Date of Award
Campus Access Dissertation
Doctor of Philosophy (PhD)
Recommendation systems are information filtering systems that deal with information overload by helping users navigate huge volumes of content on various platforms, resulting in the potential discovery of items or services they deem interesting or important. Predominantly, these systems search for candidate items through predictions based on a user or some other entity’s historical consumption sequence. The sequential nature of the recommendation task lends itself to reinforcement learning techniques for resolution.
The first part of this dissertation tackles the web service composition problem in the face of redundant services, by presenting an approach to web API recommendation for mashup development using reinforcement learning (RL). Specifically, we search for and recommend new or replacement web APIs to existing mashups in a dynamic environment, where the quality properties of the component web APIs continue to change. Our approach models mashup performance reward using external quality factors of web APIs.
The importance of tagging content for discovery and recommendation cannot be overemphasized. The second part of this dissertation involves recommending suitable tags to reap its benefits. First, we map entity descriptions to keywords, using a planned exploration technique that mimics online user behavior in a model-based reinforcement learning setup to generate description-keywords pairs. Second, using the description-keywords data, we fine-tune a BERT language model to generalize on the named entity recognition task of keywords extraction. Lastly, we perform a nearest neighbor multi-label classification task to obtain our suitable tags for recommendation.
To alleviate the cold start problem, recent advances have seen a shift towards session-based recommender systems which provide recommendations solely based on a user's interactions in an ongoing session. Recommendation coverage is a key limitation for existing approaches. In the final part of this dissertation, we propose a two-stage approach to carry out session-based recommendations. First, we train a skip-gram model with negative sampling to generate candidate items that co-occur with a given query set. Next, we use a multi-armed bandit approach to boost recommendation coverage by balancing the exploration-exploitation trade-off. Experiments with real world datasets for the above problems show that our methods outperform state-of-the art methods in the recommendation task.
Anarfi, Richard, "Reinforcement Learning-Based Semantic Search for Recommendation" (2022). Graduate Doctoral Dissertations. 725.