Let's (not) stick together

The ability to generalise towards either new users or unforeseen behaviours is a key requirement for activity recognition systems in ubiquitous computing. Differences in recognition performance for the two application cases can be significant, and user-dependent performance is typically assumed to be an upper bound on performance. We demonstrate that this assumption does not hold for the widely used cross-validation evaluation scheme that is typically employed both during system bootstrapping and for reporting results. We describe how the characteristics of segmented time-series data render random cross-validation a poor fit, as adjacent segments are not statistically independent. We develop an alternative approach -- meta-segmented cross validation -- that explicitly circumvents this issue and evaluate it on two data-sets. Results indicate a significant drop in performance across a variety of feature extraction and classification methods if this bias is removed, and that prolonged, repetitive activities are particularly affected.