Unlocking the Power of E-commerce Big Data with Deep Learning

Table of Contents

1. Introduction
2. Overview of the Marketplace
3. Miracle: The Leading Marketplace SAS Platform
4. Catalog Data and Product Information
5. Use Case 1: Product Categorization
   • 5.1 Multi-Modal Product Embedding
   • 5.2 Curation of the Catalog by Removing Duplicates
   • 5.3 Performance and Results
6. Use Case 2: Detection of Duplicate Products
   • 6.1 Challenges in Identifying Duplicate Products
   • 6.2 Machine Learning Approach for Duplicate Detection
   • 6.3 Engineering Considerations
7. Key Takeaways
   • 7.1 Multimodal Product Embeddings
   • 7.2 Frugal and Pragmatic Solutions
   • 7.3 Optimization on Spark and Onyx Runtime
8. Conclusion

Multimodal Deep Learning at Scale: Learning from Catalogs

Welcome to our presentation on the topic of multimodal deep learning at Scale. In this presentation, we will be discussing how Miracle, a leading marketplace SAS platform, leverages catalog data and deploys machine learning models to enhance product categorization and detect duplicate products within large catalogs.

Introduction

Miracle is a pioneering marketplace SAS platform that provides software solutions to manage marketplaces, catalogs, orders, and various other aspects of the marketplace ecosystem. Our data science team at Miracle has developed advanced machine learning models and techniques to analyze catalog data and improve the overall performance and user experience of marketplaces.

Overview of the Marketplace

Before delving into the details of Miracle's solutions, it is essential to understand the concept of a marketplace. A marketplace is a three-part entity consisting of a marketplace tenant, sellers, and customers. The marketplace tenant manages and sells products on the marketplace, while sellers, which can be brands or individuals, offer their products for sale. Customers browse the marketplace and make purchases, experiencing a unified front where they can buy products from multiple sellers.

Miracle: The Leading Marketplace SAS Platform

Miracle serves as a market leader in providing marketplace SAS platforms. Our software empowers marketplaces to effectively manage their catalogs, orders, and other marketplace-related operations. With over 300 marketplaces and 200,000 sellers, Miracle has established itself as a reliable partner in the e-commerce industry. Some renowned marketplaces leveraging Miracle's solutions include Kroger, Carrefour, Macy's, Bed Bath & Beyond, and even B2B customers like Airbus Helicopters.

Catalog Data and Product Information

The catalog is a database in Miracle that contains product information for all the products available on the marketplace. The data in the catalog is contributed by multiple sellers, including crucial details such as product images, titles, descriptions, category information, and more. Miracle's catalog manager ensures the representation of a single reliable catalog from diverse inputs.

Currently, Miracle's catalog holds approximately 310 million products, belonging to 140,000 different categories across various marketplaces. This diversity in product data provides a large and diverse source of data for analysis and improvement.

Use Case 1: Product Categorization

One of the significant challenges in managing large catalogs is accurately categorizing products. Incorrect categorization can lead to a poor user experience and hiding Relevant products from customers. Miracle has developed a multi-modal categorization model to address this issue.

5.1 Multi-Modal Product Embedding

The multi-modal categorization model combines text and image inputs of products to predict their respective categories accurately. By transforming the text and image inputs into embeddings and passing them through convolutional and dense layers, the model learns to predict the correct category. Additionally, resonance features of product images are used to enhance the accuracy of the categorization model.

5.2 Curation of the Catalog by Removing Duplicates

Another key aspect of catalog management is dealing with duplicate products. Identifying and removing duplicate products from a vast catalog can significantly improve the user experience and ensure Clarity in product listings. Miracle employs a machine learning approach to detect and curate duplicate products within catalogs.

5.3 Performance and Results

The multi-modal categorization model and the duplicate product detection system have demonstrated impressive performance in terms of accuracy and speed. With around 20 million parameters, the categorization model achieves inference time of approximately 50 milliseconds per prediction. Further, the use of product embeddings and multi-task learning has enabled the creation of robust models capable of handling missing data and supporting multiple languages.

Use Case 2: Detection of Duplicate Products

Detecting duplicate products within large catalogs can be a daunting task due to the sheer volume of products and the variations introduced by different sellers. Miracle has developed a highly efficient and precise machine learning-Based approach to address this challenge.

6.1 Challenges in Identifying Duplicate Products

Identifying duplicate products requires comparing each product with every other product in the same category, resulting in billions of comparisons in large catalogs. This necessitates developing robust and precise models that can handle such massive computations efficiently.

6.2 Machine Learning Approach for Duplicate Detection

Miracle's approach to detecting duplicate products relies on comparing embeddings of various product features, including text, images, and low-level descriptors. The model predicts whether two products are duplicates based on these embeddings. By leveraging deep learning techniques and decision trees, Miracle has achieved high precision in detecting duplicate products.

6.3 Engineering Considerations

Efficiently executing the duplicate detection process poses engineering challenges, including pre-processing large datasets, optimizing model training with multi-GPU support, and utilizing ONNX runtime for efficient inference. Miracle's engineering pipeline, orchestrated by Apache Airflow and executed on Databricks clusters, leverages various techniques to streamline the duplicate detection process.

Key Takeaways

In conclusion, there are several key takeaways from Miracle's approach to multimodal deep learning at scale:

1. Multimodal product embeddings play a crucial role in improving catalog data quality and allowing efficient machine learning models to be built.
2. Frugal and pragmatic solutions should be considered before resorting to brute force techniques, especially when dealing with large-scale catalogs.
3. Optimization on Spark and leveraging ONNX runtime can significantly improve the scalability and efficiency of catalog data processing and model inference.

Conclusion

Miracle's multimodal deep learning solutions have proven to enhance product categorization accuracy and improve the overall quality of catalog data. By utilizing advanced machine learning techniques and smart engineering considerations, we have successfully tackled the challenges associated with large-scale catalog management and duplicate detection. Our sophisticated models and streamlined processes contribute to the success of marketplaces and provide customers with a seamless shopping experience.