SeEn: Sequential enriched datasets for sequence-aware recommendations.

Area of Science:

• Scientific domains, including Astronomy and Chemistry.
• Information retrieval and recommender systems.

Background:

• Sequential recommendation systems, driven by deep learning (e.g., BERT4Rec), have advanced user preference modeling.
• The application of recommender systems for next-best-item prediction in scientific fields remains underexplored.
• Existing methods require adaptation for specialized scientific domains.

Purpose of the Study:

• To improve next-best-item recommendation accuracy in scientific domains.
• To adapt existing sequential recommendation techniques for scientific datasets.
• To introduce a novel dataset enrichment method to boost recommendation performance.

Main Methods:

• Adapted the LIBRETTI method for creating sequential recommendation datasets tailored to scientific fields (Astronomy, Chemistry).
• Developed a hybrid approach named sequential enrichment (SeEn) to augment existing sequential datasets.
• SeEn involves appending the 'n' most similar items after each original item in a sequence.

Main Results:

• The adapted LIBRETTI method provided a baseline for sequential dataset creation in scientific domains.
• The SeEn approach demonstrated significant improvements in recommendation accuracy.
• Chemistry datasets improved by ~7% and Astronomy datasets by ~16% in Hit Ratio and Normalized Discounted Cumulative Gain (NDCG).

Conclusions:

• Sequential enrichment (SeEn) is an effective strategy for enhancing scientific recommendation datasets.
• Improving datasets, rather than solely algorithms, can lead to substantial gains in recommendation performance.
• The findings suggest a promising direction for advancing recommender systems in specialized scientific fields.