Use a simple perceptron neural network to predict NBA mvps with tournament algorithm
This plot shows classifier performance over 300 epochs. The error value is calculated from the least squares method between the normalized mvp shares between 2 players. However classifying mvp share is very difficult given player stats as other factors such as narrative apply. However with the given data and features, this classifier should be able to predict which player is better. The following plot shows the performance given the task of classifying the better player.
Unfortunately this did not work. The training did not converge in accuracy. Although it seems the loss functions has converged seamlessly, it turns out that the model outputs a number close to 0.5, thus acts as the average middle value between all outputs resulting in minimal error.
After my original model failing, I was ensure if the data I was providing was faulty or was it the model implementation. To find out, I wrote an equivalent model using the library sci-kit learn to get an upper bound for the accuracy of my model. After writing my sci-kit model, I tested it on my dataset and the result was the same where it picked a value close to 0.5 to minimize error thus again resulting in an accuracy similiar to random chance. I fixed this by normalizing my sample data
# old data sample
sample1 = [player1stat1, player1stat2, ... , player1statn, player2stat2, ... player2statn]
def normalize(x, y):
return x/x+y
# new data sample
sample1 = [ normalize(player1stat1, player2stat1) , ... , normalize(player1statn, player2statn)]
Now after fixing the data to achieve competent results on the sci-kit model. I decided to test it on my model which was a Single perceptron trained with my own implementation of the gradient descent algorithm. These were the results.
After training the model, we simulate of tournament between players and our model evaluates which player is more likely to be MVP.
Here are the results
| Season | Prediction | MVP Winner |
|---|---|---|
| 2020-2021 (Unfinished) | Nikola Jokić | Nikola Jokić |
| 2019-2020 | James Harden | Giannis Antetokounmpo |
| 2018-2019 | James Harden | Giannis Antetokounmpo |
| 2017-2018 | Lebron James | James Harden |
| 2016-2017 | James Harden | Russel WestBrook |
| 2015-2016 | Stephen Curry | Stephen Curry |
| 2014-2015 | James Harden | Stephen Curry |
Interesting enough, this model seems to prefer James Harden alot, and suggest it favors box-score statistics over games. This is understandable since most of parameters are related to personal statistics rather than team wins. Interestingly enough the year James Harden won the MVP, my model predicted LeBron James instead. Although this model did not predict the MVP as accurately as I would have hoped, the players is suggested are definitely mvp debatable, for example James Harden was a perennial top MVP candidate for his box score stats, and this model clearly reflected that.