Benchmarking Tabular Imputation Methods for 6G Wireless Datasets

Code for the CSE3000 Research Project. This repository benchmarks seven tabular imputation methods across four missingness mechanisms and three missing rates on two DeepSense 6G scenarios, evaluated on reconstruction error, distributional fidelity, and downstream beam-prediction accuracy.

It includes a bootstrap script that builds a Python virtual environment with all the imputation libraries used in the project. Tested on Linux, macOS, and Windows via WSL (recommended).

Imputation methods covered

The environment supports the full method shortlist:

• Mean — scikit-learn (SimpleImputer)
• kNN — scikit-learn (KNNImputer)
• MICE — scikit-learn (IterativeImputer)
• SoftImpute — fancyimpute.SoftImpute (low-rank, important for CSI data)
• HyperImpute — hyperimpute package
• GRAPE — bundled inside DiffPuter at external/DiffPuter/baselines/GRAPE/
• DiffPuter — included as a git submodule at external/DiffPuter/

DiffPuter is a git submodule (not pip-installed) because it's a research codebase rather than a packaged library. The submodule is pinned to a specific commit via .gitmodules, so everyone builds against the same source.

Repository layout

src/common/        shared experiment code (amputation, metrics, datasets, tuning, ...)
src/imputers/      imputer wrappers (imputers.py, grape_imputer.py, diffputer_imputer.py)
notebooks/         tuning.ipynb, imputation.ipynb, downstream.ipynb
results/csv/       result tables the paper is built from
results/figures/   generated figures
results/           cached imputations (.npy) and tuned_params.json
scripts/           setup_imputation_env.sh
requirements/      base.txt, pyg.txt, dev.txt
external/DiffPuter DiffPuter submodule (contains GRAPE under baselines/GRAPE/)
data/              DeepSense 6G scenarios (not redistributed — see Dataset below)

Prerequisites

• Python 3.10–3.12 (with venv and pip)
• git
• A C/C++ compiler (build-essential on Debian/Ubuntu) — only needed if a prebuilt torch-scatter wheel isn't available for your torch+CUDA combo

Dataset

The DeepSense 6G V2I scenarios (Scenario 5 and Scenario 33) are not redistributed in this repository (non-commercial academic licence). Obtain them directly from https://deepsense6g.net and place them under data/ following the structure expected by src/common/deepsense_data.py.

Quick start

Clone with submodules so DiffPuter is populated:

git clone --recurse-submodules https://github.com/KenChan000/ImputationFor6GDatasets.git
cd ImputationFor6GDatasets

If you already cloned without --recurse-submodules:

git submodule update --init

Then build the environment (the script anchors to the repo root, so it works from anywhere; it also initialises the submodule if needed):

chmod +x scripts/setup_imputation_env.sh
./scripts/setup_imputation_env.sh

After setup

Activate the environment:

# Linux / macOS / WSL
source imputation_env/bin/activate

Either:

• Open the project in VS Code (run code . from inside WSL) and select the kernel "Imputation (6G Datasets)" in any notebook. The Python and Jupyter extensions handle activation for you — no manual source needed inside notebooks. Recommended.
• Or run jupyter lab and pick the same kernel.

Reproducing the results

Run the notebooks in order:

1. notebooks/tuning.ipynb — hyperparameter search (writes results/tuned_params.json)
2. notebooks/imputation.ipynb — imputation + reconstruction/fidelity metrics (writes results/csv/*_imputation_results.csv)
3. notebooks/downstream.ipynb — beam-prediction evaluation (writes results/csv/*_downstream_*.csv)

Cached imputations under results/ let the downstream step run without re-imputing.