AbstractMemory-based collaborative filtering use past activities of a group of similar users to recommend future preferences for a target user in the group. Recommender systems based
on this type of technique are prone to errors when there are too few historic interactions (e.g. rating, likes, transaction-history, visit-frequencies) between users of the system. The sparsity
in users’ historic data render the memory-based algorithm less effective at finding similar users for the personalisation process.
In contrast, model-based collaborative filtering techniques such as matrix factorisation (MF) use predictive models. In single-domain recommender systems, one problem prevalent to all techniques is the cold start problem. A cold start situation happens when there is no historical information about new users or items that have just been introduced to the system.
Several recommender techniques have used semantic knowledge extracted from additional user and item information to build profiles that reflect otherwise implicit user preferences. This semantic representation of the user is then used to find other similar users and address sparsity and cold-start problems in single domain recommender systems. Recent attempts to resolve cold-start and sparsity problems are considering cross domain collaborative filtering techniques. Cross-domain recommender systems exploit additional user and item information from domains that are unique but related to the target recommendation domain. Extending predictive models to include parameters that model the semantic similarity in user and item information across the domains constitute a genuine approach in cross-domain recommendation.
The contents of this thesis centre around the use of semantically enhanced cross-domain recommender systems as a solution to cold start and sparsity problems. The contributions to cross-domain recommender systems are in three folds. First, we investigate and analyse the performance of a cross-domain recommender model as we vary the size of intersecting user/item information from the target and auxiliary domains. Secondly, we proposed a predictive model that adds semantically related tags as additional parameters to a matrix factorisation model. Thirdly, we present a model that incorporates category similarity into a POI ranking function as contextual information for improving the performance of multi-category POI recommenders.
In our investigation, we empirically evaluate the proposed models on datasets that are sourced from different domains, specifically movies, books and several POI categories. On the one hand, the results show that semantically enriching tags in cross-domain recommender models are possible without negatively impacting recommendation accuracy. On the other hand, cross-domain recommender models that are semantically enhanced with additional latent parameters are effective in cold start scenarios and reduce the effect of sparsity on recommendation accuracy.
|Date of Award||Nov 2018|
|Supervisor||Gulden Uchyigit (Supervisor)|