DAGMM

[janfrancu]

There is no original source code for this method. Extended https://github.com/mperezcarrasco/PyTorch-DAGMM to fit datasets other than KDDCUP. Compared performance on multiple runs with the same architecture yielding following results

    │ Row │ model_seed │ niter │ precision │ recall   │ f1       │ auc      │ threshold │
    ├─────┼────────────┼───────┼───────────┼──────────┼──────────┼──────────┼───────────┤
    │ 1   │ 843        │ 8760  │ 0.742011  │ 0.545673 │ 0.628874 │ 0.76224  │ -7.29858  │
    │ 2   │ 933        │ 6860  │ 0.786819  │ 0.604782 │ 0.683894 │ 0.778152 │ -7.15721  │
    │ 3   │ 116        │ 7340  │ 0.862177  │ 0.723329 │ 0.786673 │ 0.9147   │ -6.7746   │
    │ 4   │ 627        │ 7670  │ 0.8217    │ 0.663624 │ 0.73425  │ 0.916803 │ -5.60648  │
    │ 5   │ 499        │ 4560  │ 0.916522  │ 0.827299 │ 0.869628 │ 0.930591 │ -6.95185  │
    │ 6   │ 261        │ 7710  │ 0.85944   │ 0.72409  │ 0.78598  │ 0.931925 │ -6.42293  │
    │ 7   │ 761        │ 7180  │ 0.917163  │ 0.830177 │ 0.871505 │ 0.934966 │ -6.88933  │
    │ 8   │ 797        │ 6560  │ 0.940677  │ 0.875018 │ 0.90666  │ 0.944641 │ -7.52265  │
    │ 9   │ 498        │ 4380  │ 0.902527  │ 0.798968 │ 0.847596 │ 0.947122 │ -6.89145  │
    │ 10  │ 938        │ 14660 │ 0.896952  │ 0.792594 │ 0.84155  │ 0.949002 │ -6.74864  │
    │ 11  │ 555        │ 6880  │ 0.960658  │ 0.922737 │ 0.941316 │ 0.967731 │ -6.24293  │
    │ 12  │ 724        │ 14860 │ 0.872477  │ 0.744814 │ 0.803607 │ 0.970287 │ -3.81762  │
    │ 13  │ 214        │ 5020  │ 0.968746  │ 0.939966 │ 0.954139 │ 0.985249 │ -4.8568   │
    │ 14  │ 878        │ 14930 │ 0.970904  │ 0.943317 │ 0.956912 │ 0.988188 │ -4.18334  │
    │ 15  │ 532        │ 8280  │ 0.977635  │ 0.961636 │ 0.969569 │ 0.992474 │ -4.13328  │
    │ 16  │ 854        │ 11890 │ 0.974479  │ 0.953823 │ 0.96404  │ 0.99467  │ -2.07396  │

which show that the performance in the paper is attainable, but with high variance. More details on this is in ./test/models/dagmm.jl.

[janfrancu]

Add header with acknowledgement.