Ghadami, Amirhossein2024-09-132024-09-132024-09-13http://hdl.handle.net/10393/46576https://doi.org/10.20381/ruor-30557In the rapidly evolving digital landscape, recommendation systems are pivotal in optimizing user experience and driving the economic success of businesses involved in digital commerce and content distribution. These systems intelligently suggest products and services to users by analyzing a myriad of factors such as historical behaviors, item characteristics, and individual preferences. Among the various types, hybrid recommendation systems have emerged as particularly influential. They combine multiple computational techniques to enhance the accuracy of predictions and effectively address the shortcomings inherent in single-method systems by integrating diverse data sources and algorithmic approaches. This thesis introduces the Convolutional Autoencoder Recommendation System (CAERS), which utilizes a Convolutional Autoencoder (CAE) to deeply analyze and decode content data from users and items. The primary advantage of CAERS lies in its ability to discern intricate content patterns, which significantly aids in mitigating the cold start problem - providing robust recommendations based solely on content analysis, thus enhancing user engagement right from their initial interaction with the system. Building upon the foundation laid by CAERS, this research extends into the realm of hybrid systems with the development of CAERS-CF. This model synthesizes the strengths of CAERS's deep learning-based content analysis with the traditional collaborative filtering techniques, creating a robust framework that harnesses both the content data and historical user-item interactions. This integration results in a marked improvement in recommendation accuracy, making CAERS-CF a superior choice compared to standalone models and other hybrid configurations. The culmination of this research is embodied in TriDeepRec, a novel hybrid recommendation system that further extends the capabilities of CAERS by integrating it with Neural Collaborative Filtering (NCF) and a Multilayer Perceptron (MLP). This three-pronged approach allows TriDeepRec to leverage the nuanced capabilities of CAERS for content data analysis, NCF for insight into behavioral data, and MLP for effectively combining these diverse inputs into a cohesive output. The result is a recommendation system that not only predicts user preferences with high precision but also adapts to new and evolving data patterns. To quantitatively measure the effectiveness of these innovations, this thesis employs two key metrics: Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). These metrics serve to evaluate and compare the performance of our models against established benchmarks. The results clearly demonstrate that our proposed systems offer substantial improvements in recommendation accuracy and reliability, setting new standards for hybrid recommendation systems.enAttribution 4.0 Internationalhttp://creativecommons.org/licenses/by/4.0/Recommendation SystemHybrid Recommendation SystemConvolutional AutoencoderDeep LearningCold StartThesis