Abstract

Hollywood enjoys the position of being the biggest movie producers when it comes to global recognition among movie-making industries. Despite being the biggest movie producer, it has been facing high revenue losses lately since most of the films that it has created have failed to capture viewer’s attention in the first few weeks of its release resulting in a box-office flop. It has been observed in a recent study that Hollywood is estimated to witness a loss of around 1 billion to almost 10 billion US dollars till 2020. Revenue risks have created immense pressure on movie producing stakeholders. They feel a constant pressure to come up with a formula to make a successful movie, however, to date; there are no fixed ingredients that can ensure the success of a movie. Researchers and movie producers constantly feel a need to have some expert systems which would predict the fate of the movie prior to its production with reasonable accuracy. Regardless of the difficult nature of the issue area, few researchers have created expert systems to forecast the financial success of movies using different approaches, but most of them are targeted pre-release forecasting or have low prediction accuracy. Such predictions are of a seminal nature as of their limited prediction scope, and non-ability to reduce revenue loss risk. Therefore, there is a constant demand from investors to have pre-production forecasting tools with high accuracy which can help them plan and make necessary alterations to save huge investments. In this study, we proposed eighteen new features to forecast box-office success, as soon as the quotient (director and cast) signs an agreement. This proposed forecasting time is the earliest prediction that has ever been reported in the movie forecasting literature. The decision support system ranks director and lead cast by utilizing their performances of the last 100 years (1915–2015). The processed output file is a table that ranks each director and cast into four categories based on cast experience, journalist critics, media reporting, user ratings, and revenue generated by associating movie. To produce more accurate results, learner-based feature selection is also performed to select the best subsets of features. This system is intended to be a dynamic tool, integrating further data for real-time adaptation. The system has the ability to incorporate different feature selection algorithm for the progressive improvement of movie success forecasting We demonstrate the effectiveness of extracting features and explain how they improve forecasting accuracy over existing models. The adaptive behaviour of the presented system is achieved by incorporating conceptually different machine learning classifiers, i.e. support vector machine, gradient boosting, extreme boosting classifier, and random forest. A voting system is used to make the prediction by averaging the output class-probabilities. To assess the adequacy of new features, a cross-validation test is directed. Our classification results are evaluated by using two performance measures, i.e. average per cent success rate, or within one class away from its actual prediction. The new features have achieved the most noteworthy accuracy of 85% with an expansion of a 46.43% (average per cent success rate) and 5.56% (within a class away) in comparison with other state-of-the-art feature sets.

中文翻译：

---

试生产票房成功商预测

摘要

当在电影制作行业中获得全球认可时，好莱坞享有成为最大电影制作人的地位。尽管它是最大的电影制片人，但由于其制作的大多数电影在发行后的前几周未能引起观众的注意，导致票房下滑，因此它一直面临着高额的收入损失。在最近的一项研究中观察到，好莱坞估计到2020年将损失大约10亿至近100亿美元。收入风险给电影制作业的利益相关者带来了巨大压力。迄今为止，他们一直想出一个公式来制作一部成功的电影，这一直是他们的压力。没有可以确保电影成功的固定成分。研究人员和电影制片人不断感到需要一些专家系统，这些系统可以在电影制作之前以合理的准确性预测电影的命运。不管问题领域的困难程度如何，很少有研究人员创建专家系统来使用不同的方法来预测电影的财务成功，但是其中大多数是有针对性的预发布预测或预测准确性较低。此类预测由于其有限的预测范围而具有开创性，并且无法降低收益损失风险。因此，投资者不断需求具有高精度的生产前预测工具，这可以帮助他们计划并进行必要的更改以节省大量投资。在这项研究中，我们提出了18种新功能来预测票房成功，商（导演和演员）签署协议后。该提议的预测时间是电影预测文献中已报道的最早的预测时间。决策支持系统利用过去100年（1915年至2015年）的表现对导演和领导者进行排名。处理后的输出文件是一个表格，该表格将演员的排名和演员的经历，记者评论员，媒体报道，用户评级以及与电影关联产生的收入分为四个类别。为了产生更准确的结果，还将执行基于学习者的特征选择，以选择最佳的特征子集。该系统旨在成为一个动态工具，集成更多数据以进行实时调整。该系统具有合并不同特征选择算法以逐步改善电影成功预测的能力。我们演示了提取特征的有效性，并说明了它们如何在现有模型上提高了预测准确性。通过结合概念上不同的机器学习分类器（即支持向量机，梯度提升，极限提升分类器和随机森林）来实现所提出系统的自适应行为。投票系统用于通过平均输出类别概率来进行预测。为了评估新功能是否足够，需要进行交叉验证测试。我们的分类结果是通过两项绩效指标（即平均成功率）或与实际预测相差一类的指标来评估的。