GPT-based Sequential Recommendation System
GPT-based Seq. Recommender System | CSE258: Recommender Systems, UCSD | Nov 2025 -- Dec 2025 |
- Developed a collaborative LLM-based next-item recommender that injects learned user/item embeddings as soft tokens into a full-finetuned GPT-2 backbone, combining autoregressive cross-entropy with dual context–item and user–item BPR losses on ~100K Kindle users and ~389K items under strict temporal splits.
- Outperformed strong SASRec, FPMC, and BPR baselines on the Amazon Kindle Store dataset, achieving Hit@10 = 0.558 and NDCG@10 = 0.413 on the test set using SASRec-style sampled evaluation with 1 positive + 100 negatives per query.