josalhor / si-pokemon

Introduction

This repository hosts the code for the third deliverable of the subject. The authors are:

• Joan Palau Oncins
• Josep Maria salvia Hornos

Challanges

The most challanging parts have been:

1. Analysing the attributes and determine which one's to take
2. Thinking out of the box in orther to preprocess the information in order to correct the data, improve the data or aggregate the data
3. Once the train score was above the 85% accuracy, each iteration provided few upgrades, overfitting was really easy and you gainned score in the train data and decreased in the test data.

Data preprocessing

The most difficult part regarding this activity has been by far the data preproccessing step.

Inside the pipline we have implemented several custom transformers that have allowed us to discover different combinations of attributes, group data or make data more consisten which have greatly contributed to the model accuracy.

Attribute selection

The 5 most important features being used by the models are by order the following:

1. velocity_diff_binary
2. sum_stats_diff
3. sum_attack_stats_diff
4. sum_attack_ratio_diff
5. best_attack_ratio Although here there are 5 listed features, the most important feature by far us the velocity_diff_binary

Models

Our first approach was to compare between the following models

• DecisionTreeClassifier
• RandomForestClassifier
• MLPClassifier We finally discarted the first model as it was less performant than the RandomForestClassifier and both of them take the "same" aproach. However, the DecisionTreeClassifier model granted us the possibility to see how was the classification being done with the following code:

from sklearn import tree
from sklearn.tree import DecisionTreeClassifier
from sklearn.model_selection import train_test_split, cross_val_score

from matplotlib import pyplot as plt

dtc_model = DecisionTreeClassifier(criterion="entropy", max_depth=3, min_samples_leaf=5, random_state=42)
dtc_model.fit(X_train_encoded, y_train)
'Train Score', dtc_model.score(X_train_encoded, y_train)

plt.figure(figsize=(30, 20))
tree.plot_tree(dtc_model, class_names=["False", "True"], feature_names=DropTransformer.KEEP_COLUMNS)
plt.show()

The result was seeing that the tree only cared about the speed, and that unless the depth was considerable (overfitting), other features would not influence the output of it.