WeightliftingApp ML Model

A series of programs written in Python using the Rotoworld mobile API to train a Keras deep learning model for the purpose of evaluating NBA players based on tokenized news articles.

NBA player news article training input example:

Here is a snippet of the original article:

"Hassan Whiteside scores 17 points in loss: He was invisible in the first half..."

In the first step, we clean the article, replacing names and numbers with generic tokens to avoid confusing the algorithm. We do not worry about punctuation or capitalization as it is automatically removed during tokenization.

"NAME scores points in loss: He was invisible in the first half..."

During training, we look at a large batch of sample articles (1000+) and use them to determine the most used words. By trimming this list down to the top 1000 tokens, we have a better chance of assigning meaning to these words while avoiding extraneous words that we may not have enough data on. These words, for example invisible, are replaced with 0 and ignored by the training system. Finally, in order to feed our model the same sized data every time, we pad the sequences to a length of 100, adding 0 if the article is too short or trimming the article if it is longer than 100 words.

name	scores	points	in	loss	he	was	invisible	in	the	first	half	...
1	33	20	3	178	7	22	0	3	2	65	224	...

Categorical expected output example:

"+" becomes [ 0 0 1 0 ] :

+ (positive)	- (negative)	0 (neutral)	! (injury)
0	0	1	0

      ||||||||
      ||||||||
      ||||||||
  Processed by model
    \\\\\\//////
      \\\\////
        \\//

Categorical actual output example:

[ .067 .126 .807 .000 ] becomes "+" :

+ (positive)	- (negative)	0 (neutral)	! (injury)
.067	.126	.807	.000

Current accuracy (500 training points @ 1,000 epochs, categorical crossentropy):

• Training data: 100.0% 👍
• Test data: 58.0% 😞

Current model