Human-machine interaction has become one of the most active research areas, and influenced several new paradigms of computing such as Social computing, Mobile computing, and Pervasive/Ubiquitous computing, which are typically concerned with the study of human user's behavior to facilitate behavioral modeling and prediction. Human behavioral data are usually high-dimensional time series, which need dimensionreduction strategies to improve the efficiency of computation and indexing. In this paper, we present a dimension-reduction framework for human behavioral time series. Generally, recent behavioral data are much more interesting and significant in understanding and predicting human behavior than old ones. Our basic idea is to reduce to data dimensionality by keeping more detail on recent behavioral data and less detail on older data. We distinguish our work from other recent-biased dimension-reduction techniques by emphasizing on recent-behavioral data and not just recent data. We experimentally evaluate our approach with synthetic data as well as real data. Experimental results show that our approach is accurate and effective as it outperforms other well-known techniques.