Scalabeling: Linear Slider Supported Labeling for the Classification of Streaming Data

Abstract

Supervised Machine Learning algorithms need labeled datasets in order to train models for the classification of new unlabeled data. For most purposes, especially in the field of computer vision, suitable tools to support the labeling process have already been developed. However, for some use cases convenient methods are still not available. In order to bridge this gap, we propose a new lightweight labeling tool for highly sensitive streaming data like human body data from a sensor wristband. The proposed system Scalabeling uses a linear slider as the main interface for configurable user created labels. The system records user inputs from this slider, converts them to labels and combines them with data from an asynchronously received data stream. Finally the resulting labeled dataset is saved and can be used in various machine learning systems. We implemented this tool and performed a preliminary evaluation by applying it to a real-world labeling problem from an ongoing research project. Our findings show a good user acceptance as well as a significant reduction of labour costs for labeling and preprocessing.