KDnuggets Home » News » 2019 » Oct » Tutorials, Overviews » How Bayes’ Theorem is Applied in Machine Learning ( 19:n41 )

How Bayes’ Theorem is Applied in Machine Learning


Learn how Bayes Theorem is in Machine Learning for classification and regression!



By Jaime Zornoza, Universidad Politecnica de Madrid

Header image

 

In the previous post we saw what Bayes’ Theorem is, and went through an easy, intuitive example of how it works. You can find this post hereIf you don’t know what Bayes’ Theorem is, and you have not had the pleasure to read it yet, I recommend you do, as it will make understanding this present article a lot easier.

In this post, we will see the uses of this theorem in Machine Learning.

Ready? Lets go then!

 

Bayes’ Theorem in Machine Learning

 
As mentioned in the previous post, Bayes’ theorem tells use how to gradually update our knowledge on something as we get more evidence or that about that something.

Generally, in Supervised Machine Learning, when we want to train a model the main building blocks are a set of data points that contain features (the attributes that define such data points),the labels of such data point (the numeric or categorical tag which we later want to predict on new data points), and a hypothesis function or model that links such features with their corresponding labels. We also have a loss function, which is the difference between the predictions of the model and the real labels which we want to reduce to achieve the best possible results.

Figure

Main elements of a supervised Learning Problem

 

These supervised Machine Learning problems can be divided into two main categories: regression, where we want to calculate a number or numeric value associated with some data (like for example the price of a house), and classification, where we want to assign the data point to a certain category (for example saying if an image shows a dog or a cat).

Bayes’ theorem can be used in both regression, and classification.

Lets see how!

 

Bayes’ Theorem in Regression

 
Imagine we have a very simple set of data, which represents the temperature of each day of the year in a certain area of a town (the feature of the data points), and the number of water bottles sold by a local shop in that area every single day (the label of the data points).

By making a very simple model, we could see if these two are related, and if they are, then use this model to make predictions in order to stock up on water bottles depending on the temperature and never run out of stock, or avoid having too much inventory.

We could try a very simple linear regression model to see how these variables are related. In the following formula, that describes this linear model, y is the target label (the number of water bottles in our example), each of the θs is a parameter of the model (the slope and the cut with the y-axis) and x would be our feature (the temperature in our example).

Figure

Equation describing a linear model

 

The goal of this training would be to reduce the mentioned loss function, so that the predictions that the model makes for the known data points, are close to the actual values of the labels of such data points.

After having trained the model with the available data we would get a value for both of the θs. This training can be performed by using an iterative process like gradient descent or another probabilistic method like Maximum Likelihood. In any way, we would just have ONE single value for each one of the parameters.

In this manner, when we get new data without a label (new temperature forecasts) as we know the value of the θs, we could just use this simple equation to obtain the wanted Ys (number of water bottles needed for each day).

Figure

Figure of an uni-variate linear regression. Using the initial blue data points, we calculate the line that best fits these points, and then when we get a new temperature we can easily calculate the Nº sold bottles for that day.

 

When we use Bayes’ theorem for regression, instead of thinking of the parameters (the θs) of the model as having a single, unique value, we represent them as parameters having a certain distribution: the prior distribution of the parameters. The following figures show the generic Bayes formula, and under it how it can be applied to a machine learning model.

Figure

Bayes formula

 

Figure

Bayes formula applied to a machine learning model

 

The idea behind this is that we have some previous knowledge of the parameters of the model before we have any actual data: P(model) is this prior probability. Then, when we get some new data, we update the distribution of the parameters of the model, making it the posterior probability P(model|data).

What this means is that our parameter set (the θs of our model) is not constant, but instead has its own distribution. Based on previous knowledge (from experts for example, or from other works) we make a first hypothesis about the distribution of the parameters of our model. Then as we train our models with more datathis distribution gets updated and grows more exact (in practice the variance gets smaller).

Figure

Figure of the a priori and posteriori parameter distributions. θMap is the maximum posteior estimation, which we would then use in our models.

 

This figure shows the initial distribution of the parameters of the model p(θ), and how as we add more data this distribution gets updated, making it grow more exact to p(θ|x), where x denotes this new data. The θ here is equivalent to the model in the formula shown above, and the x here is equivalent to the data in such formula.

Bayes’ formula, as always, tells us how to go from the prior to the posterior probabilities. We do this in an iterative process as we get more and more data, having the posterior probabilities become the prior probabilities for the next iteration. Once we have trained the model with enough data, to choose the set of final parameters we would search for the Maximum posterior (MAP) estimation to use a concrete set of values for the parameters of the model.

This kind of analysis gets its strength from the initial prior distribution: if we do not have any previous information, and can’t make any assumption about it, other probabilistic approaches like Maximum Likelihood are better suited.

However, if we have some prior information about the distribution of the parameters the Bayes’ approach proves to be very powerful, specially in the case of having unreliable training data. In this case, as we are not building the model and calculating its parameters from scratch using this data, but rather using some kind of previous knowledge to infer an initial distribution for these parameters, this previous distribution makes the parameters more robust and less affected by inaccurate data.

I don’t want to get very technical in this part, but the maths behind all this reasoning is beautiful; if you want to know about it don’t hesitate and email me to jaimezorno@gmail.com or contact me on LinkdIn.

 

Bayes’ Theorem in Classification

 
We have seen how Bayes’ theorem can be used for regression, by estimating the parameters of a linear model. The same reasoning could be applied to other kind of regression algorithms.

Now we will see how to use Bayes’ theorem for classification. This is known as Bayes’ optimal classifier. The reasoning now is very similar to the previous one.

Imagine we have a classification problem with i different classes. The thing we are after here is the class probability for each class wi. Like in the previous regression case, we also differentiate between prior and posterior probabilities, but now we have prior class probabilities p(wi) and posterior class probabilities, after using data or observations p(wi|x).

Figure

Bayes formula used for Bayes’ optimal classifier

 

Here P(x) is the density function common to all the data points, P(x|wi) is the density function of the data points belonging to class wi, and P(wi) is the prior distribution of class wiP(x|wi) is calculated from the training data, assuming a certain distribution and calculating a mean vector for each class and the covariance of the features of the data points belonging to such class. The prior class distributions P(wi) are estimated based on domain knowledge, expert advice or previous works, like in the regression example.

Lets see an example of how this works: Image we have measured the height of 34 individuals: 25 males (blue) and 9 females (red), and we get a new height observation of 172 cm which we want to classify as male or female. The following figure represents the predictions obtained using a Maximum likelihood classifier and a Bayes optimal classifier.

Figure

On the left, the training data for both classes with their estimated normal distributions. On the right, Bayes optimal classifier, with prior class probabilities p(wA) of male being 25/34 and p(wB) of female being 9/34.

 

In this case we have used the number of samples in the training data as the prior knowledge for our class distributions, but if for example we were doing this same differentiation between height and gender for a specific country, and we knew the woman there are specially tall, and also knew the mean height of the men, we could have used this information to build our prior class distributions.

As we can see from the example, using these prior knowledge leads to different results than not using them. Assuming this previous knowledge is of high quality (or otherwise we wouldn’t use it), these predictions should be more accurate than similar trials that don’t incorporate this information.

After this, as always, as we get more data these distributions would get updated to reflect the knowledge obtained from this data.
As in the previous case, I don’t want to get too technical, or extend the article too much, so I won’t go into the mathematical details, but feel free to contact me if you are curious about them.

 

Conclusion

 
We have seen how Bayes’ theorem is used in Machine learning; both in regression and classification, to incorporate previous knowledge into our models and improve them.

In the following post we will see how simplifications of Bayes’ theorem are one of the most used techniques for Natural Language Processing and how they are applied to many real world use cases like spam filters or sentiment analysis tools. To check it out follow me on Medium, and stay tuned!

Figure

Another example of Bayesian classification

 

That is all, I hope you liked the post. Feel Free to connect with me on LinkedIn or follow me on Twitter at @jaimezorno. Also, you can take a look at my other posts on Data Science and Machine Learning here. Have a good read!

 

Additional Resources

 
In case you want to go more in depth into Bayes and Machine Learning, check out these other resources:

and as always, contact me with any questions. Have a fantastic day and keep learning.

 
Bio: Jaime Zornoza is an Industrial Engineer with a bachelor specialized in Electronics and a Masters degree specialized in Computer Science.

Original. Reposted with permission.

Related:


Sign Up

By subscribing you accept KDnuggets Privacy Policy