'Vowpal Wabbit negative weights significance
I'm doing a feature study and I was wondering what the negative feature weights in the audit output signify. I'm currently using the contextual bandits function with the following parameters:
from vw.cb_explore_adf import CB_EXP_ADF
cb = CB_EXP_ADF('--cb_adf -q AA --audit --epsilon=0 --lambda=10', quite=True, enable_logging=True)
Here's an example of the audit output:
Incoming^category1=xqc115:61140:1:[email protected]
ASpace^category2=Mali*ASpace^category1=Mali:143860:1:[email protected]
ASpace^category3=Mali:213594:1:[email protected]
Solution 1:[1]
Negative feature weights should not signify something special. Depending on the provided arguments and input data, vowpal wabbit will try to optimize the underlying problem (cb_adf reduces to a cost-sensitive one-against-all problem which reduces to a GD optimization problem)
Sources
This article follows the attribution requirements of Stack Overflow and is licensed under CC BY-SA 3.0.
Source: Stack Overflow
| Solution | Source |
|---|---|
| Solution 1 | olgavrou |