A0287
Title: Modeling the missing mechanism in partially ranked data with adjacency-based regularization
Authors: Kento Nakamura - The Universeity of Tokyo (Japan) [presenting]
Keisuke Yano - The University of Tokyo (Japan)
Fumiyasu Komaki - RIKEN CBS (Japan)
Abstract: Ranked data appear in a wide variety of social domains such as election and consumer survey. They often comprise partial rankings that represent only a part of preferences. Partial rankings can be regarded as the result of missing from complete rankings. Although several studies have proposed estimators for partially ranked data, they ignore the modeling of a missing mechanism, which generally leads to significant bias in estimation. However, the modeling of a missing mechanism requires a large number of parameters, which leads to over-fitting when the sample size is small. To solve this trade-off, we propose the regularized estimator for both complete rankings and missing mechanisms based on an adjacency structure inherent to partially ranked data. We introduce the implementation of the proposed estimator by using a graph regularization framework and the expectation-maximization (EM) algorithm. Simulation studies and real data application indicate that the proposed estimator performs better than existing estimators under the non-ignorable missing mechanism.
