Recommender systems already are a consistent part in the life of most people regularly using the internet. They get recommendations when they shop at Amazon.com, when they watch video clips on Youtube.com, or when they listen to music on Spotify.com etc. There are still many challenges in recommender systems research, though. One challenge that is present in almost all application domains is data sparsity, i.e. missing information about items or users. In very sparse application domains, data sparsity can completely hinder the creation of recommendations. In more diverse application domains, where few items are heavily used while most items are rarely used, the popular items tend to be recommended over-proportionally often. In contrast, the niche items tend to be excluded from the recommendation lists. This thesis therefore aims to contribute to the state-of-the-art in handling data sparsity in recommender systems. Therefore, it investigates techniques to find similarities between the items solely by analysing their usage. This approach is based on the assumption stemming from context-aware computing that the users' contexts and knowledge influence their activities and, thus, are inherent in the items' usage. Hence, no additional information like content or social metadata are required to find relations between the items. For this purpose, techniques that are successfully applied in corpus linguistics to detect relations between words by analysing their usage in language are adapted to items and their usage. This way, pair-wise item relations as well as item clusters are created based on the items' usage. These usage-based item relations are then utilised in standalone and hybrid recommender systems with the goal to create suitable recommendations for as many items as possible including the rarely used ones. The discussed techniques are evaluated on four data sets, two of them were collected in web portals that support learners in finding suitable learning materials while the other two data sets were collected in web portals that recommend movies to users. The evaluation results show that by exploiting the items' usage, usage-based relations between the items can be discovered that indeed give a hint at their similarity. Furthermore, the usage-based recommender systems are able to create more recommendations in application domains holding predominantly rarely used items than the presented state-of-the-art recommendation approaches. In application domains holding heavily used items that are recommended over-proportionally often in addition to many rarely used items, the usage-based recommender systems are able to recommend more niche items than the presented state-of-the-art recommendation approaches without lowering the accuracy of the recommendations. Thus, the usage-based approaches are better suited to provide users with accurate recommendations for idiosyncratic items than the recommendation approaches presented in literature so far that do not require additional metadata either.
展开▼
