NoCASC: A Novel Optimized Cost-Complexity Pruning for AdaBoost Model Compression on Edge Computing-Based Smart Lighting

K-nearest neighbor (KNN) is an effective method for smart lighting control that promotes user comfort through the classification-integrated moving average (CIMA) method. However, KNN is a lazy learner, where if the training data size is large, it will affect the KNN model size, which contradicts the principles of edge computing-based smart lighting (EdgeSL). Our research aim is to use cost-complexity pruning as a compression model for adaptive boosting (AdaBoost) in the EdgeSL environment. In the first step, we carried out an analysis of the smart lighting data that we had used in several previous studies. We designed AdaBoost and implemented a Pareto front grid search for AdaBoost hyperparameter optimization. Then, we compared the performance of AdaBoost and KNN in smart lighting control prediction. In the last step, we leverage the cost-complexity pruning for model compression on AdaBoost. We use a novel method called normalized-complemented accuracysize curve (NoCASC) to balance between reducing the model size and maintaining accuracy loss. The test results show that grid search with Pareto front found that $max\_{d}epth =9$ and $n\_{e}stimators =30$ were the optimum hyperparameters for AdaBoost on the CIMA dataset. Our optimal AdaBoost model has a better recall than the legacy method, with a value of 0.9999. The size of the AdaBoost model after cost-complexity pruning is 281 kB. This size is $2.70 \times$ smaller than the original AdaBoost, then $6.93 \times$ smaller than KNN. AdaBoost with cost-complexity pruning is the most optimal model for CIMA in EdgeSL.