I have a binary classification problem, where each data point is multi-channel time-series, which can be represented as a matrix $T \times F$, where $T$ is the time-series length, and $F$ as the channels number.
$T$ isn't constant - each data point has different length. For classification I need to extract a fixed sized features vector from each data point. Moreover, I suspect that the information is partly embedded as short events in the data, so transformations as fourier are not very helpful.
I tried to calculate the covariance matrix of $T\times F$ to get a $F\times F$ matrix, take the upper/lower part (include the main diagonal) and flat it. After that, I used PCA to reduce dimensionality.
I'm afraid I loosing lots of important information using this transformation. Mainly because SVM gives me something near to chance level. I don't think that methods like Matrix Completion will help, maybe stretching methods will be more helpful.