Use auc score to unify the intermediate and objective values

[nzw0301]

Motivation

Resolve https://github.com/optuna/optuna-examples/issues/251.

Description of the changes

Use pre-computed booster's validation auc score rather than computing inconsistent accuracy.

Alternative ways I considered were not working by the following reasons:

• Compute AUC score using sklearn: We cannot calculate the AUC score because gbm does not have predict_prob method https://github.com/dmlc/xgboost/issues/4583, which is necessary to compute the score
• Use accuracy as the intermediate values: I suppose the accuracy metric isn't provided by metrics in lightgbm.

[y0z]

@not522 @eukaryo Could you review this PR?

[not522]

Sorry for my late response. How about using sklearn.metrics.roc_auc_score(valid_y, preds)? It seems like a more usual way of calculating AUC. (I'm not familiar with LightGBM, so please correct me if I'm wrong.)

[nzw0301]

Thank you for your suggestion! You're absolutely right and my comment:

Compute AUC score using sklearn: We cannot calculate the AUC score because gbm does not have predict_prob method https://github.com/dmlc/xgboost/issues/4583, which is necessary to compute the score

was wrong. I'll update the file by following your suggestion.