Movie Box Office Prediction Based on Multi-Model Ensembles
<p>Distribution of total box office (10,000).</p> "> Figure 2
<p>Screens in cinema lines (block) fitting curve.</p> "> Figure 3
<p>Release year distribution pie chart.</p> "> Figure 4
<p>Which day of the week is the release date?</p> "> Figure 5
<p>Stacking algorithm architecture diagram.</p> "> Figure 6
<p>Importance of features of LightGBM model.</p> ">
Abstract
:1. Introduction
2. Related Research Work
3. Establishment of Prediction Model
3.1. Data Collection
3.2. Data Feature Processing and Analysis
3.2.1. Chinese Micro-Data Processing
3.2.2. Epidemic Data Processing
3.2.3. Baidu Search Index Data Processing
3.2.4. Director, Actor, and Other Characteristic Data Processing
3.2.5. Feature Processing of Release Date Data
3.2.6. Processing of Other Data Features
3.3. Model Establishment
- Step 1:
- Firstly, the processed data are divided into a training set and test set in a ratio of 9:1.
- Step 2:
- Establish the model with LightGBM, and obtain the importance of each feature in the LightGBM model. Delete the features with importance less than three, and retain the Chinese national micro-data, epidemic data, and Baidu search index features.
- Step 3:
- Build the first layer-based learner model, and build the XGBoost, LightGBM, CatBoost, GBDT, random forest and SVR models based on deleted features. During the construction of the first layer-based learner model, the training set obtained by Step1 is divided into five folds for training by adopting the method of five-fold cross-validation, and the goodness of fit of each training is obtained.
- Step 4:
- Build the second layer meta-learner model. The second layer meta-learner selects a multiple linear regression model and connects the prediction results of the training set of each base learner in the first layer as the training set of the meta-learner for the training of the meta-learner. The goodness of fit of each fold prediction model of each base learner in the first layer is used to simulate the “precision”, and the weighted average of the prediction results of the test set is used as the test set of the meta-learner.
3.4. Experimental Results and Analysis
3.4.1. Evaluation Indicators
- (1)
- For R2, the range is [0, 1]. If the result is 0, it indicates that the model fitting effect is poor. If the result is 1, the model fits perfectly. The larger R2 value is, the better the model fitting effect is.
- (2)
- MAPE represents the relative error between the predicted value and the real value. The smaller the MAPE value is, the higher the prediction accuracy is and the smaller the error is.
- (3)
- MAE represents the average absolute error between the predicted value and the real value. The smaller the MAE value is, the better the model effect is.
- (4)
- MSE is the mean of the sum of squares of errors between the predicted value and the true value. The smaller the MSE value is, the better the model effect is.
- (5)
- RMSE is the square root of MSE, and the smaller the RMSE value, the better the model effect.
3.4.2. Model Comparison
- (1)
- XGBoost is based on the gradient boosting algorithm. Compared with other implementations of the gradient boosting algorithm, the execution speed of XGBoost is very fast. The hyperparameters of the XGBoost algorithm used in this paper are as follows: n_estimators = 200, learning_rate = 0.08, subsample = 0.8, max_depth = 5, and colsample_bytree = 0.9.
- (2)
- LightGBM, like XGBoost, is an efficient implementation of GBDT, with a faster training speed, lower memory consumption, and other advantages. The hyperparameters of the LightGBM algorithm used in this paper are as follows: boosting_type = ‘gbdt’, learning_rate = 0.1, min_child_samples = 20, n_estimators = 200, and num_leaves = 31.
- (3)
- CatBoost, which is also an improved implementation under the framework of the GBDT algorithm, is mainly characterized by efficient and reasonable processing of category features. The hyperparameters of CatBoost algorithm used in this paper are as follows: iterations = 200 and loss_function = ‘RMSE’.
- (4)
- GBDT, which is a representative algorithm of the boosting algorithm, is an optimal model obtained by iterative training with a weak learner (decision tree) and has the advantages of good training effect and difficulty in over-fitting. The hyperparameters of the GBDT algorithm used in this paper are as follows: max_depth = 5, n_estimators = 200, subsample = 0.8, and learning_rate = 0.1.
- (5)
- RF is the representative algorithm of the bagging algorithm. Sampling is carried out in the way of putting back. Sub-models are established with the sampled samples and trained on the sub-models. The hyperparameters of the RF algorithm used in this paper are as follows: n_estimators = 200 and bootstrap = True.
- (6)
- SVR is an application of the support vector machine (SVM) to regression problems. The basic idea is to fit the optimal hyperplane. The hyperparameter of the SVR algorithm used in this paper is kernel = ‘RBF’.
- (7)
- Voting can help improve the generalization ability of models and reduce the error rate of models. For the regression model, voting’s final prediction is an average of the results predicted by multiple other regression models. The single models used in the voting model in this article are XGBoost, LightGBM, CatBoost, GBDT, RF, and SVR, whose hyperparameters are shown above.
- (8)
- Averaging means the results of multiple single models are weighted and averaged. We determined the weights based on the R2 values of each model. The single models used by the averaging model in this paper are XGBoost, LightGBM, CatBoost, GBDT, RF, and SVR, whose hyperparameters are described above.
- (9)
- For the ordinary stacking model, the first layer used in this paper includes XGBoost, LightGBM, CatBoost, GBDT, RF, and SVR, and the second layer of the meta-learner is LR. The hyperparameters of the base learner are as shown above. The meta-learner uses default parameters.
- (10)
- For the enhanced stacking model, the first layer used in this paper includes XGBoost, LightGBM, CatBoost, GBDT, RF, and SVR, and the second layer of the meta-learner is LR. The hyperparameters of the base learner are as shown above. The meta-learner uses default parameters.
3.5. Example of Partial Box Office Prediction
3.6. Correlation between Epidemic Factors and Box Office
4. Summary
Author Contributions
Funding
Institutional Review Board Statement
Informed Consent Statement
Data Availability Statement
Acknowledgments
Conflicts of Interest
References
- Ma, J.; Wang, Y.; Niu, X.; Jiang, S.; Liu, Z. A comparative study of mutual information-based input variable selection strategies for the displacement prediction of seepage-driven landslides using optimized support vector regression. Stoch. Environ. Res. Risk Assess. 2022, 36, 1–21. [Google Scholar] [CrossRef]
- Ma, J.; Niu, X.; Tang, H.; Wang, Y.; Wen, T.; Zhang, J. Displacement prediction of a complex landslide in the Three Gorges Reservoir Area (China) using a hybrid computational intelligence approach. Complexity 2020, 2020, 2624547. [Google Scholar] [CrossRef]
- Wang, J.; Yan, S. Research on the commercial value evaluation model of Chinese film copyright. Contemp. Cine. 2015, 11, 73–80. [Google Scholar]
- Litman, B.R.; Kohl, L.S. Predicting Financial Success of Motion Pictures: The 80s’ Experience. J. Media Econ. 1989, 11, 35–50. [Google Scholar] [CrossRef]
- Chang, B.H.; Ki, E.J. Devising a Practical Model for Predicting Theatrical Movie Success: Focusing on the Experience Good Property. J. Media Econ. 2005, 18, 247–269. [Google Scholar] [CrossRef]
- Zhang, W.; Skiena, S. Improving movie gross prediction through news analysis. In Proceedings of the 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology, Milan, Italy, 15–18 September 2009; Volume 1, pp. 301–304. [Google Scholar]
- Han, Z.; Yuan, B.; Chen, Y.; Zhao, N.; Duan, D. An effective early movie box office prediction model based on GBRT. Comput. Appl. Res. 2018, 35, 410–416. [Google Scholar]
- Li, J.; Wang, S. Domestic movie box office predict based on grey correlation analysis and BP algorithm. Electron. World 2018, 24, 18–19. [Google Scholar] [CrossRef]
- Zheng, J.; Zhou, S. Prediction modeling of movie box office based on neural network. Comput. Appl. 2014, 34, 742–748. [Google Scholar]
- Ru, Y.; Li, B.; Chai, J.; Liu, J. A box office prediction model based on deep learning. J. Commun. Univ. China (Sci. Technol.) 2019, 26, 27–32. [Google Scholar] [CrossRef]
- Wang, Z.; Xu, M. Analysis of influencing factors of movie box office—A study based on Logit model. Inq. Into Econ. Issues 2013, 11, 96–102. [Google Scholar]
- Zhao, X.; Gao, F. Research on the influencing factors of box office of main melody films in China. Movie Lit. 2020, 20, 3–7. [Google Scholar]
- Chen, H. Statistical test of influencing factors of world movie box office. Stat. Decis. 2019, 2, 5. [Google Scholar]
- Bae, G.; Kim, H. The impact of movie titles on box office success. J. Bus. Res. 2019, 103, 100–109. [Google Scholar] [CrossRef]
- Hua, R.; Wang, S.; Xu, X. Research on the influencing factors of Chinese film box office. Stat. Decis. 2019, 35, 97–100. [Google Scholar]
- Pei, P.; Jiang, Y. Analysis of the influencing factors of domestic and foreign movie box office. Co-Oper. Econ. Sci. 2014, 2, 18–21. [Google Scholar]
- Jiang, Z.; Wu, Z.; Sun, W. The Impact of Online Word-of-Mouth on the Box Office of Domestic and Imported Animated Films: Taking 2009–2018 as an Example. Chin. J. J. Commun. 2020, 42, 16. [Google Scholar]
- Du, J.; Xu, H.; Huang, X. Box office prediction based on microblog. Expert Syst. Appl. 2014, 41, 1680–1689. [Google Scholar] [CrossRef]
- Basuroy, S.; Ravid, S.C.A. How Critical Are Critical Reviews? The Box Office Effects of Film Critics, Star Power, and Budgets. J. Mark. 2003, 67, 103–117. [Google Scholar] [CrossRef]
- Gaenssle, S.; Budzinski, O.; Astakhova, D. Conquering the box office: Factors influencing success of international movies in Russia. Rev. Netw. Econ. 2018, 17, 245–266. [Google Scholar] [CrossRef]
- Lee, S.; Choeh, J.Y. The interactive impact of online word-of-mouth and review helpfulness on box office revenue. Manag. Decis. 2018, 56, 849–866. [Google Scholar] [CrossRef]
- Baek, H.; Oh, S.; Yang, H.D.; Ahn, J. Electronic word-of-mouth, box office revenue and social media. Electron. Commer. Res. Appl. 2017, 22, 13–23. [Google Scholar] [CrossRef]
- Sun, L.; Zhai, X.; Yang, H. Event marketing, movie consumers’ willingness and box office revenue. Asia Pac. J. Mark. Logist. 2020, 33, 622–646. [Google Scholar] [CrossRef]
- Sharda, R.; Delen, D. Predicting box-office success of motion pictures with neural networks. Expert Syst. Appl. 2006, 30, 243–254. [Google Scholar] [CrossRef]
- Biramane, V.; Kulkarni, H.; Bhave, A.; Kosamkar, P. Relationships between classical factors, social factors and box office collections. In Proceedings of the 2016 International Conference on Internet of Things and Applications (IOTA), Pune, India, 22–24 January 2016; pp. 35–39. [Google Scholar]
Year | Date | Converted Date to Number | Screens in Cinema Lines (Block) |
---|---|---|---|
2015 | 31 December 2015 | 42,369 | 31,600 |
2017 | 31 December 2017 | 43,100 | 50,776 |
2019 | 31 December 2019 | 43,830 | 69,787 |
Movie Name | Release Date | Convert the Release Date to Number | Predicted Value of Screen (Block) in Cinema Line |
---|---|---|---|
The Battle at Lake Changjin | 30 September 2021 | 44,469 | 85,346.67536984 |
Wolf Warriors 2 | 27 July 2017 | 42,943 | 45,625.00855624 |
Hi, Mom | 12 February 2021 | 44,239 | 79,071.3646575 |
Ne Zha | 26 July 2019 | 43,672 | 63,895.13502012 |
The Wandering Earth | 5 February 2019 | 43,501 | 59,454.47257928 |
Date | Newly Confirmed Cases | Existing Confirmed Cases | New Asymptomatic Infection | Accumulated Cured Discharged Cases | Cumulative Death Cases | Reported Cumulative Confirmed Cases | Existing Suspected Cases | Cumulative Tracking of Close Contacts | Close Contacts under Medical Observation |
---|---|---|---|---|---|---|---|---|---|
19 August 2021 | 46 | 1866 | 30 | 88,044 | 4636 | 94,546 | 2 | 1,154,405 | 41,706 |
18 August 2021 | 28 | 1887 | 17 | 87,977 | 4636 | 94,500 | 1 | 1,153,308 | 43,156 |
17 August 2021 | 42 | 1928 | 17 | 87,908 | 4636 | 94,472 | 1 | 1,151,965 | 44,471 |
16 August 2021 | 51 | 1938 | 20 | 87,856 | 4636 | 94,430 | 1 | 1,150,082 | 45,325 |
15 August 2021 | 53 | 1922 | 29 | 87,821 | 4636 | 94,379 | 1 | 1,146,611 | 44,300 |
Movie Name | Release Date | Average Search Index 30 Days before Release | Average Search Index 7 Days before Release | Search Index on Release Day | Average Search Index 7 Days after Release | Average Search Index 30 Days after Release |
---|---|---|---|---|---|---|
The Battle at Lake Changjin | 30 September 2021 | 30,008.7 | 72,595.43 | 502,677 | 1,059,134 | 451,876.7 |
Wolf Warriors 2 | 27 July 2017 | 24,158.23 | 45,348.71 | 256,913 | 1,262,703 | 920,833.4 |
Hi, Mom | 12 February 2021 | 37,625.4 | 104,796.7 | 395,390 | 748,152.4 | 309,665.4 |
Ne Zha | 26 July 2019 | 16,633.3 | 63,845 | 283,248 | 539,412.1 | 331,797.2 |
The Wandering Earth | 5 February 2019 | 31,611.17 | 53,317 | 471,650 | 1,099,486 | 516,540.5 |
Movie Schedule | Time Range |
---|---|
New year’s day | From the end of December of each year to 3 January of the following year |
Spring Festival | Spring Festival holiday every year |
Valentine’s Day | Every year from 13 February to 14 February |
Women’s day | Every year from 7 March to 8 March |
Qingming Festival | Every year on Tomb-Sweeping Day. |
Labour Day | Every year from the end of April to 5 May |
Children’s Day | Every year from the end of May to 1 June |
Dragon Boat Festival | Dragon Boat Festival holiday every year |
Summer Vacation | Every year from the beginning of June to the end of August |
Chinese valentine’s day | Every year Chinese Valentine’s Day |
Mid-autumn Festival | Mid-Autumn Festival holiday every year |
National Day | Every year from the end of September to 7 October |
New Year movie season | Every year from the end of November to the end of December |
Ordinary | Outside the time range of other schedules |
Model | R2 | MAPE | MAE | MSE | RMSE |
---|---|---|---|---|---|
XGBoost | 0.8578 | 17.96% | 5010.3662 | 366,253,785.4177 | 19,137.7581 |
LightGBM | 0.8993 | 16.19% | 4204.9473 | 259,371,807.7103 | 16,105.0243 |
CatBoost | 0.8709 | 18.19% | 5063.2543 | 332,490,044.7863 | 18,234.3096 |
GBDT | 0.8503 | 16.41% | 4972.4056 | 385,492,351.2144 | 19,633.9591 |
RF | 0.7812 | 21.27% | 6650.2541 | 563,602,015.1131 | 23,740.3036 |
SVR | 0.6531 | 69.82% | 12,680.4396 | 893,514,540.6887 | 29,891.7136 |
Voting | 0.8420 | 20.14% | 5829.9820 | 406,872,334.5275 | 20,171.0767 |
Averaging | 0.8476 | 19.13% | 5524.3003 | 392,511,674.4714 | 19,811.9074 |
Ordinary Stacking | 0.8726 | 14.52% | 4168.6142 | 328,196,923.9439 | 18,116.2061 |
Enhanced Stacking | 0.8746 | 14.49% | 4143.7905 | 323,085,161.6079 | 17,974.5699 |
Film Title | Year | Total Box Office (10,000 yuan) | Predict Box Office | MAPE Error |
---|---|---|---|---|
47 Meters Down: Uncaged | 2020 | 4760.74 | 5009.039599 | 5.22% |
D-War | 2008 | 3100.0 | 2656.583856 | 14.30% |
Doraemon: Shin Nobita no Nippon tanjou | 2016 | 10,364.31 | 9381.766633 | 9.48% |
Magic Mirror 2 | 2018 | 2063.43 | 2563.248472 | 24.22% |
London Has Fallen | 2016 | 33,981.65 | 36,959.946839 | 8.76% |
The Great Wall | 2016 | 117,456.91 | 136,822.815602 | 16.49% |
Detective Chinatown | 2015 | 82,441.94 | 94,299.604909 | 14.38% |
Lie Detector | 2021 | 2788.05 | 3027.54925 | 8.59% |
Feature | Description | Feature | Description |
---|---|---|---|
fea_97 | Starring2 all-time total box office | fea_122 | The highest number of audience, on MaoYan, who wanted to watch the movies that leading actor3 had worked on in the past |
fea_446 | Baidu search index of PC on the 20th day after movie releasing | fea_9 | The movie’s score on Douban |
fea_561 | Newly confirmed cases 30 days after the movie releasing | fea_41 | Production company average box office in the past |
fea_67 | Release date number | fea_487 | Baidu search index of intelligent terminal on movie release day |
fea_486 | Baidu search index of intelligent terminal on the 1st day before movie releasing | fea_10 | Duration |
fea_125 | The average number of audience, on MaoYan, who wanted to watch the movies that leading actors had worked on in the past | fea_102 | Leading actor3 average box office in the past |
fea_124 | The total number of audience, on MaoYan, who wanted to watch the movies that leading actors had worked on in the past | fea_364 | Baidu search index of all end on the 1st day before movie releasing |
fea_146 | The total number of audience, on Douban, who wanted to watch the movies that leading actors had worked on in the past | fea_233 | The all-time total box office of integrated marketing company1 |
fea_310 | The average box office of the distribution company1 over the past 5 years | fea_7 | The movie’s score on MaoYan |
fea_422 | Baidu search index of PC on the 4th day before movie releasing | fea_46 | Manufacturing company average box office in the past |
fea_34 | Producer total box office in the past | fea_136 | The highest number of audience, on MaoYan, who wanted to watch the movies that leading actors had worked on in the past |
fea_69 | Main genre | fea_44 | Manufacturing company total box office in the past |
fea_397 | Baidu search index of PC on the 29th day before movie releasing | fea_35 | The total number of movies that producers worked on in the past |
fea_6 | The number of people who want to watch the movie on Douban | fea_39 | Production company total box office in the past |
fea_450 | Baidu search index of PC on the 24th day after movie releasing | fea_197 | The total number of audience, on Douban, who wanted to watch the movies that screenwriters had worked on in the past |
fea_187 | The total number of audience, on Douban, who wanted to watch the movies that director had worked on in the past | fea_82 | Average ticket price |
fea_89 | The number of genre | fea_105 | Leading actor’s average box office in the past |
fea_5 | The number of people who want to watch the movie on Taobao film | fea_19 | Director total box office in the past |
fea_516 | Baidu search index of intelligent terminal on the 29th day after movie releasing | fea_91 | Genre_2 |
fea_424 | Baidu search index of PC on the 2nd day before movie releasing | fea_49 | Distribution company total box office in the past |
fea_555 | Existing suspected cases | fea_100 | Leading actor3 all-time total box office in the past |
fea_51 | Distribution company average box office in the past | fea_40 | Production company total movie number in the past |
fea_157 | The total score of movies, on MaoYan, that leading actors had worked on in the past | fea_81 | Box office in the first week (10,000) |
fea_458 | Baidu search index of PC on the 29th day before movie releasing | fea_83 | The average number of audience in all screenings |
fea_560 | Newly confirmed cases one week after the movie releasing | fea_78 | Total number of movie screenings |
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |
© 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Ni, Y.; Dong, F.; Zou, M.; Li, W. Movie Box Office Prediction Based on Multi-Model Ensembles. Information 2022, 13, 299. https://doi.org/10.3390/info13060299
Ni Y, Dong F, Zou M, Li W. Movie Box Office Prediction Based on Multi-Model Ensembles. Information. 2022; 13(6):299. https://doi.org/10.3390/info13060299
Chicago/Turabian StyleNi, Yuan, Feixing Dong, Meng Zou, and Weiping Li. 2022. "Movie Box Office Prediction Based on Multi-Model Ensembles" Information 13, no. 6: 299. https://doi.org/10.3390/info13060299