Homsters and AltexSoft: Building Machine-Learning-based Recommender System for Real Estate Marketplace

AltexSoft’s team researched, evaluated existing data, and defined key variables that impact user choice. Data scientists considered existing user interaction data and suggested additional types of information to power a recommender engine. Variables included user interactions, metadata (location, device, browser, etc.), and other analytics. The team also helped to design data storage for this purpose and covered all data preparation and cleaning activities to remove irrelevant variables and correct incomplete or inaccurate data.

One of the biggest challenges was that most users visit the marketplace just once. To be helpful, the system must gather and analyze individual visitor data rapidly during the first session. The data science team designed an algorithm that instantly generates a user profile, which allows for tailoring property recommendations even for a single visit right after the first interactions. The engine uses a content-based filtering method meaning that it matches user profile data with specific attributes for each property profile.

As user preferences and specifics of properties change over time, the algorithm connected to a web service is capable of updating itself to account for more recent and relevant data. The engine is deployed as a web service powered by Flask, a Python-based framework. The service is used by the main web application that delivers a personalized list of properties for each visitor. In the marketplace, the recommender system achieved a 15-30 percent higher conversion rate (contacting property sellers) across different markets compared to a non-personalized search.