Explainable AI Whiteboard Technical Series: Decision Trees

0:11 today we'll be looking at decision trees

0:13 and how they identify common network

0:15 issues like faulty network cables AP or

0:18 switch health and wireless coverage the

0:21 algorithms used include simple random

0:23 Forest gradient boosting and XG boost

0:26 our example will be a simple decision

0:28 tree algorithm the essential steps in

0:30 any machine learning project involve

0:32 collecting data training a model then

0:34 deploying that model decision trees are

0:37 no different let's use a real world

0:39 example in networking decision trees can

0:41 be applied to incomplete negotiation

0:44 detection MTU mismatch among others our

0:47 example today applies the fundamental

0:49 problem of determining cable faults a

0:52 cable is neither 100% good nor bad but

0:55 somewhere in between to isolate a fault

0:58 we can create a decision tree using

1:00 features of a bad cable such as frame

1:02 errors and one-way traffic we begin a

1:05 decision Tree by asking a question which

1:07 will have the greatest impact on label

1:09 separation we use two metrics to

1:11 determine that question Genie impurity

1:13 and Information Gain Genie impurity

1:16 which is similar to entropy determines

1:18 how much uncertainty there is in a node

1:21 and Information Gain lets us calculate

1:23 how much a question reduces that

1:26 uncertainty a decision tree is based on

1:28 a data set from known results

1:30 each row is an example the first two

1:33 columns are features that describe the

1:34 label in the final column a good or bad

1:37 cable the data set can be modified to

1:40 add additional features and the

1:41 structure will remain the same a

1:43 decision Tree starts with a single root

1:45 node which is given as a whole training

1:47 set the node will then ask a yes or no

1:50 question about one of the features and

1:52 split into two subsets of data which is

1:54 now the input of a child node if the

1:57 labels are still mixed good and bad then

2:00 another yes or no question will be asked

2:02 the goal of a decision tree is to sort

2:04 the labels until a high level of

2:06 certainty is reached without overfitting

2:08 a tree to the training data larger trees

2:11 may be more accurate and tightly fit the

2:13 training data but once in production may

2:15 be inaccurate in predicting real events

2:17 we use metrics to ask the best questions

2:19 at each point and so there are no more

2:21 questions to ask then prune branches

2:23 starting at the leaves to address

2:25 overfitting the tree this produces an

2:28 accurate model with leaves illustrating

2:30 the final prediction Genie impurity is a

2:33 metric that ranges from 0 to one where

2:35 lower values indicate less uncertainty

2:37 or mixing in a node it gives us our

2:40 chance of being

2:42 incorrect let's look at a male carrier

2:45 as an example in a town with only one

2:47 person and one letter to deliver the

2:49 genie impurity would be equal to zero

2:52 since there's no chance of being

2:54 incorrect however if there are 10 people

2:56 in the town with one letter for each the

2:58 impurity would be 0.9 because now

3:01 there's a one in 10 chance of placing

3:03 the randomly selected mail into the

3:05 right

3:05 mailbox Information Gain helps us find

3:08 the question that reduces our

3:10 uncertainty the most it's just a number

3:12 that tells us how much a question helps

3:14 to unmix the labels that are known We

3:16 Begin by calculating the uncertainty of

3:18 our starting set impurity equals

3:21 0.48 then for each question we segment

3:24 the data and calculate the uncertainty

3:26 of the child nodes we take a weighted

3:28 average of their un C certainty because

3:30 we care more about a large set with low

3:32 uncertainty than a small set with high

3:34 uncertainty then we subtract that from

3:37 our starting uncertainty and that's our

3:39 information gain as we continue we'll

3:42 keep track of the question with the most

3:44 gain and that will be the best one to

3:46 ask at this node now we're ready to

3:48 build the tree we start at the root node

3:51 of the tree which is given the entire

3:53 data set then we need the best question

3:56 to ask at this node we find this by

3:58 iterating over each of these vales

3:59 values we split the data and calculate

4:02 the information gain for each one as we

4:04 go we keep track of the question that

4:06 produces the most gain found it once we

4:09 find the most useful question to ask we

4:11 split the data using this question we

4:14 then add a child to this branch and it

4:16 uses a subset of data after the split

4:19 again we calculate the Information Gain

4:21 to discover the best question for this

4:23 data rinse and repeat until there are no

4:25 further questions to ask and this node

4:27 becomes a leaf which provides the final

4:29 prediction

4:31 we hope this gives you insight into the

4:32 AI we've created if you'd like to see

4:34 more from a practical perspective visit

4:36 our Solutions page on juniper.net