Today’s world is full of sensors, social media and knowledge bases. Organizations, cities or individuals are increasingly required to make decisions leveraging the available information. Our competencies in machine learning allow us to fuse and analyse these sources of data in order to automatically make ongoing streams of optimal decisions
Machine learning aims at studying efficient algorithms which use historical data to improve their performance on tasks involving new data. It is a discipline that primarily draws itself from statistics and probability, mathematical programming and computer science. Two generic capabilities enabled by machine learning are predictive analytics (learning from historical data in order to take better decisions in the future), and knowledge generation (infer non-explicit knowledge from existing data).
Our scientific machine learning research is in the fields of:
Spatio-temporal data modeling: We investigate how to model data with complex spatial and temporal interconnections, and how to classify it automatically, make future predictions, impute missing data, and discover static patterns or dynamic trends. Examples include real-world networks such as public transport data, energy consumption data or social network contents.
Mechanism design: Sometimes also referred to as reverse game theory, it studies the design of optimal incentive systems to lead self-interested agents to a particular goal. Xerox applications of mechanism design include transport demand management (for example for on-street parking and optimal outsourcing.
Statistical relational learning: Our aim is to invent algorithms that can predict or correct values in a relational database by using the redundancies and similarities in the data. The ambition is to derive models and algorithms that generalize to a large variety of problems, including recommendation, time series prediction, text and image categorization or approximate logical reasoning. Our research in statistical relational learning has impact in Xerox’ transportation or customer care services, among others.
Knowledge transfer: We investigate models and algorithms capable of adapting existing knowledge to new target domains where either the data or the problems is different, but related to the “source domain”, with a minimum (or zero) amount of supervision. This includes transfer learning and domain adaptation techniques, with application to transportation and image data.
Sequential decision models: We study systems which require producing an optimal sequence of predictions, especially for systems which need to react to the environment or which interact with users. Our competencies in active learning, reinforcement learning and sequential optimization have applications in analysing trends or discovering topics in social media and other types of networks, design systems that optimally interact with users or find a stream of optimal decisions from data.