Utilization of Efficient Features, Vectors and Machine Learning for Ranking Techniques

Document ranking systems and recommender systems are two of the most used applications on the internet. Document ranking systems search for documents in response to a query given by the user. On the other hand, recommender systems suggest items to the users on the basis of their previously expressed preferences. Both document ranking systems and recommender systems make use of ranking techniques, since they typically present their results in the form of a ranked list. The order of the results is important because the users expect the most useful results at the top of these ranked lists. Improvements in algorithms used by document ranking systems and recommender systems, including the utilization of advanced machine learning techniques, lead to the generation of improved rankings. Moreover, advanced document ranking systems often use features collected from the documents to generate rankings. Similarly, vectors generated for the users as well as items are utilized by the recommender systems. Therefore, generation of features and vectors of good quality is instrumental for ranking techniques. This dissertation makes the following contributions to explore the improvements in ranking techniques using efficient features, vectors and machine learning: a) Creation of a feature extraction algorithm for learning to rank tasks in document ranking, b) Creation of pairwise preference vectors of ratings on items by using neural embeddings that can be utilized in machine learning tasks including recommender systems, c) Utilization of deep neural networks and transfer learning for serendipitous recommendations, d) Recommendations using ranking probabilities and non-negative matrix factorization and e) Application of neural embeddings to search for cities and tours, taking user’s travel interests into account. Keywords: Ranking, Information Retrieval, Recommender Systems, Deep Learning, Neural Embedding, Serendipitous Recommendations