Missing Values in Nonlinear Factor Analysis

Tapani Raiko and Harri Valpola Helsinki University of Te hnology, Neural Networks Resear h Centre P.O.Box 5400, FIN-02015 HUT, Espoo, Finland E-mail: Tapani.Raiko hut.fi, Harri.Valpola hut.fi URL: http://www. is.hut.fi/ ABSTRACT The properties of the nonlinear fa tor analysis (NFA) model are studied by measuring how well it re onstru ts missing values in observations. The NFA model uses a multi-layer per eptron (MLP) network for approximating the nonlinear mapping from fa tors to observations. The NFA model is ompared with linear fa tor analysis (FA) and with the self-organising map (SOM). The number of parameters in the NFA model is loser to FA than the SOM, but unlike FA, NFA is able to model nonlinear manifolds. Based on experiments with real world spee h data and Boston housing data, we on lude that the performan e of the NFA model is loser to FA.