Predicting personal information behaviors with lifelog data

The research for monitoring and recognizing personal behaviors from various digital sensors has recently been doing in a variety of fields. We address this for “lifelog” - all of the digital information about personal daily life. The research typically focuses on collecting personal lifelog, managing huge amount of lifelog data, and recognizing activities and behavior patterns from them. The methods of extracting key features and characterizing patterns would be crucial for finding meaningful information from huge and complex lifelog data. The research is a significant challenge because individual's lifelog data would be useful to provide personal life services such as healthcare. In this paper, we propose the process for predicting personal future behavior by tracing back to the past experiences. The behavior prediction process is composed of five stages. Firstly, physical activities through various sensors are collected and then, major physical activities are extracted through feature selection. Secondly, behavioral context information such as location, time and object is annotated to each activity for recognizing the behavior states more exactly. Then all sequences of physical activities with contextual information are divided into each daily set. Thirdly, behavior patterns from them are extracted by analyzing key features. After that, all daily sequences are transferred as the set of semantic activities for presenting major behavior states. Fourthly, from the set of semantic activities, based on the behavior probability to be used for the behavior prediction in next step, a sequence tree is generated. Finally, the highest predicted activities can be shown in a user interface from the query based on `Time' or `Event'. In a user interface, the functions for retrieving past and current behaviors and searching the predicted behaviors will be provided by choosing specific point in time or the specific event. Currently we are building a system for processing the proposed behavior prediction.