ScholarFinder: Knowledge Embedding Based Recommendations using a Deep Generative Model

摘要

Bold scientific research tasks today need multi-disciplinary knowledge and interdisciplinary collaborations that require finding scholars from a particular domain with relevant knowledge. Given the variety of scholars and diversity of research tasks, finding the appropriate scholar is a critically important and challenging problem for scientific communities. Prior approaches to identify scholars use supervised learning with fixed or high-level research interest tags. Such approaches make it hard to recognize scholars with specific interests, or track their changes in research interests. Hence, there is a need to investigate suitable methods to quantify scholars' expertise knowledge for matching research tasks. In this paper, we propose a novel model viz., "ScholarFinder" that uses contextual information (abstracts or publications) for embedding a scholar's knowledge in an unsupervised learning manner. Subsequently, with pre-trained knowledge embeddings, we can perform machine learning tasks such as classification, visualization or checking whether a scholar is suitable for performing particular research tasks or not. Based on our pre-trained techniques, we also provide a novel negative sampling method to overcome the issues of missing negative samples. Using a "follow-the-money" strategy, we apply our model to a large collection of NSF (National Science Foundation) grant awards dataset collected over the last twenty years that contains more than 20,000 award records (with project abstracts and names), corresponding to 15,074 scholars who received grants. We evaluate different deep learning models to see how to use pre-trained knowledge embedding for achieving optimal performance, and how our negative sampling method improves model performance. We also compare our model with state-of-the-art baseline models (e.g., XGBoost, DNN), and our results show that the ScholarFinder model outperforms those models in terms of precision, recall, F1-score, and accuracy.