SeongKu Kang

We propose CCQGen framework to generate queries with comprehensive coverage of document's concepts.

We propose ERAlign, an unsupervised and robust cross-lingual entity alignment pipeline.

We propose Taxonomy-guided Semantic Indexing (TaxoIndex) for effective academic concept matching in paper search.

We introduce a new research direction that combines knowledge distillation and continual learning for practical recommender systems.

We present a new conversational recommendation dataset named PEARL, synthesized with persona- and knowledge-augmented LLM simulators.

We propose SCRAP which optimizes its model to generate reasonings and the corresponding sentiment quadruplets in sequence.

We introduce a new multi-domain sequential recommendation method, specifically targeting the challenging scenario where recent interactions are highly sparse.

We propose DynaCor that distinguishes mislabeled instances based on the dynamics of the training signals.

We introduce a new approach to leverage dissensus of models to mitigate the popularity amplifications of a large-scale RS.

We introduce a new plug-and-play ToTER framework which improves PLM-based retrieval using a corpus topical taxonomy.

We propose Top-Personalized-K Recommendation, a new recommendation task aimed at generating a personalized-sized ranking list to maximize individual user satisfaction.

We propose a new multi-domain recommendation framework that learns various seen-unseen domain mappings in a unified way with masked domain modeling.

We propose Multi-view Feature Selection (MvFS), which promotes more balanced feature selection while mitigating bias toward dominant patterns.

We propose a new framework that compresses ensemble of heterogeneous models, reducing huge inference costs while retaining high accuracy.

We introduce a new topology-based gating module for molecular property prediction.

We introduce a new training strategy that exploits the complementarity from heterogeneous objectives for one-class collaborative filtering.

We introduce TaxoCom which recursively expands the topic taxonomy by discovering novel sub-topic clusters of terms and documents.

We propose two calibration methods for ranking model and a new unbiased empirical risk minimization framework to guide the calibration methods.

We propose GROC, a one-class classifier robust to geometrically-transformed inputs.

We introduce a new distillation strategy, distilling the preference knowledge in a balanced way without relying on any hyperparameter.

We introduce topology distillation approach that guides the student by transferring the topological structure built upon the relations in the teacher space.

We propose BUIR, a new training framework that does not require negative sampling.

We propose ProxySR which imitates the missing information of general user interest by modeling proxies of sessions.

We propose TimelyRec which exploits heterogeneous temporal patterns of user preference.

We introduce BD framework whereby both the teacher and the student collaboratively improve with each other.

We propose a new regularization method designed to maximize the effect of the ranking distillation.

We propose two methods: (1) DE for latent knowledge distillation, (2) RRD for ranking knowledge distillation.

We introduce DRE, a new framework to choose the initial seed items for new users.

We propose a new method of multi-modal feature combination for fake news detection.

We introduce a new semi-supervised learning solution that is effective when the overlapping users are exteremly limited.

- Ranked 13th among the most influential papers at CIKM 2019 (link)

We propose a new method to densify social network for recommendation.