Consider alternative weighting schemes

[magsol]

The method of setting the baseline for computing anomalies has the potential to be heavily skewed in favor of trailing eigenvalue components. See this line:

ornet/src/analysis/temporal_anomaly_detection.py

Line 94 in 6dd29a3

eigen_vals_avgs = [np.mean(x) for x in eigen_vals]

If eigenvalue components obey a classic power-law distribution (associated with most real-world data), then disproportionate weight will be assigned to the trailing components given how averages work. We may want to consider a weighted mean, whereby weight is assigned to each eigenvalue according to the variance it explains in the original data. Alternatively, we could compute a median or interquartile range (e.g., top quartile).