Baseline Model

Model used:

• Random Forest Classifier
• feature to be predicted: position

Quantitative Features

• kills
• deaths
• assists
• damageshare
• wardsplaced
• earnedgoldshare
• total: 6

Nominal Features

• year
• patch
• total: 2

Ordinal Features

• no ordinal features were used
• total: 0

Encodings

• RFC don’t need numerical data to be scaled or altered, so all quantitative features remained as is
• We used One Hot Encoding on the year and patch features
  • although year and patch are stored as numbers, they have no order or weight

Model Performance

The five different positions are model can predict are:

• bot
• jng
• mid
• sup
• top

Because the consequence of an incorrect prediction is the same for all positions, we will be using accuracy as our main evaluation metric, as opposed to precision, recall, or F1 score

Model hyperparameters:

These were arbitrarily chosen, just to give us a reference of how good the model would be without any tuning

• max_depth: None
  • None because we were curious to see how far down a decision tree would go without limiting it
• n_estimators: 300
  • 300 was set because it is a decently high number for the size of our dataset

Training Data

• accuracy score: 100%
• this means there is a very good chance the model took advantage of having no limit for max_depth and the random forest has many deep trees
• very good change this model overfit to the data

Testing Data

• accuracy score: 62.76%
• comparing this to the training data score, it is clear our baseline model was overfitting
• difference in training and testing scores: 37.24%

Conclusion

We need to tune our model’s hyperparameters so that it does not overfit