介绍 (Introduction)

One of the most fascinating historical examples of the power of crowds can be found within the pages of James Surowiecki’s “The Wisdom of Crowds,” in which a team of engineers, oceanographers, salvage crew members, and mathematicians were asked to make their best estimate on where a particular sunken submarine, the Scorpion, could be found (the Navy did not have the manpower to search the entire area and wanted a more specific guess). Individually, these guessers were wildly inaccurate, but the group’s combined averaged guess was only 220m from the actual position of the sunken submarine!

詹姆斯·苏洛维耶克(James Surowiecki)在《人群的智慧》一书中找到了最令人着迷的人群力量历史例子，其中要求一组工程师，海洋学家，救助船员和数学家做出最好的估计在哪里可以找到一艘特殊的沉没式潜水艇“蝎子”(海军没有人力搜寻整个地区，并希望作出更具体的猜测)。 个别地，这些猜测者非常不准确，但是该小组的综合平均猜测距离沉没式潜水艇的实际位置只有220m！

It has been shown repeatedly that the best way to arrive at an estimate is to ask lots of people from diverse backgrounds — the more diverse the better. How can we apply this sociological concept to machine learning?

反复表明，得出估算的最佳方法是问很多来自不同背景的人-越多样化越好。 我们如何将这种社会学概念应用于机器学习？

合奏学习 (Ensemble Learning)

Ensemble models are just a conglomerate of models that are averaged to provide a “crowd’s guess.” Just as humans have bias, so too do models carry with them inherent assumptions and bias. Averaging these out across a few models is almost guaranteed to decrease error.

集成模型只是模型的集合，这些模型的平均被用来提供“人群的猜测”。 就像人类有偏见一样，模型也带有内在的假设和偏见。 几乎可以保证在几个模型中对它们进行平均，以减少误差。

数据 (The Data)

For this example, I used some credit card fraud data from Kaggle. The first item of business was to pick an evaluation metric. I noticed that there was a huge imbalance of classes (99.8% of the data was marked as normal transaction volume, the other 0.2% were the fraudulent transactions), so the accuracy metric was out of the picture. For this type of problem, it’s better to choose precision, recall, or F1 scores. For simplicity, I chose precision.

在此示例中，我使用了Kaggle的一些信用卡欺诈数据 。 第一项业务是选择评估指标。 我注意到类别之间存在巨大的不平衡 (99.8％的数据被标记为正常交易量，其他0.2％是欺诈性交易)，因此准确性度量超出了预期。 对于此类问题，最好选择精度，召回率或F1分数。 为了简单起见，我选择了精度。

模型 (The Models)

After looking at the data a bit, I trained five models: a K-Nearest Neighbors, a Logistic Regression, a Random Forest, an XGBoost, and a Naive Bayes. The precisions I received after each one are shown in the chart below.

在看了一点数据之后，我训练了五个模型：K最近邻，逻辑回归，随机森林，XGBoost和朴素贝叶斯。 下图显示了我每次获得的精度。

Clearly, the Naive Bayes and KNN classifiers were not ideal, but the other three did quite well. Our maximum precision using an individual model would be 0.96. Keep that number in mind.

显然，朴素贝叶斯分类器和KNN分类器并不理想，但其他三个分类器表现良好。 使用单个模型的最大精度为0.96。 请记住该数字。

民主(几乎) (A Democracy (almost))

One of the ways to turn these outputs into a “crowd-like” format is to have each model vote based on what it received as an answer. Essentially, if the Random Forest algorithm predicted a 0, it will vote “0.” If all the other algorithms predicted a 1, they will all vote “1.” That means the ultimate tally is four 1’s and one 0. In this case, the final output will be a 1.

将这些输出转换为“人群状”格式的方法之一是根据收到的答案对每个模型进行投票。 本质上，如果随机森林算法预测为0，它将投票为“ 0”。 如果所有其他算法都预测为1，则它们都将投票为“ 1”。 这意味着最终的计数是四个1和一个0。在这种情况下，最终输出将是1。

I used a “weighted voting model” in which the models that performed the best have the most votes. The chart below details how many votes I gave each algorithm.

我使用了“加权投票模型”，其中表现最好的模型获得最多的选票 。 下表详细列出了每种算法给我多少票。

These numbers were determined from the relative precision of each model. Now let’s have them vote!

这些数字是根据每个模型的相对精度确定的。 现在让他们投票！

结果 (The Results)

After running the voting method, I received the following results:

运行投票方法后，我收到以下结果：

Prediction: 0, Actual: 0 → 56,868 (true negative)

预测：0，实际：0→56,868(真为负)

Prediction: 1, Actual: 0 → 2 (false positive)

预测：1，实际：0→2(误报)

Prediction: 0, Actual: 1 → 21 (false negative)

预测：0，实际：1→21(假阴性)

Prediction: 1, Actual: 1 → 71 (true negative)

预测：1，实际：1→71(真为负)

This translates to a precision of 0.973! Notice that for this data, the false negative is the costliest outcome, since this means the transaction was fraudulent but our models predicted a normal transaction.

这意味着精度为0.973！ 请注意，对于此数据，假阴性是最昂贵的结果，因为这意味着交易是欺诈性的，但我们的模型预测交易正常。

Just to prove that the crowd is better than the individual, here’s a chart of how many times each algorithm was correctly “outvoted” by the crowd (i.e. one algorithm voted incorrectly but the other algorithms voted correctly and hence “outvoted” the first).

只是为了证明人群比个人更好，这是一张图表，说明人群正确“投票”了每种算法多少次(即，一种算法投票不正确，而另一种算法投票正确，因此第一种算法“投票”)。

As expected, our Naive Bayes model was overriden the most number of times, but even the Random Forest was correctly outvoted 23 times by the other algorithms! The Naive Bayes model contributed to that outvoting as well, so it still is useful even if is not the best performer.

不出所料，我们的朴素贝叶斯(Naive Bayes)模型被覆盖了最多的次数，但即使是随机森林，其他算法也被正确地淘汰了23次！ 朴素贝叶斯(Naive Bayes)模型也对这一结果有所贡献，因此即使不是最佳执行者，它仍然很有用。

结论 (Conclusion)

The benefits of using an ensemble method include turning even bad models into important votes. This approach has very few drawbacks and should be an item in every data scientist’s toolbox. Better results could be achieved with hyperparameter optimization, but I just used the stock models for this example.

使用集成方法的好处包括将糟糕的模型变成重要的选票。 这种方法几乎没有缺点，应该成为每个数据科学家工具箱中的一个项目。 通过超参数优化可以实现更好的结果，但是我仅在本示例中使用了库存模型。

If you enjoyed this article, you can follow me for more content like this. Thanks for reading!

如果您喜欢本文，可以关注我以获取更多类似内容。 谢谢阅读！