Συστήματα συστάσεων: αντιμετώπιση αραιών δεδομένων με παραγωγή εγγράφων χαρακτηριστικών
Εφραιμίδης, Ζαχαρίας Η.
SubjectArtificial intelligence -- Mathematics ; Εξόρυξη δεδομένων ; Recommender systems (Information filtering)
The proliferation of the Internet has allowed every person to extend to a wider audience. The design and development of integrated systems that allowed users to remotely access any kind of data began. Soon redundant information was produced making systems quite dysfunctional, since most users had difficulty to find interesting data. Since then research has started for solutions which should allow further development of systems, but without discouraging their users. One of the suggested guidelines for this objective is the idea of recommender systems, which, through various techniques, try to predict the degree of acceptance of each item for each user and make appropriate recommendations. Although there is a broad range of techniques available, their basic function is associated with the system’s existing data, i.e. they are based on the features of the item or the user, and the interactions between them, in order to propose appropriate items to each user. One of the major problems of recommender systems is the data sparsity. A large proportion of the aforementioned features is missing, either because of the newcomer user, for whom the system has no memory, or due to incorrect modeling of the items, which makes their structure to consist of empty features. In order to confirm the existence of this problem an experimental solution is being developed by utilizing movies’ data from MovieLens and IMDb. Then the paper seeks to address this problem by extracting topics from the set of the features, using the latent Dirichlet allocation algorithm, and describing items as mixtures of topics.