TheWalkingDevs — Product Price Prediction

A multimodal machine learning pipeline for predicting grocery/consumer product prices from catalog text and product images. Built in Google Colab with GPU acceleration.

---

Overview

This notebook builds a price prediction model by combining multiple feature representations extracted from product catalog descriptions and images:

• TF-IDF features from product text (10,000 dimensions)
• Semantic embeddings via all-MiniLM-L6-v2 (Sentence Transformers)
• Image embeddings via EfficientNet-B0 (timm)
• Hand-crafted text features (nutrition info, certifications, category flags, packaging, etc.)

All feature matrices are horizontally stacked into a single sparse matrix and fed into a downstream regression model to predict log-transformed prices.

---

Project Structure

student_resource/
├── dataset/
│   ├── train.csv           # Training data with price labels
│   └── test.csv            # Test data for submission
├── images/                 # Product images (downloaded separately)
├── tfidf_matrix.npz        # Cached TF-IDF sparse matrix
├── semantic_embeddings.npy # Cached sentence embeddings
├── image_embeddings.npy    # Cached EfficientNet image embeddings
└── precomputed_text_features.csv  # Cached hand-crafted features

Output: test_out_updated.csv — submission file with sample_id and predicted price.

---

Dataset

The dataset is expected at /content/drive/MyDrive/student_resource/dataset/ on Google Drive and must contain:

Column	Description
sample_id	Unique product identifier
catalog_content	Product description text
image_link	URL or filename of product image
price	Target variable (train only)

Combined train + test size observed: 150,000 rows × 4 columns.

---

Pipeline

1. Setup

Mounts Google Drive and configures dataset/image paths.

2. Data Loading

Reads train.csv and test.csv, applies log1p transformation to the price column, and concatenates train and test for consistent feature engineering.

3. Hand-Crafted Text Features

Extracts 29 features from catalog_content across these categories:

• Basic stats: content length, word count, items-per-quantity (IPQ)
• Dietary claims: gluten-free, non-GMO, dairy-free
• Nutrition: grams, protein, carbs, sugar
• Certifications: USDA Organic, Fair Trade, ISO, FDA approved
• Product category: snack, beverage, supplement, bakery, frozen
• Packaging: pack size, eco-friendly, container type
• Premium signals: premium/luxury/artisan keywords, import origin
• Chocolate type: chocolate, dark, milk, white
• Brand encoding: label-encoded brand, rare brands grouped

4. TF-IDF Features

Fits a TfidfVectorizer with up to 10,000 features on the full corpus. Results are cached to tfidf_matrix.npz for reuse.

5. Semantic Embeddings

Uses sentence-transformers/all-MiniLM-L6-v2 to encode catalog_content into dense 384-dim vectors. Results are cached to semantic_embeddings.npy.

6. Image Embeddings

Uses a pretrained EfficientNet-B0 (via timm, no classification head) to extract 1280-dim image feature vectors. Missing or unreadable images fall back to a zero vector. Results are cached to image_embeddings.npy.

7. Feature Matrix Assembly

Horizontally stacks all features into sparse CSR matrices:

X = [TF-IDF (10000) | hand-crafted (29) | semantic (384) | image (1280)]

8. Submission

Generates test_out_updated.csv with columns sample_id and price (exponentiated back from log-space predictions via final_predictions).

---

Requirements

pandas
numpy
scipy
scikit-learn
sentence-transformers
timm
torch
Pillow
tqdm
google-colab          # for Drive mounting

Install in Colab:

pip install timm sentence-transformers

---

Usage

1. Upload the notebook to Google Colab.
2. Place your dataset and images under MyDrive/student_resource/ as described above.
3. Set runtime to GPU (T4 recommended).
4. Run all cells top to bottom. Cached .npy/.npz files will be reused on subsequent runs.
5. Download test_out_updated.csv from the Colab working directory.

---

Notes

• The notebook caches all heavy computations (TF-IDF, embeddings) to Google Drive. If you change the dataset, delete the cached files to force recomputation.
• The variable final_predictions (used in the submission cell) must be defined by the model training step, which is not included in this notebook excerpt — add your regression model between steps 7 and 8.
• GPU acceleration significantly speeds up both SentenceTransformer encoding and EfficientNet inference.