官术网_书友最值得收藏!

Comparing the entropy differences (information gain)

To know which variable to choose for the first split, we calculate the information gain G when going from the original data to the corresponding subset as the difference between the entropy values:

Here, S(f1) is the entropy of the target variable and S(f1,f2) is the entropy of each feature with respect to the target variable. The entropy values were calculated in the previous subsections, so we use them here:

  • If we choose Outlook as the first variable to split the tree, the information gain is as follows:

G(Train outside,Outlook) = S(Train outside) - S(Train outside,Outlook)
                                                 = 0.94-0.693=0.247

  • If we choose Temperature, the information gain is as follows:

G(Train outside,Temperature) = S(Train outside) - S(Train outside,Temperature)
                                                           = 0.94-0.911=0.029

  • If we choose Humidity, the information gain is as follows:

G(Train outside,Humidity) = S(Train outside) - S(Train outside,Humidity)
                                                     = 0.94-0.788=0.152

  • Finally, choosing Windy gives the following information gain:

G(Train outside,Windy) = S(Train outside) - S(Train outside,Windy)
                                                  = 0.94-0.892=0.048

All these calculations are easily performed in a worksheet using Excel formulas.

The variable to choose for the first splitting of the tree is the one showing the largest information gain, that is, Outlook. If we do this, we will notice that one of the resulting subsets after the splitting has zero entropy, so we don't need to split it further.

To continue building the tree following a similar procedure, the steps to take are as follows:

  1. Calculate S(Sunny), S(Sunny,Temperature), S(Sunny,Humidity), and S(Sunny,Windy).
  2. Calculate G(Sunny,Temperature), G(Sunny,Humidity), and G(Sunny,Windy).
  3. The larger value will tell us what feature to use to split Sunny.
  4. Calculate other gains, using S(Rainy), S(Rainy,Temperature), S(Rainy,Humidity), and S(Rainy,Windy).
  5. The larger value will tell us what feature to use to split Rainy.
  6. Continue iterating until there are no features left to use.

As we will see later in this book, trees are never built by hand. It is important to understand how they work and which calculations are involved. Using Excel, it is easy to follow the full process and each step. Following the same principle, we will work through an unsupervised learning example in the next section.

主站蜘蛛池模板: 柏乡县| 广东省| 德庆县| 会昌县| 黄石市| 崇州市| 美姑县| 卢湾区| 建湖县| 淳安县| 晋州市| 中宁县| 安乡县| 普兰店市| 景宁| 兴山县| 金平| 公安县| 平罗县| 奉新县| 白朗县| 阿鲁科尔沁旗| 汶上县| 梓潼县| 博客| 武陟县| 梧州市| 新平| 望都县| 金华市| 依兰县| 兴化市| 利辛县| 双流县| 高唐县| 平远县| 崇左市| 岳西县| 绥棱县| 津南区| 新巴尔虎右旗|