Sensitivity Analysis of a BERT-based scholarly recommendation system

Résumé

With the exponential growth of publicly available datasets, a scholarly recommendation system of datasets would be an essential tool in the field of information filtering. Recommending datasets to users can be formulated as a classification problem where deep learning models can be carefully trained. In such a case, when preparing training data for the learning models, one needs to consider different ratios of false and true pairs. Therefore, a sensitivity analysis is necessary. In this work, we conduct a sensitivity analysis using different class ratios on a deep learning model (BERT) for recommending datasets. We found out that our BERT-based recommender model is relatively robust using recommender metrics such as Mean Reciprocal Rank (MRR)@k, Recall@k, etc., except for the extreme class imbalance case (1:5000). Therefore, we conclude that a moderate ratio of the random negative sampling scheme, (in our case 1:10) is reasonable, sufficient and time-efficient in the recommendation system training