PUBG Finish Placement Prediction
06/01/2019
Introduction
Battle Royale games are very popular these last few years, all are different in essence but keep the same principles, survive (alone or with teammates) on very large maps using a lot of resources for defense and attack (usually with weapons).
A very precise and abbreviated definition is the one that Gamespot used in its article on Battle Royale games:
To put it simply, battle royale is a large scale free-for-all deathmatch with the goal to be the last player alive. With only one life to live, you'll have to find any weapon you can--crowbars and frying pans included, depending on the game--while keeping focused as you stay one step ahead of the competition.
For this competition, Kaggle gave a very large dataset of PUBG matches, as they explain:
PlayerUnknown's BattleGrounds (PUBG) has enjoyed massive popularity. With over 50 million copies sold, it's the fifth best selling game of all time, and has millions of active monthly players.
The team at PUBG has made official game data available for the public to explore and scavenge outside of "The Blue Circle." This competition is not an official or affiliated PUBG site - Kaggle collected data made possible through the PUBG Developer API.
The ultimate goal is to predict the position in the match that a given player will have based on a lot of features, from the number of kills to how long he walked the map.
Exploratory Data Analysis
Kaggle hosted two different versions of the data set, the first month (with the first version) I did an exploratory data analysis trying to solve some hypotheses and find trends. Since the first version of the data set is no longer available, you can not play the next Jupyter Notebook:
Prediction using LightGBM
To the final prediction, I build a model using gradient boosting algorithm LightGBM, helped by some feature engineering and data processing the final public score was 0.0506 mean average error (MAE).
Conclusions
Battle Royale games, specially PUBG, has some options to each players who want to master the game and want to reach the highest positions always. With the exploratory data analysis the conclusion is that not only the player with higher kills is the one who will survive the most. Instead, keep moving and get the more boost items seems to be even a most efficient strategy.
Finally, LightGBM model told us that is completely possible to predict the position in every match with a very good average error of 0.0506 points in the percentile.