I recently read "Enabling tabular deep learning when d ≫ n with an auxiliary knowledge graph" (https://arxiv.org/pdf/2306.04766.pdf) for one of my graduate classes. Essentially, when there are significantly more data points than features (n >> d), machine learning usually works fine (assuming data quality, an underlying relationship, etc.). But, for sparse datasets where there are fewer data points than features (d >> n), most machine learning methods fail. There's just not enough data to learn all the relationships. This paper builds a knowledge graph based on relationships and other pre-existing knowledge of data features to improve model performance in this case. It's really interesting - I hadn't realized there were ways to get better performance in this case.
