Using Multiple Instance Learning techniques to rank maize ears according to their traits

Abstract

Abstract Multiple-Instance Learning (MIL) is a sub-field of machine learning. Its main goal is to do accurate predictions on new data based on a predictive model generated from previously group of labeled bags of data, known as training data, containing many instances. MIL has many real world important applications such as image retrieval or text categorization and medical diagnosis problems. It is often difficult for crop breeders to predict yield by combining different yield components to produce better plants with superior performance. Data analysis is one area that is striving to let farmers have an idea of their expected yield pre-harvest. Accurate early yield prediction will improve agricultural strategies plan, proper resources allocation and improve management of maize ear cultivation with consequent increase in productivity. Most experiments on maize ears traits only considered ear evaluation and maize improvement without yield prediction. One of the experiments that included yield prediction was PR. NDCG measure which was developed to rank maize evaluation for Sousa Valley Best Ear Competition. The focus of this work was to make an intelligent regression models recognition and analysis by running some MIL algorithms to predict and assign real value to maize yield from randomly group N vary parameter sizes of maize ear traits and soil parameters of maize population dataset. Furthermore, this dissertation also ranked the models per result and establish a relationship between variables.