Product categorisation with TF-IDF

What this project does

This project demonstrates a compact product categorisation pipeline that maps short catalogue descriptions to a small controlled taxonomy. It trains a TF-IDF text classifier (logistic regression) and adds an explicit fallback path for low-confidence or unseen-category predictions by returning an “UNKNOWN” label and top candidate suggestions suitable for an editor review workflow. The pipeline reports held-out macro F1 and a simple measure of confusion between sibling categories in the taxonomy.

Data

The code uses a small synthetic dataset of product descriptions generated in-script to stand in for curated catalogue data. Each description is an intentionally varied natural-language phrase (e.g. “lightweight Android smartphone with 64GB”) and each row is assigned an editor-style category label from a compact taxonomy with branches such as Electronics, Home and Apparel. This is synthetic and used to illustrate development and evaluation; a real project should replace it with licensed, editor-verified product descriptions and explicit variant grouping prior to splitting.

How to build it

1. Create the synthetic description corpus and a small hierarchical taxonomy inside the code. Group near-duplicate variants before splitting to avoid leakage.
2. Split the dataset into training and held-out sets using a fixed random seed and stratification by category.
3. Train a TF-IDF vectoriser and a LogisticRegression classifier; add a probability threshold to route low-confidence predictions to an UNKNOWN path.
4. Evaluate on the held-out set: report macro F1 and a sibling-confusion metric; provide top-k candidate categories for editor review.

How to run

Run the project module as a script (Python 3.11+) or call run(seed) from another program. The script prints a JSON dictionary of numeric metrics and small text values.

How to test

Use pytest to run the provided tests. They check that run(42) returns expected keys, that metrics are in plausible ranges, that macro F1 meets a conservative floor, and that the run is deterministic for a fixed seed.

Results

Measured by running python project.py (seed 42) on 2026-10-23; the tests in test_project.py assert the floor.

Metric	Value
macro_f1	0.9845
unknown_rate	0.0296
sibling_confusion	0

Also returned (see python project.py output): top_k_example (3 items).

Limitations

This demonstration uses synthetic text data created inside the code. That means the vocabulary, phrase patterns and category balance are artificial and will not reflect a production catalogue. Also the taxonomy is small and simple; real taxonomies have many nested levels and ambiguous labels. The fallback is a simple low-confidence threshold — in production one would combine this with editorial signals, active learning and candidate-ranking calibrated on real correction logs.

Next steps

• Replace synthetic descriptions with a licence-approved corpus of editor-verified product descriptions, and group near-duplicate variants before splitting.
• Expand the taxonomy and evaluate hierarchical losses and per-branch calibration.
• Introduce an editor-in-the-loop correction UI and record corrections to improve the classifier with incremental training and uncertainty-based sampling.

---

Full case study, build steps and evaluation checks: Product categorization system on networkershome.com

Built by Networkers Home · Learn the skills in the AI Engineer Course · Part of 100 AI projects with code