Projects

Abstract: Human concept learning is semi-supervised learning (SSL) because people naturally combine a few labeled data with lots of unlabeled data to make classification decisions. Human concept learning is both static and dynamic. To accommodate…

Abstract: Location-based data mining has been extensively studied as spatial-temporal trajectory data accumulates. This field studies self-supervised pre-training to learn location embedding vectors. Pre-trained embedding vectors can use abundant unlabeled trajectory data to benefit downstream…

Abstract: Information dissemination mining, ads recommendation, and other applications require early hot event prediction. Existing methods require long-term event observation or expensive feature extraction. Hot and non-hot events have different temporal features, but early data…

Abstract: Geographical topic models are used in recommendations, user mobility modeling, event detection, and geo-tagged document mining for topical region and geographical topics. Studying effective geographical topic models ignores efficiency. However, training geographical topic models…

Abstract: Periodicity is common in temporal social networks. Understanding temporal network group behaviors requires mining periodic communities. Most temporal network community mining studies ignore community periodicity. This paper investigates finding periodic communities in a temporal…

Abstract: A general-purpose, assumption-free imputation method for multivariate missing data, fractional hot-deck imputation (FHDI) fills each missing item with multiple observed values without using artificial values. FHDI J. Im, I. Cho, and J. K. Kim,…

Abstract: Modern traffic management relies on crowd flow prediction (CFP) for many intelligent transportation services tasks. Most methods only forecast metro station entrance and exit flows, which is not enough for traffic management. Managers desperately…

Abstract: Many multi-view clustering algorithms have been proposed as real-world applications use multi-view data. Existing algorithms usually focus on the cooperation of different visible views in the original space but neglect the influence of hidden…

Abstract: Multiset membership lookup takes an item e as input and outputs a binary answer whether e is in S and, if so, the subset ID. Multiset membership lookup, more advanced than canonical membership lookup,…

Abstract: This paper discusses semi-supervised learning instance and feature co-selection. Co-selection is harder when data contain labeled and unlabeled examples from the same population. sCOs, a unified framework, efficiently integrates labeled and unlabeled parts into…

Abstract: Most advanced scientific, business, and social media applications use online analytics to efficiently execute large numbers of Aggregate Continuous Queries (ACQs). ACQs aggregate streaming data and periodically calculate max or average over a window…

Abstract: Many trajectories have user ID, location ID, time-stamp, and activity type (like “topic” in text mining) due to the widespread use of positioning devices. Existing works use latent activity types from topic models or…