Automated Classification of Student Activities Using Machine Learning Techniques
Since time is the most limited ressource students have, an efficient use of it is neccessary. To get the most out of it, time management is a crucial process which itself unfortunately counsumes too much time and therefore is often neglected by students. However, research shows that good time management results in less stress and an increased academic performance of students. The process of time management can be split into three major parts, which should be conducted periodically, e.g. each day: (i) defining goals and subgoals, (ii) prioterizing these and (iii) monitoring which goals were achieved and how much time was spent achieving them. An automatization of single parts of this process implies lower time consumption for the whole process. To assist students in time management, a system capable of automatically detecting students’ daily activities, like listening to a Lecture, is needed. Such a system would simplify part (iii) of the time management process.
Activity recognition systems can be split into two groups: (a) systems to detect simple activities like sitting, walking or running and (b) systems to detect complex activities like listening to a Lecture. Student activities refer to complex activities. Hence, they can only be detected with systems of group (b). Current activity recognition systems commonly use sensor data at personal scale to detect simple activities. These systems focus on users’ physical activities and do not consider the actual context. Nevertheless, they are able to detect simple activities like sitting, walking or running with an accuracy above 90 %. Systems bother with detecting complex activities – like student activities – instead, do not gain such a high value of accuracy. They often try to gather more than smartphone-based sensor data by deploying additional sensors resulting in increased costs of the overall system. Albeit the ubiquitous sensors, the social context of the user is not considered. For example, it cannot be used to determine the number of people around the user, nor the number of friends nearby. However, this information is crucial to derive a deep insight into the actual situation and to correctly classify a conducted activity.
The goal of this thesis is to build a system capable to automatically detect the following seven student activities: Lecture, Learning, Teamwork, Transition, Leisure, Sleeping and Job.
The goal of this thesis is to build a system capable to automatically detect the following seven student activities: Lecture, Learning, Teamwork, Transition, Leisure, Sleeping and Job. Therefor different kinds of supervised machine learning classifiers are utilized. To train the classifiers, various features of ubiquitous information sources are extracted. Beside the built-in smartphone sensors, the social graphs of users are exploited by crawling friends lists of users’ facebook accounts. The extracted features contain information about different measurement values of smartphone sensors and users’ social contacts surrounding a user during a specific activity. Due to the fact that supervised classifiers are used, a user study was conducted to collect a set of labeled instances for training. 163 students participated in a study over a time period of 28 days. Using the collected dataset the different models of the calssifiers are learned and evaluated.
The best model for detecting student acitivites is a random forest with 100 trees. The model can detect various activities with an overall accuracy about 70 %, a satisfying value for a multi-class classification problem. The evaluation shows that different kind of activities can be detected with different values of accuracy. For example, participating in a Lecture can be detected with a F1-Score of 61,5 %. Learning for a specific course of university can be detected with a F1-Score of 56,9 %. Results show, that detecting various activities of students’ daily routine can be done automatically using machine learning classifiers. Thus, a student assistent capturing executed activities and time of performance can be built. The assistent can simplify time management process. This leads to the conclusion that built models of machine learning classifiers can assist students in time management tasks, implicating less stress and an increased academic performance.
- Christian Meurisch (christian.meurisch(a-t)tk.informatik.tu-darmstadt.de)
- Benedikt Schmidt
- Immanuel Schweizer (schweizer(a-t)tk.informatik.tu-darmstadt.de)
Forschungsgebiete: Ubiquitous Knowledge Processing