有趣的是，我多次执行相同的代码，每次得到的accuracy_score都不同。我意识到在train_test splitting时没有使用任何random_state值。于是我设置了random_state=0，得到了稳定的Accuracy_score为82%。但是……然后我想尝试不同的random_state值，我设置了random_state=128，Accuracy_score变成了84%。现在我需要理解这是为什么，以及random_state如何影响模型的准确性。输出如下：1> 没有random_state:
runfile('C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic/Colab File.py', wdir='C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic'): boolean use_inf_as_null had been deprecated and will be removed in a future version. Use `use_inf_as_na` instead.[[90 22] [21 46]]0.7597765363128491runfile('C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic/Colab File.py', wdir='C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic'): boolean use_inf_as_null had been deprecated and will be removed in a future version. Use `use_inf_as_na` instead.[[104 16] [ 14 45]]0.8324022346368715runfile('C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic/Colab File.py', wdir='C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic'): boolean use_inf_as_null had been deprecated and will be removed in a future version. Use `use_inf_as_na` instead.[[90 18] [12 59]]0.8324022346368715runfile('C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic/Colab File.py', wdir='C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic'): boolean use_inf_as_null had been deprecated and will be removed in a future version. Use `use_inf_as_na` instead.[[99 9] [19 52]]0.84357541899441352> 使用random_state = 128 (Accuracy_score = 84%)
runfile('C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic/Colab File.py', wdir='C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic'): boolean use_inf_as_null had been deprecated and will be removed in a future version. Use `use_inf_as_na` instead.[[106 13] [ 15 45]]0.8435754189944135runfile('C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic/Colab File.py', wdir='C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic'): boolean use_inf_as_null had been deprecated and will be removed in a future version. Use `use_inf_as_na` instead.[[106 13] [ 15 45]]0.84357541899441353> 使用random_state = 0 (Accuracy_score = 82%)
runfile('C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic/Colab File.py', wdir='C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic'): boolean use_inf_as_null had been deprecated and will be removed in a future version. Use `use_inf_as_na` instead.[[93 17] [15 54]]0.8212290502793296runfile('C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic/Colab File.py', wdir='C:/Users/spark/OneDrive/Documents/Machine Learing/Datasets/Titanic'): boolean use_inf_as_null had been deprecated and will be removed in a future version. Use `use_inf_as_na` instead.[[93 17] [15 54]]0.8212290502793296回答：
本质上，random_state确保每次运行代码时输出相同的结果，因为每次都进行相同的精确数据分割。这主要对你的初始训练/测试分割有帮助，并且可以创建其他人可以精确复制的代码。
相同的数据分割与不同的数据分割
首先要理解的是，如果你不使用random_state，那么每次数据的分割都会不同，这意味着你的训练集和测试集每次都会不同。这可能不会造成很大的不同，但会导致模型参数/准确率等的轻微变化。如果你每次都将random_state设置为相同的值，比如random_state=0，那么数据每次都会以相同的方式分割。
每个random_state值导致不同的分割
其次要理解的是，每个random_state值都会导致不同的分割和不同的行为。因此，如果你想能够复制结果，你需要保持random_state为相同的值。
你的模型可能有多个random_state部分
第三要理解的是，你的模型的多个部分可能包含随机性。例如，train_test_split可以接受random_state，RandomForestClassifier也是如此。因此，为了每次获得完全相同的结果，你需要为模型中包含随机性的每个部分设置random_state。
结论
如果你使用random_state来进行初始的训练/测试分割，你需要设置一次并继续使用这个分割，以避免对测试集过拟合。
一般来说，你可以使用交叉验证来评估模型的准确性，而不必过多担心random_state。
一个非常重要的注意事项是，你不应该使用random_state来尝试提高模型的准确性。这样做本质上会导致模型对数据过拟合，并且对未见过的数据的泛化能力不佳。
