Abstract:
Recommendation systems are software tools that seek to suggest relevant items regarding user preferences. Preference might be any; text to read, product to buy, or anything depending on the industry. Since it’s vastly preferred by companies nowa days, it has been studied extensively. However, there are still open research areas based on performance improvements. In this thesis, a movie recommendation system is built using deep learning methods in order to see the e↵ect of deep learning on the performance of recommendation systems. Moreover, this study incorporates natu ral language processing methods since movie description texts are the main metadata used by the system. To increase the quality of recommendation, a two-layered system is built and each layer is a deep neural network-based. The first layer acts as a com plex embedding layer built on convolutional neural network architecture which takes movie descriptions as input, gives the predicted genres as output, and uses network parameters as embedding vectors of the second layer’s input. This way, each movie has more complex representation on the second layer, making system’s recommendations elaborated. The second layer is a recurrent neural network architecture, which takes embedding vectors of user’s pre-interacted movies and user ratings to that movie as sequential input and outputs the next items to be recommended. Hence, the system can promote any movie even though it has never been interacted with a user before. After having completed the implementation and testing the system’s recommen dation performance, we have shown that such architecture which fuses two DNNs for both feature extraction and recommendation purposes produces better results than the baseline method.
