League of Legends: An Exploratory Data Analysis
By: Brandon Zhao (Project Lead), Zoeb Jamal, Ryan Barney, Kushaal Madadi, Yashas Jain
Introduction
League of Legends, also known as League, is one of the most popular video games in the world. It is played by over 100 million active users every month, and is the most popular esport in the industry with an international competitive scene composed of 12 leagues. In 2019, the game regularly peaked at eight million concurrent players, and its popularity has led to tie-ins such as music videos, comic books, short stories, and an animated series, Arcane.
Developed by Riot Games, League is a Free-to-Play team-based strategy multiplayer game where two teams of five powerful champions battle in player versus player combat, each team occupying and defending their half of the map. Each of the ten players controls a character, known as a “champion”, with unique abilities and differing styles of play. While the map shown below stays the same, the different combinations between 157 champions chosen by the ten players in-game can create vastly different rhythms and feels. Whether you like dealing damage from afar, brawling at close quarters, or helping your team with utility, there’s a champion for you.
During a match, champions become more powerful by collecting experience points, earning gold, and purchasing items to defeat the opposing team. The ultimate goal of League is to destroy the other team’s base, known as “Nexus”, but it’s not easy. Your enemies will do everything they can to kill you and destroy your base. Each base has a series of turrets and waves of minions that constantly spawn, as well as neutral monsters within the Jungle that can grant points or power ups. Your team needs to clear at least one lane to get to the enemy Nexus.
General Gameplay
While games last half an hour on average, the first 10–15 minutes are undoubtedly the most important, largely due to the fact that you can make your champion stronger as you buy more items, level up, and get buffs from neutral objectives in game. You have to ensure that your team is well set up for an easy victory, but also make sure that you put the other team as far behind as possible. Given this fact, it is no surprise that the average match length decreases significantly as skill level goes up. As you go from an iron ranking to challenger, the match times go down from 30 mins to 25 mins, as better players are able to capitalize on the small advantages gained in the early game. We were able to analyze a dataset that collected statistics from the first 15 mins of diamond ranked games (one of the higher tiers) and first conducted a Principal component analysis to see if we could observe any clustering. We scaled the data and reduced the dimensionality down to two PCs and colored points by the winning team.
We noticed that the clustering pattern was extremely symmetrical, which is obvious, given what we just mentioned about how better players are able to snowball a small advantage into a win.
After seeing how significant the clustering is, we wondered what statistics are most important when it comes to winning a game. In our dataset, we had access to statistics like team gold, average level, kills, towers, etc. at 15 minutes. We first made new boolean columns to indicate whether the blue team or red team had a higher value, and then checked the correlation with the blue_win column. We found that the most highly correlated statistic was the gold level at 0.67, followed closely by average level at 0.65 and total kills at 0.649. This indicates that gold level, average level, and total kills at the 15 minute mark all have a strong, positive relationship with the eventual winner of the game.
Competitive Scene
League of Legends has one of the largest and healthiest competitive scenes in all of esports. Perhaps due to its popularity among the general population, League of Legends has managed to foster a massive community of those who not only want to be as competitive in the game as possible, but those who’d choose to form a career around it and go professional. Every year, there are a myriad of tournaments worldwide where the best players can compete for massive cash prizes. These tournaments garner a lot of attention from spectators of all levels, especially since they reflect what the highest achievable level of gameplay currently looks like. As a result, these tournaments can be a good indicator of the current meta, or state, of the game. For example, the performance and pick rate of certain champions at the professional level can show us where the meta currently stands or in what direction it may be heading. As such, examining the data of large tournaments, such as Worlds 2021, is important if we want to determine the state of the game as well as how it may affect casual and competitive play alike in the future.
Beginning with individual champions, it’s interesting to compare their performances at the tournament. Champions may be played in different positions, but given a position that a champion is currently playing in, we can easily determine how they compare to other champions that also play in that position.Taking a look at the following bar graph, which measures a champion’s KDA given its position, this becomes pretty clear. This KDA metric totals the kills, deaths, and assists by each player within a game, and it is a common metric to use when determining a player’s performance in a game. Comparing champions within the position they played, it’s clear that there were a few winners in each category. In essence, some champions simply had a higher average KDA at the tournament compared to others in their respective positions. This may provide an indication as to which champions are better, and it could give us insight as to which champions are more meta-defining than others. In other words, picking the champions that had the highest KDA could translate to being a better contributor to team success. However, this is not entirely true. This graphic alone cannot tell us how often each champion was picked, which is just as important as performance when it comes to defining the state of the game. Additionally, there’s the huge variable in player performance. This graph analyzes pro player performances, which is subject to skewing if one player is better than another, a player is not performing at their best, or a player is performing exceptionally well. For example, a few bad games or a few really great games would not be able to reflect the actual capabilities of the champion, and it might cause a champion that’s generally considered bad to appear good or a champion that’s generally considered good to appear bad. Also, differing player skill level might just mean the champion the best players pick appears the best even if, on paper, that champion isn’t agreed upon as the best.
Despite these flaws, this data does provide a good reflection of which champions did well at the tournament, which in turn affects how the game will be played in the future. The casual players who spectate the professional scene are subject to being swayed into choosing to play the champions that did the best at the tournament. After all, seeing someone do well at the tournament might end up causing many casual players to try to pick up that champion and try to replicate that player’s performance. In the future, there might be a lot more of those high KDA champions appearing in your next game of League.
Beyond individual champion performance, position and role are pivotal components to analyzing the state of the game. Starting with position, the following bar chart depicts the average kills and deaths players of each position had in games from preliminaries of Worlds 2021. As seen from the graph, the kill-death ratio varies highly with the role being played. For example, Attack Damage Carry players, often shortened as ADC players, have very high kill-death ratios per game on average, while support players have very low kill-death ratios per game on average. This demonstrates that, at the professional level, it is much more common and important to fulfill your position’s responsibilities. At a casual level, many players are guilty of being too obsessed with their own kill-death ratio no matter what position they play, but if they can maybe improve their play and win rates in the long-term by focusing more on making sure they are able to accomplish the tasks of their roles.
Next, this following graph demonstrates the pick rate of certain roles at the tournament, with larger bubbles indicating a higher pick rate. Naturally, this reveals to us what roles are most important when it comes to constructing a successful team. While this graph alone cannot tell us the success rate of a team given the roles they picked, it can tell us how professional players choose to construct their team. Similar to individual champion performance, this graph could help us determine which roles are meta-defining and which roles may become more popular in the future. As was mentioned previously, spectators tend to emulate the success of professionals, so the roles with higher pick rates may become a lot more prominent in the future state of the game. While we do not know how these roles performed, their high pick rate suggests that the professionals view them highly, which means the general population of players will end up doing the same.
Altogether, there is a lot that can be gained from observing the data at a professional level. Examining and analyzing professional performances gives great insight into how the game’s meta can be defined now and where it will be heading in the near future.
Modelling
In order to further understand the different strategies that players of League of Legends should implement in their game, we created a classification model to predict the results of League of Legends of matches from Master, GrandMaster, and Challenger ranked games. To do this, we implemented a Decision Tree Model to try and predict the results of a match. Using an initial maximum depth of 3, which refers to the number of layers within a decision tree, we were able to create a model that predicted the winner of matches very well, at an average accuracy of 94%. The results of the decision tree are shown below.
Based on the results of the importance of different features and the graph of the Decision Tree, it can be seen that the number of tower kills for the red and blue teams are the most important statistics when predicting who will win a match. This makes sense as the number of tower kills is often the determining factor in which team will be able to destroy the opposing team’s nexus first. When we increase the maximum depth of the tree to 20, while the graph of the decision tree is no longer readable, we see that the trend of number of tower kills being most important still holds.
To investigate what other features are important in predicting who will win a match, we removed the number of tower kills for the blue and red teams as features and similarly created new Decision Tree Models to predict the results of the same matches. Using the same initial maximum depth of 3, we were still able to create a model that predicted the winner of matches very well despite not having access to the number of towers killed by the teams, at an average accuracy of 93%. The results of the decision tree are shown below.
These new results show that, without information on the number of towers killed, the number of inhibitors killed and the number of enemy kills are the most important predictors for who will win a match. The results of all of these models indicate a clear strategy for League of Legends players in order to win more matches. Players should first focus on maximizing the number of enemy towers that they destroy while minimizing the number of their own towers that are destroyed. Players should secondly focus on maximizing the number of inhibitors destroyed and enemy kills that their team gets, while minimizing the number of inhibitors destroyed and deaths on their own team.
