NOTE: due to a tracking error with git, I had to restart the repos from scratch, so there are no logs for this one, sorry.
This set of programs was made as a way to visually observe the process of SVR machine learning in action - to understand how the algorithm adapts to the features as more observations are added to the training data. Since visual observation was the priority here (so we can gain an intuitive understanding of what it's actually doing), I made 3 design choices:
-I used SVR rather than SVM, as discrete output labels are harder to visually observe compared to a continuous label, which could be represented as an axis (in this case the z-axis). -I only included 2 features, continuous from 0-100% (again for visual clarity), giving 3 dimensions in total, which in addition to a colourmap made it easier for human visualisation -As specified above I chose a dataset that uses continuous labels, fixed from 0-100% to prevent axis limit changes (which would add useless space to a plot and hinder the visualisation)
the plot itself The dataset takes a set of 2341 unique universities in the world, takes 3 continuous features (teaching quality, industry income, research quality) and then uses SVR to use the former 2 parameters as features to predict the latter. The entire set is used as training data to create the model, which will then use every integer combination of the 2 features (0-100%, 0-100%, xy axes) to predict the corresponding research quality (z-axis), which will create a surface plot on how the model predicts the scores. This should hopefully appeal to humans' geometrical intuition to visually understand how the model makes its predictions. I've also included a colourmap to make it even clearer.
There is no testing data; the "test" is done by the viewer - the original data points will be plotted on the same set of axes as the plot so it's easy to visually compare the differences between the predicted output and the actual output. To facilitate this, I've added some alpha transparency to the surface so points both above and below (underestimates and overestimates resp.) can be seen by the viewer.
the animation In addition, I've also created an animation that over time, iterates through the number of data points used in training the model, from n=125 to n=1232 (roughly after this point there is no significant change in the surface) so the audience can understand visually how the number of training points influences the behaviour of the model. Also, I like how aesthetically pleasing it is to watch the surface mold and bend over time.
Note: for all programs, I've increased the number of rows and columns (rcont,ccont) interpolated on the surface plot so it appears more smooth and less polygon-y.
svr_surfaceplot.py This is the program that creates the still image plot. It's what I used as a template to build the animated plot program rather than make everything from scratch.
svr_surfimgmaker.py This is the program that creates the frames for the animated plot. I tried to animate them using FuncAnimation but it was too large to efficiently process, so I made it save each frame instead and make a dedicated program to create the videos.
videomaker.py This program takes the processed frames from the above program and makes the GIF videoclips from them in sections, which can then be combined into a single video.
frames contains all the frames used in the program. clips contains all the clips used in the program. mp4s contains all the separate mp4 videos used to make the final video.
unis is the original dataframe. mainclip in both gif and mp4 form (needed gif to post on LinkedIn), is the final animated clip.
The original data can be found from this link: https://www.kaggle.com/datasets/samiatisha/world-university-rankings-2023-clean-dataset
I included this part as a second miniproject addition to this one. I took the model and again varied it from 1 training sample to all 2341 samples, and then for each case recorded the mean error - in this case, both mean absolute difference (the mean difference between the true values and the predicted values at those points) and the root mean square difference (the mean of the squares of the differences, square-rooted) and recorded them onto a txt file (since doing this takes a long-ish time).
I then took the results from the txt file and plotted them onto 2 axes, one for each type of error, to analyse where the error is highest and lowest and what number of samples leads to these highs and lows, in order to determine the best possible number of training data samples. I also plotted quintic regression curves to approximate the error curves in each case.
In the end the lowest error using a non-trivial number of datapoints was at n=2208 for both types of error, or about 94.3% of the dataset. As to why this is the case, I honestly have no evidenced idea - perhaps something to do with how the dataframe was originally ordered, as I noticed in the animation that the data points seemed to be ordered in approximately descending order.
svr_surferrorchecker.py determines the errors at each n value, errorrec.txt records the errors, and errorplotter.py performs the plotting. All are commented and explained in the source code.