A Research On Recommender Systems Based On Autoencoders And Temporal Dynamics

In recent years,because recommendation systems provide users with information that can make better choices among countless possible options,their popularity has greatly increased.Nowadays,almost all online platforms have the application of recommendation system,including online libraries,multimedia platforms,e-commerce and so on.Most recommendation systems use collaborative filtering,content-based filtering or both.The purpose of the study is to increase the accuracy of rating predictions by combining neural network structure and temporal dynamics.Matrix factorization-based collaborative filtering models can not compute non-linear similarity due to their linear structure.In recent years,various models based on neural networks have been proposed for rating prediction.Auto Rec is the first model,which implements autoencoders for collaborative filtering.These models create more accurate rating predictions than matrix factorization models because they can compute non-linear similarity.On the other hand,users’ preferences and items’ preferability change over time.Implementation of temporal dynamics into collaborative filtering is proven to increase the accuracy of rating prediction in matrix factorization models.In this thesis,we proposed Biased Auto Rec,which is built on Auto Rec.Biased Auto Rec can compute user biases and item biases at the same time although Auto Rec can only compute either users’ bias or items’ bias.Afterward,we proposed several approaches to integrate temporal signals into the Biased Auto Rec model to merge the power of nonlinearity and temporal signals.We also conducted several experiments and the experiment results showed that the Biased Auto Rec and its temporal extensions provided more accurate rating predictions than other collaborative filtering models.Briefly,this thesis contributes several innovation points in recommendation systems.(1)A new autoencoder model is designed,which can compute nonlinearity and biases for users and items at the same time.(2)Three extensions of the model are proposed,which contain various temporal signals.(3)The programming of these models are completed in coding environment and empirical tests are conducted.(4)The predictions of these three proposed temporal models are combined to improve the overall prediction accuracy.