From the course: Introduction to Artificial Intelligence

Select the best algorithm

From the course: Introduction to Artificial Intelligence

Select the best algorithm

“

- So now you've seen three examples of supervised machine-learning algorithms. There was K nearest neighbor, regression analysis, and naive bayes. These are most often used for classifying, and then there was K means clustering which is used for unsupervised learning and clustering. Remember that each of these is like a kitchen tool. These tools are designed for something specific, but you can still be creative with how you use them. It's the same way you can use a fork to whip up eggs or a knife to pit your avocado. But as any good chef knows, you never just present one dish. Instead you're judged by the whole meal. That's why it's very common for data science teams to do something called ensemble modeling. If you're an actor or music fan, then you probably have heard the term ensemble. It's when a group performs together. It's the same thing with machine-learning algorithms. There's a few different ways to create ensembles. The most popular is bagging and stacking. Bagging is when you use several versions of the same machine-learning algorithm. Stacking is when you use several different machine-learning algorithms, then you stack them on top of one another. I used to work for a large home improvement retailer. One of their challenges was what items do they put near the checkout? You'd be surprised how much retailers earn by selling something just a few minutes before you checkout. So this was a big challenge and they wanted to create an ensemble of machine-learning algorithms. They debated which ensemble might lead to the best results. They could use bagging to try different results of the same algorithm. Then they'd see if they could improve their accuracy. This was a national retail chain so they could pull training data from stores throughout the country. So they could get data samples from random stores, then use K nearest neighbor to classify those datasets separately. Then they would aggregate those results together to see if they could come up with a larger trend. They would aggregate the insights of what people purchased right before checkup. In a sense, they were averaging out the insights to see if they could come up with a more accurate result. The retailer could also try boosting. Here instead of averaging the insights together, they'd boost the results step by step so the retailer could take a training set of their most popular items. Let's say that their best selling item was a hammer. Then they could use K nearest neighbor to see what's often bought with the hammer. Let's just say it was nails and a tool belt. Now most of us intuitively know that if someone buys a hammer, they're more likely to buy nails. But that might not help us if we want to put something near the checkout line. For that, we might want to use something like naive bayes. Remember that naive bayes is naive because it doesn't assume that predictors are correlated. So we don't assume that if you're buying a hammer, you're going to need nails. Instead it will will predict other items that are popular but might not seem related. Maybe people who buy hammers are more likely to buy chocolate bars. Mixing and matching your machine-learning algorithms will give you different insights with different results. Like any good ensemble, the accuracy of predictions will depend on the creativity of your data science team.

Contents