Ever wondered why some movies become timeless classics while others fade into obscurity? Movies are more than just entertainment: they’re windows into our cultures, trends, and shifting societal values. We dove deep into the factors that shape how movies are rated, uncovering fascinating patterns along the way. From where a movie is made to when it’s released, we found that ratings reflect so much more than just how much people liked the film. For example, movies from different countries and genres show unique rating patterns; although surprisingly, documentaries and romantic comedies seem to break the mold! Time also plays a huge role. While older films might conjure nostalgia, we found that a movie’s release date alone doesn’t tell the whole story of its success. So, we developed innovative ways to dig deeper, like measuring how diverse a cast is (in terms of gender and ethnicity) and even how closely connected a movie is to big historical moments. Using a powerful Random Forest model – think of it as a team of 100 decision-making “movie critics” – we trained our system to predict ratings with 76.4% accuracy. What stood out most? Release dates topped the list as the most important factor, but it was clear that a film’s number of actors, runtime, region and genre also shape how audiences perceive it. In short, our project doesn’t just crunch numbers, but moreover tells the evolving story of how movies resonate with audiences over time, across borders, and through different cultural lenses. And who knows? Maybe it even offers a sneak peek into what makes a movie a future favorite for the crowd.

The Good, the Bad, and… you guessed it, the Wrangled

Our work relies on data sourced from several unique platforms. General movie features, such as film runtimes, release dates, and similar metadata were collected from the CMU Movie Corpus. Whilst this movie metadata is comprehensive, many attributes lack a significant number of values for the dataset’s contained movies. Movie ratings are also missing from this data source. Hence, we turned to a couple of external sources to populate and expand our data:

1. IMDb: “The world’s most popular and authoritative source for movie, TV and celebrity content”, IMDb, provides movie ratings sourced from their users in non-commercial data dumps that are publicly available [1]. Whilst these ratings are averaged in an undisclosed manner as to limit voter bias, we consider this the best and largest possible data source to retrieve ratings from. We therefore source this project’s ratings from IMDb.

2. The Movie Database (TMDB): TMDB compiles millions of movies and their metadata tagged by their IMDb ID. Using this data frame, we were able to populate missing movie values for features in the Corpus. As well, the IMDb IDs provided with this dataset allowed correct matching of ratings to movies. These ratings were then added into the Corpus dataset based on title and release date as a unique identifier.

With movies and their metadata compiled, we further tidied up this data by dropping irrelevant attributes, such as movie poster paths, as well as movies missing data for key features, such as their country of production.

As a tangential point, our interpretation of the country feature is worthwhile explaining. Take, for example, “Indiana Jones and the Kingdom of the Crystal Skull”, which was filmed at Yale University, but then also in Brazil and Argentina. Clearly, the Corpus movie’s label of “United States of America” is in reference to the movie being produced by an American company and coming from Hollywood. So, the countries attribute is from here on out considered a movie’s location of production, which we hypothesize will impact the style and content of the movie based on regional differences in ideologies.

To conclude on our extensive data cleaning, we had to consider movies in the Corpus with more than one country label and genre listed. Why care, you might wonder? Because, for accurate rating prediction based on these features, we cannot consider their presence across more than one attribute value. Otherwise, regional and genre-specific rating trends would homogenize, and we would potentially encounter a situation in which we can predict one geographical region of movies perfectly given the exact same movies exist in another region.

We therefore opted to preserve the ~86% of movies that had solely one country listed. Furthermore, to handle movies with more than one genre, we opted to expand movies to have more than one entry in our dataset but with each row an entry per unique genre the movie has. Our reasoning for this is threefold:

1. Expanding multi-genres instead of choosing one of their listed genres to incorrectly represent all of the other genres allows us to consider their complexity as a useful feature for future work.
2. Keeping movies across their genres ensures truthful representation of the way ratings are dispersed across genres. Consider if many movies had the multi-genre of being comedies and romance movies. If comedy was then chosen as the genre to represent these movies, all of these movie ratings would go solely to the comedy genre, excluding romance as a contributor to these ratings values.
3. Expanding movie genres allowed us to identify the most relevant genres for analysis. The original Corpus dataset contains 365 unique genres, varying from single genres to six or more combined genres—an eye-wateringly broad scope for analysis. To cut this number dramatically, we chose the top 20 genres of the expanded movie genres across the Corpus, ultimately resulting in the movies to analyze with genres in: action, action/adventure, adventure, black-and-white, comedy, crime fiction, documentary, drama, family, horror, indie, musical, mystery, romance, romantic comedy, romantic drama, short, silent, thriller, world cinema.

Finally, we manually grouped movies into regions based on their listed countries to reduce the complexity of further analysis from the original 195 possible countries down to 9 regions of interest:

• North America
• Central and South America
• Western Europe
• Eastern Europe
• South Africa and Central Africa
• North Africa and Middle East
• India
• Remaining Asia
• Oceania

Understanding Movie Production and Ratings Evolution Across Time and Space

The journey into understanding movie ratings begins with a global exploration by mapping ratings across countries and tracking their changes over time. Movies in our dataset span the past century and over nine regions for analysis. Fear not though, as the dynamic map below brings clarity to this vast analysis! With just a glance, you can explore how ratings for a specific genre evolve in a region of interest. For instance, comedy ratings in Australia were strikingly stable at a low score of 2 between 1940 and 1970. After this period, ratings soared, fluctuating around an average of 5. In contrast, comedy ratings in other countries during this era were highly erratic, visualized by the flickering map of rating colors during this thirty year interval. These spatial and temporal variations reveal that both features play a pivotal role in shaping movie ratings, and are interconnected with rating variations.

Select Genre: Drama Comedy Romance Black-and-white Action Thriller Crime fiction Indie Short World cinema Action/adventure Horror Adventure Family Musical Documentary Romantic drama Mystery Romantic comedy Silent

To bring clarity to rating fluctuations over geographies and time, the below global race of the frequency of movie production across our top 20 genres can be examined. This race reveals fascinating trends, such as how some genres, like drama, show a steady increase in production over time. We hypothesize this may saturate audiences and dampen enthusiasm for similar movies ratings. Hence, this race highlights that there is surely a time component of movie production frequency and thereby ratings that must be considered in our further analysis of movie ratings.

To further explore the dynamics of movie’s production, the below histograms provide genre-specific insights of production frequencies. With this, genres including comedy, romance, action, and mystery can be seen as having bimodal production patterns. Peaks emerged prominently between the 1990s and 2000s, coinciding with the technological revolution and the rise of digital media. Another notable peak appears in the 1920s, often associated with the film industry’s resurgence during the post-WWI “Roaring Twenties.” These dual peaks suggest that historical and technological contexts play a crucial role in shaping production trends, which we hypothesize may in turn impact ratings. This temporal variation at the genre-specific levels underscores the importance of considering genres alongside movie production regions and release dates for further ratings analysis.

Select Genres:

Movies Ratings by Genre: A Deeper Dive

Having established that movie production depends on both region of production and release date, we now turn our attention to another vital variable: genre. The heat map below presents the distribution of ratings for the top 20 genres, grouped into bins for easier visualization. This approach allows us to clearly observe differences in public perception and reception across genres.

While many genres like action and black-and-white films show their highest rating counts in the 6.5-7 bin, other genres deviate significantly. Documentary and world cinema films, which stand out as the most non-fictional genres in the top 20 our analysis considers, dominate higher rating bins: documentaries peak in the 7.5-8 range, while world cinema leads in the 7-7.5 bin. Conversely, genres known for their polarizing nature, such as horror and thriller, show distributions skewed toward lower rating bins, reflecting divided audience opinions.

Even within genres with similar peak bins, variance tells an intriguing story. For instance, black-and-white movies have a tightly clustered distribution, with 64% of ratings falling within plus-minus 0.5 rating values of the 6-6.5 bin. Meanwhile, only 48% of action movie ratings occupy this same range, indicating a broader spread and more diverse audience reception, despite the same central bin between both genres’ ratings.

These differences underscore that ratings are not only time- and region-dependent but also profoundly genre-specific. This insight highlights the importance of genre as a variable in our deeper analysis of movie ratings, setting the stage for our main analysis of the interplay of movie genre with other factors such as production location and release date on ratings.

Select Genre:

Select First Region:

Select Second Region:

To uncover deeper insights into movie ratings, we’ve created an intuitive visualization that lets you compare ratings of a specific genre between two regions over the decades. This tool plots the mean rating for each decade, from the 1950s to the 2000s, for the selected genre and region, complete with 95% confidence intervals for each data point. To avoid confusion, the line colors are independent of previous genre-based color schemes, as they represent regional comparisons. Importantly, no data point is plotted for a given region and decade unless at least three movies exist in that genre for that period.

With nine possible regions and 20 genres to explore, this visualization offers an impressive 720 combinations of comparative plots. Don’t worry, though! We’ve streamlined the analysis by using Western Europe as a baseline for comparison (since it’s where we’re based). This narrows the scope to eight regional comparisons across 20 genres, reducing the analysis to 160 plots—still comprehensive but more manageable. Even better, we’ve carefully reviewed each one and are excited to present a curated summary of the most compelling patterns and insights:

Regional Dynamics:

• North America often underperforms relative to Western Europe across all top 20 genres, with the most notable instance being horror movies.
• India displays a dramatic decline across multiple genres over decades, often starting above Western Europe but dropping significantly by the 1980s or later.
• Eastern Europe frequently outperforms Western Europe in ratings, notably for dramas and crime fiction movies.
• Oceania closely follows Western European trends in multiple genres, including dramas, comedies, short films, and mysteries.

Genre-Specific Observations:

• Indian drama ratings dropped by a value of 1.43 in IMDb rating, from the 1950s to the 2000s, mirroring broader trends across Indian cinema. Central and South America lack data for drama films but align in shape with Western Europe’s trend in mean ratings from the 1950s to the 2000s.
• Eastern Europe, once above Western Europe by a statistically significant amount of over 1.5 in IMDb rating, has converged in ratings over time.
• North Africa and the Middle East have seen a decline and convergence with Western Europe by the 2000s.
• Romance films do not show significant regional disparities in ratings.
• Central and South America saw a dramatic increase in action movie ratings post-1980s, contrasting with a consistent decline in North America and India.
• Eastern Europe consistently outperforms Western Europe in crime fiction movies, with a noticeably statistically significantly higher value of mean rating in the 1980s.

Key Patterns:

• Indian films transitioning from outperforming to underperforming relative to Western Europe is a notable pattern, particularly in drama, comedy, musical, and mystery genres.
• Eastern Europe’s comparative mean ratings for crime fiction movies are significantly higher than other regions, a trend not observed for Eastern Europe in other genres.
• Significant data gaps exist for African regions, particularly Central and South Africa, across almost all genres except family and horror, where production was sporadic.

Our visual analysis reveals a fascinating insight: within any given region, movie ratings vary significantly depending on the genre. To confirm this observation, we ran a one-way ANOVA test on each region’s movies, grouped by genre. The results showed statistically significant differences in mean ratings for nearly all genres, with the exceptions of documentary and romantic comedy, which exhibited more uniform ratings.

Taking this analysis a step further, we examined whether movie ratings differ across regions for a single genre. Using the Kruskal-Wallis test, which allows comparison of movie ratings across genres with different distribution shapes, we uncovered significant differences in movie ratings by region for most genres. Interestingly, Eastern Europe was the sole exception, where ratings appeared consistent regardless of the genre. These findings confirm that both genre and region are pivotal factors influencing movie ratings. By recognizing these patterns, we reinforce the importance of including these features in predictive models to better understand and anticipate audience ratings based on readily available movie attributes.

Finally, this statistical rigor was extended to assess the impact time (release date) on movie ratings to confirm this feature’s usage as a predictive ratings feature. An ANOVA test, examining the interaction between decades, genres, and regions, confirmed statistically significant differences in ratings across time for all top 20 genres and nine geographical regions. While this interaction might sound complex, all that this is telling us is that we can confidently state that ratings vary meaningfully across time, genre, and location . Hence, their prediction must be at the minimum based on these attributes as predictive features.

Next, we analyzed the slopes of rating trends over decades for each genre and region. By focusing on the top five genres across the nine regions, we reduced the analysis to 45 slope comparisons. Of these, 21 showed statistically significant changes in mean ratings over time, demonstrating that for many genre-region combinations, ratings evolve dynamically across decades. To probe deeper, we employed autoregressive integrated moving average (ARIMA) models to test whether release dates alone could predict ratings. The results revealed that release dates, while important, are not sufficient predictors in isolation. Instead, release dates must be combined with genre and region to capture the full complexity of movie ratings. This finding solidifies the case for incorporating all three features into our predictive model to achieve accurate and meaningful insights.

Characters Data for Additional Insights

Given the need for increased features to accurately predict movie ratings, let’s turn the spotlight to the stars themselves: the actors. They are, after all, the heart of every great movie. We asked ourselves, who are these actors, where do they come from, and most importantly, does considering actors as characteristics of movies really matter in impacting movie ratings? To dive into these questions, we explored the movie character dataset. This analysis began by examining the number of actors in each film and then their diversity in terms of ethnicity and gender.

The More the Merrier?

We began by looking at the distribution of the number of actor’s in a movie and their potential impact on ratings. Unsurprisingly, in our dataset, the number of actors in a movie rarely exceeds 20, given the size limits of filming sites and more likely budget constraints. But does having more actors increase the chances of higher ratings? At a Pearson correlation of 0.0355 and a p-value < 0.001, the relationship between the number of actors and ratings is weak yet significant. This leads us to consider actor number as a feature to include in our rating prediction model.

To further inspect actor characteristics in movies, we move onto considering the demographics of their casts. By enriching the provided character data, we managed to populate 77.17% of our movie’s characters, and thereby actors, data with nationalities and ethnicities. Analyzing the former first, we sought to understand if regions typically cast from their own talent pool as to indicate if movies from those regions are diverse for actor cultural backgrounds.

From the above visualization, we can see that most regions show a strong preference for in-region talent. In fact, Asia, India, and Eastern Europe favor local actors overwhelmingly, with over 90% of their casts being from the same region. Even North America, known for Hollywood's global reach, leans toward in-region actors, though not as strongly. Compare this to Bollywood, which remains almost exclusively centered on Indian actors, emphasizing its cultural specificity/target audience. This preference for in-region talent could impact movie ratings as audience's may be influenced by their cultural and regional connection and favoritism to the cast. However, in-region casting alone is not a strong predictor for movie ratings, and a more granular approach with ethnicities is employed.

The ethnicities we were able to populate of our movies’ actors can be categorized into cultural groups: North Americans, Oceanians, Native Americans, West and East Europeans, Indians, Remaining Asia, South Americans, North Africans and Middle Easterns. This clustering reflects not only geographical boundaries but also the cultural and historical dynamics that shape the cinematic landscape.

Taking a closer look at ethnicities within regions, a first observation is the widespread presence of Western Europeans and North Americans across other regions, appearing in nearly every region’s movies by over 1% of all ethnicities except in India. Another striking pattern is that the most dominant ethnicities per region is almost always that of the region. That being said, the size of this in-region ethnicity component varies per region, leading us to conclude that a movie irrespective of their region of production may be composed of numerous actor ethnicities. Some metric to represent this ethnic diversity is required, but before we reveal what that score is, let’s briefly turn to the actor genders.

The above dynamic pie chart demonstrates that the gender proportions per movie are far from balanced. Male actors have almost always dominated the stage at a consistent rate, and so we also need to capture the gender diversity of a movie to properly reflect the movie’s characteristics.

Creating Cast Diversity Features

Knowing now that films come with varying degrees of actor diversities in terms of ethnicity and gender, a quantity to capture this is required. We used the Shannon Diversity Index (SDI), a measure of the entropy of distributions. This metric was applied to both gender and ethnicity of the casts per movie, capturing two important aspects: the richness, or the number of distinct ethnicities or genders represented, and the evenness, or how balanced the representation is across these groups. For example, a cast evenly spread across five ethnicities scores higher than one dominated by a single group. This approach offers a structured way to evaluate how well movies reflect global diversity.

To finally return these metrics back to our mission of rating predictions, we tested both diversity scores against our movies average ratings. For both scores, we observed a weak negative correlation, suggesting a meaningful but potentially non-linear relationship. To explore these relationships further, we grouped our movies by their cast ethnicity index and performed an ANOVA test, yielding a p-value of 0.025. This tells us that ratings vary significantly over cast ethnic diversities, leading us to use this metric in our final predictive model on average ratings. Similarly, for the gender diversity score, an ANOVA test passed over our movies grouped by this index resulted in an even stronger p-value of 0.0001, highlighting a significant relationship between gender diversity and ratings. The significance in the distribution of movie ratings being inconsistent across movies with varying ethnic and gender diversities in cast pushes us to use these metrics for further analysis.

Spoken Languages and Movie Ratings

In a final consideration of diversity as a predictor of movie ratings, we focused on the number of spoken languages in movies. By retrieving a movie’s number of spoken languages, we were surprised to find a clear clustering of multilingual films around higher audience ratings. Notably, Sicilian and Haryanvi languages correspond to the highest ratings, defying conventional expectations. Regression analysis over average movie ratings and number of spoken languages revealed a weak yet positive correlation between the two. To indicate if multilingual films resonate more with audiences, we finally split the movies into high (above 15) and low multilinguality groups and compared their average ratings with an ANOVA test. This revealed a significant p-value of 0.0174, leading us to include spoken language count as a feature in our final predictive model.

Two Final Scores as Rating Predictive Features

To further uncover the nuanced factors influencing movie ratings, we crafted a novel feature: the Historical Proximity Score. This artificial feature assigns a value between -1 and 1 to each movie, based on the temporal closeness of its release to the closest historically significant event in the production region’s history. Given the global scale of our analysis, each event considered cannot be listed here, however, these events generally classify as being wars and conflicts, political transformations, economic events, social and cultural milestones, or international relations and treaties. A magnitude close to 1 signifies a movie released during the same year as such an event, while this value linearly falls off over a five-year period from the date of occurrence to a movie’s release date. A sign is then given to the score based on if the event is considered positive or negative. By integrating this feature, we aim to investigate whether the weight of history leaves its mark on audience perception.

The final feature that we included for predicting a movie’s characteristics to predict its audience ratings was a movie’s Genre Complexity Score. This feature quantifies the intricacy of a movie’s genre composition, assigning a score equal to the number of genres it spans from our top 20 list of genres considered. A straightforward single-genre film earns the baseline score of 1, while a genre-blending epic boasting 10 genres achieves the maximum score of 10. Like so, we aim to include the richness of a movie’s genre and its influences on audiences.

The Final Predictor

Our Random Forest

Having considered features with identifiable differences in their values on ratings, as well as composing unique features for movies, we can finally uncover the hidden dynamics between movie characteristics and audience ratings using a Random Forest classification model. Drawing from the carefully curated dataset, we trained a model to predict binned movie ratings, a rounded, categorized version of the IMDb average audience ratings. This 0.5 increment rating value is our model’s target variable, representing a simplified metric of audience ratings that allows for clearer classification.

The Random Forest Classifier is a robust and interpretable ensemble method for supervised learning that is able to handle complex relationships between features. Before training, categorical features like genres, regions, and actor counts are one-hot encoded to make them digestible for the model. This ensures that no valuable information is lost in translation.

An 80-20 train-test split was used to divide the data into training and testing subsets. The Random Forest is then trained on 58,430 samples, fine-tuning its predictive capabilities through ten-fold cross-validation. This step not only validates the model’s performance but also ensures resilience to overfitting. Despite challenges like imbalanced data, the model achieves a solid mean cross-validation score of 75.26% and an accuracy of 76.4% on unseen test data, a testament to its reliability.

Feature Importances: Unveiling the Stars

One of the Random Forest’s strengths lies in its ability to reveal feature importances by the information each contains at separating movies into their rating classes. In taking these information importances, we are able to rank our analyzed features at their predictive power on movie average ratings. In 1st place is the movie release date! The movie's release date is the most informative predictor of a movie's rating, emphasizing the cultural and temporal relevance of films.This also reinforces our early intuition that temporal insights, though initially considered alone, can gain decisive predictive power when strategically combined with the right features. In 2nd place, and slightly surprisingly, is the number of actors in a movie. This aligns well with our initial actor-based analysis, reaffirming a significant yet likely non-linear relationship between the number of actors and movie ratings. This conclusion reinforces the appeal of star-studded casts in a movie, polarizing audience ratings and hence providing clear distinction between a movie's rating class based on the number of actors involved. Then in 3rd and final spot of our podium of ranking is the film runtime. This demonstrates that the pacing of movies is critical in dividing a movie between achieving one rating over another. A movie’s length, therefore, is a crucial factor in shaping audience engagement.

Beyond the expected, diversity metrics emerge as significant players in classifying movie ratings as well. Naively, we’d expect a movie’s revenue to hold a lot of information about a movie’s rating. However, with the gender and SDI scores above the revenue feature for information contained at predicting a movie’s rating, we must consider that movie ratings are not linked to the box office sales. Instead, this demonstrates movie ratings are not distinguishable by their revenues such that movies amassing large crowds are not those with noticeably different ratings from smaller production films. In addition, both the gender and ethnicity SDI falling at the middle of the ranking of features in predicting movie ratings underlines the growing importance of representation, hinting that inclusivity resonates strongly with modern viewers. Finally, other features, such as "Cast Diversity" and "Historical Events," demonstrate subtle yet impactful contributions in movie ranking classification. Together, these factors create a rich and multi-faceted picture of what influences audience ratings.

A Data-Driven Lens on Cinema

This model does more than predict ratings—it provides a lens to interpret the many forces that shape a film’s reception. By analyzing features ranging from representation and timing to financial metrics and thematic elements, we bridge the gap between quantitative data and the subjective experience of moviegoers. The insights in preparing our Random Forest classifier for movie ratings can therefore empower filmmakers, critics, and audiences alike to appreciate cinema not just as entertainment but as a reflection of societal values and cultural trends.