Browsing by Author "Peng, Yu-Hsiang"

Now showing 1 - 3 of 3

Associative Classification for Human Activity Inference on Smart Phones
(2014-01-01)
Peng, Yu-Hsiang
;
Njoo, Gunarto Sindoro
;
Li, Shou-Chun
;
Peng, Wen-Chih
;
資訊工程學系
With the population of smart phones, the general trend of human activity inference is prospering under a powerful computation capabilities on modern phones. Such an assistant make users life more convenient and help them prevent from unnecessary interferences. In conventional research, the activity inference problem is considered a classification instance, so in this paper we propose an association-based classifier framework (ACF) that aims at exploring the correlation among collected sensor data. Each data consists of multiple sensor readings with a label, e.g., dining, shopping, working, driving, sporting, and entertaining. Note that ACF caters to the discrete data; as a consequence, the continuous sensor readings are needed to be transformed to some discrete groups. Therefore, we propose an Interval Length-Gini Discretization (LGD) method which considers the groups and misclassified cases to obtain the best hypothesis for a given set of data. After an appropriate discretization, we propose one-cut and memory-iteration-based approach to select a set of useful sensor-value pairs for reducing the model size by removing redundant features and guaranteeing an acceptable accuracy. In the experiments our framework has a good performance on real data set collected from 50 participants in eight months, and a smaller size than the existing classifications.
Machine Learning Hyperparameter Fine Tuning Service on Dynamic Cloud Resource Allocation System - taking Heart Sounds as an Example
(2018-01-01)
Peng, Yu-Hsiang
;
Chuang, Chia-Chuan
;
Wu, Zhou-Jin
;
Chou, Chia-Wei
;
Chen, Hui-Shan
The hyperparameters tuning of machine learning has always been a difficult and time-consuming task in deep learning area. In many practical applications, the hyperparameter tuning directly affects the accuracy. Therefore, the tuning optimization of hyperparameters is an important topic. At present, hyperparameters can only be set manually based on experience, and use Violent Enumeration, Random Search or through Grid Search to try and error, lack of effective automatic search parameters. In this study, we proposed a machine learning hyperparameter fine tuning service on dynamic cloud resource allocation system, which leverages several mainstream hyperparameter tuning methods such as flyperopt and Optunity. In the meanwhile, various tuning methods are measured and compared by example application in this work. Finally, we dedicated actual case - Heart Sounds, and then tested it. In order to verify that the system service can not only automate the task of tuning, but also break through the limitation of the number of adjustable parameters. Furthermore the proposed hyperparameter fine tune system makes optimization process more efficient.
On Selecting Feature-Value Pairs on Smart Phones for Activity Inferences
(2014)
Njoo, Gunarto Sindoro
;
Peng, Yu-Hsiang
;
Hsu, Kuo-Wei
;
Peng, Wen-Chih
;
交大名義發表
Prior works have been elaborated on activity inferences from context information sensed on smart phones. Most of sensing computations are performed on CPU of smart phones. Thus, Sensor Hub is designed to avoid CPU involvement. However, Sensor Hub has several limitations, such as limited memory space and computation power. Since activity inference is a classification problem, prior works have already proposed some classifiers on smart phones. However, if one would like to build a classifier model in Sensor Hub, one challenging issue is to reduce the model size. One approach to reduce the model size of classifiers is feature selection. Feature selection reduces the model size by removing features in the feature set. Nonetheless, because sensor space is limited, removing features could reduce accuracy of classifier algorithm significantly. Therefore we explore feature-value selection concept, which considers the value rather than the feature to reduce the model size while preserving accuracy of classifiers. In this paper, we propose three feature-value selection methods, which consider confusion and redundancy among the feature-value. Due to the nature of feature-values, discretization of sensor data is important. We design a discretization method, LGD (Length Gini Discretization) and compare it with another method, MDLP (Minimum Description Length Discretization), to discretize sensor data using confusion metric to choose the cut point. Extensive experiments are conducted to evaluate our proposed feature-value selection methods. Feature-value removal allows us to reduce up to 80% of the model size and maintain average accuracy performance to 86%.