Reading Time: 4 minutes

Good morning. Andrey and David here, filling in this week for Zank.

Today, we’re unveiling the newest version of our travel web app, affectionately called TRECS, short for “travel recommendations”.

How does it work? Start by entering one of nearly 10,000 places on Earth (and one not on Earth. Can you can find it?) Next, you’ll see up to 50 locations just like it. Click on the TRECS tag descriptions at the top to customize your list of places and uncover special aspects of the destination you searched for.

Who’s it for? TRECS is for adventurers who want to sift through the vast number of places in the world and find the perfect next destination.

Why did we build it? Apart from being data scientists and knowing that we can turn the several data sources we have into a data product, we noticed there was a T-Rex size hole in the travel market. It’s just not easy at all to figure out where you want to go next. Do you pick up a travel guide? Talk to a friend who went and stayed with a friend on a remote island in Thailand? Do you whip out a globe and see what fate has in store for you? All interesting ideas, but pretty unlikely you’ll find exactly what you wanted since no one resource has a good amount of information about the entire world. Until TRECS, the best apps we could find only helped you figure out what you wanted to do once you decided where to travel. Further, those recommendations are based on user ratings, which are often highly skewed or, even worse, bought.

Let’s test it out: Andrey, our Head of Data Science, here. I’ve long been an avid traveller. Known for its world-class art scene and architecture, Venice is one of my favorite places to visit. Unfortunately, just as local Venetians do, I get quickly tired of huge crowds. Recently, I have been feeling picky and am in the mood for something just as artistic, but far less popular. When searching for Venice on TRECS, I see that Italian cities like Florence, Rome and Naples are among my top recommendations, which is hardly a novel discovery for an art aficionado. However, when I click on “painting” and “art” tags, I’m surprised to find that Sarajevo — “Bosnia and Herzegovina’s urban and artistic hub” — is my top destination.

Are you also an art and culture fan? Do you fancy seeing a contemporary art exhibition in an abandoned nuclear bunker (running until October 26, 2019)? Be quick to check out these links to get even more art and design inspired articles and blog posts about the unexpectedly vibrant city of Sarajevo.

How did we build it? We assembled a massive dataset from many sources that required extensive cleaning. From there, we used state-of-the-art Natural Language Processing (NLP) techniques such as: Term-Frequency Inverse-Document-Frequency (TF-IDF), Latent Semantic Indexing (LSI), and deep neural network embeddings such as those from BERT. We combined several models together with some filters such as climate, and tweaked the results until we were satisfied with the recommendations. Since it’s based on unsupervised learning, there’s no right or wrong. In other words, we’d love your feedback.

At BDS, we know how to move from idea, to data gathering, to V1+ production to create data products that bring joy to users. We hope you love TRECS. Don’t forget to send us your travel photos!

You can find TRECS here.

Of Interest

We Can do Better
Why do only 13% of data science projects, or around one out of every 10, actually make it into production? That’s one of the topics tackled by Deborah Leff, CTO for data science and AI at IBM. This is a huge dilemma with A.I. It’s very difficult to teach the production part in school because it’s generally out of the scope of what can be accomplished in an academic setting. And since most companies struggle to deploy meaningful, impactful models, it’s a difficult skill to learn. It’s not about money or company size. It’s about access to data, compute resources, collaboration, know-how and it’s got to be driven from the executive staff. There are a heck of a lot of stars that need to align to achieve effective data science. We recommend hiring A.I. professionals who have done it before and understand the pitfalls. Read the article here.

Recommenders vs. Product – Who Wins?
There’s no doubt that personalized product offerings are beneficial to customers and the companies that serve them. After all, personalization is a massive part of A.I. Netflix spends a fortune each year on personalization. In fact, many years ago (2006) they offered $1M to anyone who could make there recommendations just 10% better. It turns out there were a lot of flaws with the rationale that making their recommendations better would be useful. Netflix figured this out, and in 2009 when a team finally won the Netflix Challenge, Netflix dolled out the $1M then promptly refused to use the algorithm. It was simply too complex. There’s a pace to business and a pace to science. Science needs to keep up and to serve the right purpose (be product driven). This article shows what happens as business outpaces personalization.

Give me jeans not shoes
From the ever wonderful StitchFix blog, comes this look at natural language processing (NLP). How would you (as a computer) untangle the sentence, “Give me jeans not shoes”? If we follow normal NLP steps and remove “stop words” (words that generally have no semantic meaning) we might end up with “Give jeans shoes”, removing “me” and “not”. But immediately the sentiment is gone.

They use the BERT model mentioned above to capture the nuance of their customer requests. It works surprisingly well and if that’s your thing, this technical article is worth a read.