Project 4 – Matt Sloan¶

Introduce the problem¶

Introduce the project. What is the problem you are trying to solve? What questions are you trying to find answer

https://www.kaggle.com/datasets/cisautomotiveapi/large-car-dataset/download?datasetVersionNumber=4 (warning: large file)
from https://www.kaggle.com/datasets/cisautomotiveapi/large-car-dataset

Create ipynb template for website -darkmode html command, set low jpg, ignore warnings, default imports numpy pandas traintest¶

I’m looking to see if there are specific specific features that tend to make vehicles more fuel efficient. We all know that having a smaller engine inherently means higher mpg’s… or does it? What other factors tend to be at play here? Does one manufacturer tend to specialize in this? Bonus… what cars tend to have the best price/mpg ratio?

What is clustering and how does it work?¶

With supervised machine learning, you always have a target that you’re looking for and you will have an output from your particular algorithm to compare to.
Clustering is unsupervised machine learning. In other words, you will NOT have a target to compare to. It does not predict anything. Instead, clustering is useful for breaking down your data samples into groups.
It might be helpful in some cases to categorize based on overall similarities.

In order to explain what the k-means algorithm does, imagine that one could plot one data-point vs another. If these were numerical values, you could plot all of these on the same plane. If we were to look at phone battery life vs phone price, we could end up with a plot that has several obvious groups, or clusters, of phones plotted. K-means takes a number, k, to determine how many clusters to look for. Then it tries to find a combination of k number of phones in k groups, with each group only containing the phones that were closest to it. This is a simple example, but most datasets have more than 2 variables!

Introduce the data¶

Where did you find it? What is the data about (include links)? What are the features (with additional explanations if not already self-explanatory from the name itself)?

I was looking for mpg data for vehicles that was also paired with msrp. I decided to use the dataset from fueleconomy.gov for mpg values and this… 5gb dataset from kaggle. for vehicle info.
It appears to be a sample of AutoDealerData.com of vehicle listings from Illinois from June 2018 until 2020.
I will probably remove duplicates in an attempt to save my RAM from annihilation.
This dataset has 156 features and ~5.7 million entries. The features that I am most interested in are:

• msrp
• brandName
• modelName
• bodyClass
• DisplacementCC
• Doors
• DriveType
• EngineCylcinders
• EngineManufacturer
• MakeID
• Manufacturer
• Model
• ModelYear
• Turbo
• TransissionSpeeds
• TransmissionStyle

However I will be utilizing most of the remaining features as well.¶

And from the fueleconomy.gov dataset, I will be using:

• comb08 – combined MPG from fueltype 1 (I will be assuming the primary fuel type is normally used)
• make – manufacturer (division)
• mfrCode – 3-character manufacturer code
• model – model name (carline)
• year – model year
• displ – engine displacement in liters

I intend to use these variables to accurately append the combined mpg to the vehicle

Data Understanding/Visualization¶

Use methods to try to further understand and visualize the data. Make sure to remember your initial problems/questions when completing this step. While exploring, does anything else stand out to you (perhaps any surprising insights?)

How does this step relate to your modeling?

Let us start out by loading the file.

I had some issues trying to read it in chunks, so I chose which columns were the most important and loaded only those. Most columns had a large quantity of missing values anyways 🙁

Nulls?

Collinearity?

Pre-processing the data¶

What pre-processing steps do you follow? Explain why you do each pre-processing step.

First I want to remove colms with a large number of missing values. Let’s remove columns missing 10% or more.

Now we’ll rid of rows with missing values

Now since our goal is to classify vehicle features that are associated with higher mpg values, I will bin mpg into sets of 5 with higher values being higher mpg values

Finally, before we merge our df’s let us remove some duplicate rows.

Let’s check it out 🙂

Also our mpg df has high collinearity between cylinders and displacement… surprise! (90%!) So let’s see if we can reduce some of that collinearity by replacing cylinders with displacement per cyclinder – dpc

70% is better I suppose…

Time to add our mpg values to our main df.

1. make sure brand name/manufacterur and model name are same case so we can match them
2. rename our mpg_df columns so that we can inner join

Modeling (Clustering)¶

What model(s) do you use to try to solve your problem? Why do you choose those model(s)?

I will use random forests to predict mpg of a vechile. I believe that engine displacement will have high multicollinearity so I will try it both with and without.
As for clustering: I will be using k-means clustering to add another feature to the dataset in an attempt to increase the accuracy of the random forest model. My goal with this is to see which feature sets are consistently associated with higher mpg values.

Let us get a baseline model!

1. we’ll dummy variable every ‘object’ field
2. standardize our data
3. train test split 80/20
4. rf with n_estimators = 30
5. compute our baseline score

Let’s add our dummies

Let’s standardize our data now.

82% isn’t bad! Let’s try some linear regression.

Let’s see if we can improve this with clustering now…
Let’s start with finding the optimal number of clusters

There is an elbow around 30-35. After trying k = 35 and k = number of mpg bins
I have decided to use half the number of bins so that it is easier to visualize.

We are seeing how mpg vs year is clustered

Here we are looking at mpg vs displacement/cyclinder is clustered

Time to add a variable for the clusters back into x and then try our rf again!

Well, slightly worse… So let’s optimize this Random Forest using grid search!

I suppose the best RandomForest accuracy is destined to be ~82%.

Storytelling (Clustering Analysis)¶

I was not able to use clustering in a meaningful way here, not yet. I was hoping that the features of the vehicles could be grouped and therefore give another datapoint that would ‘group’ vehicles together.

Impact Section¶

Discuss the possible impact of your project. This can be socially, ethically, etc. It cannot be something like "our project has no impact" or "our project has no negative impact." Even the most well-intentioned projects could have a negative impact. We will not be checking for "right" or "wrong" answers, but showing your critical thinking.

One positive outcome I was hoping for would be the ability to look at a vehicle and pinpoint a few features of that vehicle to make a decision on whether or not it would be worth buying. For example, if you were at a dealership and saw a car you liked, you could compare a list of it’s features, like engine manufacturer, body stile and displacement per cylindern and then have a good feeling on whether or not the high or low miles per gallon rating is worth it.

One negative possibility could have been that manufacturers knew exactly which features tend to lead to higher or lower miles per gallon and then raise the price and hurt the average consumer.

References¶

Include any references.

• https://www.kaggle.com/datasets/cisautomotiveapi/large-car-dataset
• https://fueleconomy.gov/feg/download.shtml

Code¶

Make sure to include your code somehow as well. You can do this in various ways. For example, you can include code snippets throughout your writing and/or include a link to your Jupyter notebook file.

https://github.com/matthewasloan95/p4-clustering