A Hybrid Approach for Automatic Model Recommendation

摘要: One of the challenges automating machine learning applications is automatic selection an algorithmic model for a given problem. We present AutoDi, novel and resource-efficient approach selection. Our combines two sources information: metafeatures extracted from data itself word-embedding features large corpus academic publications. This hybrid enables AutoDi to select top-performing algorithms both widely rarely used datasets by utilizing its types feature sets. demonstrate effectiveness our proposed on dataset 119 179 classification grouped into 17 families. show that can reach average 98.8% optimal accuracy algorithm in 49.5% all cases.