Abstract:
The prediction accuracy of recommendations given by recommender systems and algorithms has
been a major concern in recent years with a high prediction accuracy being the first goal for any
recommendation. Research on evaluating recommender systems shows that algorithms in this
area are still deficient in prediction accuracy but recent works prove that modeling with temporal
dynamics improves the degree of recommendation accuracy. Recommendations presume the
presence of users or items in the matrix. They are invariably based on similarities of users and/or
items in the user-item matrix of a system, user profiles, and rating information. The major
difference in operation of the recommendation algorithms is in the way the algorithms analyze
data sources to develop notions of affinity between users for use in identifying well matched
pairs. There is limited work focused on the temporal absence as an indicator of preference or
concept drift: and hence a factor for inclusion in the recommender algorithms. This thesis has,
through live data from users, determined the factors and aspects that lead to absence and
indicators of absence from a social media perspective; these values are then measured and fitted
into a model that predicts a user‘s absence duration. The deployment of this modeling is
expected to indicate concept drifts, enhance affective communication and product marketing. A
working temporal absence definition is realized and using Sparse Matrix manipulation with
linear regression, the model is trained using gradient descent algorithm to predict future absence
time and durations.