Monthly Seminars – 2022 – DKM department – IRISA

Speaker: Farouk Toumani (LIMOS, Clermont Ferrand)
Title: Reasoning in description logics with variables: beyond matching and unification
Abstract: Description logics (DLs) are a family of logic-based knowledge representation languages used to represent knowledge of an application domain in a structured and well-understood way while providing inference procedures for reasoning about the represented knowledge. DLs form a very active research area, spanning the last 40 years, and which led to a rich variety of description formalisms, with differing expressive power, employed in various application domains (e.g., information systems and databases, natural language processing, semantic web, etc). Standard inference problems (e.g., subsumption, satisfiability and instance problem) as well as their computational properties (decidability, connection between complexity and expressivity) are now well understood. After a brief introduction to description logics nonstandard inferences, this talk will be devoted to presenting description logics augmented with (concept or role) variables. Concepts with variables (also called patterns) have been introduced in description logics since the midnineties and led to a highly interesting research stream on the so-called nonstandard reasoning, specifically, matching and unification. The talk will discuss new semantics for variables and highlights new inference problems which can be viewed as generalization of matching and unification.

Speaker: Charlotte Truchet (TASC, University of Nantes)
Title: Randomization of solutions in constraint programming
Abstract: Constraint programming provides generic techniques to efficiently solve combinatorial problems. In this talk, I will present an ongoing work on constraint sampling. The question is: is it possible to sample combinatorial problems in a generic way, using a constraint solver, and without making the computation time explode? I will present an algorithm, inspired from Meel’s method on SAT, to add randomly chosen hashing constraints to the problem, in order to split the search space into small cells of solutions. By sampling the solutions in a cell, it randomly generates solutions without revamping the model of the problem. We ensure the randomness by introducing a new family of hashing constraints: randomly generated tables. We implemented this solving method using the constraint solver Choco-solver. The quality of the randomness and the running time of our approach are experimentally compared to a random branching strategy. We show that our approach improves the randomness while being in the same order of magnitude in terms of running time.

Speakers
- Dominique Lavenier (CNRS, Fr) data storage on DNA
- Cynthia Liem (TU Delft, N) validation and validity in data processing pipelines
- Michèle Sebag (CRNS, Fr) causal explanations
- Julia Stoyanovich (NYU Center for Data Science, USA) building data equity systems

Speaker: Marie-Christine Rousset (SLIDE, LIG, Grenoble)
Title: Some ongoing work on building interpretable explanations for AI algorithms on tabular or graph data
Abstract: AI systems use sophisticated algorithms that apply to personal data for developing more and more decision making applications that directly impact humans. Both for social acceptability and for ethical purposes, it is of utmost importance to make the decisions of AI systems interpretable by humans and also to provide guarantees of privacy protection. In this talk, I will present some ongoing work that we are conducting in the Grenoble MIAI chaire “Explainable and Responsible AI” for building interpretable explanations of AI algorithms. In a first part, I will summarize experimental results that we have obtained on building local and global explanations for predictions of microcredit default learned by black-box models from a tabular dataset. In the second part, I will present our ongoing work on explaining privacy risks detected by a graph-based reasoning algorithm used to check incompatibility between privacy and utility policies expressed as queries. in this setting, queries are interpreted as logical formulas over a common schema and the explanation is based on the construction of a small synthetic graph data illustrating a possible entailment between graph patterns.

Speaker: Rémy Eyraud (LabHC CNRS, University Jean Monnet, Saint-Etienne)
Title: Hankel Matrix & Weighted Automata: From spectral Learning to Spectral Distillation
Abstract: This talk focuses on the Hankel matrix and Weighted Automata. After introducing these two notions and the main algorithm that links them, I will detail the spectral learning algorithm, a machine learning approach that comes with theoretical guarantees and interesting practical results. Extensions will be discussed with a particular insight on the extraction of weighted automata from any black box trained on (categorial) sequential data. If time allows it, a use case of an easy to use toolbox will be shown.

Speaker: Sihem Amer-Yahia (CNRS, Univ. Grenoble Alpes)
Title: Fairness on Online Labor Markets
Abstract: Online labor markets are increasingly becoming a destination for work. These marketplaces include freelancing platforms such as Qapa and MisterTemp’ in France, and TaskRabbit and Fiverr in the USA. On those platforms, workers can find temporary jobs in the physical world such as moving furniture, or in the form of virtual micro-gigs such as helping with designing a website. I will present the results of a study of fairness on those platforms, and discuss the design of a model to enforce fairness in the Future of Work.
Bio: Sihem Amer-Yahia is a Silver Medal CNRS Research Director and Deputy Director of the Lab of Informatics of Grenoble. She works on exploratory data analysis and fairness in job marketplaces. Before joining CNRS, she was Principal Scientist at QCRI, Senior Scientist at Yahoo! Research and Member of Technical Staff at at&t Labs. Sihem is PC chair for SIGMOD 2023 and vice president of the VLDB Endowment. She currently leads the Diversity & Inclusion initiative for the database community.