Open VincentWtrs opened 5 years ago
I thought I had successfully tackled the issue by addressing cause 1. and reversing the order. However I reloaded the package with the unchanged original development branch and not the one introducing the fix. At first I tested with 10% outliers and mu_outlier = 10 and tests were actually successfull. Then testing with 20% outliers, they were not. TPR on outliers was very low, and FPR was very high.
Two new tangential issues:
Also I am testing with the IC option turned on (e.g. EBIC) but with the default hyperparameter grid construction (a big grid). This is clearly taking too long for useful testing, however it minimizes the chance of issues due to missing good lambda values.
Update: Re-training with outlier_mu = 10 and 10% outliers. The standard procedure using IC (EBIC) works well when not giving lambdas and alphas sequences. When specifying my own, it seems to fail.
Things to investigate further
When running models with higher amounts of outliers (say: 10%) or higher. The algorithm fails to detect the outliers. The original version does a much better job, although not perfectly. I need to track the root cause of the issue.
Potential causes: