Frequency parameter of robust PCA for anomaly detection

Question

I am using the R implementation of robust PCA here for anomaly detection.

I have a vector of time series data, and a vector of dates. The algorithm works fine when the length of the vector is a multiple of the frequency of the seasonality - as stated in the documentation:

If X is a vector it will be cast to a matrix of dimension 
 frequency by length(X)/frequency

However, my seasonality is very long (yearly), while my data is collected daily. If I set the frequency to 365, I get the following error (as my vector's length is not divisible by 365):

    ggplot_AnomalyDetection.rpca(AnomalyDetection.rpca(df$cnt, 
       frequency=365, dates=df$dt)) + 
       ggplot2::theme_grey(base_size = 25)
    Error in data.frame(X.transform = (as.vector(i$rpca$X)) * 
     i$sd + i$mean,  : 
      arguments imply differing number of rows: 3650, 3384
    In addition: Warning message:
    In matrix((j - j.global.mean)/j.global.sd, nrow = frequency) 
  :
    data length [3384] is not a sub-multiple or multiple of the 
 number of rows [365]

Do I have to trim values from my vector so that it's divisible by 365 (and wait a whole year to get enough data to run the algorithm for the past year) or can I fill in dummy values?

Answer 1

Indeed the periodicity requirement in calling this function is a bit restrictive. The following link mentions that you can specify a periodicity of 1 in such a case as a work around.

That way your data would be divisible by the frequency and the function should work.

With regards to whether this is statistically sound/ideal, I wouldn't comment on that.

https://github.com/Netflix/Surus/issues/7