As a real estate guy, sometimes I look over to my data scientist’s screen and believe what he is doing is some sort of magic — it’s like pure computer science sorcery.
I mean, they’re building a chatbot that can communicate using human language and learn from the conversations it’s having.
It’s like me every time I have a conversation with my data science team I learn something about artificial intelligence and I’d like to share that with you. So straight from the minds of 3 PhD’s in statistics, mathematics and computer science – here’s how AI is being applied to real estate.
Artificial Intelligence (AI)
Let’s start at the top, with AI. Generally speaking most true techies don’t actually call what they are doing “AI” – they break it down into a specific technology like:
- Deep learning
- Natural Language Processing
- Predictive analytics
- Machine learning
- Supervised learning vs. unsupervised learning
AI is when a computer can do things humans need intelligence to do.
Think about all the decisions that go into assessing the value of a property – that could be one example of AI, but generally that’s a predictive analytics or deep learning model.
Analytics play a huge role in AI. In fact, I think that analytics is truly what shapes AI. Think about the analytics from your business. Your users viewed some properties, liked a few more, opened your emails with more properties then scheduled a showing for another property- all great analytics that can be used to power AI.
Most Automated Valuation Models (AVMs) are built with predictive analytics – i.e. lots of data describing previous home sales used to predict the value of a new similar property.
The Zestimate is a good example of this, built using a Random Forests model. This model is basically the sum of a huge number of decisions made across multiple decision trees (and lots of trees make a forest).
Deep learning is probably the closest thing we have to true AI. Some techies argue that deep learning is just computer scientists’ dreams and wishes but to the true nerds, deep learning is where the fun begins.
Today, my team explored a Neural Network model they built using thousands of property descriptions. A neural network is often thought of as a network of decisions made intelligently, like a human mind. But that’s just the media’s portrayal – here’s how they actually work:
In our case we took some thousands of listing descriptions and turned every word into a vector and plotted all the vectors in a huge multi-dimensional space. Next, we wanted to figure out which sentences and entire remarks were similar by examining the distances between words in a description in this multi-dimensional space.
Then we found cluster’s of words and descriptions near each other and we fed the model some words like “open concept” to find other words that were similar like “spacious” or “airy”.
To visualize, we automatically created a word cloud (shaped like our logo) that plots similar words near each other and more common words larger – check it out:
We even tried the word “hookup” for fun (since we are college kids) and our model returned “wine”, “bedroom” and “cellar”…which makes me wonder, what are agents writing about?
Needless to say, neural networks are fun and powerful models, but they are extremely picky and take millions of data points to become properly tuned.
Supervised and Unsupervised Learning
Just like some of us humans need supervision so to does AI. In fact, most of the workload for this type of AI is done by humans. At Structurely, our team has hand tagged tens of thousands of listing descriptions and another tens of thousands individual sentences in listing descriptions.
This data is fed into a model like TF-IDF (Term Frequency – Inverse Document Frequency) which is often thought of as “Old Faithful” in data science, to pick out and summarize listing descriptions based on the most important words – like “open concept” or “big backyard” and forget irrelevant words overly used by agents like “beautiful”.
Unsupervised learning tells the computer to learn something on its own just from the data. It’s very similar to deep learning because both require millions of data points. Companies like X.ai who build build bots to help you schedule meetings take the unsupervised approach by feeding millions of emails about meetings to their models.
What Does All This Actually Mean?
To be honest, sometimes I have no idea. AI is becoming a buzzword in most industries and rightfully so – it’s buzz-worthy work. However, as the geeks of real estate we need to make sure the industry is well informed of what AI actually is and what it can and cannot do.
Because AI has the opportunity to impact the most important aspect of real estate: relationships.