Amazon ML Challenge, 2025
Problem statement: Smart Product Pricing Challenge
In e-commerce, determining the optimal price point for products is crucial for marketplace success and customer satisfaction. Your challenge is to develop an ML solution that analyzes product details and predict the price of the product. The relationship between product attributes and pricing is complex — with factors like brand, specifications, product quantity directly influence pricing. Your task is to build a model that can analyze these product details holistically and suggest an optimal price.
In e-commerce, determining the optimal price point for products is crucial for marketplace success and customer satisfaction. Your challenge is to develop an ML solution that analyzes product details and predict the price of the product. The relationship between product attributes and pricing is complex — with factors like brand, specifications, product quantity directly influence pricing. Your task is to build a model that can analyze these product details holistically and suggest an optimal price.
Data Description:
The dataset consists of the following columns:
- sample_id: A unique identifier for the input sample
- 2. catalog_content: Text field containing title, product description and an Item Pack Quantity(IPQ) concatenated.
3. image_link: Public URL where the product image is available for download.
Example link — https://m.media-amazon.com/images/I/71XfHPR36-L.jpg
To download images use `download_images` function from `src/utils.py`. See sample code in `src/test.ipynb`.
4. price:Price of the product (Target variable — only available in training data)
Dataset Details:
Training Dataset: 75k products with complete product details and prices
Test Set: 75k products for final evaluation
Output Format:
The output file should be a CSV with 2 columns:
1. **sample_id:** The unique identifier of the data sample. Note the ID should match the test record sample_id.
2. **price:** A float value representing the predicted price of the product.
Note: Make sure to output a prediction for all sample IDs. If you have less/more number of output samples in the output file as compared to test.csv, your output won’t be evaluated.
Constraints:
1. You will be provided with a sample output file. Format your output to match the sample output file exactly.
2. Predicted prices must be positive float values.
3. Final model should be a MIT/Apache 2.0 License model and up to 8 Billion parameters.
EXECUTIVE SUMMARY
Our approach integrates multimodal learning by combining textual and visual information to predict product prices. We generated embeddings using the CLIP model for both text and images, then applied a gated fusion mechanism to effectively merge the two modalities. The fused embeddings were used to train an XGBoost regressor, enabling robust price prediction based on comprehensive product understanding.
PROBLEM ANALYSIS
a) We approached the pricing challenge as a multimodal regression task, leveraging both product descriptions and images to capture semantic and visual factors influencing price. Our EDA revealed significant data skewness, strong textual correlations with price, and valuable visual cues from product images. These insights guided us toward using CLIP-based embeddings and a gated fusion approach for robust multimodal learning.
Key Observations:
● Price Distribution: Highly skewed with several outliers — required log transformation and robust modeling.
● Text Features: Brand names, quantity, and specifications showed strong correlation with price; text length was moderately predictive.
● Image Features: Consistent presence across data; resizing and normalization were essential for uniform representation.
● Multimodal Strategy: Combining text and image embeddings offered richer feature representation than using either alone.
b) Approach Type: Hybrid Multimodal Learning
Core Innovation
Our approach adopts a hybrid multimodal learning framework that integrates both textual and visual modalities to accurately predict product prices. Since the pricing of a product often depends on both descriptive attributes (like brand, quantity, and specifications) and visual characteristics (like packaging, design, or perceived quality), we leveraged the CLIP model to generate embeddings for both the text descriptions and product images. This allowed us to represent each product in a shared embedding space where semantic and visual features complement each other.
After generating these embeddings, we employed a gated fusion mechanism to combine the two modalities effectively. The gated fusion approach acts as a learnable filter that determines how much importance should be given to textual versus visual features for each product instance. This ensures that the model dynamically adapts to cases where one modality might be more informative than the other — for example, text-heavy descriptions for some products and visually distinct images for others. This was a key innovation in our design, as simple concatenation or averaging of embeddings often fails to capture the nuanced relationships between modalities.
For the downstream regression task, we experimented with three modeling techniques — Multi-Layer Perceptron (MLP), Random Forest Regressor, and XGBoost Regressor. While linear models such as Linear Regression performed poorly due to the nonlinear and complex interactions between multimodal features, nonlinear learners showed significant improvements. Among them, XGBoost Regressor emerged as the most robust and consistent performer. Its ability to handle skewed price distributions, manage outliers effectively, and model nonlinear relationships made it ideal for our dataset. The MLP performed reasonably well but required careful tuning and was more sensitive to overfitting, while the Random Forest struggled with generalization in comparison.
Our final pipeline thus involves:
1. Preprocessing and cleaning of textual descriptions.
2. Generation of multimodal embeddings using the CLIP model.
3. Fusion of text and image embeddings via gated fusion.
4. Training an XGBoost Regressor on the fused embeddings for price prediction.
Get Nishant sharma’s stories in your inbox
Join Medium for free to get updates from this writer.
This hybrid multimodal approach allowed us to combine the interpretability and robustness of tree-based methods with the deep representational power of transformer-based embeddings, resulting in a strong, generalizable model for price prediction across diverse product categories.
Architecture Overview
The model is a multimodal regression pipeline that predicts product prices by combining text and image features. Text descriptions and product images are encoded using CLIP ViT-B/32 into 512-dimensional embeddings, which are fused through a gated mechanism to form a unified representation. PCA optionally reduces dimensionality, and an XGBoost regressor predicts log-transformed prices. The pipeline includes normalization, missing-image handling, and modular end-to-end design for robust performance.
Model Components
Text Processing Pipeline
Preprocessing steps:
● Clean text to remove encoding artifacts (e.g., “–”, “’”).
● Unescape HTML entities.
● Remove labels/prefixes like “Bullet Point 1:”, “Item Name:”, “Value:”, “Unit:”.
● Normalize whitespace and remove extra newlines.
● Fill missing text with empty strings.
Model type:
CLIP ViT-B/32 (text encoder component) for generating embeddings.
Key parameters:
● Maximum token length (implicit in CLIP processor, usually 77).
● Text embedding dimension: 512.
● L2 normalization applied to embeddings.
● Gated fusion weight (scalar or vector) for combining with image embeddings.
● Image Processing Pipeline
Preprocessing steps:
● Load images safely from file paths.
● Convert images to RGB format (3 channels).
● Resize images deterministically to 224×224 pixels using bicubic interpolation.
● Replace missing/corrupt images with placeholder black images.
● Normalize pixel values as per CLIP preprocessing.
Model type:
CLIP ViT-B/32 (image encoder component) for generating embeddings.
Key parameters:
● Image embedding dimension: 512.
● Batch size for processing: 32 (with per-row fallback).
● L2 normalization of embeddings.
● Gated fusion weight (scalar or vector) for combining with text embeddings.
Model Performance METRICS
● SMAPE Score: 0.523
● MAE: 0.592566
● RMSE: 0.755101
● R2: 0.354190
Conclusion
The model is a multimodal regression pipeline that predicts product prices by combining text and image features. Text descriptions and product images are encoded using CLIP ViT-B/32 into 512-dimensional embeddings, which are fused through a gated mechanism to form a unified representation. PCA optionally reduces dimensionality, and an XGBoost regressor predicts log-transformed prices. The pipeline includes normalization, missing-image handling, and modular end-to-end design for robust performance.
Team : RepoX
Github Link: https://github.com/Nishant05092/RepoX_Amazon_ML_Challange
EDA Link: https://colab.research.google.com/drive/1VfdWPkwDeKeONWNsf6zUuy7ZldEfl2Hm?usp=sharing
