Hello! I understand and fully agree, that in many fields the Machine Learning methods get us better results for comparison with Traditional Statistical methods. But what is advantage to use Machine Learning methods (e.g., one-class chart) instead of regular Hotelling chart in MSPC (Multivariate Statistical Process Control) ? Main reason to use one-class chart is that "Hotelling's T2 chart assumes that observations follow a multivariate normal distribution. But in many tasks the distribution concerning the observations is not known". That's fair, but please, explain me, why we could not use averaging (e.g., to use sample as mean of 5…10 observations) and these average values will be according normal distribution, isn't it?
Sergey,
There has been a general trend of using ML as a hammer for everything when it is not. The use of a particular ML technique will vary depending on the problem. Each problem has a particular underlying distribution that one has to figure out based on the gathered samples(also most ML and statistical techniques assume a stationary distribution when I think it is more the exception than the rule and there are more advanced ml techniques for these). One assumption that you must test is whether your problem follows a particular behavior (follows a Gaussian distribution, etc.). This will dictate if you use MSPC or other technique
Dear
İ found an interesting article and sent you.
http://www.analyticsvidhya.com/blog/2015/07/difference-machine-learning-statistical-modeling/
http://www.eeanalytics.com/wp-content/uploads/2013/11/machinelearning.pdf
Regards, Shafagat
Dear Shafagat, thanks a lot for your very fast answer! You are fully right, attached articles are very interesting, but my concrete question was about Machine Learning using on MSPC. What is your opinion about this? Regards, Sergey.
Artifical Intelligence and Machine Learning of Statistics vs Traditional Statistical Methods
I like the comparison as thought-provoking
I to be formed before the rise of computer science at current levels (early 90s of the last century), clinical and educational researcher rather deal with solving a current problem that involves lack of knowledge of some phenomenon that currently can only assume (hypothesizing) and take it to design and mathematical modeling and how to analyze and evaluate the results with the Traditional Statistical Method.
Of course being a professor of Medicine, and Medical new forming and promoter in the formation of new researchers in clinical and teaching prefer to learn methodologies,them Traditional Statistical Method.
This approach does not limit computer usage (hardware and software as updates and "edge" technology) and pre-programmed biostatistics (STATA, EPISTAT, SPSS, Excel Plus) for analysis of results of our research, allowing handle packages properly databases, relate to specific fields have combined data information and express all in graphical and numerical form at high speed. (Tables and Charts in 2 and 3 dimensions)
Once you mastered the traditional statistical method and be experts at it, will be able to use Artificial Intelligence to keep learning and getting better handling and quality, the results of their research. This also enables them to become experts in Estimation Theory, displaying only already prepared data, expressed in summary measures and managed graphically.
Go directly to learn Statistics by Computer Science and Artificial Intelligence, the risk that students both novices and experts commit many errors, in case of failure or breakdown support (hardware) and operational programs (software) run these computers.
On the gold side, they would know nothing of Estimation Theory (which is the means by which errors are detected in the development and management of data, and therefore the inference thereof). Besides his work he would stop for failing to develop and analyze information not meet Traditional Statistical Methods.
Following example of "truism" would know not to use traditional mathematical operations (addition, subtraction, division, multiplication) without the use of a simple or scientific calculator.
Sincerely
Dr. Jose Luis Garcia Vigil
Dear Dr. Jose Luis Garcia Vigil, thanks a lot for your detailed answer! Regards, Sergey.
Sergey,
There has been a general trend of using ML as a hammer for everything when it is not. The use of a particular ML technique will vary depending on the problem. Each problem has a particular underlying distribution that one has to figure out based on the gathered samples(also most ML and statistical techniques assume a stationary distribution when I think it is more the exception than the rule and there are more advanced ml techniques for these). One assumption that you must test is whether your problem follows a particular behavior (follows a Gaussian distribution, etc.). This will dictate if you use MSPC or other technique
Your question misses the point. The question should be what are the best methods to analyze data to answer the question asked in the research. The analysis of data is neither statistics nor machine learning. The analysis of data is determining if the data have sufficient information to inform you of what you need to know. The analysis of data requires that you prepare for taking data in a manner that will answer your question, i.e., the data are relevant to the question and contain no confounding contributions. The analysis requires that there is sufficient data to show that an important distinction can be made. Requirements for this distinction should be made before collecting data. The data are then analyzed mathematically to establish if the distinction was made and the distinction is of sufficient quality. Machine learning or statistics may be part of the process, but must be shown to be applicable to the question asked and the data collected.
Dear Arturo Geigel and Joseph L Alvarez, thanks a lot for your answers! Regards, Sergey.
- Badges
- Science topic