Data AI Lab
We are committed to uncovering data-centric solutions to complex, real-world challenges by harnessing the power of artificial intelligence. Our research areas span a multitude of problem levels, ranging from application-based solutions such as social network analysis, recommender systems, and anomaly detection, to theoretical constructs within machine learning and probabilistic inference.
Our research group is actively searching for talented and enthusiastic students. Please check this document for more information.
Research Overview
Graph Neural Networks in the Wild
How can we architect deep neural networks for use with graphs and social networks? How can we effectively train graph neural networks in situations where the available data is insufficient or noisy? Graph neural networks (GNNs) extend traditional neural networks to accommodate dynamic, non-grid data structures. By integrating GNNs with probabilistic inference and self-supervised learning, we enable them to tackle real-world problems that come with demanding constraints.
Keywords: Probabilistic Inference, Inductive Learning, PU Learning, Markov Graphs
Machine Learning on Time Series
How can we predict future values of time series? How can we discern the correlations between multiple time series variables? Our surroundings are teeming with time series data, from sensor outputs and prices, to electricity consumption, all of which invariably impact each other. Our goal is to unveil the hidden relationships between multiple time series variables and leverage this knowledge to enhance the performance of machine learning algorithms, particularly in the realm of finance.
Keywords: Transformers, Self-attention, Stock Price Prediction, Causal Inference
Recommender Systems
How can we recommend items considering both the personal preferences of individual users and the global trend within a system? Personalized recommendation is one of the most crucial topics in data mining, and is crucial for the success of online marketing, social media, and streaming services. Leveraging our expertise on graphs and time series, we aim to build a multi-modal recommender system capable of encompassing both spatial and temporal relationships between entities.
Keywords: Personalized recommendation, Ranking, Bipartite graphs, Locality
Self-supervised Anomaly Detection
How can we identify anomalies that have never been observed before? How can we train a model without any labeled data? Self-supervised learning (SSL) is the future of machine learning because it circumvents the substantial costs associated with labeling and curating refined data. Our research explores the application of SSL in anomaly detection scenarios, where labeled data is typically scarce, placing a specific emphasis on the creation of efficient data augmentation algorithms.
Keywords: Data Augmentation, Internal Performance Measures, Model Selection