Deep learning — the next step for Recommendation systems?

Chances are if you are reading this, you have heard the phrase deep learning in one way or another. Deep learning is a term that is casually used in context with big data, machine learning, artificial intelligence (AI), and recommendation systems. This all sounds very cool, right? But what does it stand for? And what does it mean for the future of recommendation systems like Yusp?

What is deep learning?

Deep learning is a class of machine learning algorithms that uses artificial neural networks that have multiple layers of data. What this means is that the computer can analyse huge amounts of unlabelled data and structure it in a meaningful way without prior explicit instructions. This is achieved by deconstructing the data set layer after layer.

(reference: http://deeplearning.stanford.edu/wiki/images/4/40/Network3322.png)

In practice, using deep learning the computer can look at the picture of two kittens playing in the grass and actually describe the picture as two kittens playing in the grass.

(reference: http://on-demand.gputechconf.com/gtc/2015/presentation/S5630-Piotr-Teterwak.pdf)

The road to AI

Deep learning gets us closer and closer to achieving true AI. Imagine the possibilities, a well built deep learning program can identify patterns in virtually any kind of data set. This can improve the efficiency of voice search algorithms for your personal assistant softwares like Google Now, Cortana and Siri. Deep learning is already utilized to identify sound patters to drive recommendations for Spotify.

Deep learning can also improve speech-to-text transcription solutions that we are increasingly relying on for text input. Could the day come soon when keyboards are replaced by little bone conduction headsets? The technology is soon there, as with many groundbreaking innovations, social structures cannot keep up.

Neural processing can also be used for spam filtering and eCommerce fraud detection. Facebook is already using a form of deep learning to recognize the faces of users and their friends and probably to filter out sexual or other forms of prohibited pictures. Of course, deep learning still has a way to go and requires human supervision but as a system that is always learning and improving it is only a matter of time.

And better recommendations

And of course deep learning presents very exciting possibilities for personalization solutions. As Gravity R&D’s CEO, Domokos Tikk pointed out, deep learning opens up a plethora of opportunities for recommendation systems. Indeed Gravity R&D is always pushing the boundaries when it comes to developing the next generation of personalization solutions for our clients.

Besides analysing content metadata (CBF), user-item interaction data (CF), and context (CARS), one can take into consideration the content itself when modelling. This could mean serving even more accurate product recommendation by accounting visual similarity by analysing the product pictures themselves.

Deep learning can also help with the cold-start problem. We used RNNs for session-based recommendation in our latest research for ICLR 2016, taking into account the sequential information of user interaction for modelling. This helps in cold-start problem because it is applicable even if we have no user history, just using the current session data in the modelling approach.

Indeed the future is bright with better product recommendations using deep learning. And best of all, these advancements will not only be available to our enterprise clients. With our recommendation solution Yusp eCommerce players of all shapes and sizes can take advantage of Gravity R&D’s cutting edge innovations.