realtor.com Tech Blog - Engineering and Technology blog at Move, Inc.

Smarter Search: How Realtor.com Uses Natural Language Processing to Help Visitors Find Their Dream Home

Consumers come to R ealtor.com looking for the perfect place to call home. In addition to knowing the neighborhood they’d like to be in, they’ve probably envisioned what the home they’re buying should look like and what amenities they prefer, for example, a 2-car garage or a heated private pool. But it’s not always easy to sift through thousands of listings to find that perfect home. Our data science team at Realtor.com is trying to ease the search process, and enable home shoppers to hone in on the listings that best fit their wish list.

Here’s an example: A home buyer goes online to search for a home in San Jose, Calif. with a private pool. They specify the price, property type, beds, and baths, then select the keyword “pool” from the dropdown menu. On the backend, the site applies the search criteria and filters and serves up all available properties that fit the parameters, and has the word “pool” somewhere in the listing. That’s all there is to it, right?

Not necessarily. While those results will include homes with private pools, they could also include listings of properties that say things such as “community pool”, “pool view,” or “room for pool” – results that we would call “false positives.” For a home buyer searching for a property with a private pool, this can be a frustrating experience – which goes against our goal to deliver a customized experience that is simple, frictionless, and enjoyable.

The reason this happens has to do with data sources. Listings on Realtor.com come from local MLSs (multiple listing services) and when a broker is adding a new listing, they can choose from more than 90 property attributes. This data may not always be consistent or complete, and as a result, search results can be inaccurate or incomplete. To help alleviate this, the data science team uses data enrichment and creates tags based on the property description. However, this can lead to listings being tagged incorrectly when context about the sentence containing the keyword is lacking.

So what’s the solution? The data science team at Realtor.com tackled the problem with a multi-class classification model that uses natural language processing (NLP) to identify dozens of attributes of a home with a high degree of granularity. Let’s take a look at how they did it.

A Home for Every Realtor

At Realtor.com, we pride ourselves on providing the most complete and accurate realtor data in the industry. We’re able to do this through our relationships with the National Association of Realtors (NAR), multiple listing services (MLSs), real estate brokerages, and the REALTORS® themselves. In this article, we explore how we manage and use this data to help connect REALTORS® with home buyers.

Profile Data Collection

We manage, organize, and combine data from an array of sources, including the NAR and many various MLSs and real estate brokerages throughout the United States. As NAR’s official site, we provide a page on Realtor.com for every registered real estate agent and office that the organization provides us via their data feed. The pages include REALTORS® names, contact information, credentials, languages spoken, and other important details. This information allows us to present a complete listing of REALTORS® nationwide to our end users while also providing agents and offices with a meaningful online presence.

We also receive feeds from most MLS associations throughout the nation. Every agent we access via these feeds can sign up for a dashboard where they can claim their NAR profile. When they claim their profile, we can then combine the listing data that we ingest from MLSs with the data from the NAR, making for a rich online profile that we provide without the need for the REALTOR® to provide any additional editing.

Many brokerages also send us feeds with updates on their agents and offices, helping to enrich the profiles further and keep the information up to date.

The 2021 virtual Grace Hopper Celebration of Women in Computing

The Grace Hopper Celebration (GHC) is the world’s largest conference of women in technology and is designed to bring the research and career interests of women in computing to the forefront. Realtor.com sent 15 employees to GHC this year because we are committed to empowering women in technology and expanding the diversity of the engineering workforce. I was fortunate to be one of the Realtor.com nominees to attend the Grace Hopper Celebration of Women in Computing. This was my first time attending GHC, though I have attended other conferences for women in the past. GHC offered a significant number of events and panels for attendees at all career levels and with different technical expertise. At times it was hard to decide which sessions to attend and sometimes challenging to get into sessions due to high popularity and demand.

Enhancing CX through Identity Resolution: A Q&A with Anne Hunt, Author of “Idiosyncratic Ontologies in Industrial Applications”

At Realtor.com, we’re all about the customer experience. Leveraging cutting-edge data science techniques and models, we are continually improving and optimizing the experience people have with our website and services. A keystone of that experience is making a visit to our site personalized and guided so that visitors find what they’re looking for quickly and have the information they need to make decisions about buying, selling, or renting a home.

But this experience doesn’t just happen magically. Behind the scenes are many advanced data science models — and experienced data scientists who create them.

We recently had the opportunity to interview Anne Hunt, VP of Product at Realtor.com, about her work on representing idiosyncratic ontologies in industrial applications that she presented at FOIS 2021, the 12th International Conference on Formal Ontology in Information Systems. Anne developed a new design pattern that can be used to support the customer experience Realtor.com.

Data Science Round-Up: Four Data Scientists Share What It’s Like to Work at Realtor.com

Ever wonder how Realtor.com tailors its online experience and marketing outreach to meet the needs of different types of visitors? Look no further than the Data Science organization, which is on a mission to create relevant, personalized experiences for our buyers and sellers.

Within the organization are four teams, each focused on solving a different challenge:

Consumer: How do we learn as much as possible about our consumers?
Monetization: How do we leverage machine learning to make our business more successful?
Discovery: How do we enhance our product to better engage users?
Experimentation: How do we use testing to inform our strategy and improve our performance?

We recently interviewed four of our crackerjack data scientists from these teams about their current projects and what they love about working at Realtor.com. Here’s what they had to say.

High Performance NodeJS Microservices in AWS

Three Simple Tips to get the most out of AWS in a REST service

Out of the box, the AWS SDK for NodeJS performs reasonably well, but subjecting services to high load can often reveal a few deficiencies. Specifically, intermittent connection timeouts and response latency can trigger alerts and ultimately frustrate end users. The following strategies not only mitigated these issues, but unexpectedly boosted performance across the board for a key service in realtor.com which averages ~400 requests per second.

NodeJS Exports for Beginners

Export modules may be something a programmer is interested in, in scenarios where they discover common cross project functionalities and would like to refactor duplicated code into an independent export project shared between projects. This tutorial is meant to give a quick start to creating NodeJS export modules for both new NodeJS developers working with JavaScript, TypeScript or NestJS framework, looking to create an export module. We will go through simple cross repository imports and exports of modules and classes. The first part will demonstrate JavaScript exports, the second part will be for exporting TypeScript, and the last part will show how to export services for NestJS framework using NestJS modules.

Evolving Personalized Recommendations using Match Score

For most buyers, finding a dream home is a daunting task. At realtor.com, we’re helping users along in their journey using a system called “Match Score”, which evaluates the relative importance of any home from the user’s perspective. The user’s Match Score is derived from a variety of features inferred from the user’s search history such as average home price, beds, baths, and lot square foot. As the user interacts with homes on realtor.com, a score is computed in real-time that will estimate the user’s preference for a particular home.

Match Score is a generic personalization model that can be used to personalize many aspects of the website. For example, Similar Homes is a prominent machine learning module at realtor.com which enables users to find homes similar to the homes they are looking for. The Similar Homes Machine Learning model which recommends similar homes will now be personalized for every user using Match Score. Match Score will also drive the notification system in ranking the potential candidate homes for each user.

You might want to get notified of the personalized best homes in the market!

The lifecycle of machine learning projects is always iterative to collect data, train, and serve the model. In this blog post, we describe the end to end design and implementation of the Match Score and the challenges faced along the way.

Our self-hosted to MongoDB Atlas journey

This article describes our self-hosted to hosted migration of several key databases. We started by self-hosting MongoDB on EC2 instances and transitioned to MongoDB Atlas.

Scaling mobile development at realtor.com with Flutter

An overview of our Flutter journey and why we decided to use Flutter in our native mobile apps.