Exploratory Search Using Interactive Visualization Techniques

This thesis studies exploratory search of large datasets using machine learning and human computer interaction techniques. It is often that the user wants to search for something but cannot formulate the exact query or the keywords which would help him to reach the most relevant search results. One might also want to address these issues about a particular topic by gathering information after each search. We worked on extending an existing interactive exploratory search system known as SciNet. SciNet allows the users to provide relevant feedback to the system using interactive user interface. The system allows the users to direct their search query using interactive intent modeling. The users obtain relevant results by giving personalized feedback to the system through a radar based layout. Our aim is to make SciNet work for a large news data set and add new features which allow the users to explore and investigate the news articles. We try to visualize the entire collection of news articles stored in the system using neighborhood embedding and display it as an interactive map to the user. The locations of the search results are displayed on the map using markers. The users can explore the articles by clicking on markers. They can also select areas of the map where search results are located, which would enable them to view a list of most relevant unigrams in an area and are able to select relevant unigrams to boost the query. This serves as an additional feedback mechanism. We performed user experiments with twenty users to compare the performances of the original SciNet and the new extended system. The user experiments clearly showed that the extended system performs better than the original one. We also took feedback from the participants of the experiments in a form of a questionnaire, which showed that the extended system improves the overall user experience. We can further improve the performance of the new system by adding more features like tagging different regions of the map with descriptive keywords and using distributed computing based algorithms which would allows us to incorporate more data from different domains