How Globo.com is Automating Metadata Extraction for Video Recommendation
Felipe Ferreira, Globo.com, Brazil
Globo.com is Grupo Globo’s Internet company and has some of the largest portals in Brazil (G1, GE, GShow and Videos). Thousands of massive video content is produced daily on a wide range of subjects. At Globo.com video recommendation systems have increasingly gained a key role in the personalization and contextualized distribution of multimedia content across digital on-demand video streaming platforms such as Globoplay. One of the natural challenges that comes with the large amount of videos produced daily is the creation of metadata for these content in order to enrich the recommendation algorithms in a scalable way. We will talk about how automatic metadata extraction through computer vision and NLP techniques has been explored to support the creation of algorithms for video content-based recommendation.
Explainable Machine Learning for a Better Understanding of Search Engines at Elo7
Junior A. Koch, Elo7, Brazil
Elo7 is the largest Brazilian marketplace for handcrafted products. Search is a key tool for a user to find products at Elo7. Hence, the main goal of our search system is to optimize relevance. Today, this is accomplished through the boost of manually specified fields inside our search retrieval engine. However, since there are many fields and many different scores, understanding the search results is not trivial. To make sure the search engine behaves as expected, with the least possible side-effects, we developed an Explainable Machine Learning model using the novel method SHAP (SHapley Additive exPlanations), which is based on cooperative game theory. The model was able to show how products features used by the search engine contribute as coalition for ranking and relevance of the retrieval set.
How Buzzmonitor is Using Facebook and Instagram Data to Improve Customer Experience for Global Brands
Fábio Leal, Buzzmonitor, Brazil
Buzzmonitor is the leading software tool for Social Media in Latin America and Europe. On average, over 4000 posts or comments are added to our database every second, and over 8000 interactions are sent each day by agents that use our platform. In this talk we will discuss how Buzzmonitor uses social media data along with machine learning and other recommendation techniques to improve customer experience for the companies that we serve, such as Coca-Cola, Rede Globo, Casas Bahia, Renault, UFC and hundreds of others.
Since its foundation in 1957, Magazine Luiza has grown from a single physical store to more than a thousand stores, several distribution centers and omnichannel presence. As one of the most relevant retailers in Brazil, Magalu has earned trust of more than 20 million customers for which it offers a item catalog of more than 10 million products. In order to understand, personalize and simplify these consumers’ purchase decisions, Magalu have been deploying several recommender models over the years, both from industry and academia. This talk presents how recommender systems were used in the past, how Magalu currently personalize customer experience and what challenges are expected for the next few years.
Challenges and Research for a Real-Time Recommendation in a Dynamic Marketplace Environment
Tiago Motta, OLX, Brazil
OLX is the largest online classifieds marketplace in Brazil. We help millions of Brazilians to sell and find products and services every day. Unlike the majority of the existing recommendation systems, most of the time we have unique items that can be consumed by only one user. Besides, we have a flow of more than a thousand new ads per minute. Those singular differences give us many peculiar challenges from mainstream solutions. In this talk we are going to explain how our recommendation system works, trade-offs of our approach and our newest research lines to face those challenges.
The Problem of Matching Scarcity: When RSs Don’t Connect the Edges in Recruitment Services
Fernando Mourão, SEEK, Brazil
Connecting candidates and jobs to promote real placement opportunities is one of the most impacting and challenging scenarios for Recommender Systems (RSs). At SEEK, we work to provide such personalized solution for millions of people in different countries. A major concern for us when building RSs is ensuring placement opportunities for all candidates and jobs on the system as soon as possible. Indeed, long waiting periods cause financial damages to both sides. We refer to these scenarios where candidates or jobs suffer from the absence of matching in the system as the Problem of Matching Scarcity (PMS). This talk introduces the PMS, discussing the reasons we consider PMS as a recurring threat to recruitment services, and our efforts to identify, characterize and mitigate it on real scenarios.
Food is a very personal choice. At iFood, we are obsessed about Customer Experience and want to make food discovery in the platform seamless and a delight to the consumer. We’ll also talk about our Journey, Learning and Challenges in building the Food Recommendation System using Sagemaker and Databricks. Do you know that feeling that an app knows exactly what you want and the marketing team must know you so well because that push notification read your mind and every list seems like it’s been tailor-made for you? Yep, it actually was.
How Virtus is Using Recommender Systems to Solve Sofware Engineering Problems
Mirko Perkusich, Virtus, Brazil
Virtus is a Research, Development and Innovation Center at the Federal University of Campina Grande. We have over 200 people working on over 30 concurrent projects in several domains ranging from Information Technology, Communication, and Automation. A plethora of data is created daily regarding software artifacts, including requirements, source code, test cases. Along with this, there is an increasing challenge to manage the projects and resources, motivating the use of intelligent tools to support managers. Given this, to maintain market competitiveness, improving its efficacy and efficiency, using the data of past projects to support decision-making has become a priority at Virtus. One of the challenges that come with analyzing a large-volume, mostly unstructured data, is to extract the features of the software artifacts produced daily to enrich the recommendation algorithms. This talk focuses on our experience in using tags and categories to support the creation of algorithms for recommending non-functional requirements for a given project. It discusses the faced challenges, our solutions, and the challenges expected for the next few years.