reefzehavi/readme_2

Assignment 2: Classification, Regression & Clustering

Student Name: Reef Zehavi Date: November 23, 2025

---

1. Overview

In this assignment, I applied various Machine Learning techniques to predict corporate bankruptcy using the 'US Company Bankruptcy Prediction' dataset. The project is divided into three parts:

1. Classification: Predicting whether a company will go bankrupt (Binary Target).
2. Regression: Predicting the Net Income (Continuous Target).
3. Clustering: Grouping companies based on financial similarities (Unsupervised).

---

2. Part 1: Classification (Bankruptcy Prediction)

I compared a Baseline model (Logistic Regression) with an Advanced model (Random Forest). The data was highly imbalanced, so I used class_weight='balanced'.

Results:

• Logistic Regression (Baseline):
  • Recall (Class 1): 0.77 (Very High) - Caught most bankruptcies.
  • Precision: Low (0.16) - Many false alarms.
• Random Forest (Advanced):
  • Recall (Class 1): 0.22 (Low) - Missed many bankruptcies.
  • Accuracy: 96% (High, but misleading due to imbalance).

Conclusion: While Random Forest had better accuracy, Logistic Regression was arguably better for this specific task because it successfully identified most of the companies at risk (High Recall), which is crucial for bankruptcy prediction.

---

3. Part 2: Regression (Predicting Net Income)

I attempted to predict the Net Income (X1) based on all other financial features.

Results:

• Linear Regression:
  • R2 Score: 0.99 (Perfect fit).
  • RMSE: Very low.
• Random Forest Regressor:
  • R2 Score: ~0.68.
  • RMSE: Higher errors.

Conclusion: The relationships in financial statements are often linear (e.g., Income = Revenue - Expenses). Therefore, the simple Linear Regression model vastly outperformed the complex Random Forest model, which struggled to capture the exact linear formula.

---

4. Part 3: Clustering (K-Means)

I used K-Means to cluster companies into 3 groups based on financial data (without labels) and visualized it using PCA.

Findings:

• Cluster 0 & 1: Contained "safe" companies with almost zero bankruptcy cases.
• Cluster 2: The largest cluster, which contained 100% of the bankruptcy cases (3,040 companies).
• Insight: The unsupervised algorithm successfully separated "very safe" companies (Clusters 0 & 1) from the rest. However, it could not separate bankrupt companies into their own unique cluster; they were mixed within the general population of Cluster 2.

---

5. Final Thoughts

This assignment demonstrated that:

1. Simpler is sometimes better: Linear Regression beat Random Forest in regression.
2. Imbalance matters: Accuracy is not a good metric for rare events like bankruptcy; Recall is key.
3. Unsupervised limits: Clustering can group similar companies but isn't enough on its own to isolate failures without labels.