Explaining Usual Machine Learning Suspects.

I usually find myself in a dead end explaining machine learning models to business people as my chosen examples are always skewed towards mathematical modeling and scientific intuitions.

Given it is a hard task to look for real-life analogies, I proposed myself a fun exercise to look for small day-to-day examples to the following selected models.

1. Linear regression :

An example of use would be that of a parking lot service. You start with a basic fee which increases continually following a certain rate. The basic fee represents the intercept describing the initial state before the service set off. Depending on whether you're having a membership, using the parking area on week-ends or on holidays, that rate may vary with respect to these variables.

2. Logistic regression :

The model can be personified into a court judge whose judgment is based on facts, who consults with solicitors and other judges and gives voice to witnesses. Finally, after thoroughly looking upon the matter, they hand down the verdict stating that a suspect is either guilty or not.

The output of logistic regression is formalized through a probability which could describe a degree of certainty with the judgment. The more implacable the facts, the heavier the sentence. Likewise, the lesser the facts, the more likely the suspect will be acquitted. Interestingly enough, the verdict varies from a judge to another, as it is also about their decision threshold and perspective of justice.

3. Support Vector Machines :

If you are a loyal Jurassic Park fan, this example might be of some relevance. Think of your data classes as dinosaur species. As Ingen did for its amusement park, the idea is to build electroshocking enclosures to keep the animals as far away as possible from the pens ( sufficient security margin ) and to maintain each species confined within its territory. However, some of the species require additional logistical actions to be shepherded. For pterodactyls, birdcages must be installed thus the need to go into another separative dimension.

4. KNN ( K-Nearest Neighbors ):

There is a saying going that everyone is the average of the five persons spending most of their time with. You can simply extrapolate to k persons instead, for the argument's sake. When a person is placed at the constant and immutable vicinity of k other individuals, they will likely be sharing the same characteristics, thus belonging to the same group. This makes sense at a more abstract level speaking for general observations.

5. K-Means :

You want to try the latest Spotify recommendation yet you are not sure which genre you are dealing with and you certainly never heard of the band beforehand. 

You first check out the length of the track; 17 minutes, so it is either a classical music cover, a new progressive metal hit, or electronic music. You launch the track and you get right afterwards a clear sense of the beat; whether it is aggressive, rhythmical, or more of relaxing. 

The next few seconds give you convincing indicators as to which of your numerous playlists you shall add it to. 

Fun part of all that is that you didn't even name them after their genres, but rather after which moment of the day you would likely listen to them the most ( during working time, while working out, during yoga time, before sleep .. ). However, you still managed to group them.

6. Boosting algorithms :

You are desperately looking for a truthful and reliable fitness program. You check all the different set of videos on Youtube looking for your dearly virtual instructor.

You would first trust a moderately experienced coach with showable results and satisfied clients. You would then do more research and find out that his program is filled with reviewable instructions by one of his peers. You take notice and go on with more experienced specialists that provide the necessary adjustments. Equipped with your sheer sense of criticism, you are confronted with a continuous accumulation of recommendations that you progressively take into account with different proportions, leading you to the most appropriate combination of recommendations.

7. Neural Networks:

Say you are raising a toddler in your family. He slowly but surely works his way through letters, words and objects. His learning process follows a certain pattern. 

He sees something in his surroundings, a table for instance. In his head, it is just an object he recognizes using a few descriptors. He sees another object, say a chair. He is not that capable of telling the difference. His thinking process is more of a blackbox. 

His parents bring along the missing piece of the puzzle. They help him designate each object by its name. They keep on naming him objects according to this approach with more chairs and tables along the way. 

Growing up with enough data in his mind, the toddler knows, when in front of a four-legged furniture with a rough or smooth texture and other characteristics, how to synthesize all the perception he has of the object and refer to it by its name. 

Neural Networks function in a way that resembles in many aspects how a toddler is introduced to the real world through perception.

8. Decision Trees :

If you have ever played online with Akinator the incredible diviner, you already know what decision trees are all about. Let us reproduce a live game scenario of me versus Akinator. I will be thinking of a fictional character and he will try to guess it. Since I am huge fan of The Office, I will be thinking of Michael Scott. So here come the questions :

Akinator > Is your character a real person?

No.

Akinator > Does your character speak Japanese?

No.

Akinator > Is your character from a game?

No.

Akinator > Is your character a male?

Yes.

Akinator > Is your character a YouTuber?

No.

Akinator > Is your character older than 35 years old?

Yes.

Akinator > Is your character from a movie?

No.

Akinator > Is your character European?

No.

Akinator > Is your character from an horror game?

No.

Akinator > Does your character live in America?

Yes.

Akinator > Does your character have a son?

No.

Akinator > Does your character kill the mother of a hero?

No.

Akinator > Is your character associated with a Superhero?

No.

Akinator > Is your character from a cartoon?

No.

Akinator > Is your character from a TV series? ( Akinator starts to encircle my guess ) ..

Yes.

Akinator > Is your character afraid of sunlight?

No.

Akinator > Does your character have a brother?

No.

Akinator > Is your character bad?

No.

Akinator > Is your character a policeman or a detective?

No.

Akinator > Is your character linked to comedy?

Yes.

Akinator > Has your character appeared in the TV show 'Friends'?(Very close !)

No.

Akinator > Does your character have white skin?

Yes.

Akinator > Is your character from Happy Monster Band?

No.

Akinator > Does your character work in an office?( Almost !!)

Yes !

Akinator > I think of Michael Scott !

Nailed it !!

When Akinator asked if my character was a real person at first, it eliminated a whole spectrum of celebrities, politicians, musicians, actors so he would not be dealing with them. Same thing when he asked if my character was a girl so he would know if he should erase all female characters. Decision trees are no different than a series of asked questions in order to rid of meaningless possibilities and lead to optimal results with ideally as few questions as possible.

Conclusion :

I hope the examples shown above can assist you into smoothly introducing your business peers to your realm of competence. The examples remain purely subjective and evidently far away from being the best out there. I will not hesitate to share additional content unless my limited imagination fails me again :) .