LEE- Bilgisayar Mühendisliği Lisansüstü Programı

Bu topluluk için Kalıcı Uri

http://hdl.handle.net/11527/19206

Gözat

Konu "Algoritmalar" ile LEE- Bilgisayar Mühendisliği Lisansüstü Programı'a göz atma
  • Öge
    Hybrid reciprocal recommendation with advanced feature representations
    (Graduate Institute, 2021) Yıldırım, Ezgi ; Öğüdücü, Şule ; 674767 ; Department of Computer Engineering
    Over the last few decades, with the rise of online web services such as Facebook, Amazon, and Netflix, Recommender Systems (RecSys) have taken an indisputable place in our daily lives. The application domain of RecSys has an extensive range from e-commerce to online advertisement that aims to suggest to users the right contents matching their preferences, and it is not limited to one-way interacting platforms. In some challenging application domains, RecSys are developed to serve multiple users at each decision, to mutually satisfy the accompanying parties. Where a matching problem occurs and the satisfaction of both parties is the key to success, those recommender systems are called Reciprocal Recommenders (Rec2) in literature. Differing from traditional one-directional recommendation problems, the reciprocal recommendation has more adversity to overcome, which outlines its characteristics. In this study, based on gradual research, we first seek the key points of a strong recommender system, and then, by the learned lessons from this part, focus on the reciprocal recommendation. For this purpose, we first seek answers to these questions in a general recommender system: • How can auxiliary data affect recommendation quality? • How can we easily integrate different data sources and different approaches to empower a recommender system? Then, in the second part, we shift our research focus towards reciprocal recommendation and try to answer the following research questions: • How can we effectively solve reciprocal recommendation problems without detriment to system performance? • How can we avoid vagueness of recommendations and explain conceptual associations of requested and offered characteristics? In recent years, deep learning has gained indisputable success in computer vision, speech recognition, and natural language processing. After its rising success in these challenging areas, it has been studied on recommender systems as well, but mostly to include content features into traditional methods. In the initial part of this thesis, we introduce a generalized neural network-based recommender framework that offers an easy-to-use platform to combine different data sources, approaches, and methods into a single recommender system. This framework, Neural Hybrid Recommender (NHR), also allows us to exploit the same data sources to find out more elaborate information by different learning functions. In our experiments, we have worked on item prediction problems, however, with a single change on the loss function, the framework can be used for rating prediction problems as well. To evaluate the effect of such a framework, we have tested our approach on benchmark and not yet experimented datasets; movie reviews and job applications of job-seekers from an online recruitment platform. The results in these real-world datasets show the superior performance of our approach in comparison with the state-of-the-art deep learning methods in Click-Through-Rate (CTR) prediction. With the use of auxiliary data in different forms, NHR models perform better than collaborative filtering methods that depend on interaction data only. On the movie recommendation task, based on the average of a group of experiments, NHR models achieve 2.03% relative improvements on HR@10 score and 2.51% on NDCG@10 over the most successful baseline used in the evaluation. With the same setup, the improvements on the job recommendation task become even higher; 2.60% and 2.91% on HR@10 and NDCG@10, relatively. Having more promising results on job recommendation with auxiliary data is since this task is far more complex than the movie recommendation task due to the multi-variate socio-economic dependencies in job applications. Our further experiment that investigates the effect of predictive factors, which define the predictive capability in neural networks, also verifies that. Increasing the model complexity without changing the other parameters did not deteriorate the success of models in job recommendation because complex problems are less prone to over-fit, which can usually result from high model complexity. In the latter part of this thesis, we propose a multi-objective learning approach for online recruiting. Online recruiting and online dating are the most known reciprocal recommendation problems. However, the reciprocal recommendation has gained little attention in the literature due to the lack of public datasets. We aim to resolve this shortage in our study. Since the satisfaction of both candidates and companies is indispensable for successful hiring as opposed to traditional recommenders, online recruiting should respect to expectations of all parties and meet their common interests as much as possible. For this purpose, we integrated our multi-objective learning approach into various state-of-the-art methods, whose success has been proven on similar prediction problems, and we achieved encouraging results. We propose one of the prominent architectures as a prototype of our multi-objective learning approach, however, our approach applies to any recommender system employing neural networks as its final decision-maker. Our multi-objective prototype has achieved 12.15% lower LogLoss and 6.37% higher AUC than its single-objective counterpart. Besides the predictive performance, our multi-objective approach has reduced the training and testing times by half. This speedup contributes to overcoming the time constraint that complex models suffer from, so critical in the era of deep learning. Furthermore, our prototype offers explainable recommendations thanks to its Factorization Machines (FM) component. Since explainability has recently gained importance with the global changes and for ethical reasons, we have paid special attention to the selection of our base model for prototyping. Consequently, our prototype offers the reasoning behind the recommendations, so that companies can use it when requested or needed. The explainable recommendation can create a transparent hiring process and so a fair and trustworthy environment for job-seekers. This can increase the turnover rate of users and thereby help to alleviate sparsity.