Feature Space Selection for Cyber-Physical System based on Improved Feature Space Partitioning Tree

Abstract

Cyber-Physical System (CPS) is a new generation of the digital systems based on dynamic interdependencies and convergence between the physical world and cyberspace. There are various data mining techniques are used to handle the data in CPS. A novel deep learning inference runtime called DeepRT was designed to support predictable inference performance both spatially and temporally. It comprised of two main processes like feature pre-processing system and deep learning system. In the feature pre-processing system, data collected from the mobile devices and sensors were unwrapped. Then, the feature elements were used as the input of the deep learning engine which classifies the input data and infers results. Nevertheless, the feature scope which has insufficient training data usually has a much higher error rate. Hence, the performance of DeepRT is degraded due to feature space lacking. So in this paper, Improved Feature Space Partitioning Tree with DeepRT (IFSPT-DeepRT) is proposed to decrease the feature scope lacks in the training data. Initially, IFSPT-DeepRT splits the feature space into number of partitions using classification and regression tree (CART) which constructs a tree for feature partitioning. Also, a new splitting criterion called weighted gain in Gini index is introduced for selecting the splitting points in the tree. The feature space partition is evaluated using Shapley Feature Importance (SFI) measure and selects the feature space partition which has high SFI value. It is given as input to the DeepRT to classify the input data and infer results.