Analysis on Factors Contributing to Animes Popularities

Hi, my name is Jason, and my name is Apollo. Our final project for CSE 163 is on analyzing factors contributing to popular anime. You can access our full report at the link provided, but first and foremost, why does this matter? Most of us probably heard of or watched anime, such as Pokémon, Case Closed, and more. While they are bring us joy, anime production companies are affected by the current coronavirus situation. The production company for Asteroid in Love, which was founded 60 years ago, delayed their anime episode airing for the first time in history.

Therefore, in times of limited resources, 00:34 - we want to provide a guide that may help them predict which animes has the most potential to become popular, so they can prioritize resources on those to make high quality animes with maximum return. To achieve that goal, we’ll first visualize the historical trend of genres for popular anime. We’ll first use the pandas library to filter our dataset that contains information about all animes ever produced, group them by season and select the top animes, then use matplotlib library to show the trend. Secondly, we’ll train a machine learning model with scikit-learn to predict scores for an anime given relevant information such as genres, episodes count, duration, and more; then analyze which metrics are used in making such predictions to decide which factors influence anime popularities the most. For the first question, we produced the following 2 stacked area charts for genres of top anime.

As you may observe in our plot, Comedy, Action, 01:27 - Youngster, Drama, and School are the 5 most common genres. While there’s no genre that appears to gain more popularity over the past 5 years, genre supernatural, which is colored red in the first graph, was popular in recent years but audiences are starting to become less interested in those as the band shrinks in width significantly. We also plotted the popular genres of top 10 anime throughout the anime history. genre comedy is consistently bright, indicating it is popular for all times, as expected. Topics around space and giant robots was popular in the past, but no more.

Instead, we can 01:57 - see sparkles of genre game, and genre polygynous and polyandrous relationships as they become more popular. LGBTQ+ related anime remains a niche culture, which could be why Asteroid in Love, the anime we discussed before, was delayed. For research question 2, we use a decision tree to make predictions. As you can see, by going through along the tree, the model can derive as well as predict a possible score for an anime with given information. The order of the factors it considers is duration, source, and number of episodes.

In addition, source will be a huge influence of predicting the 02:29 - result, because animes came from a manga is usually because that manga has already achieved a huge popularity. However, this leads to several drawbacks for the model we’ve trained and have a bad prediction. For duration and episodes, MyAnimeList doesn’t have enough information for these two for those upcoming animes. Therefore, we also trained an adjusted model without duration and episodes. Here’s the prediction our model makes for animes in 2020 Spring, the upcoming season.

The final season of 02:59 - My Youth Romantic Comedy Is Wrong, As I Expected has a predicted score of 8.05 with the adjusted model. This is similar to scores it had for the previous two season, which are 8.09 and 8.28, respectively. As we calculated the mean squared error, or MSE, for our trained model, the MSE for testing dataset is much higher than that for the training dataset, indicating that our model is overfitted. Thus, to make improvements, we can adjust hyper-parameters to optimize the model, such as number of features and the depth of the decision tree. Additionally, we want to include information like staffs, licensors, and studios in a way that doesn’t reinforce existing biases and is fair to first- time employees and companies.

Lastly, we could also gather information and train models for other anime 03:51 - sites such as bilibili from China and niconico from Japan, for a more complete picture of anime popularity in different regions. Thank you! .