🐍 Python → Machine Learning — Hands-On Journey

Every chart in this repo was generated by code I wrote myself. No tutorials copied. No AI-generated code. Real datasets. Real results.

---

👩‍💻 About Me

I'm Nidhi Zalavadiya — a Python developer from Junagadh, Gujarat with 2+ years building production automation systems. I'm now applying that engineering mindset to Machine Learning — writing clean, readable code and measuring every result.

I don't just run algorithms. I ask: why did this model perform better? What does this feature importance tell me about the data?

---

🔬 What I've Built & Learned

📊 Student Score Analysis — Pandas + Matplotlib/Seaborn

Real dataset. 4 different chart types. Data cleaning + visualization pipeline.

What I did:

• Loaded and cleaned a student dataset with city, Python score, and ML score columns
• Built 4 chart types in one figure: bar chart, line plot, scatter plot, histogram
• Used groupby() to find average scores per city — Surat students scored the highest
• Identified a correlation between Python score and ML score using scatter

Code skills demonstrated: pandas, matplotlib.pyplot, seaborn, groupby, subplots.

---

🌸 Iris Dataset — EDA & Feature Visualization

Classic ML dataset. Used scatter plots to visually confirm class separability before modeling.

What I did:

• Loaded Iris dataset from sklearn, converted to a DataFrame
• Built side-by-side scatter plots: Sepal features vs Petal features
• Key insight: Petal features separate the 3 classes cleanly — Sepal features overlap. This means petal measurements will be more important features in any classifier.

Code skills demonstrated: sklearn.datasets, seaborn.scatterplot, feature analysis thinking

---

📈 Linear Regression — California House Price Predictor

Regression on real data. Visualized actual vs predicted. Analyzed which features drive price.

What I did:

• Trained Linear Regression on California Housing dataset (20,000+ samples)
• Plotted Actual vs Predicted — model follows the trend but struggles at high prices (common in linear models)
• Analyzed feature coefficients: AveBedrms and MedInc are the strongest positive predictors; Latitude and Longitude are negative (location effect)

Key learning: A straight line can't capture non-linear price jumps — this is exactly why we need tree-based models.

Code skills demonstrated: LinearRegression, train_test_split, coefficient analysis, matplotlib

---

🎯 Logistic Regression — Titanic Survival Classifier

Binary classification. Confusion matrix analysis. Understanding the precision vs recall tradeoff.

What I did:

• Built a Logistic Regression classifier on Titanic dataset
• Achieved ~82% accuracy — correctly predicted 90 non-survivors and 55 survivors
• Analyzed confusion matrix: 19 false negatives (predicted not survived, actually survived) — in a real scenario, this type of error matters more than false positives
• Visualized survival by gender: females survived at much higher rates (women & children first policy confirmed in data)

Key learning: Accuracy alone is misleading. A model predicting "nobody survives" gets 61% accuracy on Titanic — but it's useless. The confusion matrix tells the real story.

Code skills demonstrated: LogisticRegression, confusion_matrix, classification_report, seaborn.heatmap

---

🌳 Decision Tree — Titanic (max_depth=4)

Visualized the full decision tree. Understood how the model actually splits data.

What I did:

• Trained Decision Tree with max_depth=4 to prevent overfitting
• Visualized the full tree — first split is on sex <= 0.5 (encoded: male=1, female=0)
• The tree confirms what history tells us: females (sex=False) go right → most survived. Males (sex=True) go left → most didn't survive.
• Analyzed feature importance — sex dominates at 0.58, pclass second at 0.20

Key learning: Decision Trees are explainable — you can trace exactly why a prediction was made. This matters in real business applications (banking, healthcare, legal).

Code skills demonstrated: DecisionTreeClassifier, plot_tree, feature_importances_, Gini impurity understanding

---

🌲 Random Forest — Feature Importance & Model Comparison

Ensemble learning. Compared 4 models on same dataset. Understood when Random Forest beats Decision Tree.

What I did:

• Trained Random Forest on Titanic — feature importance shifted significantly vs Decision Tree
• In Decision Tree: sex = 0.58 dominant. In Random Forest: fare, age, sex are all ~0.24-0.26 — more balanced, more realistic
• This shows Random Forest is less biased toward one feature — it samples features randomly across trees

Model Comparison on Titanic Survival:

Model	Accuracy
KNN	~72%
Decision Tree	~80%
Random Forest	~80%
Logistic Regression	~81% ✅ Best

Key learning: More complex ≠ always better. Logistic Regression outperformed Random Forest here because Titanic survival has strong linear relationships (sex, class). Random Forest wins on complex, non-linear data.

Code skills demonstrated: RandomForestClassifier, feature_importances_, model comparison, choosing the right model

---

🛠️ Tech Stack Used

---

📁 Repo Structure

Python-Fundamentals/
│
├── 🐍 Python Core (Day 01–13)
│   ├── day01.py – day06.py          # Data types, OOP, functions
│   ├── day03_noon_2sum.py            # LeetCode: Two Sum
│   ├── day05_leetcode_*.py           # LeetCode: Palindrome, Contains Duplicate
│   ├── day04_student_management_system.py
│   └── review_week01.py             # CLI Grade Manager (Week exam)
│
├── 📊 Data & ML (Day 09–13)
│   ├── day09_iris.py                 # EDA — Iris scatter plots
│   ├── day10_linear_regression.py    # House price predictor
│   ├── day11_logistic_regression.py  # Titanic — Logistic Regression
│   ├── day12_decision_tree.py        # Titanic — Decision Tree + visualization
│   ├── day13_random_forest.py        # Titanic — RF + model comparison
│   └── Student_Data_Analysis.py     # Custom dataset — 4-chart EDA
│
├── 📂 images/                        # All output charts (auto-generated by code)
├── 📂 models/                        # OOP model classes
└── 📂 utils/                         # Helper utilities

---

🔮 What's Next

• Model evaluation deep dive — Cross-validation, GridSearchCV, ROC-AUC
• KMeans Clustering — Customer segmentation project
• Neural Networks — MNIST digit classifier with Keras
• NLP — Sentiment analysis on product reviews
• HuggingFace — Load and run a pre-trained model
• LangChain — RAG app deployed on Streamlit
• Capstone: End-to-end ML project with domain data (ERP/Financial)

---

🤝 Connect

---

"I don't just run models. I understand what they're telling me about the data."