Autotune xgboost: reg:linear vs reg:squarederror

chrisn · October 21, 2019, 8:46pm

TVM appears to use “reg:linear” or ‘rank:pairwise’ for it’s loss function when doing autotune:

dmlc/tvm/blob/master/python/tvm/autotvm/tuner/xgboost_cost_model.py#L102


        'max_depth': 3,
        'gamma': 0.0001,
        'min_child_weight': 1,


        'subsample': 1.0,


        'eta': 0.3,
        'lambda': 1.00,
        'alpha': 0,


        'objective': 'reg:linear',
    }
elif loss_type == 'rank':
    self.xgb_params = {
        'max_depth': 3,
        'gamma': 0.0001,
        'min_child_weight': 1,


        'subsample': 1.0,


        'eta': 0.3,

But XGBoost itself seems to have deprecated reg:linear:

Should the first line be changed to ‘reg:squarederror’?

moderato · March 2, 2020, 12:06am

Agree, seen the same thing.