DSLR

Hogwarts student records: 1,600 labelled training students, 13 course scores, 4 houses; the official 400-student test set is unlabelled.

Solo

1. Reimplemented a pandas-style describe from scratch: count, mean, standard deviation, quartiles with interpolation, variance, skewness and kurtosis — no NumPy or pandas.
2. Implemented logistic regression from scratch: sigmoid, log-loss, gradient and z-score standardisation, with batch, stochastic and mini-batch gradient descent as interchangeable variants.
3. Built the one-vs-all multiclass scheme: one binary classifier per house, prediction by argmax over the four sigmoid probabilities, weights exported to JSON.
4. Wrote the three exploratory visualisations by hand with Matplotlib (histograms, scatter plots, a coloured pair plot) for feature selection.

One-vs-all logistic regression

Four independent binary classifiers, one per house, and a student is assigned the house with the highest sigmoid probability.

Three selectable gradient-descent variants

Batch, stochastic and mini-batch, chosen at runtime, so the optimisation strategies can be compared.

Manual z-score standardisation

Every feature is standardised by hand before training and prediction, so scores on very different scales stay comparable.

On the 1,600 training students the model reproduces the house labels with 97.9 % accuracy (34 errors). This is a training-set score, measured on the same data it was trained on — there is no held-out test set.

• No held-out evaluation: the official test set is unlabelled, so the only accuracy is measured on the training set itself — not a generalisation score.
• No regularisation and no early stopping; a fixed learning rate and hand-tuned iteration counts.
• The shipped model uses all thirteen features; the feature-selection code is left commented out.
• Minimal cleaning: non-numeric or missing values are replaced by zero rather than imputed.