C-loss is robust for zero-count.

<p>The observed zero values do not reflect real gene expression, which can be seen as outliers. When we use C-loss on zero count, small punishment (small loss value) can be imposed on the error. In contrast, <i>L</i><sub>2</sub> loss impose a larger punishment for the same error. This is to say, C-loss is more robust than <i>L</i><sub>2</sub> loss. Since the observed high-expressed gene expressions (non-zero count) are usually accurate, we use <i>L</i><sub>2</sub> loss to measure the error.</p> <p>(PDF)</p>