Cost-sensitive regression learning on small dataset through intra-cluster product favoured feature selection

Massive regression and forecasting tasks are generally cost-sensitive learning problems with asymmetric costs between over-prediction under-prediction. However, existing classic methods, such as clustering feature selection, subject to difficulties in dealing small datasets. As one of the key challenges, it is difficult statistically validate importance features using traditional algorithms (e.g. Boruta algorithm) owing insufficient available data. By leveraging information intra-cluster (item group similar attributes), we propose an product favoured (ICPF) selection algorithm select based on filtering method (specifically our study). The experimental results show that ICPF significantly reduces number dimensions selected set improves performance learning. misprediction cost decreased by 33.5% (linear-linear function) 32.4% (quadratic-quadratic after adopting algorithm. In addition, advantage robust other models, random forest XGboost.