DSLR — DataScience x Logistic Regression

A from-scratch implementation of logistic regression to sort Hogwarts students into their houses, inspired by the 42 school project. The classifier is trained with one-vs-all logistic regression using gradient descent, and comes with data exploration tools built without high-level statistical libraries.

Project structure

.
├── datasets/
│   ├── dataset_train.csv   # Labeled training data
│   └── dataset_test.csv    # Unlabeled test data
├── describe.py             # Statistical summary (reimplements pandas describe)
├── histogram.py            # Score distribution per house
├── scatter_plot.py         # Most correlated feature pair
├── pair_plot.py            # Full feature pair plot
├── train.py                # Train the logistic regression model
├── predict.py              # Predict house from trained weights
└── requirements.txt

Installation

pip install -r requirements.txt

Usage

1. Describe the dataset

Prints count, mean, std, min, quartiles, max, range, and IQR for every numeric feature — computed without pandas.describe().

python describe.py datasets/dataset_train.csv

2. Visualize the data

Histogram — shows score distributions per house to identify homogeneous features:

python histogram.py datasets/dataset_train.csv

Scatter plot — automatically finds and plots the two most correlated features:

python scatter_plot.py datasets/dataset_train.csv

Pair plot — full pairwise feature grid colored by house:

python pair_plot.py datasets/dataset_train.csv

3. Train the model

Trains a one-vs-all logistic regression and saves weights to weight.csv.

# Batch gradient descent (default)
python train.py gradientdescent datasets/dataset_train.csv

# Stochastic gradient descent
python train.py SGD datasets/dataset_train.csv

# Mini-batch gradient descent (batch size 64)
python train.py minibatch datasets/dataset_train.csv

4. Predict

Applies the saved weights to a dataset and writes predictions to houses.csv. If the dataset includes the true house labels, also prints accuracy.

python predict.py datasets/dataset_test.csv weight.csv

How it works

• Preprocessing: drops low-information features (Arithmancy, Care of Magical Creatures), fills missing values with the training mean, then standardizes all features.
• Training: one sigmoid classifier is trained per house (Ravenclaw, Slytherin, Gryffindor, Hufflepuff) via gradient descent. Three optimizers are available: full batch, SGD, and mini-batch.
• Prediction: each student is assigned the house whose classifier returns the highest score (argmax over the four sigmoid outputs).

Dependencies

• pandas
• numpy
• matplotlib
• seaborn
• scikit-learn (accuracy score in predict.py only)