Personalization on the internet by Deep Learning
An offshoot of machine learning, Deep Learning adopts various approaches to tackling the primary and most important goal of AI research: getting computers to model our world to the extent that they become capable to acquire something akin to what we humans call intelligence.
All deep learning approaches share a very basic trait on a conceptual level. Raw data is interpreted through multiple processing layers with the aid of DL algorithms. The output of the previous layer is taken as an input for each of the subsequent layers and a more abstract representation of it is created. Consequently as more and more data is fed into the correct algorithm, the easier it becomes for more and more generalized rules and features to be inferred in relation to a given scenario and, therefore, the smarter it gets at handling new, similar situations.
Two prominent examples of DL are Google Translate’s science-fiction-like “Word Lens” function which is powered by a deep learning algorithm, as also Deep Mind’s recent AlphaGo victory. However, the triumphant algorithm AlphaGo is not a pure neural net, but a hybrid, combining deep reinforcement learning and tree search – one of the foundational techniques of classical AI.
Complex computational problems such as image classification or natural language processing cannot be solved easily by simple algorithms; they are adequately addressed by Deep Learning. Yet the current business uses of DL are quite limited. Current best practices in many industries could soon be disrupted by those that currently leverage machine learning and unleash the unexploited potential for DL and deep learning based approaches. With Google’s former head of AI, John Giannandrea taking over the company’s search department, many articles in the recent past have deliberated about how DL is going to revolutionize search, and how the entire field of Search Engine Optimization is going to be radically transformed.
It is pretty certain that deep learning will be the next quantum leap in the field of personalization. Due to its proven potential it has been well established that Personalization drives sales, increases engagement and improves overall user experience and is an increasingly important focus area for businesses ranging from e-Commerce stores to publishers and marketing agencies. If we consider data to be the fuel for personalization, it follows that recommender systems are its engine. Advances in Personalization algorithms have a powerful impact on the online experiences of users across domains and platforms.
Let’s take a peek at three specific areas where deep learning can complement and improve upon the existing recommender systems.
A standard task for recommender systems would be item-to-item recommendations, for instance the e-Commerce store or publisher site recommends a similar product or piece of content that is similar to the one presently being viewed by the user. This could be handled with the use of metadata: the typical data source is user interactions which Amazon uses and results in something like “users who bought this item also bought…” logics. But, in a large percentage of real life situations values are not assigned systematically or are entirely missing; the poor quality of metadata is a recurring problem. While meta-tags may be perfect, this type of data only represents the actual item only indirectly and in less detail than, for example a pictorial. With the help of deep learning, the actual, intrinsic properties of the content including images, video and text can be weaved into the recommendations with the aid of DL and would be less reliant on manual tagging. Extensive interactional histories, reliance on manual tagging can be obviated by DL and item-to-item relations could be based on a much more comprehensive picture of the product.
