Hyeongjun Yang

Hyeongjun Yang

Auto-generated profile, based on publicly available information

Hyeongjun Yang

Department of Computer Science, Yonsei University, Seoul, Republic of Korea

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 April 2023
Verified
Conference Paper
Open access
HKHeeseon Kim
Heeseon Kim
HYHyeongjun Yang
Hyeongjun Yang
KLKyong-Ho Lee
Kyong-Ho Lee
Conversational recommender systems (CRS) aim to acquire a user’s dynamic interests for a successful recommendation. By asking about his/her preferences, CRS explore current needs of a user and recommend items of interest. However, previous works may not determine a proper action in a timely manner which leads to the insufficient information gathering and the waste of conversation turns. Since they learn a single decision policy, it is difficult for them to address the general decision problems in CRS. Besides, existing methods do not distinguish whether the past behaviors inferred from the historical interactions are closely related to the user’s current preference. To address these issues, we propose a novel Hierarchical policy learning based Conversational Recommendation framework (HiCR). HiCR formulates the multi-round decision making process as a hierarchical policy learning scheme, which consists of both a high-level policy and a low-level policy. In detail, the high-level policy aims to determine what type of action to take, such as a recommendation or a query, by observing the comprehensive conversation information. According to the decided action type, the low-level policy selects a specific action, such as which attribute to ask or which item to recommend. The hierarchical conversation policy enables CRS to decide an optimal action, resulting in reducing the unnecessary consumption of conversation turns and the continuous failure of recommendations. Furthermore, in order to filter out the unnecessary historical information when enriching the current user preference, we extract and utilize the informative past behaviors that are attentive to the current needs. Empirical experiments on four real-world datasets show the superiority of our approach against the current state-of-the-art methods.

1 January 2025 in IEEE Access
Verified
Journal Article
Open access
DKDonghyun Kim
Donghyun Kim
HYHyeongjun Yang
Hyeongjun Yang
SHSeokju Hwang
Seokju Hwang
KYKyuhwan Yeom
Kyuhwan Yeom
SMShim Midan
Shim Midan
KLKyong-Ho Lee
Kyong-Ho Lee

24 August 2024
Verified
Conference Paper
Open access
KYKyuhwan Yeom
Kyuhwan Yeom
HYHyeongjun Yang
Hyeongjun Yang
GPGayeon Park
Gayeon Park
MJMyeongheon Jeon
Myeongheon Jeon
YKYunjeong Ko
Yunjeong Ko
KLKyong-Ho Lee
Kyong-Ho Lee
Numerous large-scale knowledge graphs (KGs) fundamentally represent two-view KGs: an ontology-view KG with abstract classes in ontology and an instance-view KG with specific collections of entities instantiated from ontology classes. Two-view KG embedding aims to jointly learn continuous vector representations of entities and relations in the aforementioned two-view KGs. In essence, an ontology schema exhibits a tree-like structure guided by class hierarchies, which leads classes to form inheritance hierarchies. However, existing two-view KG embedding models neglect those hierarchies, which provides the necessity to reflect class inheritance. On the other hand, KG is constructed based on a pre-defined ontology schema that includes heterogeneous relations between classes. Furthermore, these relations are defined within the scope of those among classes since instances inherit all the properties of their corresponding classes, which reveals structural similarity between two multi-relational networks. Despite the consideration to bridge the gap among two-view KG representations, existing methods ignore the existence of structural similarity between two-view KGs. To address these issues, we propose a novel two-view KG embedding model, CISS, considering Class Inheritance and Structural Similarity between two-view KGs. To deal with class inheritance, we utilize class sets, each of which is composed of sibling classes, to learn fine-grained class representations. In addition, we configure virtual instance-view KG from clustered instances and compare subgraph representations of two-view KGs to enhance structural similarity between them. Experimental results show our superior performance compared to existing models.

30 January 2025 in IEEE Transactions on Knowledge and Data Engineering
Verified
Journal Article
Closed access
HYHyeongjun Yang
Hyeongjun Yang
DKDonghyun Kim
Donghyun Kim
GPGayeon Park
Gayeon Park
KYKyuHwan Yeom
KyuHwan Yeom
KLKyong-Ho Lee
Kyong-Ho Lee
Unlike the traditional recommender systems that rely on historical data such as clicks or purchases, a conversational recommender system (CRS) aims to provide a personalized recommendation through a natural conversation. The conversational interaction facilitates capturing not only explicit preference from mentioned items but also implicit states, such as a user’s current situation and emotional states from a dialogue context. Nevertheless, existing CRSs fall short of fully exploiting a dialogue context since they primarily derive explicit user preferences from the items and item-attributes mentioned in a conversation. To address this limitation and attain a comprehensive understanding of a dialogue context, we propose CoreSense, a conversational recommender system enhanced with social commonsense knowledge. In other words, CoreSense exploits the social commonsense knowledge graph ATOMIC to capture the user’s implicit states, such as a user’s current situation and emotional states, from a dialogue context. Thus, the social commonsense knowledge-augmented CRS can provide a more appropriate recommendation from a given dialogue context. Furthermore, we enhance the collaborative filtering effect by utilizing the user’s states inferred from commonsense knowledge as an improved criterion for retrieving other dialogues of similar interests. Extensive experiments on CRS benchmark datasets show that CoreSense provides human-like recommendations and responses based on inferred user states, achieving significant performance improvements.


6 November 2026
Verified
Conference Paper
Open access
JCJinho Choi
Jinho Choi
HBHye-Yoon Baek
Hye-Yoon Baek
SPSang-Hyun Park
Sang-Hyun Park
SKSang-June Kim
Sang-June Kim
HYHyeongjun Yang
Hyeongjun Yang
BOByungkook Oh
Byungkook Oh
Knowledge graph question answering (KGQA) increasingly relies on graph-based retrieval-augmented generation (GraphRAG). However, graph retrievers are trained under path-level weak supervision, which treats every triple on an answer-reaching path as equally informative. Retrieved subgraphs thus contain spurious triples that reach the answer without causally supporting it, leaving noise filtering to the reasoner. We propose Ca usality-Guided S purious E vidence Pruning with G Flow Net (CaseFlow), which internalizes spurious-evidence pruning into the retriever by reformulating retrieval as a flow-guided pruning process. Repurposed from a candidate generator into a pruner, the GFlowNet removes spurious evidence while retaining answer-supporting causal structure. To make this pruning causally informed and tractable, a relation-conditioned disentanglement encoder supplies a causal-relevance proxy for the chunk-level pruning prior, and adaptive relation-coherent chunks bound the action space. CaseFlow outperforms strong GraphRAG retrievers on the WebQSP, CWQ, and MetaQA-3hop benchmarks, with the largest gains in complex multi-hop reasoning. Our code is available at https://github.com/ChoiPi2/CaseFlow