Important ins which are not in a common articles.

1. You can sort a learning data set as you wish. So what? Hm... For different orders you will get different entropy at each key on the set. Those prediction realized by a tree could be very different depending on the order of elements in the training data set. So what the order to choose? It is simple. We have at least one attribute on the learning data which is absent on a predicting data set. We need to sort a learning data by a target feature values. Those we will have respectful to a predicting feature entropy by each key in the learning set. It has sense. The more existing feature dependent to the predicting feature the lesser entropy it produces because for the predicting feature we have an index with guaranteed 0 entropy.
2. Tree degeneration with recursion. The problem is that we have limited number of feature-keys and in common case we have much more (possibly unlimited) number of items in the learning data set. If an algorithm recursively started a branch building it will recursively build single branch at each node. The keys be used before all the data will be analysed. In this way the tree is becoming a linked list. Same thing occurs with multi-threading (with Python). To solve the issue it is possible to store a leafs during all the learning process.