Amir Konigsberg and Ron Asherov, General Motors R&D, USA

One of the basic foundations for many recommender systems is the assumption of preference consistency and transitivity of choice. In this paper we challenge this assumption and argue that it should be revised. We also provide a method by which recommender systems can estimate preference reversals and choice intransitivity. Our general approach is to incorporate variants of choice-behavior such that recommender systems incorporate decision-sensitive factors within choice-sets that tend to influence decision making. After establishing why recommender systems ought to consider preference reversal phenomena, we provide a high-level outline of a mechanism for automatically discovering and predicting preference reversals. Thereafter we show how the discovered or predicted preference reversals can help in generating choice sets that 1) are better for the users all things considered (by being in line with their general preferences as exhibited over time) and 2) help users make better decisions.

Intransitive choice; preference reversals; recommender systems; rational choice