Skip to main content

Predicting the NBA playoffs using neural networks (sorta)

It's my first week at W&B, here's what I learned....
Created on December 9|Last edited on March 16
Note: I'm coming from a mar/ad-tech background, so bare with me

The project

I found this blog post by Riccardo Di Sippio about using ML to predict the 2020 NBA playoffs.

💡
The model was trained using season data of teams over various categories between 1990 and 2020

Team
G
MP
FG
FGA
FG%
3P
3PA
3P%
2P
2PA
2P%
FT
FTA
FT%
ORB
DRB
TRB
AST
STL
BLK
TOV
PF
PTS
Year
4
5
6
7
8
9
Run set
14

After getting the notebook set up and getting the data piped in, I experimented with tinkering with the number of layers in the model. With more Time, I'd leave these number up to sweeps
def make_dnn_model(n_input_features, n_possible_outcomes):
x_in = tf.keras.Input(shape=(n_input_features,))
x = tf.keras.layers.Dense(128, activation='relu')(x_in)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.Dense(32, activation='relu')(x)
x = tf.keras.layers.Dropout(0.2)(x)
x_out = tf.keras.layers.Dense(n_possible_outcomes, activation='softmax')(x)
return tf.keras.models.Model(x_in, x_out, name="DNN_regresson")

Run set
9



Run set
14





First Impressions of W&B

Things that are Awesome 🤩

  • Incredibly easy to get working out of the box, really great inherent metadata pulled
  • Really great docs
  • Everything is incredibly snappy (although the data was small)


Things I struggled with

  • Once I got in the app, lots and lots of features, it's almost overwhelming
List<Maybe<File<(table)>>>