CN104834969A - Film evaluation prediction method and system - Google Patents

Abstract

The invention discloses a film evaluation prediction method and system. The method comprises the steps of firstly selecting characteristics of film data set for training through a sequential forward selection method, and conducting de-noising, pre-processing and extracting of a characteristic vector; secondly collecting rating information corresponding to each film, and conducting normalization processing of all the ratings to obtain a rating distribution vector; thirdly, training a maximum interval rating distribution model based on the film characteristic vector and the rating distribution vector, and conducting deformation optimization to obtain a final parameter model for film rating prediction; and fourthly, extracting a characteristic vector of a film to be released, obtaining a vector by using the obtained parameter model for rating distribution prediction, conducting normalization processing of the vector, and obtaining the predicted result of the rating distribution of audience after the film is released. According the invention, the model for film rating distribution prediction can be quickly and effectively trained, and the accuracy of prediction is substantially higher than that by using existing methods.

Description

A kind of film method of evaluation and forecast and system thereof

Technical field

The present invention relates to pattern-recognition and machine learning, particularly a kind of film method of evaluation and forecast and system thereof.

Background technology

Film evaluation and foreca is the master data information of not showing film according to, and as performer, director, distributing and releasing corporation, budget etc., after estimating this movie show, the public is to its film scoring distribution.Cinematic industry is that global a, scale reaches 10,000,000,000 dollars of mature industries.Thousands of portions film all can be shown every year in the whole world, and successfully film is very few.For film investor and publisher, growing shooting cost and the competitive environment be growing more intense make the investment risk of film greatly increase; For spectators, the advertisement of covering the sky and the earth and marketing methods make to select a film really liked and be worth seeing and become more and more difficult.Therefore, at movie show in earlier stage, or even preparing the stage, reliably predicting that the view of the public to this film becomes extremely important.It makes investment with can helping investors' rationality, and spectators also can be helped to select the film really liked and be worth seeing.

Summary of the invention

Goal of the invention: for problems of the prior art, the present invention proposes a kind of film method of evaluation and forecast, drastically increases optimal speed and precision of prediction.The invention allows for a kind of film evaluation and foreca system.

Technical scheme: the present invention proposes a kind of film method of evaluation and forecast, comprises the steps:

1) obtaining the cinematic data collection for training, selecting feature by the system of selection of sequence forward direction;

2) to step 1) in the feature of all cinematic data collection that obtains carry out denoising, pre-service;

3) from step 2) extract proper vector in all movie features of obtaining;

4) collect the score information that every portion film is corresponding, all scorings are done normalized and obtain distribution vector of marking;

5) based on step 3) in the movie features vector sum step 4 that obtains) in the scoring distribution vector that obtains, training largest interval scoring distributed model;

6) to step 5) the scoring distributed model that obtains carries out distortion optimization, obtains finally marking the parameter model of forecast of distribution for reflecting front filmgoer;

7) proper vector of not showing the cinematic data of pending scoring forecast of distribution is extracted, and use step 6) in the scoring forecast of distribution parameter model that obtains calculate a vector, finally be normalized this vector, what after obtaining this movie show, spectators marked and distribute predicting the outcome.

Described step 1) feature selection approach is the system of selection of sequence forward direction, the feature of selection comprises show time, director, performer, writes a play, dubs in background music, film types, distributing and releasing corporation, duration, language, show country, budget.

Described step 2) denoising is carried out to cinematic data and pretreated concrete grammar is: setting threshold value θ, when the occurrence number of feature value is greater than θ, this eigenwert is effective; When being less than θ, this eigenwert is merged into eigenwert other.

Described step 3) extract the concrete grammar of proper vector and be: for discrete data, each value of feature is split as the independent feature of one dimension; For continuous data, calculate maximal value and the minimum value of this feature value of data centralization, and all values are deducted minimum value simultaneously, then divided by maximal value.

Described step 4) by scoring number, normalized done by grading system to all scorings of film obtain distribution vector of marking, namely by this film scoring number that is a certain grade divided by the total number of persons of marking to this film.

Described step 5) use each grading system of sigmoid function representation, its representation is:

$f_j\left(x\right)=\frac{1}{1+e^{-(w_{j}x+b_j)}}---\left(1\right)$

Meanwhile, utilize largest interval method establishing target function, the minimum value of getting objective function builds scoring distributed model, and its objective function is expressed as:

J＝∑ _ijl _ij(w _j,b _j)+λh(W,b) (2)

In formula (1) (2), x represents film, and i represents i-th movie samples, and j represents a jth grading system, W and b is parameter to be optimized, w _jand b _jbe respectively W and b jth row, d is true scoring, and λ is the parameter of artificial setting, l _ij(w _j, b _j)=(d _ij-f _ij) ²,

Described step 6) objective function that formula 2 represents is out of shape, former objective function is indirectly optimized by the upper bound optimizing former objective function, and use kernel method on this basis, obtain finally for reflecting the parameter model of front film scoring forecast of distribution, its objective function is expressed as:

J′＝∑ _ijl′ _ij(w _j,b _j)+λh(W,b) (3)

Wherein, W and b is parameter to be optimized, w _jand b _jbe respectively W and b jth row, d is true scoring, and λ is the parameter of artificial setting, $l_{\mathrm{ij}}^{\′}={(-\frac{1}{d_{\mathrm{ij}}}-(w_j{x}_i+b_j))}^2,h(W,b)=\frac{1}{2}{\left|\right|W\left|\right|}^2.$

The present invention also proposes a kind of film evaluation and foreca system, comprises characteristic vector pickup module, scoring distribution statistics module, Application of Parametric Model Forecasting module and appraisal result prediction module; Wherein characteristic vector pickup module is for selecting the feature of cinematic data collection, and carries out denoising, pre-service and characteristic vector pickup to the feature obtained; Scoring distribution statistics module for collecting score information corresponding to every portion film, and is done normalized to all scorings and is obtained distribution vector of marking; Application of Parametric Model Forecasting module trains scoring forecast of distribution parameter model based on the proper vector of training cinematic data collection with scoring distribution vector; Appraisal result prediction module does not show the scoring distribution of film based on the scoring forecast of distribution Application of Parametric Model Forecasting of the proper vector and training of not showing film.

Beneficial effect: the method and system that the present invention proposes can train the model for the prediction of film scoring distributed feed-back quickly and efficiently, utilize the method to carry out the prediction of scoring distributed feed-back to new film of not showing, existing method is significantly led in its accuracy.

Accompanying drawing explanation

Fig. 1 is film method of evaluation and forecast process flow diagram of the present invention;

Fig. 2 is film of the present invention scoring distribution example.

Embodiment

Below in conjunction with specific embodiment, illustrate the present invention further, these embodiments should be understood only be not used in for illustration of the present invention and limit the scope of the invention, after having read the present invention, the amendment of those skilled in the art to the various equivalent form of value of the present invention has all fallen within the application's claims limited range.

As shown in Figure 1, film method of evaluation and forecast comprises the steps:

Step S1, the cinematic data collection for training is obtained by the open API of IMDb, feature is selected by the system of selection of sequence forward direction, comprise show time, director, performer (leading role, supporting role etc.), write a play, dub in background music, film types, distributing and releasing corporation, duration, language, show the information relevant to film such as country, budget, amount to totally 8462.

The system of selection of sequence forward direction is a kind of method of feature selecting, it specifically can with reference to Kittler J.Featureselection and extraction [J] .Handbook of pattern recognition and image processing, 1986:59-83.

Step S2, carries out denoising, pre-service to the feature of all cinematic data collection obtained in step S1.Because cinematic data is too sparse, a lot of feature value (performer as nameless in some, director, playwright, screenwriter, distributing and releasing corporation etc.) occurrence number is very few, such meeting impact prediction effect, therefore by setting threshold value θ, when a certain feature value occurrence number is greater than θ effectively, eigenwert other is merged into when being less than θ.Concrete, to director, performer, write a play, dub in background music, film types, distributing and releasing corporation, language, show the features such as country, setting threshold value θ is 10, and namely occurrence number is more than or equal to the individual features value of 10 just effectively, otherwise replace by other value, wherein θ value is determined by the mode of cross validation.

Step S3, extracts proper vector in all movie features obtained from step S2.For discrete type feature (as director, performer, write a play, dub in background music, film types, distributing and releasing corporation, language, show country etc.), each value of feature is split as the independent feature of one dimension, as using each performer as one-dimensional characteristic, for the film of being taken part in a performance by this performer, its characteristic of correspondence dimension value is 1, otherwise is 0; For continuous type feature, calculate maximal value and the minimum value of this feature value of data centralization, and all values are deducted minimum value simultaneously, then divided by maximal value, ensure that its value is between [0,1] with this.Finally obtaining proper vector is 2045 dimensions.

Step S4, give every portion film one scoring distribution vector, but not the average mark that single, more completely can express spectators like this to the view of film and viewpoint, its concrete methods of realizing is as follows: all scoring records obtaining the every portion film collected from step S1 from Netflix website, are designated as wherein I _irepresent i-th movie samples, S _irepresent the scoring set to i-th film; To every portion film, scoring number is done normalized by grading system and is formed scoring distribution by all scorings of this film, and namely mark as the number of a certain grade is divided by the total number of persons of marking to this film to this film, effect as shown in Figure 2.

In this step, S _irepresent all scoring set of i-th film, known grading system number M simultaneously, thus obtain category set L={L ₁, L ₂..., L _m, make S _ijrepresent that all scorings are the scoring set of a jth grade, thus obtain here Y _ijrepresent Y _ia jth element (i.e. the description degree of a jth grading system of i-th sample film).

Step S5, uses movie features vector and scoring distribution thereof as training set, training largest interval scoring distributed model.The each grading system of this model sigmoid function representation, utilizes largest interval method establishing target function simultaneously, carrys out Optimization Solution scoring distributed model by the minimum value of getting objective function;

In this step, for a jth grading system, the representation of its sigmoid function is:

$f_j\left(x\right)=\frac{1}{1+e^{-(w_{j}x+b_j)}}---\left(1\right)$

Utilize largest interval method establishing target function, objective function is expressed as simultaneously:

J＝∑ _ijl _ij(w _j,b _j)+λh(W,b) (2)

In formula (1) (2), x represents film, and i represents i-th movie samples, and j represents a jth grading system, W and b is parameter to be optimized, and d is true scoring, and λ is the parameter of artificial setting, l _ij(w _j, b _j)=(d _ij-f _ij) ², $h(W,b)=\frac{1}{2}{\left|\right|W\left|\right|}^2.$

Step S6, because sigmoid function is not linear function, comparatively difficult to the direct Optimization Solution of objective function J, and cannot kernel function be used, therefore we are out of shape objective function J, indirectly optimize former objective function by the upper bound of optimization object function again, thus train the parameter model obtaining may be used for reflecting front film scoring forecast of distribution, drastically increase optimal speed and precision of prediction simultaneously.

In this step, by l _ijbe deformed into following form carry it into J to obtain:

J′＝∑ _ijl′ _ij(w _j,b _j)+λh(W,b) (3)

Its meaning of parameters is identical with (1) (2).

Due to l ' >=l, therefore J ' >=J, so can indirectly optimize J by optimization J ', (optimization method specifically can with reference to Matilde S ' anchez Fern ' andez, Mario de Prado-Cumplido, Jer ' onimoArenas-Garc ' ia, and Fernando P ' erez-Cruz.SVM multiregression for nonlinearchannel estimation in multiple-input multiple-output systems.IEEE Trans.SignalProcessing, 52 (8): 2298 – 2307,2004.)

Step S7, by the cinematic data of pending scoring forecast of distribution through the denoising of step S2 and pre-service, then the method for step S3 is used to extract the proper vector of film, then the parameter model of training out in step S5, S6 is used to calculate a vector, then vector is normalized, the vector finally obtained be to spectators after this movie show mark distribution predicting the outcome.

What the present invention taked is 10 times of cross validation methods, experimental result have employed common several indexs for measuring two distribution similarity, comprise Euclidean, Sorensen, Squared, KL divergence, Intersection, Fidelity, wherein more novel obvious results fruit is better for front four indexs, and the larger explanation effect of latter two index is better.Contrast algorithm comprises MSVR (the reference Matilde S ' anchez Fern ' andez through aftertreatment, Mario dePrado-Cumplido, Jer ' onimo Arenas-Garc ' ia, and Fernando P ' erez-Cruz.SVMmultiregression for nonlinear channel estimation in multiple-input multiple-outputsystems.IEEE Trans.Signal Processing, 52 (8): 2298 – 2307, 2004.), the SVR of aftertreatment, LBFGS_LLD is (with reference to X.Geng and R.Ji.Label Distribution Learning.In:Proceedingsof the 2013International Conference on Data Mining Workshops (ICDMW ' 13), Dallas, TA, 2013, pp.377-383.), IIS_LLD is (with reference to X.Geng, K.Smith-Miles, Z.-H.Zhou.Facial Age Estimation by Learning from Label Distributions.In:Proceedingsof the 24th AAAI Conference on Artificial Intelligence (AAAI ' 10), Atlanta, GA, 2010, pp.451-456.), CPNN, AA-KNN.Experimental result shows, the method that we propose is all much effective than additive method in any one evaluation index.

Euclidean

Sorensen

Squared

KL

Intersection

Fidelity

OurMethod

.1587th0026

.1564th0026

.0887th0026

.0921±.0035

.8436±.0035

.9764±.0035

SVR

.1734±.0035

.1723±.0023

.1040±.0023

.1059±.0023

.8277±.0023

.9722±.0023

MSVR

.1843±.0023

.1814±.0023

.1084±.0023

.1073±.0030

.8186±.0030

.9710±.0010

LBFGS_LLD

.1853_LLD10

.1814_LLD10

.1176_LLD10

.1265_LLD10

.8186_LLD10

.9683_L0012

IIS_LLD

.1866LD0012

.1828LD0012

.1195LD0054

.1288LD0054

.8172LD0044

.9676LD0044

CPNN

.2209LD0044

.2153LD0044

.1625LD0044

.1826LD0044

.7847LD0150

.9551LD0150

AA-KNN

.1917ND0150

.1899ND0150

.1246ND0150

.1274ND0150

.8101ND0150

.9664ND0150

Table 1

Claims (8)

1. a film method of evaluation and forecast, is characterized in that, comprises the steps:

1) obtain the cinematic data collection for training, and select feature by the system of selection of sequence forward direction;

2) to step 1) in select after the movie features that obtains carry out denoising, pre-service;

3) from step 2) extract proper vector in the movie features that obtains after process;

4) collect the score information that every portion film is corresponding, all scorings are done normalized and obtain distribution vector of marking;

5) based on step 3) in the movie features vector sum step 4 that obtains) in the scoring distribution vector that obtains, training largest interval scoring distributed model;

6) to step 5) the scoring distributed model that obtains carries out distortion optimization, obtains finally marking the parameter model of forecast of distribution for reflecting front filmgoer;

7) proper vector of not showing the cinematic data of pending scoring forecast of distribution is extracted, and use step 6) in the scoring forecast of distribution parameter model that obtains calculate a vector, finally be normalized this vector, what after obtaining this movie show, spectators marked and distribute predicting the outcome.

2. film evaluation and foreca system as claimed in claim 1, is characterized in that, described step 1) feature selected comprises show time, director, performer, writes a play, dubs in background music, film types, distributing and releasing corporation, duration, language, show country, budget.

3. film evaluation and foreca system as claimed in claim 1, is characterized in that, described step 2) denoising is carried out to cinematic data and pretreated concrete grammar is: setting threshold value θ, when the occurrence number of feature value is greater than θ, this eigenwert is effective; When being less than θ, this eigenwert is merged into eigenwert other.

4. film evaluation and foreca system as claimed in claim 1, is characterized in that, described step 3) concrete grammar that extracts proper vector is: for discrete data, each value of feature is split as the independent feature of one dimension; For continuous data, calculate maximal value and the minimum value of this feature value of data centralization, and all values are deducted minimum value simultaneously, then divided by maximal value.

5. film evaluation and foreca system as claimed in claim 1, it is characterized in that, described step 4) by scoring number, normalized done by grading system to all scorings of film obtain distribution vector of marking, namely by this film scoring number that is a certain grade divided by the total number of persons of marking to this film.

6. film evaluation and foreca system as claimed in claim 1, is characterized in that, described step 5) each grading system of scoring distributed model sigmoid function representation, its representation is:

$f_j\left(x\right)=\frac{1}{1+e^{-(w_{j}x+b_j)}}---\left(1\right)$

Meanwhile, utilize largest interval method establishing target function, carry out Optimization Solution scoring distributed model by the minimum value of getting objective function, its objective function is expressed as:

J＝∑ _ijl _ij(w _j,b _j)+λh(W,b) (2)

In formula (1) (2), x represents film, and i represents i-th movie samples, and j represents a jth grading system, W and b is parameter to be optimized, w _jand b _jbe respectively W and b jth row, d is true scoring, and λ is the parameter of artificial setting, $l_{\mathrm{ij}}(w_j,b_j)={(d_{\mathrm{ij}}-f_{\mathrm{ij}})}^2,h(W,b)=\frac{1}{2}{\left|\right|W\left|\right|}^2.$

7. film evaluation and foreca system as claimed in claim 6, it is characterized in that, described step 6) objective function that formula 2 represents is out of shape, former objective function is indirectly optimized by the upper bound optimizing former objective function, and use kernel method on this basis, obtain finally for reflecting the parameter model of front film scoring forecast of distribution, its objective function is expressed as:

J′＝∑ _ijl′ _ij(w _j,b _j)+λh(W,b) (3)

Wherein, W and b is parameter to be optimized, w _jand b _jbe respectively W and b jth row, d is true scoring, and λ is the parameter of artificial setting, $l_{\mathrm{ij}}^{\′}={(-\frac{1}{d_{\mathrm{ij}}}-(w_j{x}_i+b_{\text{j}}))}^2,h(W,b)=\frac{1}{2}{\left|\right|W\left|\right|}^2.$

8. a film evaluation and foreca system, is characterized in that, comprises characteristic vector pickup module, scoring distribution statistics module, Application of Parametric Model Forecasting module and appraisal result prediction module; Wherein characteristic vector pickup module is for selecting the feature of cinematic data collection, and carries out denoising, pre-service and characteristic vector pickup to the feature obtained; Scoring distribution statistics module for collecting score information corresponding to every portion film, and is done normalized to all scorings and is obtained distribution vector of marking; Application of Parametric Model Forecasting module trains scoring forecast of distribution parameter model based on the proper vector of training cinematic data collection with scoring distribution vector; Appraisal result prediction module does not show the scoring distribution of film based on the scoring forecast of distribution Application of Parametric Model Forecasting of the proper vector and training of not showing film.