Movie Data Analysis and a Recommendation Model

Hang Yang, Yiqing Pei, Zhihui WangAtlantis 2022 · DOI · OpenAlex

TL;DR

This paper presents a data-driven framework to mitigate the high financial risks inherent in the movie industry. By analyzing decades of historical data from TMDB, the authors developed a weighted recommendation model based on cosine distance to help producers identify "cautionary tales" or "success guides" for their upcoming projects based on budget, genre, and casting.

What Problem Does This Paper Address?

The film industry is notoriously high-risk; even experienced filmmakers struggle to predict whether a project will be a box-office hit or a financial disaster. While the "Internet age" has provided a wealth of data, translating that data into actionable insights for pre-production—such as setting a budget, choosing a genre, or hiring a director—remains a challenge.

The authors aim to bridge the gap between raw historical data and strategic decision-making. They seek to answer: Which genres are currently trending? How do budget and popularity correlate? And most importantly, how can we mathematically identify "similar" movies to provide producers with a benchmark for potential performance?

Methods

The researchers implemented a multi-step pipeline involving data visualization and a custom recommendation engine:

  1. Data Preprocessing & Standardization: To handle the massive time span (1969–2016), the authors focused on the "earning rate" (revenue divided by budget) rather than raw box office numbers to account for inflation and varying production scales.
  2. Vectorization: Both categorical data (genres, cast, directors) and numerical data (scores, earning rates) were converted into binary vectors (0 or 1).
  3. For numerical variables like "score" and "earning rate," they used the 75th percentile as a threshold. If a movie's score exceeded 6.8 or its earning rate exceeded 4.34, it was assigned a "1" in the vector, marking it as "excellent."
  4. Cosine Distance Metric: To find the similarity between two movies ($A$ and $B$), the model calculates the cosine distance. This measures the "affinity-disaffinity" relationship in a multi-dimensional feature space.
  5. Weighted Similarity Formula: Recognizing that some factors are more important than others, the authors developed a weighted distance formula: $$TotalDistance = 0.8 \cdot genreDist + directDist + castDist + wordsDist + 2 \cdot earnDist + 1.3 \cdot scoreDist$$ Notably, financial performance (earning rate) is given the highest weight (2.0), followed by critical reception (1.3), ensuring the recommendations prioritize successful precedents.

Datasets

The study utilizes data from TMDB (The Movie Database), covering movie information from 1969 to 2016. The dataset includes two primary components:

  • Movie Metadata: Title, genre, budget, revenue, release date, and keywords.
  • Credits: Detailed information on the cast and directors.

Key Insights and Ideas

  • The Popularity-Quality Gap: The analysis revealed that highly-rated genres (History, War, Music) are often not the most "popular" in terms of audience volume. Conversely, Adventure and Animation dominate popularity but sit in the middle of the pack for average ratings.
  • The 1992 Inflection Point: Visualization of genre trends shows a massive surge in movie production and genre diversity starting around 1992, marking a period of industry prosperity.
  • Budget as a Quality Proxy: The heat map analysis showed a strong correlation (0.73) between budget and revenue. The authors interpret this not just as "spending money to make money," but as a reflection of higher production values (better scripts, effects, and cast) that attract larger audiences.
  • Engagement is the Best Predictor: The strongest correlation found was between popularity and vote count (0.78), suggesting that audience engagement is a more reliable indicator of financial success than critical score alone.

Key Results

  • Genre Profitability: Animation, Adventure, and Fantasy were identified as the most "high-stakes" genres—requiring the highest budgets but also yielding the strongest profitability and revenue.
  • Model Validation: Using "Avatar" as an input, the model successfully recommended thematically and financially similar blockbusters like Aliens, The Abyss, and Terminator 2: Judgment Day.
  • Strategic Recommendations: The paper concludes that while large studios should focus on high-budget Adventure/Action films, small and medium-sized companies can find high rates of return in Comedy and Drama, which require significantly lower initial investment.

Limitations and Future Work

The authors acknowledge several areas for improvement:

  • Precision: The current model uses a relatively simple binarization (0/1) for vectorization, which may lose nuance in the data.
  • Concreteness: The recommendations provide a list of similar movies but do not yet offer "concrete" predictive figures for a new movie's expected revenue.
  • Algorithm Complexity: While cosine distance is efficient, the authors suggest that more advanced machine learning or neural network approaches could improve the precision of the recommendations in future iterations.

Summary updated

Abstract

The high-risk nature of the movie industry has kept companies looking for better plans to make and invest in a film. With the deepening of Internet applications, it's possible to clearly visualize and demonstrate the overview of the film market and the performance of different movie genres with their unique features in the market. The box office is an important indicator to measure movie performance, and it to some extent reflects the economic development of modern society. So, explore the rules of box office and its correlation with other features of the film. Moreover, to establish a recommendation model, by vectorizing and attaching values to characteristics of the films, the distance between two movies can be calculated to find similar movies. Thus, suggestions on preparing budget, choosing genre and casting could be given to the film producer based on these historical movie data so as to achieve higher audience reviews and financial returns.
BibTeX
@inproceedings{yang2022movie,
  author = {Hang Yang and Yiqing Pei and Zhihui Wang},
  title = {{Movie Data Analysis and a Recommendation Model}},
  booktitle = {Atlantis},
  year = {2022},
  doi = {10.2991/978-94-6463-030-5\_74},
}

Loading related content...

Figure