Read NG video about machine learning system construction recommendations, feel very practical, recorded as a lecture notes.
The first is the process of machine learning system construction:
Ng Recommendation method: The first fast implementation of a possible is not very perfect algorithm system, cross-validation, draw the learning curve to learn the problem of the algorithm, is high bias or high variance details see this blog introduction: Bias and variance in machine learning application
The most important step: Error analysis, manual test of the algorithm error learning samples, find out what type of algorithm in the case of mistakes! Then there are several experiences in dealing with such errors.
Below is an example of a spam message system:
Anti-spam system found that the most common mistake is steal passwords so focus on the experience to solve this part of the problem! This will be more efficient.
There is also a small trick: learning algorithms we need to have numerical quantification criteria to evaluate the algorithm!
In class Ng mentioned whether the use of stemming in the spam Email, the best way is to experiment two times respectively and not applicable, through the evaluation of numerical quantization criteria to choose whether to use the stem extractor.
The video then specifically mentions the need for error analysis validation on the cross validation set to see the following topics:
Machine Learning System Construction