Aggregation Bug Fixed & Support Provided for Example Datasets

[Garen-Wang]

When directly running experiments according to the sample, error occurs due to passing a dict as a parameter of function agg, which is now deprecated by pandas.

Just changing its type to list can solve this problem. So there is no need to use pandas==0.25.

---

Jan 29 upd:

If directly run experiments of example datasets, actually feature importance and AUC cannot be calculated correctly.

For instance, the AUC will always be 0.5 and the feature importance of heart dataset will be like this:

   feature_name  split  gain  gain_percent  split_percent  feature_score
0           age      0   0.0           NaN            NaN            NaN
1           sex      0   0.0           NaN            NaN            NaN
2    chest-pain      0   0.0           NaN            NaN            NaN
3    bp-resting      0   0.0           NaN            NaN            NaN
4   cholesterol      0   0.0           NaN            NaN            NaN
5    bs-fasting      0   0.0           NaN            NaN            NaN
6   ecg-resting      0   0.0           NaN            NaN            NaN
7        hr-max      0   0.0           NaN            NaN            NaN
8           eia      0   0.0           NaN            NaN            NaN
9       oldpeak      0   0.0           NaN            NaN            NaN
10    k-oldpeak      0   0.0           NaN            NaN            NaN
11      vessels      0   0.0           NaN            NaN            NaN
12         thal      0   0.0           NaN            NaN            NaN

Actually we need to modify one of the parameters of algorithm LightGBM, called min_data, to a factor of the number of instances in our own dataset. Now the feature importance will become normal:

   feature_name  split        gain  gain_percent  split_percent  feature_score
7        hr-max     75  173.631106     11.917672      23.291925       0.198796
4   cholesterol     48   76.677835      5.263005      14.906832       0.120137
0           age     48   70.926419      4.868240      14.906832       0.118953
2    chest-pain     14  396.282254     27.199977       4.347826       0.112035
9       oldpeak     39  125.718957      8.629084      12.111801       0.110670
12         thal     14  290.041550     19.907839       4.347826       0.090158
11      vessels     20  174.509817     11.977985       6.211180       0.079412
3    bp-resting     27   37.454598      2.570804       8.385093       0.066408
1           sex     13   35.071794      2.407254       4.037267       0.035483
6   ecg-resting     10   17.174318      1.178809       3.105590       0.025276
8           eia      8   29.409258      2.018589       2.484472       0.023447
10    k-oldpeak      6   30.023399      2.060743       1.863354       0.019226
5    bs-fasting      0    0.000000      0.000000       0.000000       0.000000

Then we can expand our scalability to other datasets.