Regarding machine learning, today seems to be a topic that everyone is familiar with.But a few years ago, when Elizabeth first joined JMP, he was still full of curiosity and interest in exploring these areas.Today we will start the journey of machine learning and exploration together.
The author of this article is Elizabeth Claassen, the original text is published in the JMP global user community JMP Community, friends interested in the original text can also be directly View original .
When he joined JMP in 2014, Elizabeth had just earned a PhD in statistics from a school with a long and distinguished history of agricultural experiments and hybrid modeling. She used to do small experimental design, maximizing information through sensible planning, technical analysis of mixed models, and optimized output.The simplest example is the split zone design.This is true whether the scientists working with her come from animal science, entomology, food science or education.
Shortly after joining JMP, the heat of big data reached its peak.What followed was the rise of machine learning in a few years.It is not surprising that the rise of machine learning lags behind big data.So what should you do after collecting the data?Of course, it is analyzed using machine learning tools.
The formal education that Elizabeth received at school did not specifically relate to the field of machine learning, but after she started working as a statistical developer, these were the core elements that needed her to master quickly.So, when she took the first step in a new theme, she began to learn about the machine learning on the Wikipedia website.
“Machine learning is an interdisciplinary field that uses statistical techniques to enable computer systems to learn from the data (for example, to gradually improve the performance of a particular task) without the need for explicit programming.”
Does this sound unbelievable? How can a computer system "learn" anything without a clear programming? "We haven't implemented Skynet yet.Elizabeth thought.So she continued to dig deeper.
The Encyclopedia Britannica defines machine learning as: "Machine learning in artificial intelligence (computer science) is a discipline that focuses on the realization of computer software autonomous learning.”
After Elizabeth learned more about machine learning, she was more interested in autonomous learning of software."There is no explicit programming" means that once passed to the algorithm, the algorithm does not have any further instructions from humans to do the rest.
Based on the above definitions and the cognition of the terms she uses, Elizabeth recognizes that machine learning is the buzzword of current artificial intelligence, meaning computer algorithms that make decisions, predictions, or classifications based on data.
Supervised and unsupervised learning
So, can JMP implement machine learning? The answer is yes! Technically, JMP has machine learning capabilities since its first release in 1989.This shows that JMP has always been at the forefront of the times.Next, let's take a closer look at the algorithms that everyone considers when talking about machine learning.
Often, machine learning algorithms are divided into different types - supervised and unsupervised learning.
- For supervised learning, the algorithm uses both predictive variables (X columns) and reaction variables (Y columns). The training data achieves "supervision" by telling the algorithm what the expected value of the predicted variable for the given value should be.
- For unsupervised learning, the algorithm uses only predictive variables, and any pattern is determined by them only.Without using any "truth" (response), in fact many of us don't know the truth.
JMP What categories are machine learning algorithms in?
Supervised learning
|
Unsupervised learning
|
Decision tree*
|
Cluster
|
Neural network
|
Self-organizing map
|
Bootstrap Forest Law*
|
Correlation Analysis
|
return
|
Singular value decomposition
|
K nearest neighbor*
|
|
Simple shell*
|
|
Features with an asterisk (*) are limited to JMP Pro.
Seeing this, many people may be surprised by the "return" in the list, thinking that this is not machine learning in the conventional sense.But in fact it is the most direct form of supervised learning. JMP 1 The version has Fit Y by X Regression function, JMP 3 Increased Cluster After that, more machine learning options continue to appear in JMP In the new version. Those familiar with one or several algorithms may notice similarities between them.Many models can produce amazing prediction accuracy, but it can be difficult to interpret the coefficients of the model or Demarcation The meaning of the point.For example, the prediction results can be derived by tracking the nodes of the decision tree, but there is no inherent meaning behind where these nodes are selected. we do not know class The parameters in the neural network model are estimated at " real world ” What does it mean, but class Neural networks are one of the best predictive machine learning models.
In contrast, regression models provide users with an estimate of the parameters that have an inherent meaning.If the coefficient is positive, we know that increasing the value of this predictor by one unit will make Reaction value Add one equal to system The amount of the value.This and Galit Shmueli The "interpretation or prediction" that the doctor said is related.In this trade-off, the best fold Sincere One of the programs is Fit Model The generalized regression function of the platform.As JMP Statistical developer Clay Barker As said, use Lasso or Elastic Net Prediction performance can be improved, and this model can still be well explained.
But is there a lack of interpretability? Maybe, maybe not.
- An interpretable model is not necessary if the user is focused on achieving the most accurate predictions of future events.
- Conversely, if the user wants a more specific approach " Improve their scores ” , you can choose an interpretable model.
All in all, the key to any model is that you can put the output of the model into action.
Still need human intervention
Even if computer algorithms can do all the great things, manual analysis is still needed to handle more advanced problems.
such as:
What data should be included or excluded from the analysis?
- Some algorithms have the ability to select important variables, but if the model does not give the computer a variable at first, even if the variable is highly predictable, the computer will not automatically select it.
- Where is the point of diminishing returns?
If the model is too complex to run in the time it takes, then it has no use.Maybe a less complex model can provide almost as good results and can run fast.A well-known example is the winner of the Netflix Awards.The performance of the winning model is almost better than anyone expected, but it cannot be run in an immediate environment.Because Netflix needs to advise users what to watch next, but the amount of data and processing speed required by the model makes it impossible to do it in time.
- Is there an inherent bias in the model?
Obviously, computer algorithms do not have the bias that humans have, but algorithms may produce models that are highly correlated with the characteristics of the bias.For example, in a mortgage, there are some characteristics that are illegal when making a loan decision.However, there are several legitimate factors that are highly correlated with illegal factors.For example, even if the race is not directly part of the data, the algorithm may generate a model that almost completely rejects a particular race, and this model is likely to have legal consequences.
- When should I update the model?
The purpose of modeling is to generate some output that allows us to take action, so once the model is put into action, the model needs to be updated.These updates are required because the implementation changes will affect the future input of the model.If you continue to use the old model, there is a risk of extrapolation beyond the boundaries of the original data.Therefore, only humans can answer the question of which point the model should be updated.
That's why, despite the popularity of machine learning, human intervention is needed to better build models and update models to ensure that accurate models serve our work more effectively.
In the next article, we will further introduce two machine learning algorithms: k nearest neighbors and simple shells, and through practical cases to see how to use these two methods to analyze some interesting data in daily life.
