preorder4MLC: Pre-Order Based Multi-Label Classification

In multi-label classification (MLC) each instance can carry several labels at once (a song that is both happy and relaxed — as in the emotions dataset; an email tagged both business and legal — as in the enron dataset, both used in this paper). This repository introduces a method that predicts labels by first learning how labels compare rather than deciding each label in isolation — and that can abstain on labels it is unsure about instead of guessing.

📄 Code and full experimental results for: Robust Multi-Label Classification via Preference Learning — Vu-Linh Nguyen, Xuan-Truong Hoang, Sébastien Destercke, Cassio de Campos, Van-Nam Huynh (Nguyen and Hoang contributed equally). Machine Learning (Springer), 2026. <TODO: DOI once assigned>

Abstract. We explore how multi-label classification (MLC) tasks can be cast into order structure learning, exploiting the rich structure of orders to improve and robustify MLC. We formalise the transformation of MLC into an order structure learning and prediction task, study the prediction of Bayes-optimal order structures, and run experiments in settings where orders are especially beneficial: robust MLC under noisy and imbalanced labels, and MLC prediction with partial abstention.

What it does

The method is built around Bipartite Ordered Preference Orders (BOPOs). Instead of predicting each label independently, it works in three steps:

1. Learn pairwise preferences. For every pair of labels, a calibrated classifier estimates the probability of their relative ordering.
2. Search for the best order. Per instance, an integer linear program (ILP) combines those pairwise probabilities into a single coherent preference order — a pre-order or a partial-order — that minimises an expected loss (Hamming or Subset 0/1).
3. Derive a prediction. The order is turned into one of three outputs: a plain binary vector, the preference order itself, or a partial-abstention vector that marks uncertain labels as "abstain".

An optional height constraint on the order yields eight inference algorithms in total (pre-/partial-order × Hamming/Subset × height 2/∅). We compare against four standard baselines: BR, CC, CLR, ECC.

              ┌──────────────────────┐
   training → │ K(K−1)/2 pairwise    │ → pairwise probabilities pᵢⱼ
              │ calibrated classif.  │
              └──────────────────────┘
                         │
                         ▼
              ┌──────────────────────┐
              │ per-instance ILP     │ → preference order
              │ search (cvxopt+GLPK) │   (pre- or partial-order)
              └──────────────────────┘
                         │
                         ▼
              ┌──────────────────────┐
              │ derive prediction    │ → BinaryVector
              │                      │   PreferenceOrder
              │                      │   PartialAbstention
              └──────────────────────┘

Quick start

# Install (editable, so scripts/ can import the package)
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
pip install -e .

# Run one dataset end-to-end
python scripts/train.py    --dataset emotions --results_dir results/run-dev
python scripts/evaluate.py --dataset emotions --results_dir results/run-dev

To reproduce every number and figure in the paper, see REPRODUCE.md — it covers the full pipeline, expected wall times, and the exact environment.

Datasets

Ten multi-label datasets are used: chd_49, emotions, scene, yeast, water_quality, humanpseaac, gpositivepseaac, plantpseaac, viruspseaac, and enron.

The first nine ARFFs are bundled under data/, so the pipeline runs end-to-end right after pip install. enron.arff (K=53) is downloaded separately — see REPRODUCE.md §2. CLI keys match preorder4mlc.config::ConfigManager.DATASET_CONFIGS.

Partial abstention

A partial-abstention prediction is a vector in {0, 1, −1} where −1 means "abstain" on that label. This lets the model stay silent where it is uncertain instead of forcing a 0/1 call. Two pairs of metrics capture the trade-off — recovery (did the abstentions cover the truth?) and abstention rate (how often did it abstain?):

ŷ = [1, 0, 1, −1, 0, −1]      y = [0, 0, 1, 1, 0, 0]

AREC = (0 + 1 + 1 + 1 + 1 + 1) / 6 = 4/6      # −1 counts as covering {0,1}
AABS = 2 / 6                                   # fraction abstained

Full definitions (AREC, AABS, REC, ABS) live in preorder4mlc.evaluation_metric.

Repository layout

preorder4mlc/            # Library package
├── config.py            # Run configuration + dataset registry
├── datasets4experiments.py   # ARFF loading, k-fold splits, label noise
├── base_classifiers.py  # Pairwise / calibrated classifier factory
├── estimator.py         # Uniform interface over RF / ETC / XGBoost / LightGBM
├── inference_models.py  # PredictBOPOs (BOPOs + BR / CC / CLR baselines)
├── searching_algorithms.py   # ILP search for pre- and partial-orders
├── training_orchestrator.py  # Training loop over learners × folds × algorithms
├── evaluation_metric.py # Example-, label-, ranking-, abstention-metrics
└── utils/               # Summaries, statistical tests, figures

scripts/                 # CLI entry points
├── train.py / evaluate.py          # BOPOs + CLR / BR / CC
├── train_ecc.py / evaluate_ecc.py  # ECC baseline
└── smoke_predict_bopos.py          # Behavior-preservation smoke test

data/                    # 9 bundled ARFFs (enron downloaded separately)
results/                 # Per-fold CSVs + aggregated tables
run.sh                   # End-to-end reproduction driver
REPRODUCE.md             # Step-by-step reproduction recipe

Citation

If you use this code, please cite the paper (see CITATION.cff):

@article{nguyen2026robust,
  title     = {Robust Multi-Label Classification via Preference Learning},
  author    = {Nguyen, Vu-Linh and Hoang, Xuan-Truong and Destercke, S{\'e}bastien and de Campos, Cassio and Huynh, Van-Nam},
  journal   = {Machine Learning},
  publisher = {Springer},
  year      = {2026},
}

License

Released under the MIT License.