This paper proposes a novel recursive approach for orthogonal frequency division multiplexing (OFDM) channel estimation over highly mobile channels without any priori channel statistics. The key idea is to decompose each objective channel into the known component and the rest for estimation. The known component is the output of the prior iteration except that the first one is attained in linear time-varying (LTV) model. In each iteration, the rest channel component, obtained by subtracting the known channel component from the original channel, is estimated in generalized complex exponential basis expansion model (GCEBEM). Since the rest component varies much milder than the original channel, both modeling error and inter-carrier interference (ICI) are significantly reduced in GCEBEM-based estimation. Monte Carlo simulations demonstrate the superiority of the proposed approach with a few iterations over the conventional methods under highly mobile conditions.