MC-TRCM

MC-TRCM: Missingness-Aware Recursive Fusion for Depression Benchmarks

MC-TRCM (Modality-Conditioned Temporal Recursive Context Model) is a missingness-aware recursive fusion model for public depression-related mobile and wearable feature-view benchmarks. The current repository keeps the five public dataset parsers used to define the benchmark framing: StudentLife, DepreST-CAT, PSYCHE-D, Depresjon, and OBF-Psychiatric.

Repository Layout

• src/preprocess/: dataset download, extraction, parsing, split generation, and leakage checks.
• src/models/: MC-TRCM, displayed tabular/neural baselines, data loaders, and training entry points.
• src/evaluation/: shared metrics, validation-selection helpers, and comparison utilities.
• scripts/: paper-level experiment orchestration, calibration, uncertainty, and table utilities.
• configs/: dataset, model, training, split, and protocol configuration files.
• tests/: protocol and model smoke tests.
• reports/: lightweight audit notes generated by preprocessing and split checks.

Raw datasets, intermediate tables, checkpoints, prediction caches, and generated paper artifacts are intentionally not tracked. Obtain each public dataset from its original source and prepare it locally before running experiments.

Setup

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Typical Commands

After the required public datasets have been downloaded and prepared:

python -m src.preprocess.run_available_parsers
python -m src.preprocess.generate_splits
python -m src.models.run_baselines --datasets deprest_cat psyche_d
python -m src.models.train_mctrcm_v2 --datasets deprest_cat psyche_d --model-config-path configs/model_configs/mctrcm_final.json
python scripts/run_core_mctrcm_protocol.py --stage summarize