Name
Abstract | ||
|---|---|---|
Feature selection is a crucial pre-processing step in machine learning and data mining. A popular approach is based on information theoretic measures. Most of the existing methods used low-dimensional mutual information terms that are ineffective in detecting high-order feature interactions. To fill this gap, we employ higher-order interactions for feature selection. We first relax the assumptions of MI-based methods to allow for higher-order interactions. A direct calculation of the interaction terms is computationally expensive. We use four-dimensional joint mutual information, a computationally efficient measure, to estimate the interaction terms. We also use the ‘maximum of the minimum’ nonlinear approach to avoid the overestimation of feature significance. Finally, we arrive at an effective feature selection method that makes use of higher-order interactions. To evaluate the performance of the proposed method, we compare it with seven representative feature selection methods, including RelaxMRMR, JMIM, IWFS, CIFE, MIFS, MIM, and reliefF. Experimental results on eighteen benchmark data sets demonstrate that higher-order interactions are effective in improving MI-based feature selection. |
Year | Venue | Field |
|---|---|---|
2018 | PAKDD | Data mining,Data set,Nonlinear system,Feature selection,Computer science,Artificial intelligence,Mutual information,Machine learning |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
11 | 4 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Xiaochuan Tang | 1 | 4 | 0.70 |
| Yuan-Shun Dai | 2 | 1357 | 98.96 |
| Yanping Xiang | 3 | 157 | 21.73 |
| Liang Luo | 4 | 2 | 4.10 |