ברכה שפירא

In e-commerce, incomplete product data presents a significant challenge, particularly in online marketplaces where small businesses and individual sellers often lack the resources to provide full product details. Missing data in product items-whether structured fields like feature-value pairs or unstructured text such as titles and descriptions-can hinder product search, recommendations, and overall marketplace functionality. Addressing these data gaps is essential for enhancing the efficiency and user experience of e-commerce platforms. This paper introduces low-cost machine learning approaches for completing missing textual features in such items, offering an alternative to computationally expensive large language models (LLMs). We propose two cost effective methods: the first extracts data from unstructured fields like item titles or descriptions, while the second employs a Nearest Neighbors method to impute missing values based on similar items. Both methods are evaluated on real-world datasets across diverse product categories, including sports trading cards, motor parts, and computers. Our experiments demonstrate that these low-cost approaches can achieve performance comparable to LLMs, often at a fraction of the computational cost, making them a viable option for large-scale e-commerce platforms.We also demonstrate that completing missing data not only improves data quality but also enhances key tasks such as search optimization and matching items to catalog products, which are critical for e-commerce platforms.