Data Cleaning and EDA

Question: Do tanks/bruiser have higher damage mitigated per minute compared to tanky supports?

This section is dedicated to describing how we cleaned and pre-processed the data.

Here’s our initial data:

print(raw.head())

This data has too many unnecessary columns that aren’t needed such as first dragon, first baron, golddiffat20 which won’t answer our question. Thus, we dropped the majority of the columns. After dropping columns, we filtered out the incomplete data using the “completness” column, and then dropped all NA.

Here’s our cleaned data:

print(cleaned_df.head())

We then created a new column, “class” and “class_actual”, which shows whether or not a champion is a tank/bruiser or a tanky support. To determine the class, we filtered for champions with high damage mitigated per minute and amounts of wards placed, and compared it to “class_actual” which is a manually mapped class to compare accuracy.

accuracy = (cleaned_df['class'] == cleaned_df['class_actual']).mean()
accuracy
0.7817412624393058

Univariate Analysis:

This bimodal curve distribution tells us that there are two distinct groups of gold-earners within this dataset
Makes sense because certain champions and positions are expected to earn gold on different levels

Bivariate Analysis:

Because of the variance in stats for each position, we can have some confidence that this will be a good metric for predicting positions
It is clear that top and jungle positions take the most damage while bottom and support take the least

Aggregation:

this shows us that tank/bruiser and tanky support have similar levels of dmpm
hard to be confident in true levels because there are significantly less tanky supports seen
this is why we are going to run a hypothesis test, to see how significant the differences in dmpm are

df.pivot_table(index = 'class', values = 'damagemitigatedperminute', aggfunc = ['mean', 'sum'])

This table shows the same thing as the graph above but in table-form

Assesment of missingness:

A lot of the data is missing in general which is shown in the column “datacompleteness”, which we already filtered for completeness. This means that any missing data left is NMAR, for example, ‘goldat25mins’ has some missing values which can be expected because not all games last 25 mins long. Thus it’s NMAR.

Tanks/Bruisers tend to destroy towers more often than tanky supports, becaues that is one of the jobs of a top laner. One column that depends on another is towers and gamelength, which the longer the game happens, the more towers are destroyed:

nmar_col = 'towers'
other_col = 'gamelength'
miss = raw.dropna(subset = [other_col])[[other_col, nmar_col]].assign(missing = raw[nmar_col].isna())
observed_diff = miss.groupby('missing')[other_col].mean().diff().iloc[1]
reps = 500
diffs = list()
for _ in range(reps):
    shuf_df = miss.assign(shuffled = np.random.permutation(miss[nmar_col]))
    diff = shuf_df.groupby('shuffled')[other_col].mean().diff().iloc[1]
    diffs.append(diff)
fig = px.histogram(np.array(diffs))
fig.add_vline(x = observed_diff, line_color = 'red')
fig.show()
print((np.array(diffs) <= observed_diff).mean())

.498

Tanks/Bruisers destroying towers more often than tanky supports should not affect Barons, as Barons require a team to kill. One columns that does not depend on another is towers and barons, as they are unrelated in missingness:

nmar_col = 'towers'
other_col = 'barons'
miss = raw.dropna(subset = [other_col])[[other_col, nmar_col]].assign(missing = raw[nmar_col].isna())
observed_diff = miss.groupby('missing')[other_col].mean().diff().iloc[1]
reps = 500
diffs = list()
for _ in range(reps):
    shuf_df = miss.assign(shuffled = np.random.permutation(miss[nmar_col]))
    diff = shuf_df.groupby('shuffled')[other_col].mean().diff().iloc[1]
    diffs.append(diff)
fig = px.histogram(np.array(diffs))
fig.add_vline(x = observed_diff, line_color = 'red')
fig.show()
print((np.array(diffs) <= observed_diff).mean())

0.0

LoL Position Classification Data Analysis

Data Cleaning and EDA