The activity pattern analysis methodology is presented to detect different types of search tasks during information searches. The observable data from search task sessions include different page-related activities, such as seeing or revisiting a content page or search engine query result page, as well as low-level cognitive processing features from eye-movement tracking such as eye fixation periods. A search task session is defined as a sequence of interactions of these activities and is represented as a Markov chain.

The activity data captured from two search domains are analyzed using the clustering distribution analysis, run-length encoding (RLE) of repeated activities and the compression sizes, and Markov transition graph analyses, such as graph density, number of edges, or maximum clique size of the graph. The analysis shows that the graph edge numbers and RLE are good at detecting search tasks in both domains.

Unlike many previous search behavior studies, it is significant to view a task session as the activities and interrelated activity patterns, not just as independent actions. The results show that this activity pattern representation and analysis can detect different search tasks, but it is not clear whether this approach performs better than other approaches. It is also difficult to claim that all of the information-gathering activity patterns a user performs for search tasks are considered. In order for the approach to be valid for a personalized search, the activity pattern analysis for task detection or prediction should be scalable without manual components.