Modeling Superhero Alignment

The superhero universe of Marvel and DC comics is composed of a vast, amalgamated character data set of constructed characters. With the help of fan-edited scorecards of superhero stats (intelligence, strength, speed, durability, combat) and a large labeling of superhero powers, this study searches for profile insights of ‘good’ vs ‘bad’ aligned characters.

Question: What might be distinguishing superpower qualities of ‘good’ vs ‘bad’ types? If you’re not familiar with the comics, pause and consider the latest well-known superhero film series: “X-Men” and “Avengers”. Both commonly feature a most powerful supervillain supported by a small cadre of villains whose plans are thwarted by a larger coordinated coalition of good guys. Alternatively, in individual superhero films like “Spider-Man” and “Batman”, you might have a well-rounded good guy go up against a series of low-dimensionally-skilled bad types. How might these impressions from popular superhero films bear out in the broader comic book data?

Gathering a feel for the stat data, we see that ‘bad’ superheroes are on average more powerful than ‘good’ superheroes.

Additionally, ‘bad’ superheroes contain a larger population of “very strong” profiles that need to be defeated collectively by the ‘good’ guys. Further, the ‘bad’ superheroes have a low population of “weak” types - they shouldn’t be too easy to be defeated.

Individual stat categories also show a uniform distribution and not heavily correlated with other stat categories.

If we further include superpower labeling and impute the mean for characters with missing stats, we can build a model to seek insights. Initially, we perform a random forest model and obtain feature importance for targeting ‘Alignment’.

The stat categories dominate Random Forest split importance for our model accuracy. If we improve our model with XGBoost, we benefit from hyper-tuning of our feature parameters.

Permutation importance has high error.

Now, examining Partial Dependence Plots reveals the accuracy of superhero impression discussed earlier from films.

As seen earlier from ‘good’ vs ‘bad’ superhero distributions, the “very strong” characters weight the predictions toward ‘bad’.

Moderate individual stats push a character towards ‘good’ until too much individual power reverses the trend towards ‘bad’.

Related posts:

Plenty of evidence is out there demonstrating the therapeutic potential of cannabinoids in the management of difficult to treat pain, even to synergize with a concomitant opioid medication regime…

Every human being on the planet has a deep seated desire to be “known”. A craving to be understood and accepted. Let me introduce you to one of the most life-changing pleasures of prayer — the fact…

Last week I got an assignment that was due yesterday. Before I got this assignment, I had decided that I’ll be completing my assignments the day I’m assigned one. But I kept on saying to myself…