5G User Prediction

Predicting whether a telecom user is a 5G user from billing, traffic, activity, package and regional features. This is a highly class-imbalanced binary classification problem evaluated with AUC.

A course group project (Artificial Intelligence). Implemented and compared three models: Logistic Regression, Random Forest and LightGBM.

Task

Given 58 anonymised features (20 categorical cat_* and 38 numerical num_*) plus an id, predict the binary target (1 = 5G user, 0 = non-5G user). The dataset has 800,000 samples, no missing values, and the positive class makes up only 1.32% of them. The key challenges are therefore class imbalance and weak linear signal.

Approach

All models share an identical stratified 80/20 train/test split and are evaluated with 5-fold cross-validation on the training set.

Model	Type	Categorical encoding	Imbalance handling
Logistic Regression	linear baseline	one-hot + standardisation	class_weight='balanced'
Random Forest	bagging ensemble	ordinal encoding	class_weight='balanced'
LightGBM	gradient boosting	native categorical	scale_pos_weight

cat_12 has ~110k unique values and behaves like a high-cardinality continuous variable, so it is kept as a numeric feature instead of being one-hot encoded.

Results

Model	CV AUC	Test AUC	Test AP
LogisticRegression	0.8776	0.8740	0.0874
RandomForest	0.9122	0.9132	0.1386
LightGBM	0.9036	0.9029	0.1383

Both tree-based ensembles clearly outperform the linear baseline. This matches the EDA finding that feature–target correlations are weak (max |r| ≈ 0.12) and the discriminative signal is largely non-linear. Given the severe imbalance, the PR curve (AP) is more informative than ROC for the positive class.

Project Structure

.
├── 5g_user_prediction.ipynb      # main notebook: EDA, training, comparison
├── solution.py                   # same pipeline as a runnable script
├── requirements.txt
├── README.md
└── output/
    ├── figures/                  # EDA and result plots
    ├── submission.csv            # best-model predictions (generated, git-ignored)
    └── cv_results.csv            # cross-validation summary (generated, git-ignored)

train.csv is not included — it is large (282 MB) and should be placed in the project root before running.

Requirements

• Python 3.10+
• See requirements.txt

pip install -r requirements.txt

Reproducibility

# place train.csv in the project root, then
python solution.py          # trains all models, writes output/
jupyter notebook 5g_user_prediction.ipynb   # interactive version

All randomness is seeded (random_state=42), so every figure and result file under output/ is fully reproducible.