Towards bookmark folder recommendation using simulated user-guided classification, and clustering techniques

Abstract

Webpage bookmarking involves users storing their favourite webpages into an organised structure called a bookmark hierarchy or repository. We developed a user-guided bookmark folder recommendation system to aid users in maintaining their repository’s organisation. Our system builds a flat-cluster semantic model of the hierarchy using Latent Semantic Analysis. This evolves alongside users through their indirect guidance whenever a bookmark is created. Clusters model folders containing pages using a mean centroid representation. If the repository is small or empty, LSA is performed on the user’s history instead. Using the model’s centroids, it recommends up to three folders most similar to the page. If no folder’s similarity is above a threshold, the system suggests creating a new one. We also developed a folder restructuring user-tool using DBSCAN for folders containing random webpages or multiple pages from different topics, thus aiding recommendation. The recommendation system was evaluated using the SemEval 2013 Task 11 corpus. We simulated the creation and evolution of multiple hierarchies, performing recommendation before adding each bookmark. We modified a K-NN algorithm to give up to three recommendations, for system comparisons. We also used a new metric, User Cost Index, that measures recommendation quality with respect to the ‘cost’ for a user to fix it. The restructuring tool was evaluated using the same corpus. We reached our aims through a system requiring no external data that creates and maintains an evolving, user-guided model for recommendation. Using LSA we achieved up to 67.6% overall accuracy, which rises up to 72.4% when the repository’s size increases. We also discovered that using a static LSA parameter on an evolving hierarchy deteriorates accuracy by up to 4.4%. Meanwhile, the folder restructuring evaluation showed that its positive results are not always consistent for word sense based clustering. These tools were incorporated into a browser addon. Whilst it has little extra functionality, we showed that developing an intelligent bookmark management addon is possible.