https://www.kaggle.com/c/telstra-recruiting-network

• Prepare confusion matrix

• 1 Hidden layer with 10 units
• Learning rate: 0.0003
• Softmax in output layer
• Log features are normalized by biggest value in all records.
• 37 hours of training.
• Looks like over fitting. On training set the score is 0.5484

• Log features are counted as boolean value (Column volume is not used)
• Still over fits. Training stopped after score on training set: 0.557506
• Training time: 6 minutes

Testing different combination of features with cross validation score

• With all feature score: 0.6287 (train: 0.5297)
• Log features only, score: 0.6497
• Without 'severity_type', score: 0.6287 (train: 0.5363)
• Without 'resource_type', score: 0.6430 (train: 0.5213)

• Number of hidden layers (1 and 2 tested)
• Type of activation function (ReLu learn faster and over fits. sigmoid learn slower)
• Hidden units from 3 to 100 (5-10 looks like optimal value)
• Use of PCA to reduce dimension.

• ReLu units performs better then sigmoid
• Log features without volume column
• PCA reduces features from 454 to 296 (99% variance)
• L2 Regularization

• Switching to logistic units
• Adding second hidden layer. Doesn't improve score
• More then 5 hidden units (in single layer) over fits network
• Using PCA makes over fitting bigger problem