Deep learning is a subfield of machine learning and it comprises several approaches to tackling the single most important goal of AI research: allowing computers to model our world well enough to exhibit something like what we humans call intelligence.
On a basic conceptual level, deep learning approaches share a very basic trait. DL algorithms interpret the raw data through multiple processing layers. Each of these layers takes the output of the previous one as its input and creates a more abstract representation of it. As a result, the more data is being fed into the right algorithm, the more general are the rules and features that it’s able to infer in relation to a given scenario and, therefore, the apter it gets at handling new, similar situations.
Google Translate’s science-fiction-like „Word Lens” function is powered by a deep learning algorithm and Deep Mind’s recent Go victory can also be attributed to DL – although the triumphant algorithm AlphaGo isn’t a pure neural net, but a hybrid, melding deep reinforcement learning with one of the foundational techniques of classical AI — tree-search.
Deep learning is an ample approach to tackling computational problems that are too complicated to solve for simple algorithms, such as image classification or natural language processing. However, its current business uses are very limited. It is quite possible that a large portion of the industries that currently leverage machine learning hold further unexploited potential for deep learning and DL-based approaches can trump current best practices in many of them. For instance, one could read several articles in the past couple of months about how DL is going to revolutionize search, with Google’s former head of AI John Giannandrea taking over the company’s search department (and how this could potentially transform the field of SEO, as a whole).
Deep learning fueled recommender systems – the future of personalization
We are pretty sure that deep learning is going to be the next big leapfrog ahead in the field of personalization as well. Personalization constitutes more and more an area of focus for businesses ranging from eCommerce stores to publishers and marketing agencies due to its proven potential to drive sales, increase engagement and improve overall user experience. If data is the fuel of personalization, than recommender systems are its engine. The advances in these algorithms have a profound effect on the online experiences of users across domains and platforms.
Here we look at three specific areas where deep learning can complement and improve existing recommender systems.
Incorporating the content into the recommendation process
Item-to-item recommendations represent a standard task for recommender systems. This means, when an eCommerce store or publisher site recommends another product or piece of content that is similar to the one currently being viewed by the user. One typical approach to tackling this task is based on metadata (another typical data source is user interactions that fuel the Amazon-like “users who bought this item also bought…” logics). However, the poor quality of metadata is a recurring problem in a large percentage of real life situations: values are missing or are not assigned systematically. Even if meta-tags are perfect, such data only represents the actual item much more indirectly and in less detail than a picture of it, for instance. With the help of deep learning, the actual, intrinsic properties of the content (images, video, text) could be incorporated into recommendations. Using DL, item-to-item relations could be based on a much more comprehensive picture of the product and would be less reliant on manual tagging and extensive interactional histories.
A good example of incorporating the content into a recommender system is what Spotify was looking into in 2014, in order to make its song recommendations more diverse and create an improved personalized experience for its users. The music streaming service uses a collaborative filtering method in its recommendation systems. But Sander Dieleman a Ph.D. student and intern at Spotify saw this as their biggest flaw, as such an approach that relies heavily on usage data inevitably underrepresents hidden gems and lesser known songs of upcoming artists – the holy grails of music discovery. Dieleman, therefore, used a deep learning algorithm that he taught on 30-second excerpts from 500,000 songs to analyze the music itself. It turned out, that successive layers of the network learn progressively more complex and invariant features of the songs, as they do for image classification problems. In fact, “on the topmost fully-connected layer of the network, just before the output layer, the learned filters turned out to be very selective for certain subgenres”, such as gospel, Chinese pop or deep-house. In practice, this means, that such a system could effectively make music recommendations based on solely the similarity of songs (an excellent feature for assembling personalized playlists). We do not know for a fact, whether or not Spotify incorporated these findings into its algorithm, but it was nevertheless an intriguing experiment.
Tackling the cold-start problem
The cold-start is the arch-enemy of recommendation systems. It can affect both users and items. For users, the cold-start means when the system has limited or no information on the customer’s behavior and preferences. The item cold-start represents the lack of user interactions with the data upon which item-to-item relations can be drawn (we still have the metadata, though, but that won’t often suffice for truly fine-tuned recommendations). The item cold-start is an obvious domain for the aforementioned content-based approach as it makes the system less reliant on transactional and interactional data.
However, creating meaningful personalized experiences for new users is a much trickier problem that cannot necessarily be solved by simply gathering more information on them. It is quite typical – especially in the case of eCommerce sites or online marketplaces with wide product portfolios – that customers visit a website with completely different goals over time. First they come to buy a microwave, but the next time they’re looking for a mobile phone. In this scenario, the data gathered in their first session is not relevant to the second.
An intriguing approach to tackling the user cold-start problem is session based or item-to-session recommendations. This roughly means that instead of relying on the whole interactional history of customers, the system splits this data into separate sessions. The model capturing the users’ interests then builds on session-specific clickstreams. Through this approach it is quite possible that future recommender systems will not rely so heavily on elaborate customer profiles built over months or even years, rather they’ll be able to make reasonably relevant recommendations after the user’s been clicking away on the site for a while.
This is an area that is yet rather poorly researched, but possibly holds tremendous opportunity for enhancing personalized online experiences. Gravity R&D’s researchers working on the EU funded CrowdRec project recently co-authored a paper that describes a Recurrent Neural Network (RNN) approach to providing session-based recommendations. This is the first research paper that seeks to employ deep learning for session based recommendations and their results show their method significantly outperforms currently used state-of-the-art algorithms for this task.
The Four Moments of Truth
The four moments of truth are the brief time periods when customers make their decisions based on the company’s communication and the available information provided by them. These decisions are heavily influenced by long-term, personal preferences, and brand loyalty, but momentary impressions are also major factors. A deep learning-fueled approach to influencing customers during these Moments of Truth could lead to further, novel insights about the intrinsic human decision process.
We know, for example, that beautiful product pictures can drive sales (whole industries are built around making photos of rental rooms or food). But it would be interesting to assess through a DL-based image analysis approach what exactly are the visual characteristics of a product image that have significant positive effects on sales.
The above list is far from exhaustive. Personalization is no doubt one of the strongest imperatives today in the internet industry as a whole and deep learning almost certainly holds tremendous potential in this area. Therefore, businesses that aim to remain on the cutting edge need to keep an eye out for advancements in the field.
Balázs Hidasi Ph.D., Head of Data Mining at Gravity R&D and CEO, Domonkos Tikk, Ph.D. were expert sources for the article.
Like this article? Subscribe to our weekly newsletter to never miss out!