerical value with each user
Expectation difference.
(5) novelty is to recommend more unexpected winner for active user but the weighing apparatus of the things consistent with its interest orientation
Amount standard.
For the calculation of novelty, it is as follows that we can provide mathematical formulae for user u：
Wherein, k (i) pays close attention to article i customer volume.
And then obtain for all users and novelty quality and the proportional representation of evaluation size
It is as follows：
From formula as can be seen that novelty represent be things popularity numerical value expectation.Popularity then refers to some
The number of users that things is evaluated.
(6) the abundant degree of suggested design is each user's suggested design number average value tried to achieve.
Main flow such as Fig. 1 institutes of the improvement multi-objective genetic algorithm proposed by the present invention replaced based on grid and difference
Show, whole flow process designs initialization population and Clustering and selection with multiple target mesh generation mode and safeguarded with difference replacement operator
The distributivity and convergence of population.Multiple target mesh generation mode is broadly divided into design initialization population, selection operation, intersect behaviour
Work, mutation operation, Clustering and selection and difference replace five parts of operator, wherein, multiple target mesh generation mode designs initialization
Population main flow is as shown in Fig. 2 Clustering and selection and difference replacement operator main flow are as shown in Figure 3.Selection, intersection, variation
Three operating processes are consistent with the operating process of NSGA-II algorithms.
The implementation process of the present invention is described in detail with reference to Fig. 4.Embodiments of the invention are with the technology of the present invention
Implemented premised on scheme, give detailed embodiment and specific operating process, but protection scope of the present invention
It is not limited to following embodiments.
Embodiment is from film personalized recommendation problem to multiple target innovatory algorithm GDNSGA-II presented herein and base
Collaborative filtering (UserCF) in user, the collaborative filtering (ItemCF) based on article, content-based recommendation are calculated
Method (Content) and NSGA-II are tested and compared.The algorithms of different under same experiment condition is can be seen that by comparing
In the performance state of processing same problem.
Wherein, based on user the score value that collaborative filtering (UserCF) is seen a film according to user, calculates two
Similarity between two users, and select sorted from big to small with the similarity of user to be recommended before K user, it is similar with these
Number of degrees value is weight, and the film that the film seen and scored using these users has not been seen to user to be recommended is predicted
Scoring.Main flow is as shown in Figure 5.
The score value that collaborative filtering (ItemCF) based on article is seen a film according to user, calculates article two-by-two
Between similarity, and select sorted from big to small with the similarity of article to be recommended before K article, with these similarity numerical value
For weight, the film that the film seen and scored using these users has not been seen to user to be recommended is predicted scoring.
Main flow is as shown in Figure 6.
Content-based recommendation algorithm (Content) calculates the phase between each film according to the belonging relation of film types
Like degree size, and the film seen to user to be recommended scores higher film, according to similar degree size pair
The film do not seen is predicted scoring.Main flow is as shown in Figure 7.
In this embodiment, for NSGA-II and GDNSGA-II, to use most phase in user's similarity matrix
K near user calculates the corresponding prediction scoring sum of N portions film numbering, using K most close in article similarity matrix
Article calculates the corresponding prediction scoring sum of N portions film numbering and calculates similar movies using film native similarity matrix and obtains
The corresponding prediction scoring sum of N number of film numbering arrived is as three optimization object functions.For using above-mentioned three kinds of sides
Single portion's film prediction scoring that formula is calculated is recorded, and is calculated same film next time and is numbered corresponding film scoring then not
Need to calculate, directly read score value summation.Purpose is the numbering combination for solving N number of different films, i.e., each
The gene position of individual is the numbering of a film, uses diThe numbering of i portions film to be recommended is represented, individual UVR exposure form is：<d1,
d2, di。。。dN>, using real coding mode, coding range is the scope where film is numbered and is integer form, keeps compiling
Film numbering is orderly in code and does not repeat, using a kind of combination as the different films in N portions to be recommended.
The explanation of each detailed problem involved in the inventive technique scheme is provided in detail below：
S1 carries out individual UVR exposure, initialization data, and setup parameter
The individual represents the numbering of N number of film, and wherein N represents the film number for needing to recommend；Use diRepresent to need for i-th
The film to be recommended numbering, in order to applied to the TOP-N for solving the problems, such as personalized recommendation field, be numbered and be combined using N number of film,
Individual UVR exposure form is：<d1, d2, di。。。dN>, using real coding mode, coding range be scope where film is numbered simultaneously
And be integer form；Population Size is initialized as popszie by the initialization data, and it is big that each offspring produces popsize
Small population；The setup parameter includes：Crossover probability Pc is set as 0.9, mutation probability Pm is 0.1, individual lengths are N,
Film in coding is kept to number in order and do not repeat, using a kind of combination as the different films in N portions to be recommended.
S2 multiple target mesh generation modes initialize population
S2.1 randomly generates s times of population scale number n interim population P, i.e. population P individual amount for s × n, this
In s be decimal system positive integer；
S2.2 calculates each individual in interim population P the numerical value fit of each target in problem to be optimized, is expressed as
Numerical value of i-th of individual on t-th of object function；
S2.3 carries out non-branch to each individual in interim population P according to the good and bad relation between each target value fit
Calculating with sequence layering, and the level ci where each individual is marked, it is expressed as the level where i-th of individual；
S2.4 calculates crowding distance di to each individual in population P, is expressed as i-th using population P as complete or collected works space
The numerical value of body crowding distance；
S2.5, using each individual non-dominated ranking level ci and crowding distance di as two targets, is commented in population P
Valency individual between non-dominant relation come select it is optimal before (s × n)/2 individual composition candidate collection P1, wherein, n is population
Scale number.
S2.6 according in candidate collection P1 numerical value a little in each target, find the maximum on each optimization direction
Value and minimum value, and numerical value is normalized.
S2.7 divides the space into K grid according to the normalization maxima and minima in S2.6, and will belong to correspondence
The numbering k of the individual mark grid of grid.
S2.8 first time individual choices：Number M of the statistics containing all grids a little, and calculate the centre bit of these grids
The coordinate put, if M be more than population scale n, from M grid select n grid, then calculate this n grid point and
The distance of center, selects point composition initial population nearest in each grid, terminates the operating process；Planted if M is equal to
Group scale n, then calculate the point of each grid to the distance of center, select nearest point just to constitute initial population, terminate
The operating process；If M is less than population scale n, the point of each grid is calculated to the distance of center, nearest point is selected
As first time selection set, and continue executing with S2.9 second of individual choice of progress.
S2.9 carries out second according to the difference u between the individual amount of population scale number and first time individual choice
Body is selected, if u<W, then randomly choose u grid in all w grids for containing a little, randomly choosed from each grid
One point merges with the point set selected for the first time before, initial population is generated, if u>W (potentially possible rare occasion),
Then selection random point is gathered as second of selection in w grid is randomly choosed, and continues the step and carries out third time selection, directly
Untill reaching population scale.
S3 selection operations
Individual choice is carried out in the way of championship, i.e., randomly chooses k, k from this popszie individual<Popszie
Individual, it is n/2 (rounding) that k is taken here, and an optimum individual is chosen from this k individual.
Selection standard is with number of plies S where non-dominated rankingiWith crowding distance numerical value diAs two targets, according to non-branch
With winning relation, each individual is compared, relatively winning individual is more excellent individual.
S4 crossover operations
Crossover operation uses SBX crossover operators, it is assumed that when former generation (the t generations in evolutionary process) two individuals to be intersected
For XA t、XB t, α is to intersect the parameter (span is 0~1) being related to, then XA t+1、XB t+1Two individuals produced for the next generation.
Form is as follows：
S5 mutation operations
To the gen in evolutionary process for population PgenIn any individual pi=(pi1, pi2…piN), i ∈ { 1,2 ...
N }, mutation operation is participated in probability P m：Produce a decimal r ∈ [0,1], and a random integers j ∈ [1, N]；Make pi, j=
Lj+r* (uj-lj), to colony PgenEnter row variation and produce new population PgenNew。
S6 Clustering and selections replace Local Search with difference
A S6.1 generation, which is evolved in the population Pnew after terminating, n individual, each individual coordinate in each dimension
Numerical value is xim, represents the value of the individual in m dimensions, calculates the numerical value fit of each target in problem to be optimized, table
Numerical value of the individual on t-th of object function is shown as, wherein, i is each individual sequentially numbering；
S6.2 carries out non-dominant to each individual in population Pnew according to the good and bad relation between each target value fit
The calculating of sequence layering, the level ci where each individual of mark, and the crowding distance di of each individual of calculating；
Coordinate value xims and non-dominant layering ranking ci and crowding distance of the S6.3 in each dimension according to where individual
Di numerical value, is clustered to the individual in the population using adaptive minimum spanning tree clustering algorithm, obtains k classification, made
For the close individual collections of k feature { U1, U2 ..., Uk }；
Individual in each close to feature S6.4 set { U1, U2 ..., Uk }, each dimension according to where individual
On coordinate value xim and each target numerical value fit calculate two-by-two individual between similarity pij, wherein, i, j is each
The sequentially numbering of individual；
S6.5 calculate class in two-by-two individual between similarity pij, and calculate each set in average similarity Pk；
S6.6 is calculated each average similarity Pk in gathering, and calculates the individual amount for needing to retain in each class
nk；
The individual amount nk that S6.7 retains according to each class of calculating, selection non-dominant layering ranking and crowding distance numerical value
Poor (n (Uk)-nk) individual and the individual away from preferable leading surface, wherein, n (Uk) is the individual amount of k-th of set；
The solution numerical value in each individual dimension that S6.8 is selected to these random replacement in smaller range, formed it is multigroup can
To replace the alternative set for being chosen individual, and by the individual in these alternative set according to non-dominated ranking and crowding distance
Selection mode selection optimum individual is compared with original individual, is replaced if better than original individual, otherwise, retains former
There is individual.
The defect individual that S6.9 passes through these in each set of replacement operation merges with the individual collections retained, again
The new population Pnew of composition, continues executing with follow-on evolutionary process.
S7 end conditions judge
If algebraically as defined in reaching obtains satisfied result, terminate and output result, otherwise turn S3 steps.
S8 enters next heredity circulation
By population PgenNewThe initial population evolved as the next generation, proceeds S3 steps.
It is, based on the main process flow steps for improving multi-objective genetic algorithm for just giving cluster to trim, to use the algorithm above
Flow is applied to the recommendation list that film personalized recommendation problem obtains film.
The experimental result of the explanation present invention is explained in detail below：
In order to prove validity of the method for the invention in film personalized recommendation problem, GDNSGA- is respectively adopted
II (method in the present invention) and UserCF, ItemCF, Content and NSGA-II are to the TOP-N in film personalized recommendation
Problem is optimized, wherein, N be recommend in a combination film number (distinguish here value be 3,6,9,12,15,18,
21、24、27、30).Experimental result is as shown in table 1 (wherein overstriking font is the preferable experimental data of performance).
Table 1 respectively recommends method F mediations rate, diversity and novelty contrast table
As shown in Table 1, under the conditions of different recommendation list length, for traditional recommendation method, Content exists
Effect is preferably, very poor in accurate aspect of performance result in terms of diversity, is also performed poor in terms of novelty；UserCF is accurate
Property aspect effect preferably, and be all average in performance in diversity and the aspect of novelty two；ItemCF does well in terms of novelty,
And it is general in the performance of accurate aspect of performance, show poor in terms of diversity.Above-mentioned three kinds of conventional recommendation methods are all just for phase
Close the information in terms of stressing to be predicted and excavate, the result tried to achieve is carried the deviation of uniformity, may be a or two side
Face performance is preferable, but have ignored to otherwise consideration, comprehensive comprehensively to recommending problem to handle.
For the recommendation method based on multi-objective genetic algorithm, under the conditions of different recommendation list length,
GDNSGA-II is superior to NSGA-II in above three index, and this is due to that GDNSGA-II uses multiple target mesh generation side
Formula and Clustering and selection replace operator with difference and NSGA-II are improved so that GDNSGA-II convergences and distributivity side
The performance in face is lifted, GDNSGA-II than NSGA-II in terms of the detection of suggested design it is better.
For the assembled scheme of recommendation, using GDNSGA-II (present invention in method) and UserCF, ItemCF,
(wherein overstriking font is preferable for performance as shown in table 2 for the combination experimental result that Content and NSGA-II algorithms are obtained
Experimental data).
Table 2 respectively recommends the scheme of method to enrich degree contrast table
As shown in Table 2, reinforcements of the OTNSGA-II (method in the present invention) in terms of distributivity is used so that comparing NSGAII
It can obtain more various solution, and comprehensive be pushed away the characteristics of examined user's history behavior and film native itself
Recommend so that recommend more comprehensively and personalized.For these solutions that GDNSGA-II is obtained, from multiple-objection optimization
From the perspective of be all non-bad, i.e., be equivalent good and bad relation each other, businessman can carry according to different solutions
Recommend combination for more various film, user can be more suitable for the electricity of oneself current interest according to different solution selections
Shadow combination is watched so that recommendation results are more accurate and practical.
Claims (2)
1. the film recommendation method for improving multi-objective genetic algorithm is replaced based on grid and difference, it is characterised in that：This method bag
Include following steps,
S1 carries out individual UVR exposure, initialization data, and setup parameter
The individual represents the numbering of N number of film, and wherein N represents the film number for needing to recommend；Use diRepresent that i-th of needs is pushed away
The film numbering recommended, in order to applied to the TOP-N for solving the problems, such as personalized recommendation field, be numbered and be combined using N number of film, individual
Coding form is：<d1, d2, di；；；dN>, using real coding mode, coding range is the scope where film is numbered and is
Integer form；Population Size is initialized as popszie by the initialization data, and each offspring produces popsize sizes
Population；The setup parameter includes：Crossover probability Pc is set as 0.9, mutation probability Pm is 0.1, individual lengths are N, are kept
Film numbering is orderly in coding and does not repeat, using a kind of combination as the different films in N portions to be recommended；
S2 multiple target mesh generation modes initialize population
S2.1 randomly generates s times of population scale number n interim population P, i.e. population P individual amount for s × n, and s is here
One decimal system positive integer；
S2.2 calculates each individual in interim population P the numerical value fit of each target in problem to be optimized, is expressed as i-th
Numerical value of the individual on t-th of object function；
S2.3 carries out non-dominant row to each individual in interim population P according to the good and bad relation between each target value fit
The calculating of sequence layering, and the level ci where each individual is marked, it is expressed as the level where i-th of individual；
S2.4 calculates crowding distance di to each individual in population P, is expressed as i-th of individual and gathers around using population P as complete or collected works space
Squeeze the numerical value of distance；
S2.5 is in population P, using each individual non-dominated ranking level ci and crowding distance di as two targets, evaluates individual
Non-dominant relation between body come select it is optimal before (s × n)/2 individual composition candidate collection P1, wherein, n is population scale
Number；
S2.6 according in candidate collection P1 numerical value a little in each target, find maximum on each optimization direction and
Minimum value, and numerical value is normalized；
S2.7 divides the space into K grid according to the normalization maxima and minima in S2.6, and will belong to correspondence grid
Individual mark grid numbering k；
S2.8 first time individual choices：Number M of the statistics containing all grids a little, and calculate the center of these grids
Coordinate, if M is more than population scale n, n grid is selected from M grid, the Dian Yu centers of this n grid are then calculated
The distance of position, selects point composition initial population nearest in each grid, terminates the operating process；If M is advised equal to population
Mould n, then calculate the point of each grid to the distance of center, select nearest point just to constitute initial population, terminate the behaviour
Make process；If M is less than population scale n, the point of each grid is calculated to the distance of center, nearest point conduct is selected
Selection set for the first time, and continue executing with second of the individual choice of progress of Step 9；
S2.9 carries out second of individual according to the difference u between the individual amount of population scale number and first time individual choice and selected
Select, if u<W, then randomly choose u grid in all w grids for containing a little, one randomly choosed from each grid
Point merges with the point set selected for the first time before, initial population is generated, if u>W, potentially possible rare occasion, then with
Machine selects selection random point in w grid to gather as second of selection, continues the step and carries out third time selection, until reaching
Untill population scale；
S3 selection operations
Individual choice is carried out in the way of championship, i.e., randomly chooses k, k from this popszie individual<Popszie
Body, takes k to be rounded for n/2 here, and an optimum individual is chosen from this k individual；
Selection standard is with number of plies S where non-dominated rankingiWith crowding distance numerical value diIt is excellent according to non-dominant as two targets
Victory relation, is compared to each individual, and relatively winning individual is more excellent individual；
S4 crossover operations
Crossover operation uses SBX crossover operators, it is assumed that when t generations two individuals to be intersected in former generation evolutionary process are XA t、
XB t, α is that the parameter value scope that intersection is related to is 0~1, then XA t+1、XB t+1Two individuals produced for the next generation；Form is such as
Under：
<mrow>
<msubsup>
<mi>X</mi>
<mi>A</mi>
<mrow>
<mi>i</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>=</mo>
<msubsup>
<mi>&alpha;X</mi>
<mi>A</mi>
<mi>i</mi>
</msubsup>
<mo>+</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<mi>&alpha;</mi>
<mo>)</mo>
</mrow>
<msubsup>
<mi>X</mi>
<mi>B</mi>
<mi>i</mi>
</msubsup>
</mrow>
<mrow>
<msubsup>
<mi>X</mi>
<mi>B</mi>
<mrow>
<mi>i</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>=</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<mi>&alpha;</mi>
<mo>)</mo>
</mrow>
<msubsup>
<mi>X</mi>
<mi>A</mi>
<mi>i</mi>
</msubsup>
<mo>+</mo>
<msubsup>
<mi>&alpha;X</mi>
<mi>B</mi>
<mi>i</mi>
</msubsup>
</mrow>
S5 mutation operations
To the gen in evolutionary process for population PgenIn any individual pi=(pi1, pi2…piN), i ∈ { 1,2 ... n }, with general
Rate Pm participates in mutation operation：Produce a decimal r ∈ [0,1], and a random integers j ∈ [1, N]；Make pi, j=lj+r*
(uj-lj), to colony PgenEnter row variation and produce new population PgenNew；
S6 Clustering and selections replace Local Search with difference
A S6.1 generation, which is evolved in the population Pnew after terminating, n individual, each individual coordinate values in each dimension
For xim, the value of the individual in m dimensions is represented, the numerical value fit of each target in problem to be optimized is calculated, is expressed as
Numerical value of the individual on t-th of object function, wherein, i is each individual sequentially numbering；
S6.2 carries out non-dominated ranking to each individual in population Pnew according to the good and bad relation between each target value fit
The calculating of layering, the level ci where each individual of mark, and calculate the crowding distance di of each individual；
Coordinate value xims and non-dominant layering ranking ci and crowding distance di of the S6.3 in each dimension according to where individual
Numerical value, is clustered to the individual in the population using adaptive minimum spanning tree clustering algorithm, obtains k classification, be used as k
Feature close individual collections U1, U2 ..., Uk }；
In individual in each close to feature S6.4 set { U1, U2 ..., Uk }, each dimension according to where individual
The coordinate value xim and numerical value fit of each target calculates the similarity pij between individual two-by-two, wherein, i, j is each individual
Sequentially numbering；
S6.5 calculate class in two-by-two individual between similarity pij, and calculate each set in average similarity Pk；
S6.6 is calculated each average similarity Pk in gathering, and calculates the individual amount nk for needing to retain in each class；
The individual amount nk that S6.7 retains according to each class of calculating, selection non-dominant layering ranking and crowding distance numerical value are poor
(n (Uk)-nk) individual and the individual away from preferable leading surface, wherein, n (Uk) is the individual amount of k-th of set；
The solution numerical value in each individual dimension that S6.8 is selected to these random replacement in smaller range, being formed multigroup can replace
Change the alternative set for being chosen individual, and by selection of the individual in these alternative set according to non-dominated ranking and crowding distance
Mode selects optimum individual to be compared with original individual, is replaced if better than original individual, otherwise, retains original
Body；
The defect individual that S6.9 passes through these in each set of replacement operation merges with the individual collections retained, reformulates
New population Pnew, continues executing with follow-on evolutionary process；
S7 end conditions judge
If algebraically as defined in reaching obtains satisfied result, terminate and output result, otherwise turn S3 steps；
S8 enters next heredity circulation
By population PgenNewThe initial population evolved as the next generation, proceeds S3 steps；
It is the main process flow steps for the improvement multi-objective genetic algorithm replaced based on grid and difference above, uses the algorithm flow
The recommendation list of film is obtained applied to film personalized recommendation problem.
2. according to claim 1 replace the film recommendation method for improving multi-objective genetic algorithm based on grid and difference,
It is characterized in that：The method for calculating ideal adaptation angle value in population is as follows,
In this application scene, calculating for the individual relevance grade of multi-objective optimization question, it is necessary to calculate multiple object functions,
Then the number of plies where calculating each individual with the method for non-dominated ranking, angle value is virtually applicable using the number of plies as individual；
Three object functions are referred here to, refer respectively to use K user most close in user's similarity matrix to calculate N
Portion's film numbering is corresponding to be predicted scoring sum, uses K portions film most close in article scoring similarity matrix to calculate N portions electricity
The corresponding prediction scoring sum of shadow numbering and the N number of film obtained using film native similarity matrix calculating similar movies are compiled
Number corresponding prediction scoring sum；Recorded for the single portion's film prediction scoring calculated using above two mode
Come, calculating the corresponding film scoring of same film numbering next time need not then calculate, and directly read score value summation i.e.
Can；
Wherein, non-dominated ranking refers to being asked for multiple-objection optimization for famous multi-objective Optimization scholar Pareto propositions
Inscribe the winning mode between two solutions：If a solution PiCorresponding all desired values are better than another solution PjIt is corresponding all
Desired value, it is believed that PiBetter than Pj；If two solution PiAnd PjDo not possess such good and bad relation between corresponding all targets, then
It is non-dominance relation to think both；Individual is selected in such a manner, the non-dominant being selected first in whole population
Body set C1As the 1st layer, then proceed the non-dominant individual collections C for selecting to obtain in remaining all individuals2For
2 layers, go on always until finding last individual collections Cm, wherein m is total number of plies；
Wherein, the first generation evolution before, the score value that the algorithm is seen a film according to user, calculate two two users between
Similarity, and select sorted from big to small with the similarity of user to be recommended before K user；During evolution, the algorithm with
Above-mentioned K user similarity numerical value is weight, and the N portions film to be recommended of active user is calculated using the score data of these users
Prediction scoring, and sum, it is used as first aim numerical value；
Meanwhile, before first generation evolution, the score value that the algorithm is seen a film according to user, calculating is two-by-two between film
Similarity, and select sorted from big to small with the similarity of user to be recommended before K film；During evolution, the algorithm with
Above-mentioned K film similarity numerical value is weight, and the N portions film to be recommended of active user is calculated using the score data of these users
Prediction scoring, and sum, it is used as second target numerical value；
Meanwhile, before first generation evolution, the algorithm calculates similar between each film according to the attribute value between film
Spend size；During evolution, the film that the algorithm has been seen with user to be recommended scores as weight, calculates active user's
The prediction scoring of N portions film degree of correlation to be recommended, and sum, as the 3rd target value, so as to form multiple target electricity
Shadow recommends the functional value calculation of problem；
Recorded for the single portion's film prediction scoring calculated using above two mode, next time calculates same
Film is numbered corresponding film scoring and need not then calculated, and directly reads score value summation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710381920.6A CN107256241B (en) | 2017-05-26 | 2017-05-26 | Movie recommendation method for improving multi-target genetic algorithm based on grid and difference replacement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710381920.6A CN107256241B (en) | 2017-05-26 | 2017-05-26 | Movie recommendation method for improving multi-target genetic algorithm based on grid and difference replacement |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107256241A true CN107256241A (en) | 2017-10-17 |
| CN107256241B CN107256241B (en) | 2021-06-25 |
Family
ID=60027750
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710381920.6A Active CN107256241B (en) | 2017-05-26 | 2017-05-26 | Movie recommendation method for improving multi-target genetic algorithm based on grid and difference replacement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107256241B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107203590A (en) * | 2017-04-24 | 2017-09-26 | 北京工业大学 | Method is recommended based on the personalized film for improving NSGA II |
| CN108280181A (en) * | 2018-01-23 | 2018-07-13 | 成都信达智胜科技有限公司 | The immediate processing method of network data |
| CN108804586A (en) * | 2018-05-27 | 2018-11-13 | 北京工业大学 | The personalized film that fusion grid deposits the multiple-objection optimization of dominant strategy recommends method |
| CN111104764A (en) * | 2020-01-21 | 2020-05-05 | 湖南科技大学 | Structured grid optimization division method for alternating current motor rotor conducting bar thermal analysis model |
| CN111143685A (en) * | 2019-12-30 | 2020-05-12 | 第四范式（北京）技术有限公司 | Recommendation system construction method and device |
| WO2020210974A1 (en) * | 2019-04-16 | 2020-10-22 | 江南大学 | High-quality pattern mining model and method based on improved multi-objective evolutionary algorithm |
| CN114245316A (en) * | 2022-01-24 | 2022-03-25 | 浙江正泰中自控制工程有限公司 | UWB positioning-based base station deployment optimization method and system |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101489298A (en) * | 2009-01-08 | 2009-07-22 | 东南大学 | Discrete speed cross-layer power distribution method based on mobile multicast system |
| CN102622649A (en) * | 2012-03-07 | 2012-08-01 | 南京邮电大学 | Comentropy-based improved evolutionary multi-objective optimization method |
| US20130110751A1 (en) * | 2011-10-31 | 2013-05-02 | Taif University | Computational device implemented method of solving constrained optimization problems |
| CN104880991A (en) * | 2015-03-18 | 2015-09-02 | 重庆大学 | Energy-efficiency-oriented multi-step numerical control milling process parameter multi-objective optimization method |
| CN105337315A (en) * | 2015-10-21 | 2016-02-17 | 温州大学 | Wind-light-storage battery supplementary independent micro power grid high dimension multi-target optimization configuration |
| CN105868281A (en) * | 2016-03-23 | 2016-08-17 | 西安电子科技大学 | Location-aware recommendation system based on non-dominated sorting multi-target method |
| CN106651100A (en) * | 2016-10-12 | 2017-05-10 | 华南理工大学 | Internet-of-Vehicles optimal vehicle-mounted monitoring point-based air quality evaluation system and method |
-
2017
- 2017-05-26 CN CN201710381920.6A patent/CN107256241B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101489298A (en) * | 2009-01-08 | 2009-07-22 | 东南大学 | Discrete speed cross-layer power distribution method based on mobile multicast system |
| US20130110751A1 (en) * | 2011-10-31 | 2013-05-02 | Taif University | Computational device implemented method of solving constrained optimization problems |
| CN102622649A (en) * | 2012-03-07 | 2012-08-01 | 南京邮电大学 | Comentropy-based improved evolutionary multi-objective optimization method |
| CN104880991A (en) * | 2015-03-18 | 2015-09-02 | 重庆大学 | Energy-efficiency-oriented multi-step numerical control milling process parameter multi-objective optimization method |
| CN105337315A (en) * | 2015-10-21 | 2016-02-17 | 温州大学 | Wind-light-storage battery supplementary independent micro power grid high dimension multi-target optimization configuration |
| CN105868281A (en) * | 2016-03-23 | 2016-08-17 | 西安电子科技大学 | Location-aware recommendation system based on non-dominated sorting multi-target method |
| CN106651100A (en) * | 2016-10-12 | 2017-05-10 | 华南理工大学 | Internet-of-Vehicles optimal vehicle-mounted monitoring point-based air quality evaluation system and method |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107203590A (en) * | 2017-04-24 | 2017-09-26 | 北京工业大学 | Method is recommended based on the personalized film for improving NSGA II |
| CN108280181A (en) * | 2018-01-23 | 2018-07-13 | 成都信达智胜科技有限公司 | The immediate processing method of network data |
| CN108804586A (en) * | 2018-05-27 | 2018-11-13 | 北京工业大学 | The personalized film that fusion grid deposits the multiple-objection optimization of dominant strategy recommends method |
| CN108804586B (en) * | 2018-05-27 | 2021-08-06 | 北京工业大学 | Multi-objective optimization personalized movie recommendation method fusing grid optimization strategy |
| WO2020210974A1 (en) * | 2019-04-16 | 2020-10-22 | 江南大学 | High-quality pattern mining model and method based on improved multi-objective evolutionary algorithm |
| CN111143685A (en) * | 2019-12-30 | 2020-05-12 | 第四范式（北京）技术有限公司 | Recommendation system construction method and device |
| CN111143685B (en) * | 2019-12-30 | 2024-01-26 | 第四范式（北京）技术有限公司 | Commodity recommendation method and device |
| CN111104764A (en) * | 2020-01-21 | 2020-05-05 | 湖南科技大学 | Structured grid optimization division method for alternating current motor rotor conducting bar thermal analysis model |
| CN114245316A (en) * | 2022-01-24 | 2022-03-25 | 浙江正泰中自控制工程有限公司 | UWB positioning-based base station deployment optimization method and system |
| CN114245316B (en) * | 2022-01-24 | 2024-06-04 | 浙江正泰中自控制工程有限公司 | Base station deployment optimization method and system based on UWB positioning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107256241B (en) | 2021-06-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107256241A (en) | The film recommendation method for improving multi-objective genetic algorithm is replaced based on grid and difference | |
| CN106844637A (en) | Method is recommended based on the film for just giving cluster to prune improvement multi-objective genetic algorithm | |
| Li et al. | Bi-goal evolution for many-objective optimization problems | |
| Kuo et al. | Application of particle swarm optimization to association rule mining | |
| CN104794367B (en) | Medical treatment resource scoring based on hidden semantic model is with recommending method | |
| Plata-González et al. | Evolutionary-based tailoring of synthetic instances for the Knapsack problem | |
| CN107391659A (en) | A kind of citation network academic evaluation sort method based on credit worthiness | |
| CN107368540A (en) | The film that multi-model based on user's self-similarity is combined recommends method | |
| CN109740722A (en) | A kind of network representation learning method based on Memetic algorithm | |
| CN106407379A (en) | Hadoop platform based movie recommendation method | |
| CN115374106A (en) | Intelligent data grading method based on knowledge graph technology | |
| Czajkowski et al. | Steering the interpretability of decision trees using lasso regression-an evolutionary perspective | |
| CN112734510B (en) | Commodity recommendation method based on fusion improvement fuzzy clustering and interest attenuation | |
| CN107133321A (en) | The analysis method and analytical equipment of the search attribute of the page | |
| CN107562871A (en) | Social network influence power user based on social capital sorting out value has found method | |
| Duan et al. | MOBA game item recommendation via relation-aware graph attention network | |
| CN104462480B (en) | Comment big data method for digging based on typicalness | |
| Lange et al. | Frequency-aware similarity measures: why Arnold Schwarzenegger is always a duplicate | |
| Xu et al. | Discover dependency pattern among attributes by using a new type of nonlinear multiregression | |
| López-Herrera et al. | Applying multi-objective evolutionary algorithms to the automatic learning of extended Boolean queries in fuzzy ordinal linguistic information retrieval systems | |
| Daghouri et al. | Information system performance evaluation and optimization using AHP and TOPSIS: Construction industry case | |
| CN106096718A (en) | A kind of normalization PSO Stochastic Optimization Algorithms and collaborative filtering recommending method | |
| Sen | Complexities of social networks: A Physicist's perspective | |
| Wattanapornprom et al. | Application of estimation of distribution algorithms for solving order acceptance with weighted tardiness problems | |
| Yang et al. | A fast top-k group skyline query method based on skyline layer |
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 |