Identifying differentially expressed genes and behavior patterns from time series microarray data has been a challange. Such identification is crucial to our understanding of the temporal regulations of the affecting factors. This not only provides knowledge about the inter-gene relationships over time but also helps in studying the genetic change that a cell undergoes under differential conditions. Methods exist for modeling the long time series data (>8 time points) and have shown to be successful at identifying differentially expressed genes and, in some cases, patterns of interest. However, this still is a challange for short time series data due to fewer time points and still fewer samples making modeling impossible. Moreover, traditional methods such as $t$-test, although used extensively, are not justified theoretically. Other mean difference based methods do seem to work in practice but it is not clear why. We present a general theoretical framework, by extending the Hilbert-Schmidt Independence Criterion (HSIC) based feature selection, for analyzing short time series data that can perform both, gene identification and pattern discovery, in a unified manner. This obviates the need for clustering based approaches studying an exponential number of profiles of time behavior of genes.