Multi-Objective News Recommender Systems

Abstract

There are some serious limitations of the “most-popular” or most-emailed Top-N news recommender systems (NRS) – widely used by various media sites. For instance, these systems have a self-reinforcing nature, the (N+1)th article in a list is often unduly penalized and these recommended lists are easily susceptible to manipulation. The self-reinforcing nature comes from the fact that once an article makes it into a Top-N list, it gains even more popularity simply by virtue of being in such a prominent list. Prawesh & Padmanabhan (2012) propose a class of probabilistic Feedback based NRS (FNRS) to address the aforementioned limitations of “most popular” Top-N NRS. In general, in these models the recommendation probability of an article with clicks n is proportional to f(n) = ny, y ∈ ℝ is called feedback parameter of the model. Of particular interest is the case y > 1, because in these range of values, an optimization problem can be formulated that helps us to determine the optimal level of feedback parameter y. The optimization problem has two objectives for minimization, namely: (a) accuracy-loss and (b) distortion created by an FNRS (with exponent y) over time. These two objectives were transformed into a single objective-function (Prawesh & Padmanabhan, 2012) for optimization, using apriori the weights assigned to these objectives.