#MachineLearning lesson of the day: Working with a #gradientboosting model, I got no traction cross-validating hyperparameters like tree depth and number; but different evaluation metrics (e.g. SMSE vs. MAE etc.) had a major impact. Have you tried this?
IMHO #DNN get all the press because they do sexy human jobs like seeing and processing language. But in the business world of tabular data, #gradientboosting is where the real revolution is happening! #xgboost #catboost #lightgbm #DataScience
@carl24k Agree tree are often just ad useful and easier ti train. Actually, I don't see why crossvalidating on tree number would lead to interesting patterns. More is better, so it's tradeoff of performance and time. Depth, you could overfit, but many trees hell make it a bit robust to that.
@carl24k completely agree. We tend to work under the mantra, that it is better to know one algorithm really well, rather than bumbling around with 20 different algorithms. Our go-to is lightgbm - stable with very little need to hyperparameter optimisation.
Leaves us the freedom to focus on feature engineering, where the real performance impact is.
@philip_jarnhus totally agree that feature engineering is where the real action is! I wrote a #DataScience book on customer #Churn but the book is like half feature engineering! Check it out and if you are interested I can comp you an ebook copy… https://www.fightchurnwithdata.com/
@carl24k Thank you! I will have a look at it 