The following special sessions are going to be part of the main programme of ICANN 2017. During the submission phase, you will be able to select whether you would like your contribution to considered for one of them.
S01: Self-assessment in advanced machine learning models
Organisers: Lydia Fischer (Honda Research Institute Europe and Bielefeld University, Germany), Cesare Alippi (Politecnico di Milano, Italy and USI, Switzerland), Barbara Hammer (Bielefeld University, Germany).
More information: https://www.techfak.uni-bielefeld.de/~bhammer/selfassessment.html
Self-assessment refers to the ability of a machine learning model to judge the security of its classification. It constitutes a crucial requirement in safety critical applications or whenever a human observer has to validate the given classification such as driver assistance systems, predictive maintenance of plants, or medical classifications; further, self-assessment constitutes one crucial property for classifier fusion. Probabilistic models can naturally be enhanced by the notion of model confidence, and popular deterministic models such as the support vector machine, can be accompanied by an efficient estimate of its confidence. However, while these technologies perform reliably for classical batch classification, their applicability is limited for complex machine learning scenarios such as online learning models, learning scenarios which are subject to drift, heterogeneous models, models which involve complex structured data, or interactive models which incorporate human expertise. The special session aims for contributions connected to the following non-exhaustive list of topics:
- confidence estimation in online learning techniques,
- efficient rejection strategies based on adaptive confidence estimations,
- self-assessment for heterogeneous classifier models,
- interactive machine learning classification models,
- self-assessment models for complex or structured data,
- robust self-assessment for changing costs,
- reliable self-assessment in the presence of concept drift,
- applications of machine learning techniques in safety critical or interactive scenarios.
S02: Learning from data streams
Giovanna Castellano (Dept. Computer Science, University of Bari, Italy)
Francesco Masulli (Dept. Computer and Information Sciences, University of Genova, Italy) francesco.masulli\AT/unige.it
Nowadays a wide range of real-world scenarios yield data streams, i.e. collections of data being generated continuously either over time or over space. E-commerce and banking transactions, weather forecasting recordings and sensor data, customer reports and network traffic records are common examples of data streams produced every day.
Accordingly there is an urgent need of methods capable to handle and analyze streams of data that are usually vast in volume (or possibly infinite), high-dimensional and changing dynamically. Analysis of data streams requires learning algorithms that are specifically designed not only to handle very large data sets but also to adapt continuously and automatically to smooth evolutions (drifts) and abrupt changes (shifts) in the data distribution.
The main objective of this special session is to discuss the potential of learning techniques in challenging scenarios involving prediction and classification tasks in the realm of data streams.
The special session is intended to collect novel ideas and share different experiences in the field of learning from data streams. Submission of papers covering topics in theoretical and applied learning techniques for data streams are encouraged. Possible topics include (but are not limited to):
- Incremental learning
- Online/Adaptive learning
- Semi-supervised learning
- Ensemble methods
- Dynamic/Incremental Clustering
- Frequent pattern mining