Deep Learning with IMDb Movie Ratings and Genres
Color Trends over the Last 150 Years

IMDb Rating Descriptive Statistics

The movie rating and genre prediction algorithms were trained on a dataset of 39.2K movie posters found on Kaggle and from the IMDb datasets website. The movies date back from 1874 to the present day (2022), IMDb ratings are on a scale of one to ten, and span across 26 different genres. In our original dataset, 73.2% of the movies had several genres (example: Action | Comedy). I cleaned the data so there was only one genre associated with each movie to simplify the classification problem.

The main motivations of this project were to explore transfer learning in computer vision applications while exploring new machine learning tools, data visualization, and web frameworks. In particular, the fast.ai Python library for transfer learning, Huggingface spaces to host machine learning applications, the Gradio API to create flexible data flows, RESNET18 a pre-trained model which already learned rich feature representations for a wide range of different object categories, the Bootstrap web framework for a highly iterable responsive web user-experience, and Chart.js for responsive charts built on top of JavaScript. Using these tools we can build a shareable, self-contained dashboard that incorporates an end-to-end deep learning model.

One method to view all movie posters from a bird's eye view is to count the number of unique pixel colors and create a color distribution spectrum across several dimensions. I used Python and the PIL library to open each poster, count each pixel color and convert the pixel to a Red, Green, Blue color coordinate (0-255, 0-255, 0-255). RGB with coordinates (0, 0, 0) would be black and (255, 255, 255) would be white. This gives us 256³ color combinations or 16.7M colors. After scanning the pixels of all 39.K posters we have 10.7M unique colors.

Three issues arise when we have over 10M unique colors to analyze:
1. A full-HD monitor is only capable of displaying 1920 x 1080 = 2.0M pixels at once.
2. Some colors only appear once out of 10M unique colors which would be challenging to visualize in a graphic.
3. We cannot sort nor aggregate colors based on their RGB coordinates. RGB coordinates that are close together will not necessarily result in a smooth color gradient as seen in Figure 3.

To address the above issues, I decided to group the colors into a primary parent color, which would reduce the number of unique colors down from 10M to approximately 700. I converted each RGB coordinate with its CIELAB color space equivalent to find the closest primary parent color. The CIELAB coordinate is designed to approximate human vision. The three coordinates are: L* = lightness (0 is dark and 1 is light), A* = red-green value and B* = yellow-blue value. Colors which have CIELAB coordinates that are similar to eachother (small Euclidean distance) are similar in color to the human eye. For example, a pixel ■ with a blue CIELAB color space of (54, 8, -54) would aggregate to a blue parent color #006de0 ■ with a parent CIELAB color of (46, 9, -67). In contrast, their RGB values are (81, 127, 222) and (0, 109, 224) respectively. I used the SciPy KDTree algorithm to identify for each of the 10M color coordinate, which was the closest parent CIELAB coordinate (~700 colors). The KDTree algorthim used is described by Maneewongvatana and Mount 1999 ^B. By reducing and aggregating the unique color combinations we can address the first two issues. I used the Colour library in Python to identify the order of colors which would approximate an aesthetically pleasing color gradient which can be seen in Figure 4 below, where the ~700 colors now smoothly transition from black to white to the full colors of the rainbow. We can see that 75.5% of all the colors used span the black and white spectrum. We can now examine over 39.K movies in a single visualization over several dimensions such as time (decade), genre, and rating.

Color Spectrum of All Movies from 1874-2022

Next let's look at the color spectrum across each decade. I excluded movies from 1870 and 2020 since they had less than three movies per decade. We can see four general trends when we look at the color spectrum across time.

1910-1950, The Dawn of Color: The invention of the four-color wet printing process, known as CMYK (Cyan, Magenta, Yellow, and Key (black)) gave way for producers to create posters with a nearly unlimited color palette at low costs. We can see this in our Figure 5 color spectrum starting from 1910. This timeline also corresponds with when the Eagle Printing Ink company first introduced this technique to print colors back in 1906. ^C

1970-2010, Trending Towards Darkness: We can see a trend towards darker movie posters during the last few decades. In 1910, the darker palettes occupied only 28.3% of the total color spectrum while in 2010 the darker palettes occupied 58.2%.

'Some movie theorists believe this is related to the popularity of action, thrillers, and sci-fi of the past decade may have led to a darker and more "masculine" palette.'

1990-2010, Orange and Blue: Another trend is the increased use of Blue and Orange across movie posters. In 1910, blue and orange colors only occupied around 7.1% of the total color spectrum versus 12.2% in 2010. Our analysis also aligns with the observations found in many videos and articles. The articles discuss how blue and orange are on opposite sides of the color wheel which makes them complimentary colors, causing them to pop in the poster image.

'There's a very simple answer for why there's always so much orange and blue in movies, and it all starts with skin tone. Human skin tones generally fall somewhere in the orange spectrum.'

We can see the average rating decrease from 7.17 to 6.24 from its peak in the 1920s to 2010. Perhaps this is a macro version of the same trend we see with movies and TV show ratings, which tend to decrease over time.

It is fairly common for many movies and TV shows to have a relatively high rating when they first appear, only to have the rating go down over the following months. People who are more eager/interested in the title (and therefore more likely to be inclined to give it a good rating) usually go see it as soon as it opens or airs. The initial rating therefore tends to be heavily skewed towards the high end of the scale. Over time, the title is seen by a wider audience of casual viewers and attracts votes from people who may not necessarily be fans of the genre or who have different expectations and may therefore have a less favorable opinion.

There appears to be two ideas we can draw from this for our own movie poster:
1. We can consider adopting a 1920s color theme to maximize our IMDb rating to rekindle the novelty factor as this was the era that had the highest average IMDb rating.
2. We can avoid following the trend to dark colors >58.2% and the use of orange and blue colors >12.2%, as these trends correlate with decreasing IMDB ratings.

In the above color spectrum, we can see that the Horror genre consists of 60.1% of black pixels and 11.0% bright colors (red, orange, yellow, etc.) while the Family genre consists of 31.3% black pixels and 33.5% bright colors. We can see an increase in percentage of bright colors as we transition into the light-hearted genres.

What is the psychology of dark colors and how does it impact movie poster colors? Alan Costall and Tom Clarke, from the Department of Psychology at the University of Portsmouth explored the symbolism of colors in their 2006 paper Emotional Connotations of Color: A Qualitative Investigation.

The responses obtained for black were consistent with the research of Adams and Osgood and Wexner who reported that black is bad, strong, and inactive as well as powerful and masterful. About 69% of our participants regarded it as a symbol of evil, malice and death. 69% of the participants reported that both orange and yellow were "happy" colors.

Color psychology is the study of colors as a determinant of human behavior. The article discusses how ancient Egyptians and Chinese civilizations used colors as a method of treatment as early as 2000 BC. In the 20th century, Carl Jung pioneered the field of color psychology exploring the meanings of color emotionally. This demonstrates our long cultural relationship with colors and how different colors can elicit a range of emotions.

Next, we can split color spectrum into by IMDb ratings into three equal buckets Low (Less than 3.3), Mid (3.3-6.6), and High (Greater than 6.6). For the low rated movies, the dark colors occupy around 58.2% of the total spectrum. For high rated movies, the dark colors occupy 35.2% of the total color spectrum. This is a result of the Horror and Action genres which have an average rating of 4.89 and 5.72 respectively. In our prior analysis we saw that these two genres contain >55% dark colors.

We can see that the average IMDb rating for the Animation and Documentary genre is 6.84 and 7.28 respectively. We can try to create an Animation or Documentary genre poster to maximize our chances in achieving a high IMDb rating.

It appears we cannot simply judge a movie from its poster alone as there is a much larger signal coming from the movie's audio and visual, the dialogue, actors, the social media buzz, among other dimensions not included in our analysis. What is interesting is that there are signals within the poster which we can use to identify genres based on past poster designs, and we have been able to confirm some long-term trends in the posters which others have written about independently. For future iterations I would like to explore improving the accuracy of the machine learning algorithms and explore other dimensions such as number of votes, social media feeds, or trailers. Other considerations include using different JavaScript libraries such as D3, or Python's DataShader for visualizations of large datasets.

I challenge you to create a movie poster and try to beat my score!
Please send your works of art to me at hirohoso@.