🛍️ Product Category Classification Prediction using AI/ML Models

I propose a robust hierarchical machine learning pipeline that classifies e-commerce products into both top-level and fine-grained bottom-level categories. This is based on multimodal metadata (text, categorical, numeric). The pipeline leverages advanced feature engineering, dimensionality reduction, and boosting-based classification techniques.

---

📌 Project Highlights

• Task: Multimodal product categorization (top and bottom categories)
• Input Data: Cleaned metadata (titles, descriptions, tags, categorical fields)
• ML Techniques: TF-IDF, One-Hot Encoding, TruncatedSVD, XGBoost, LightGBM, SMOTE
• Evaluation: Accuracy, Macro F1, Weighted F1
• Visualization: Confusion matrices, t-SNE embeddings, category distributions

---

🔧 Pipeline Overview

Here’s a clean academic-style flowchart of the full pipeline:

---

📂 Data Preprocessing

The project started by aggregating over 350+ Parquet files (train + test), performing cleaning and enrichment on product metadata.

import glob
import pandas as pd

files = glob.glob("/data/train/*.parquet")
etsy_train_df = pd.concat([pd.read_parquet(f) for f in files])

# Cleaning text fields
def clean_text(text):
    text = re.sub(r'[^a-zA-Z0-9\s]', '', text.lower())
    return " ".join([lemmatizer.lemmatize(w) for w in word_tokenize(text) if w not in stop_words])

etsy_train_df['description_clean'] = etsy_train_df['description'].apply(clean_text)

---

✨ Feature Representation

I process three types of features:

• TF-IDF on combined text features (title, description, tags)
• One-Hot Encoding for fields like material, room, recipient
• Standard Scaling for numerical metadata (lengths, color ids)

Then I perform dimensionality reduction using TruncatedSVD:

from sklearn.decomposition import TruncatedSVD
X_text = tfidf_vectorizer.fit_transform(etsy_train_df['combined_text'])
svd = TruncatedSVD(n_components=300, random_state=42)
X_text_svd = svd.fit_transform(X_text)

---

🎯 Stage 1: Top Category Classification

I first predict the top category using XGBoost with 5-Fold Cross Validation:

model = xgb.XGBClassifier(tree_method='gpu_hist', n_estimators=300, max_depth=15)
scores = cross_val_score(model, X_train_svd, y_top, cv=5, scoring='f1_macro')

✅ Confusion Matrix for Top Category

📊 Top Category Distribution

🌐 t-SNE Visualization of Top Category Embeddings

---

🧠 Stage 2: Bottom Category Prediction

For each predicted top category, I train a dedicated classifier (XGBoost/LightGBM) using the same reduced feature space plus top category probabilities as features.

for top_cat in unique_top_categories:
    X_top = X_train_enhanced[y_top_train == top_cat]
    y_bottom = y_bottom_train[y_top_train == top_cat]
    clf = lgb.LGBMClassifier(...)
    clf.fit(X_top, y_bottom)

🧩 Top-Bottom Category Relationship Heatmap

This shows how predicted top categories relate to bottom categories:

---

📉 Evaluation Results

🔢 Bottom Category Distribution

---

💡 Visual Insights

t-SNE Embeddings with Top Labels

---

🔍 Challenges & Future Work

• Imbalance: Many bottom categories have <50 samples. SMOTE helps but has limits.
• Scalability: Training hundreds of bottom classifiers (some with 400+ classes) is time and memory intensive.
• Feature drift: Low signal in some text categories like tags or shape.

🔮 In future work, I aim to:

• Replace TF-IDF + SVD with Sentence-BERT
• Experiment with end-to-end transformers like HierBERT
• Introduce Meta-Learning for cross-category generalization

---

📁 Model & Dataset Access

Due to NDA constraints, the Etsy dataset cannot be shared. For access, please request permission from the original data provider.

⚠️ Dataset not available due to NDA restrictions with Etsy Inc.

🔗 Model Files

• Top Model (XGBoost with CrossValidation): Download
• Bottom Models (per top): Download

---

🧠 Summary

This project delivers a scalable and interpretable hierarchical product categorization system. By breaking down the classification task into top and bottom predictions, I align with the way marketplaces organize products and demonstrate strong performance across diverse product types