چكيده به لاتين
Recommender systems have rapidly grown in proportion to digital applications development. Collaborative filtering is one of the best methods in recommender systems which is working based on users and items similarities. One of the main challenges of recommender systems’ algorithms is data sparsity. This challenge would be so crucial if we are going to consider another dimension, time, besides two dimensional data, user and item. Users’ behavior change through time duration and environmental state. In this dissertation, we are going to present a dynamic recommendation model for recommender systems. We use time dimension as an independent dimension for solving dynamic recommendation. Sparsity becomes worse when we use time dimension. Introducing a dynamic solution is the main challenge of this research. In this dissertation, we have presented a blueprint design of recommendation systems which includes a few phases. Each phase, target the sparsity challenge from a different point of view. In the first phase of this structure, the co-clustering algorithm is used on user-item matrix to create smaller matrices which their users and items are similar while these matrices benefit from much less sparsity. In the second phase, we use initialized latent matrices of users and items for clustering matrices which leads to a more convergent model. In the third phase, we set these matrices in main user-item matrix and solve a new dense matrix via a collaborative filtering algorithm like matrix factorization. In the last phase, we create a new user-time matrix and then using the multi-linear regression technique we solve the user-item matrix. Our experiments are performed in any state of the phases on MovieLense and EachMovie datasets. Results demonstrate better results versus competitive models. We have evaluated our model based on RMSE value to show significant reduction of prediction error rate when we use the time dimension. But, if we are going to recommend items to the users based on the time, we have to use time dimension after co-clustering and solve small user-item-time tensor via tensor decomposition and then obtain the main tensor. Solving this dense user-item-time tensor we can recommend an item based on the time aspect, although, the accuracy of this recommendation is somehow less than the previous experiment in the MovieLense dataset, but we have obtained the time feature in our new recommendation.
