Guang Wei Yu, Senior Data Scientist, Layer 6, MScAC CS UofT Title: Cold-Start Recommendations at Scale Abstract: Recommender systems are becoming a large part of our modern life. From movies, music, to gadgets and news, with the ever expanding amount of items available in every perceivable domain, we increasingly rely on these systems to serve users with a relevant subset that we think they will like. In this talk, I discuss challenges in modern industrial recommender systems and our contribution from Layer6 AI. First, I will describe recent advances addressing the cold-start problem, where we have new users or items that we are seeing for the first time. Then I will highlight how we tackle this problem at Layer6 AI, using a simple twist on the idea of Dropout. Second, I will highlight our work on frameworks to extend state-of-the-art approaches to large scale problem commonly found in real world data. An important consideration in every industrial recommender system application is scalability, as computational and memory cost pose real constraints that must be satisfied. Therefore, we always strive to balance performance with scalability at Layer6 AI. -- Mohammad Islam, Senior Data Scientist, Wattpad, MSc CS Ryerson Title: Session Based Recommendations at Wattpad Abstract: At Wattpad, we have observed that a majority of the reading activities happen in sequence and they vastly vary by time-of-the-day or namely: user session type. To address this, we designed a recommender system which can model users temporal dynamics and predict what story a user is going to read next based on users sequential behaviour in a particular session. -- Putra Manggala, Data Scientist at Shopify, MSc Mathematics McGill Title: On Personalization for Merchants at Shopify: Heterogeneity of Items, Business-centric Explanations, and Temporal Aspects Abstract: Over than 500,000 merchants of all sizes and types use the Shopify platform to obtain a single view of their business and manage their stores across multiple sales channels, including web, mobile, social media, marketplaces, brick-and-mortar locations, and pop-up shops. Our merchants range from the niche stores to large brands such as Tesla, GE, and Kylie Cosmetics. Throughout the development of their business, merchants are faced with an overwhelming number of store customizations, business decisions, third party apps and contractors. It is critical for every merchant to perform the best sequence of actions, based on this large set of heterogeneous items, that is heavily personalized for their current sales volume, current target customers, business type, location, language, staff behaviours, and many other dimensions. Aside from the identification of the optimal sequence of actions and items to power recommendations, the delivery of these recommendations must be supported with clear business-centric explanations that also reduce the barrier to entry for the merchants to take action. There is a variety of delivery mediums for these personalized recommendations, such as an interactive feed of content, an app store, and email, to name a few. The choice of medium affects the user experience design of the recommendation presentations and explanations. This talk will describe a few of the recommender systems we have built for these different mediums. I will discuss 1) how the systems treat the heterogenous nature of the item sets, 2) the temporal aspects of the recommendations, such as the consideration of periodical external events that affect the merchants and the personalized timing between different recommendations, and 3) personalized methods for doing business-centric explanations. -- Javier Moreno, Senior Data Scientist, Rubikloud, PhD, Mathematics (UIUC) Title: Restricted Recommendations And Other Problems From Retail Personalization Abstract: On a weekly basis, Rubikloud delivers individualized recommendations for tens of millions of customers. Our clients, however, need these recommendations to fulfill a varying array of requirements and restrictions. This makes out-of-the-box recommender systems vastly insufficient. This talk will present an overview of our current solution as well as some of the main related problems we are trying to solve at this moment.