Keynote speakers

Biography :

- Leman Akoglu

- Luca Maria Aiello

- Stefan Kramer

LEMAN AKOGLU

Unsupervised Model Selection in Outlier Detection: The Elephant in the Room (pdf)

Outlier mining has a large literature containing numerous detection algorithms, as it finds high-stakes applications in numerous domains including finance, cybersecurity, surveillance, to name a few. However, given a new detection task, it is unclear how to choose an algorithm to use, nor how to set its hyperparameter(s) (HPs) in unsupervised settings. HP tuning is an ever-growing problem with the arrival of many new deep learning based outlier detectors. While they have appealing properties such as task-driven representation learning and end-to-end optimization, deep models come with a long list of HPs. Surprisingly, the issue of model selection in the outlier mining literature has been "the elephant in the room"; a significant factor in unlocking the utmost potential of (deep) models, yet little said or done to systematically tackle the issue.

In this talk, I will first quantitatively demonstrate the HP sensitivity of deep outlier detectors from various families based on a large-scale evaluation study. Next I will present a couple of new directions we have taken to tackle the unsupervised outlier model selection problem, including meta-learning based solutions that transfer "experience" from historical tasks for model selection for a new task, and scalable hyper-ensemble modeling that fully bypasses/obviates model selection.

LUCA MARIA AIELLO

Coloring Social Relationships (pdf)

Social relationships are the key determinant of crucial societal outcomes, including diffusion of innovation, productivity, happiness, and life expectancy. To better attain such outcomes at scale, it is therefore paramount to have technologies that can effectively capture the type of social relationships from digital data. NLP researchers have tried to do so from conversational text but mostly focusing on sentiment or topic mining, techniques that fall short on either conciseness or exhaustiveness. We propose a theoretical model of 10 dimensions (colors) of social relationships that is backed by decades of research in social sciences and that captures most of the common relationship types. We trained a deep-learning model to classify text along these ten dimensions, and we reached performance up to 0.98 AUC. By applying this tool on large-scale conversational data, we show that the combination of the predicted dimensions suggests both the types of relationships people entertain and the types of real-world communities they shape. 

We believe that the ability of capturing interpretable social dimensions from language using AI will help closing the gap between the oversimplified social constructs that existing social network analysis methods can measure and the multifaceted understanding of social dynamics that has been developed by decades of theoretical research.

STEFAN KRAMER

35 Years of 'Scientific Discovery: Computational Explorations of the Creative Processes': From the Early Days to the State of the Art (pdf)

It was 35 years ago, in April 1987, when the first book on computational models of scientific discovery was published: "Scientific Discovery: Computational Explorations of the Creative Processes" by Pat Langley, Herbert Simon, Gary Bradshaw, and Jan Zytkow contained a comprehensive account of systems for discovering quantitive empirical laws as well the discovery of qualitative and structural models, and marked an important milestone in a new branch of AI. Since then, methods for equation discovery, symbolic regression and the automation of science have been developed and refined, with many interesting problems remaining. Currently, deep neural networks (DNNs), representation learning, explainable AI (XAI), graph neural networks (GNNs), and many other technical innovations are bringing new elements into the field. At the same time, progress in the natural and life sciences is increasingly made by (and often requires) methods from AI and ML to produce models with high predictive and explanatory power.

In the talk, I will review progress in the field, applications from the natural and life sciences as well as a new test environment, with many options for extensions, that frames machine discovery as a reinforcement learning problem.