Preview mode. Browse milestones and edit starter code here. Sign in to run and submit.
0/8EDA
Python
intermediate~6h8 milestones
The problem
A product team needs a classifier for a sensor-derived feature set where the decision boundary is nonlinear. A linear model is the honest baseline, but it will not hit production quality. Your job is to build a multilayer perceptron path from raw table to evaluated network, without pretending the linear baseline was enough.
Your task · Explore the Feature Table
Quantify shape, class balance, and feature scale before modeling.
Implement explore_nn(train_df) returning:
row_count
n_features: columns excluding label
n_classes: distinct label values
positive_rate: mean of label, rounded to 5 places
feature_std_mean: mean of per-feature standard deviations (label excluded), rounded to 5 places
Constraints
▪No external downloads. train.csv is your only training data
▪Use scikit-learn in the sandbox (numpy, pandas, sklearn)
▪Hidden labels are scored server-side
▪Scale features with statistics fit on training data only
▪Keep training inside the sandbox time budget
Graded on
Majority baseline: accuracy >= 0.60
Logistic baseline: accuracy >= 0.65
MLP: accuracy >= 0.85, F1 >= 0.80, ROC-AUC >= 0.90
Regularized MLP still clears accuracy >= 0.85 and ROC-AUC >= 0.90