Personalized and adaptive recommendation
Date of Issue2016
School of Computer Engineering
Recommender systems have become an important research area. Plenty of algorithms have been proposed to solve the recommendation problems in various scenarios. This thesis presents the personalized and adaptive recommendation algorithms developed for three emerging application scenarios: personalized location recommendation, top-N recommendation with users' implicit feedback, and drug-target interaction recommendation. Personalized location recommendation is an important application in the rapid growing location-based social networks. Its objective is to recommend a user a list of locations that she may prefer but has not visited before. This thesis proposes a novel personalized location recommendation method, namely Instance-Region Neighborhood Matrix Factorization (IRenMF). To improve recommendation accuracy, IRenMF exploits two levels of geographical neighborhood characteristics of users' check-in data generated in location-based social networks: (a) instance-level characteristics, i.e., nearest neighboring locations which tend to share more similar user preferences; and (b) region-level characteristics, i.e., locations in the same geographical region which may share similar user preferences. More generally, many newly emerging recommendation tasks are usually formulated as top-N recommendation problems based on users' implicit feedback instead of explicit feedback. Here, explicit feedback refers to users' ratings to items, while implicit feedback is derived from users' interactions with items, e.g., number of times a user plays a song. This thesis proposes a boosting recommendation algorithm, named Adaptive Boosting Personalized Ranking (AdaBPR), for top-N recommendation using users' implicit feedback. In the AdaBPR framework, multiple homogeneous component recommenders are linearly combined to create an ensemble model, for better recommendation accuracy. The component recommenders are constructed based on a fixed matrix factorization algorithm by using a re-weighting strategy, which assigns a dynamic weight distribution on the observed user-item interactions. In addition, this thesis also presents a study that adopts personalized recommendation technique for the drug-target interaction (DTI) recommendation task in the drug discovery process of pharmaceutical sciences. A novel recommendation approach, called Neighborhood Regularized Logistic Matrix Factorization (NRLMF), has been proposed to suggest a list of potential DTIs. More specifically, NRLMF focuses on modelling the interaction probability of a drug-target pair using logistic matrix factorization. Because the positive observations (i.e., the observed interacting drug-target pairs) are already experimentally verified, they are more trustworthy than the negative observations (i.e., the unknown pairs). NRLMF assigns higher importance levels to positive observations. To further improve the accuracy, NRLMF exploits the local structure of DTI data by neighborhood regularization. Extensive experiments on real datasets have been performed to demonstrate the effectiveness of the proposed methods, compared with state-of-the-art methods.