Stable and reliable predictive accuracy of robust weighted slope one under profile injection attacks

Abstract

The Slope One class of algorithms have been shown to lead, although being relatively simple, to accuracies that are very close to the more commonly utilised memory-based Collaborative Filtering (CF) algorithms. In addition, this class of algorithms is highly scalable in comparison to memory-based algorithms. A recently observed phenomenon are profile injection attacks on recommendation algorithms that tend to increase (push attack) or decrease (nuke attack) the recommendations of an item depending on the intentions of the attacker. Model-based algorithms have performed well under such attacks and are therefore preferred over memory-based CF algorithms despite their lower accuracy. Previous work showed that the recommendation ranking of items using the Robust Weighted Slope One (RWSO) algorithm is fairly stable under profile injection attacks compared to the Weighted Slope One (WSO) and the Improved Slope One (I-SO) approaches. In this paper, it is shown that, under profile injection attacks, the predictive accuracy of RWSO is more stable and hence more reliable than the predictive accuracy of WSO and I-SO.