It is our first major project of Machine Learning. Since 2011, Geographica works with OneBookShelf, a company of digital content from the United States, doing custom development (web and mobile) for its different online shops specialised in the entertainment industry, with more than 500,000 users: DriveThruRPG, DriveThruCards, DriveThruComics, DriveThruFiction, DMSGuild, WargameVault and CurrclickReto.
These portals are leaders in the US in digital entertainment content sales, e-commerces with more than 500K users around the world that need to be constantly on the latter to maintain its position as leader.
OneBookShell, in an effort to ensure steady growth, it has a team of R&D that is devising new ways to improve the user experience, which becomes an increase in sales. It was through Edge Entertainment, a company which has strong trade relations with Geographica, how they contacted our team to turn those ideas into reality.
In 2011, Geographica began collaborating with OneBookShell, designing applications and doing custom development, however, a few months ago, we raised a new challenge that has made us grow with them:
The development of a recommendations system for users that were able to learn from past purchases of the user to make suggestions according to its tastes, thus increasing the chances of the sale, said: our first major project of Machine Learning. A major challenge in one of the most demanding technological and commercial sectors.
In response to your needs, we developed an application based on Machine Learning for implementing an intelligent, adaptive system with the ability to learn the preferences and habits of users are. Using a recommendation system, sales and sales frequency are increased for user, because the user experience is improved.
In fact the options provided by the Machine Learning to e-commerce are endless. On the one hand, it allows an analysis of user behaviour more in detail what would be done through normal analytical, which is an information with great courage, because let us see what works and why not, what are they looking for users and even how long it takes to find it.
On the other hand, as long as the system is able to recognize natural language, it can do an analysis of how users feel in relation to a particular product by the comments left and opinions.
Furthermore, as we advanced above, once learned preferences and behaviour patterns in the web user, the system will be able to offer customers things that it knows in advance and therefore, things that are easier to buy. This is an innovative system that will take OneBookShell portals to the next level.
Our solution has been searching for factors that are latent in products, nothing like categories or attributes of these, removing the possible bias included to introduce the product in the system. At the same time we found these factors, we looked how users react to them, so when recommending those products just look better relationship have regarding its factors.
This method, complicated a priori, is solved with matrix factorization, which allows us to work with widely scattered data, because the interaction of each user for each product is low.
In the implementation we use Python, host language with R in movement of Machine Learning and Data Science. With him, we get writing and ease of maintenance, and a blistering speed running, as the core of the different libraries used is written in C and C ++, so it cannot be faster.
In addition, we completed our recommendation system with Redis as cache system, and Docker, for easy and non-intrusive deployment.
The results are very encouraging and the system will continue to escalate, providing an outstanding competitive advantage to OneBookShell over other companies.
Header image: Boy looking at store window display of toys, between 1941 and 1942 (The Library of Congress). Flickr The Commons.