A property recommendation engine that learns what a buyer actually wants

Client Industry

Real estate · property portal & consultancy

Key Challenge

Too many listings, filters that miss genuine preference

Core Tech Stack

Collaborative filtering · Content-based ML · Python

Deployment Model

Hybrid recommender behind the portal search

SoftmaxAI developed a hybrid property recommendation system for a portal that caters to buyers, renters, sellers, landlords, and agents. The system uses a structured preference survey to find out exactly what a user is looking for — such as location, type of property, budget, and any essential features — and then a content-based model evaluates each listing based on these stated preferences, providing the user with a match score that they can see and ask questions about. At the same time, a memory-based collaborative filtering approach looks at engagement data to find other users who have similar tastes and then shows the properties those similar users have interacted with. The two types of signal are combined together so that the recommendations take into account both the things a person says they want and the properties that people with similar tastes have actually chosen.

• More listings than anyone can evaluate. Since a portal covering a market on a large scale yields more candidates than any individual is ever going to look at, it is the ranking and not the catalogue that decides which items are found.
• Filters force false trade-offs. Hard filters make a user decide in advance which requirement to abandon, and silently hide the property that missed one criterion but was otherwise ideal.
• Stated preference is incomplete. The things that people say they want and the things that they in fact respond to are continually different — a system which is based merely on a form will confidently continue to recommend the wrong thing.
• New users have no history. Pure collaborative filtering has nothing to work with on a first visit, which is exactly when a property portal needs to prove its worth.

The engine operates via a two-branch pipeline in which the listings together with the stated preferences are fed into a content-based match model and the engagement signals are fed into a collaborative filtering model, after which the two scores are combined to produce a single explained ranking.

• Spec before code. We defined the survey, the match-score formula, how engagement feeds back, and the cold-start behavior before building — so the recommendation logic was a decision, not an emergent property of the data.
• Content-based scoring for stated needs. Each listing is assessed in light of the user's stated requirements, which results in a ranked list together with an explicit match score instead of a simple binary filter pass.
• Collaborative filtering for revealed taste. By using memory-based filtering on engagement signals, the system identifies users whose responses overlap and then surfaces the properties those users engaged with — the mechanism that recommends the property a user would not have described but still wants.
• Cold start solved by design. When a user first uses the system, the survey provides them with useful recommendations straight away; the collaborative signal comes into play as engagement builds up rather than being a necessary condition.

• Ranked results instead of a filtered list. The properties are presented to the users in order of how well they fit, with the most suitable one appearing at the top rather than being located somewhere within a long list of results.
• A match score users can interrogate. It is only because each recommendation indicates how it meets the specified requirements that a near-miss becomes understandable rather than appearing to be arbitrary.
• Near-misses stay visible. Using scoring instead of filtering means that a listing will still appear even if it fails to meet one criterion but performs well in all other areas — whereas a hard filter would have deleted it.
• Recommendations improve with use. The more a user browses, the better the portal becomes at matching their tastes since engagement supports the collaborative model.
• Useful from the first session. The preference survey means a new user gets relevant results immediately, without waiting to accumulate a history.