Розробка рекомендаційної системи для торгівлі уживаними товарами на основі методів обробки природної мови

Abstract

This paper presents a pragmatic architecture for a recommendation system tailored to consumer-to-consumer (C2C) e-commerce platforms dealing with used goods. It addresses the «cold start» problem of unique, single-quantity items by using a content-based approach. The system leverages pre-trained Natural Language Processing (NLP) models to generate dense item embeddings based on listing metadata. Instead of relying on complex collaborative filtering that requires extensive interaction logs, user profiles are built dynamically through a time-decayed weighted average of viewed item embeddings, separated by item categories. A two-stage pipeline performs approximate nearest neighbor search followed by a heuristic ranking stage that incorporates business logic, such as item freshness and price deviation. This approach ensures immediate discoverability of new items without requiring massive datasets for model training.