Current Focus

2025 Goals

• GNN-based approach for search and recommendation

  • Use past sequential history of items as user representation
  • Transformer-based reranking
  • Transformer-based user encoding for ANN retrieval

• Multi-task, multi-purpose embeddings

  • For retrieval and reranking
  • Across various services (jobs, courses, skills)

Research

• Optimizing LLM explanations based on implicit feedback

  • How to optimize an LLM to provide better recommendation explanations by fine-tuning on implicit feedback?

• Replacing BM25

  • How to design a search system that matches BM25 performance at cold start and gradually improves with more data, without dropping below BM25 performance?