Recall text
→ frozen SentenceTransformer (sentence-transformers/all-MiniLM-L6-v2)
→ fixed-size embedding (dim 384)
→ OneVsRestClassifier(LogisticRegression)   ← this repository
→ predicted risk labels

The transformer encoder (sentence-transformers/all-MiniLM-L6-v2) is not fine-tuned and is not included in this repository. It is loaded at inference time via the sentence-transformers library. Only the trained classifier head (classifier.joblib), per-label thresholds (thresholds.json), and label metadata (label_columns.json) are stored here.

Task

Supervised multi-label text classification.

Given a structured text constructed from brand name, product description, and recall reason, the model predicts one or more of three risk categories. A record may carry a single label or multiple labels when several risk types are described.

Input Format

Build the input string using this template before encoding:

Brand: {brand_names}. Product: {product_description}. Recall reason: {recall_reason_description}.

Example:

Brand: Example Brand. Product: Dry dog food. Recall reason: May be contaminated with Salmonella.

Output Labels

Label	Description
`PATHOGEN_CONTAMINATION`	Bacterial or pathogen-related contamination (Salmonella, Listeria, E. coli, etc.)
`CHEMICAL_OR_NUTRITIONAL_RISK`	Chemical, mycotoxin, heavy metal, feed additive, vitamin, or mineral-level risks
`PHYSICAL_OR_QUALITY_ISSUE`	Foreign material, labeling, packaging, import, inspection, or process-control issues

Label order used in all arrays: label_columns.json.

Per-Label Prediction Thresholds

Thresholds were tuned on the validation split and are stored in thresholds.json.

Label	Threshold
`PATHOGEN_CONTAMINATION`	0.50
`CHEMICAL_OR_NUTRITIONAL_RISK`	0.50
`PHYSICAL_OR_QUALITY_ISSUE`	0.40

If no label clears its threshold, the label with the highest probability score is returned as a low-confidence fallback.

Usage

python

import json
import joblib
import numpy as np
from sentence_transformers import SentenceTransformer

# Load artifacts
clf        = joblib.load("classifier.joblib")
thresholds = json.load(open("thresholds.json"))["thresholds"]
labels     = json.load(open("label_columns.json"))

# Build input text
text = (
    "Brand: Example Brand. "
    "Product: Dry dog food. "
    "Recall reason: May be contaminated with Salmonella."
)

# Encode with frozen transformer
encoder   = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
embedding = encoder.encode([text])           # shape (1, 384)

# Predict
proba           = clf.predict_proba(embedding)[0]   # shape (3,)
predicted       = [l for i, l in enumerate(labels) if proba[i] >= thresholds[l]]
low_confidence  = False
if not predicted:
    predicted      = [labels[int(np.argmax(proba))]]
    low_confidence = True

print(predicted, low_confidence)

The src/predict.py script in the source repository wraps this pipeline and outputs clean JSON.

Validation Results (model selection)

Four classifiers were trained and compared on the validation split (15 examples). Logistic Regression was selected because it achieved the best validation macro F1 while also supporting predict_proba and per-label threshold tuning.

Model	Micro F1	Macro F1	Weighted F1	Hamming Loss	Subset Accuracy
Dummy baseline	0.516	0.232	0.348	0.333	0.467
Logistic Regression	0.941	0.933	0.950	0.044	0.867
Linear SVC	0.903	0.857	0.893	0.067	0.800
Random Forest	0.941	0.933	0.950	0.044	0.867

Final Test Results

The selected model was evaluated once on the held-out test set (15 examples). No training, threshold tuning, or model selection changes were made after inspecting test results.

Overall

Metric	Value
Micro F1	0.875
Macro F1	0.841
Weighted F1	0.881
Hamming Loss	0.089
Subset Accuracy	0.800

Per-Label

Label	Precision	Recall	F1	Support
`PATHOGEN_CONTAMINATION`	1.000	1.000	1.000	8
`CHEMICAL_OR_NUTRITIONAL_RISK`	1.000	0.750	0.857	4
`PHYSICAL_OR_QUALITY_ISSUE`	0.600	0.750	0.667	4

The test set contains only 15 rows, so metrics are sensitive to 1–2 examples. Results should be interpreted as assignment-scale evidence, not robust production performance estimates.

Training Data

103 labeled records from official public pet food recall and safety-alert portals (FDA, EU RASFF, UK FSA, Canada CFIA, openFDA). Labels were assigned using transparent rule-based keyword patterns. 14 uncertain records were excluded rather than forced into a label.

Split	Rows
Train	73
Validation	15
Test	15

See the companion dataset repository for full split details and source attribution.

Repository Contents

File	Description
`classifier.joblib`	Trained `OneVsRestClassifier(LogisticRegression)`
`thresholds.json`	Per-label prediction thresholds (validation-tuned)
`label_columns.json`	Ordered list of output label names
`model_config.json`	Training provenance and selection metadata
`test_metrics.json`	Final held-out test metrics
`per_label_metrics.csv`	Per-label precision / recall / F1 on test set

Limitations

Small training set — 73 training rows; model may not generalise beyond this vocabulary and source mix.
Small test set — 15 rows; per-label metrics are sensitive to 1–2 examples.
Rule-based labels — not manually annotated by domain experts.
Consolidated taxonomy — three labels combine six more specific risk types.
Source heterogeneity — records from five sources with different terminology.
Frozen encoder — the transformer is not fine-tuned on recall text.
Not a safety authority tool — the model predicts text categories, not product safety.

Educational Disclaimer

This model is for educational purposes only.

It does not provide veterinary advice, legal advice, product safety certification, official recall interpretation, or medical or nutritional recommendations.

Always consult official recall notices and qualified professionals for authoritative information about pet food safety.

ShurongSR/pet-food-recall-risk-classifier

Pet Food Recall Risk Classifier

Model Description