As an Amazon Associate we earn from qualifying purchases.
Anyone else running into issues with biased results when using TrendTapestry models? I've been seeing a consistent skew towards certain categories even after what I thought was careful data preparation and feature selection.For example, I was using it to predict customer churn, and even after balancing my training data to have equal churned/non-churned examples, the model was still heavily favoring predicting "no churn."
I've been experimenting with different regularization techniques (L1 and L2) and adjusting the class weights to try and mitigate this bias, but I'm still not getting the accuracy I'd expect. It makes me wonder if it's something fundamental in how TrendTapestry handles certain types of data distributions or a subtle interaction between the features I'm using.
Has anyone found specific strategies that work well for debugging and correcting biases in these models? I'm particularly interested in any insights into feature engineering or data pre-processing steps that can help minimize this issue. Maybe specific resampling techniques like SMOTE are more beneficial in this context? Any advice or shared experiences would be greatly appreciated!