Aion Player Stress Detection

Overview

This project focuses on automatic detection of player fatigue and stress using in-game behavioral data from the MMORPG Aion.
Using decision tree classifiers, we model whether a player shows signs of stress based on gameplay metrics, with special attention to class imbalance, false negatives, and model interpretability.

The motivation is practical: prolonged gameplay can negatively impact mental health, and regulations on digital wellbeing increasingly require early detection and preventive intervention. An interpretable model such as a decision tree allows actionable decisions while remaining auditable.

---

Project Stages

1. Dataset Preparation

The raw dataset contains gameplay metrics for thousands of players, including missing values and unlabeled cases.

Key steps:

• Load and inspect raw gameplay data.
• Preserve all variables (no feature removal).
• Remove only rows with invalid or missing labels where required by the model.
• Separate labeled players from unlabeled ones for later inference.
• Produce a cleaned dataset suitable for supervised learning.

Output:

• dades_aion_preparat.csv – cleaned and model-ready dataset.

Dataset Summary

Aspect	Value
Game	Aion
Players	9,721
Features	25 (numerical)
Target variable	class (0 / 1)
Missing labels	Yes (NaN)
Class imbalance	Severe (class 1 minor)

---

2. Building the Initial Decision Tree

A baseline DecisionTreeClassifier is trained using:

• Default hyperparameters.
• criterion="entropy" to maximize information gain.

Evaluation focuses not only on accuracy, but on:

• Precision, recall, F1-score
• False Positive Rate (FPR)
• False Negative Rate (FNR)

This stage establishes a reference point and exposes overfitting and sensitivity issues, especially for the minority class (stressed players).

---

3. Modifying the Decision Tree

To reduce overfitting, the decision tree is constrained by:

• Limiting maximum depth (max_depth).

This forces simpler decision rules and improves generalization, at the cost of reduced sensitivity to stress cases.
The trade-off between model complexity and recall is explicitly analyzed.

---

4. Comparing Algorithms

A DummyClassifier with strategy most_frequent is used as a baseline.

Purpose:

• Demonstrate why accuracy alone is misleading.
• Highlight the impact of severe class imbalance.
• Show that trivial models can achieve high accuracy while completely failing at stress detection.

This comparison validates the decision tree as a meaningful model rather than an illusion of performance.

Model Comparison Table

Model	Recall (Class 1)	FNR	FPR	Overfitting
Decision Tree (full)	High	Medium	Low	High
Decision Tree (depth=5)	Medium	Medium	Medium	Low
Dummy (most_frequent)	0.00	1.00	0.00	None
Optimized Decision Tree	Highest	Lowest	Low	Low

---

5. Model Optimization

The final stage improves the decision tree without changing the algorithm.

Techniques applied:

• Stratified train/test split.
• Feature scaling for consistency.
• Hyperparameter tuning using GridSearchCV.
• Optimization driven by F1-macro, not accuracy.

The optimized model achieves:

• Better balance between FPR and FNR.
• Improved detection of stressed players.
• Reduced overfitting.
• Applicability to previously unlabeled players.

---

Conclusion

This project demonstrates that:

• Decision trees are viable and interpretable tools for stress detection in gaming data.
• Class imbalance is the dominant challenge and must guide evaluation.
• Accuracy is insufficient; FNR and recall are critical in health-related contexts.
• Controlled model complexity and proper optimization significantly improve reliability.

The resulting model can be used as a pre-screening tool to prioritize manual review and preventive actions in digital wellbeing systems.

---

Repository Structure

data/
├── dades_aion.csv
├── dades_aion_preparat.csv

src/
├── decision_tree.py
├── optimitzacio.py
README.md