A multimodal recommendation system for real estate listings

Abstract

This study presents a computational framework designed to generate reliable embeddings while addressing challenges posed by the dynamic nature of property listings and limited research in the real estate market. We consider challenges like the recent rise in listing prices for real estate platforms. Our framework includes three key components: (i) a graph embedding model to encode sequential user behavior, (ii) feature generation utilizing listing meta-data, and (iii) performance optimization through listing segmentation by clustering. To assess our framework’s effectiveness, we used two offline methods: clustering performance and retrieval performance evaluation. In the retrieval phase, precision and normalized discounted cumulative gain (NDCG) were employed, while visualization and three clustering metrics assessed clustering performance. A comparative analysis highlighted enhancements from listing segmentation, with a 12.11% increase in precision (totaling 73.54%) and a 3.72% improvement in NDCG (totaling 67.07%). Online A/B tests using user click-through-rate evaluations demonstrated significant increases of 64.36% and 72.67% compared to two existing rule-based recommendation systems. These results underscore our framework’s capacity to effectively discern similarities and distinctions among real estate listings.