Zhengwei Qiu - Ph.D Transfer Talk - 3rd June 2011

Video Category: 
Transfer Talk
Zhengwei Qui Transfer Talk


TITLE : Detecting and Utilising Important moments for Personal Information Management


ABSTRACT: Increasingly in recent years, we are capturing life experience digitally using a myriad of mobile devices. Technology has progressed to the point where we are now in a position to capture all of our waking lives digitally in a process called lifelogging. In the past, such lifelogging was typically done by manually maintaining a diary. As humans we can easily distinguish important events from more routine or mundane events in our lives and we can record important events in diaries, if we wish.


In this work we introduce the concept of moment detection to detect the point at which an important event is taking place and we propose that it is possible to determine automatically the importance of these moments. We incorporate different types of sensors, both environmental and wearable sensors, environmental sensors to capture aspects of the environment around us, while wearable sensors can capture digitally aspects of life experience as they happen to us. We propose that it is possible to analyse these sensor streams and to develop semantic virtual sensors to annotate these streams. In addition, by examining these annotations, we propose that we can segment the lifelog into logical life moments (experiences). In addition, by employing machine learning and information retrieval techniques, we propose that it is also possible to detect which moments are important and which are not. In this report, we describe how we are using smartphones and SenseCams to gather users' information. We describe the six different types of contextual metadata that we gather and we discuss how to segment life activity into a set of moments, annotate them and detect their importance. We evaluate a number of approaches to important moment segmentation and annotation; Supervised Machine learning (such as SVMs), Unsupervised Machine learning (Clustering) Information retrieval-based and sequential pattern mining approaches.