4 things I learned at KDD 2018
This year I went, for the first time, to the conference on Knowledge Discovery and Data (KDD 2018) in London. I come from an academic background, so I find conferences particularly useful to spot new trends in machine learning research and keep up to date with what’s happening in the community. Unfortunately I was only able to stay for a couple of days, though, as it’s all go at DigitalBridge! It was definitely worth it.
The most interesting things I gathered from the conference were:
Deep Learning is here to stay (we kind of suspected it). I went to several tracks, including commerce and profiling, or graph and social networks, and deep learning was literally everywhere. However, the most impressive thing for me was to see the exceptionally wide variety of applications. Deep learning has been traditionally used to solve computer vision problems such as object recognition and detection, but at KDD this year it’s been used to tackle diverse and multiple tasks, such as: understanding emotions behind a user’s sentence on social media, predicting the intensity of a cyclone, detecting fake news, checking infrastructure quality in Africa, helping with the transcription of documents in the Vatican Secret Archive, and even composing music! It was quite astonishing to see that brilliant researchers are managing to find a deep learning model or technique that helps in these areas.
The field of deep learning is definitely evolving very quickly, but there is still a lot to discover and understand; in fact, one of the most common complaints about deep learning is the fact that we don’t understand what’s actually happening behind the scene. That’s why there is a whole new branch of research that attempts to interpret deep learning models. Dr. Yan Liu, from University of South Carolina, gave a very interesting talk that included a discussion on interpretable machine learning models during the Deep Learning day.
The newest trend is deep learning on graphs. Up until a few years ago, if you had a variable-size input you would use recurrent neural networks (RNNs). However, these kinds of networks have been specifically designed for time sequences. This means that if your input doesn’t have a fixed size, but is not a sequence either, then RNNs are probably not the best solution to your problem. That’s where graphs come in your aid. In fact, the latest techniques allow you to apply deep learning on graph structured data. This is very useful in chemistry and topology problems, as well as on problems that concern social networks, but guess where it can also be applied? You got it! On home renovation! Here at DigitalBridge we're exploring these new methodologies right now. Stay tuned for news!
There is a whole new world in terms of personalised recommendations. At DigitalBridge we’re very interested in these kinds of techniques, as we’re working to create a tool that can generate smart suggestions for our customers based on their preferences. I found the following works particularly interesting because they approach different problems/offer original solutions in the context of recommendation:
”Buy It Again: Modeling Repeat Purchase Recommendations” from Amazon, on repeat purchase recommendation. You know when you’ve just bought a very expensive Nikon camera and a week later Amazon suggests that you may want to buy another one? This paper is about fixing that (hopefully). In fact, it’s about suggesting to buy again only products that are likely to be bought repeatedly, such as grocery, diapers and so on. The paper is written very clearly, and shares with the reader the basic intuition that led the authors to choose a specific model. This is one of the few papers I read that does not use deep learning.
“Product Characterisation towards Personalisation” from ASOS, on recommending fashion products. This work deals with partially labeled datasets; in fact, the authors’ claim is that in order to make effective recommendations, there should be a full understanding of the products. Since at ASOS they often have items with missing attributes, the paper addresses the problem of predicting them using an effective multi-modal multi-task architecture. They then propose a hybrid approach (collaborative filter + content-based) to recommend fashion products that make use of the attributes predicted at the previous stage. The paper is clearly written and cites a number of relevant previous works. I did miss an experiment showing that recommendations using predicted attributes were more effective than recommendations without predicted attributes.
“Real-time Personalization using Embeddings for Search Ranking at Airbnb” from Airbnb, tackles the problem of recommendations in two sided marketplaces, where both buyers and sellers should be made happy. Airbnb market has also other peculiarities that make the problem quite challenging, for example the fact that users travel only 1 or 2 times a year, so bookings are quite sparse, and that a user rarely books the same accommodation twice. This paper proposes a real-time personalisation system, based on users’ behavior during the same session (i.e., clicking on listings or skipping them). At the same time, they also propose to learn preferences of user types from users’ long term history. It is very nice to know that this solution works in real-time and it’s been launched to production!
The latest advances make me think that we’re a step closer to a world where Netflix will finally recommend me something different from cheap imitations of “My Bestfriend’s Wedding” (which I love by the way). On our side, we’ll make sure to get closer and closer to a world where customers will be recommended a worktop that will fit their space, will go with the kitchen they’ve chosen and that they’ll absolutely love, all in just a few seconds!
I was really impressed with the overall experience of KDD. The best thing, though, was that it showcases a mixture of research-based scientific works and more application-oriented techniques. This means it doesn’t matter if you’re a hard core mathematician or a data analyst who’s excited to use new frameworks; this conference is for everyone interested in data science and it is particularly beneficial to companies that want to venture into the world of artificial intelligence. The food isn’t bad either.