The main goal of the ODD workshop is to bring together academics, industry and government researchers and practitioners to discuss and reflect on recent outlier mining challenges.
The ODD (2013) workshop focused on outlier detection and description, with particular emphasis on descriptive methods that could help make sense of the detected outliers. Next, ODD^2 (2014) extended the focus areas to outlier detection and description under data diversity, with emphasis on challenges associated with mining outliers in heterogeneous data environments (graphs, text, streams, metadata, etc.). ODDx3 (2015) focused on the translation of real world applications to different outlier definitions, highlighting the challenges associated with the variety of outlier definitions defined in theoretic models and used in a multitude of application domains. The ODD 4.0 (2016) broadened the scope to industrial challenges (e.g. known from Industry 4.0 initiatives) for on-demand computation, visualization, and verification of outliers in industrial settings.
This year, our workshop is motivated by the need for new means to de-construct the black-box nature of outlier detection methods to offer solutions for predictions to be interpreted, adopted, trusted, and safely used by decision makers in mission-critical applications. Roughly speaking, by de-construction we mean the process of tracing the contribution of each input to the output (for a given example) and evaluate to which extent a particular input would move the output due to inherited variations.
The glossary definitions of the word deconstruct include “analyze (a text or a linguistic or conceptual system) by deconstruction, typically in order to expose its hidden internal assumptions and contradictions and subvert its apparent significance or unity” and “reduce (something) to its constituent parts in order to reinterpret it”. This is exactly what the ODD v5.0 workshop focuses on in the context of outlier mining, that is, identifying the constituent parts of a detection model to expose its hidden/underlying reasoning to flag an outlier.
In short, ODD v5.0 (2018) aims to increase awareness of the community to the following topics on outlier mining:
-How can we (verbally or visually) explain the reasoning behind the decisions of various outlier detection models?
-What is the extent to which we can draw causal (i.e. beyond descriptive) explanations to the emergence of outliers?
-What techniques can be used for identifying root causes and generating mechanisms of outliers for diagnosis and treatment?
-How can we leverage interactions with human experts to mine outliers?
-How can we incorporate complex user feedback for outlier detection?
-How can we employ novel deep learning models for outlier detection?
-How can we create an ensemble of outlier detectors that is interpretable?
-How can we apply recurrent models to outlier detection in complex data such as graph or text data streams?
-How can we design explanation techniques for complex detectors such as deep models as well as ensemble detection methods?