权重_列值归一化

Question

I use a TensorFlow canned estimator to predict game actions from situations favourizing best scores (there are no "good" or "bad" answers but an answer ib better than another depending the situation). This reality is represented by scores for each situation / action sample.Scores are included in train_data and used as weight and passed as weight column in the estimator.

1 ) I tried without weight column : the model fits about 70 % with good answers. The model is correct for main examples but is not optimal for marginal examples (about 20 % remaining...)

2 ) I tried with weight = exp(-score). Model fits about 45 %, is better dealing with marginal examples but have a very poor accuracy with main examples.

I think there is a problem with my weights scale [0.0001 - 5003.98] with 0.0001 for the best result and 5003.98 for worth. What is the good way to regularize it ??

I read the formula weight = (score - mean(score)) / std(score) could be a solution but doesn't seems work

model = tf.estimator.LinearClassifier(feature_columns=feature_columns,
                                    optimizer=tf.train.AdamOptimizer(learning_rate=0.001, beta1= 0.9,beta2=0.99, epsilon = 1e-08,use_locking=False),
                                    weight_column=weights,
                                    n_classes=10,
                                    label_vocabulary=Action_vocab,
                                    model_dir='./Models/ActionPlayerModel20/',
                                    loss_reduction=tf.losses.Reduction.SUM_OVER_BATCH_SIZE,
                                    config=tf.estimator.RunConfig().replace(save_summary_steps=10))

weight = exp(-score)

Answer 1

Solved the weight problem by doing this (scaling):

 scaledWeight = (weight-weightMin)/(weightMax-weightMin)

It did not completly resolve my problem (Better but not optimal accuracy). So, I tried both feature normalization and standardization : good accuracy with normalization but quick overfitting ; very good accuracy and slow overfitting with standardization.

Standardization code :

    tf.feature_column.numeric_column("Round",dtype=tf.dtypes.float32, normalizer_fn=lambda x: (x - roundMean) / roundStd),