Dipak Meher

• Designed an LLM-based pipeline for knowledge graph construction from complex legal narratives, addressing long-range coreference, entity disambiguation, and reference normalization.
• Developed a three-stage coreference system using a type-specific prompt cache to resolve plural mentions, role shifts, and alias ambiguity, mitigating context-window limitations and loss-in-the-middle issues in LLMs.
• Achieved a 45.21% reduction in node duplication and 32.22% decrease in legal noise over GraphRAG and CORE-KG.

• Developed a modular LLM-driven framework integrating type-aware coreference resolution and domain-specific prompting to construct coherent knowledge graphs from human smuggling legal cases.
• Introduced sequential entity extraction and legal-specific filtering to reduce attention drift, suppress procedural noise, and improve entity-type accuracy in graph construction.
• Achieved a 33.28% reduction in node duplication and 38.37% reduction in legal noise over GraphRAG.

Quantifying Cross-Domain User Behavior Consistency with LLMs

• Proposed the first LLM-based framework to quantify user behavior consistency across domains via item feature extraction, sentiment analysis, and behavior alignment.
• Conducted an empirical study across 12 domain pairs and 4 state-of-the-art CDR models, showing current methods underutilize behavior consistency.
• Observed strong consistency in similar domains (Books–Movies: 0.87) and lower consistency in dissimilar domains (Electronics–Food: 0.23).

LLMs for Transfer Learning in Recommender Systems

• Co-authored an LLM-driven framework for cross-domain recommendation, improving personalization in sparse-data cold-start and warm-start scenarios.
• Benchmarked six baselines (TGT, CMF, EMCDR, PTUPCDR + DisenCDR, UniCDR) and evaluated prompt strategies with/without target-domain examples.

Vietoris-Rips Complex for Cross-Domain Cold-Start Recommendation

• Implemented and benchmarked baseline models (TGT, CMF, EMCDR, PTUPCDR) for cross-domain recommendation (SDM 2024).
• Developed VRCDR using Vietoris–Rips Complex and Area-based Triangulated Embedding (ATE) to model users’ niche source-domain preferences as geometric structures.
• Achieved up to 20% improvement over SOTA methods under extreme cold-start conditions.