Web Search Behavior Analysis And Prediction Based On Eye-tracking

At present, the Web has become a primary information source just like newspapers, radio, and TV. The main way to obtain online information is Web search, and its tool is the search engine. With the popularization of the Web, the Web search or search engine industry is flourishing and has gained a lot of attention from the research community. Investigating Web search behavior, that is, knowing how users interact with search engines, is helpful not only for improving search engines and increasing the efficiency of information retrieval, but also for designing a more user-friendly way of interaction and enhancing search experiences.Traditional studies of Web search behavior often use questionnaires, interviews, and log analyses. The questionnaire and interview can describe the interaction between users and search engines to some extent, but the results are indirect. The log analysis can obtain interaction data including queries and clicks so that the results are direct. However, how users view search results before an interaction is unknown. The emerged eye-tracking device provides an excellent tool for solving this problem. Eye-tracking devices can record users’ eye movement data during visual activities like reading, browsing, and searching. The data are real-time and procedural, such as where users look and for how long it lasts. By analyzing these data, users’ visual behavior during viewing the search results can be revealed directly and objectively.This dissertation uses an eye-tracking device to analyze user’s Web search(including webpage search and image search) behavior, and proposes a gaze-based Web search behavior prediction approach through establishing user behavior models and utilizing machine learning algorithms.This disseratation finds that users in webpage search have blindness towards user recommendations. The user recommendations are shown on the search engine results page as additional information, and an eye-tracking experiment is conducted to observe the changes of user behavior during webpage search tasks. Results show that participants’ behavior under no recommendation, passive recommendation, and active recommendation don’t have significant differences, becasuse they often ignore the user recommendations during the search process, generating the phenomenon of inattentional blindness.This disseratation finds that users have a specific pattern during image search, which is named as a ‘podium’ pattern. Through an eye-tracking experiment, participants’ behavior during image search tasks are observed, including the general characteristics and the influences of the type of search tasks and the presentation order of search results. Results reveal that when viewing and selecting search results in the same row, participants have a priority order of ‘center-left-right’, namely the podium pattern; the type of search tasks significantly influences participants’ behavior in which they are more busy and less satisfied in search tasks with general information needs; the presentation order of search results doesn’t influence participants’ behavior significantly as they have a position bias which is consist with the podium pattern.This disseratation proposes a gaze-based approach for predicting Web search behavior. The raw eye movement data are transformed into two data forms that describing the pattern of visual behavior: histogram and sequence. Through establishing user behavior models and utilizing machine learning algorithms, the behaviors of Web search are predicted. The experiments in webpage search and image search not only demonstrate the effectiveness of the proposed approach, but also suggest that the approach can deal with the problem that traditional methods can’t: utilizing good abandonment as feedbacks.