N-semble: neural network based ensemble approach
- PDF / 2,130,013 Bytes
- 9 Pages / 595.276 x 790.866 pts Page_size
- 10 Downloads / 193 Views
ORIGINAL ARTICLE
N‑semble: neural network based ensemble approach Rishith Rayal1 · Divya Khanna2 · Jasminder Kaur Sandhu2 · Nishtha Hooda2 · Prashant Singh Rana2
Received: 3 October 2016 / Accepted: 17 August 2017 © Springer-Verlag GmbH Germany 2017
Abstract Output can be predicted from experimental or achieve data by using machine learning models like random forest, artificial neural network, decision tree and many more models. Each model has its own limitations and advantages. To improve model’s accuracy, outcome of multiple models can be combined for prediction. The way of combining the predictions of different models is the key to increase the overall accuracy. In this work, a new approach is discussed to create an ensemble model on a regression dataset which overcomes the limitation of classical ensemble approach. Artificial neural network is trained in a special way to ensemble the predictions of multiple models. The comparison between N-semble and classical model is performed on various evaluation measures and it is concluded that N-semble outperforms. Keywords Supervised machine learning · N-semble · Regression · Neural network ensemble · Correlation · Validation
* Rishith Rayal [email protected] Prashant Singh Rana [email protected] 1
Department of Information and Communication Technology, ABV-Indian Institute of Information Technology, Gwalior, MP 474010, India
Computer Science and Engineering Department, Thapar University, Patiala, Punjab 147004, India
2
1 Introduction Machine learning is a branch of artificial intelligence. Estimations, predictions, decision making through machine learning models are playing an important role in human life. From business analysis to medical field, from research and development to food production, they play a critical role in the daily basis. For instance, machine learning models are used in food processing to filter quality. There are many algorithms like support vector machine, random forest, bayesian, decision tree, etc. In a classical ensemble model, weighted average of different models’ predictions are combined to get final prediction [9]. Weights are manually assigned and weighted average approach increases the accuracy of hybrid model. There are limitation of classical ensemble approach i.e., manual assignment depends upon the user and it may prone to error in assigning weights which can decrease the accuracy of ensemble model, due to the manual assignment in the weighted average process, it may limit the accuracy gain through ensemble model. As stated before, every model has its own limitations and ensemble approach is used to overcome these limitations. Rather than assigning weights manually, data predictions need to be analyzed and processed to assign suitable weights or proper relationship between them. So the predictions, obtained through models need to be considered as raw data and the ensemble process should be done as similar to model training. In recent years, methods like iDNA-KACC-El [2] and iDHS-EL [3] are used for grid search to f
Data Loading...
