Supervision : Ludovic DENOYER

Co-supervision : ARTIÈRES Thierry

Machine learning under budget constraints

This thesis studies the problem of machine learning under budget constraints. More specifically, we propose to focus on the cost of the information used by the system to predict accurately. Most methods in machine learning usually defines the quality of a model as the performance (e.g accuracy) on the task at hand, but ignores the cost of the model itself: for instance, the number of examples and/or labels needed during learning, the memory used, or the number of features required to predict at test-time. We propose in this manuscript several methods for cost-sensitive prediction w.r.t. the quantity of features used. The goal is to define models that learn to interact with the examples in order to optimize a trade-off between a good prediction and a low features' acquisition cost. We present three models that learn to predict under such constraint. The first model is a static approach applied to cold-start recommendation, where the subset of features to acquire is the same for all examples. We then define two adaptive methods that adapt their acquisition strategy depending on what is currently known about the example to classify, thus allowing for a better trade-off between accuracy and features' cost. We rely on representation learning techniques with recurrent neural networks architectures. In the last part of the thesis, we propose to study the problem of active-learning, where one aims at constraining the amount of labels used to train a model. We present our work for a novel approach of the problem using meta-learning, with an instantiation based on bi-directional recurrent neural networks.

2014-2018 Publications

• 2018
  • G. Contardo, L. Denoyer, Th. Artières : “A Meta-Learning Approach to One-Step Active Learning”, (2018)
• 2017
  • G. Contardo : “Machine learning under budget constraints”, thesis, phd defence 07/10/2017, supervision Denoyer, Ludovic, co-supervision : Artières, Thierry (2017)
  • G. Contardo, L. Denoyer, Th. Artières : “A Meta-Learning Approach to One-Step Active-Learning”, International Workshop on Automatic Selection, Configuration and Composition of Machine Learning Algorithms, vol. 1998, CEUR Workshop Proceedings, Skopje, North Macedonia, pp. 28-40, (CEUR) (2017)
• 2016
  • Th. Artières, G. Contardo, L. Denoyer : “Recurrent Neural Networks for Adaptive Feature Acquisition”, 23^rd International Conference on Neural Information Processing (ICONIP 2016), vol. 9949, Lecture Notes in Computer Science, Kyoto, Japan, pp. 591-599, (Springer) (2016)
  • A. Ziat, G. Contardo, N. Baskiotis, L. Denoyer : “Learning Embeddings for Completion and Prediction of Relationnal Multivariate Time-Series”, European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning - ESANN, Unknown, Belgium (2016)
  • G. Contardo, L. Denoyer, Th. Artières : “Sequential Cost-Sensitive Feature Acquisition”, IDA 2016, Unknown, (2016)
• 2015
  • A. Ziat, G. Contardo, N. Baskiotis, L. Denoyer : “Car-traffic forecasting: A representation learning approach”, Workshop MUD, Mining Urban Data 201, Unknown, (2015)
• 2014
  • G. Contardo, L. Denoyer, Th. Artières, P. Gallinari : “Learning States Representations in POMDP”, International Conference on Learning Representations, ICLR 2014, Banff, Canada (2014)
  • G. Contardo, L. Denoyer, Th. Artières : “Representation Learning for cold-start recommendation”, ICLR, Banff, Canada (2014)