3 ways to improve your Machine Learning results without more data

That’s a matter of domain knowledge, insight into your data, and experience with ML.Find the best hyperparamteters you can and you might just boost your model’s performance by a few percentage points!Scikit learn has a nice hyper-parameter search module.Ensemble MethodsEnsembling is the ML technique of combining the predictions of multiple models at once..The idea is that the combined knowledge of these models will give a more accurate final result than the knowledge of any single one of them..It’s a very intuitive technique used very commonly in Kaggle competitions.To build an ensemble, simply train multiple different ML models on the same data for the same task..At inference time, you will apply all of the models to your input individually..If your task is classification, you can combine the results using a simple per class voting scheme or take the prediction with the highest confidence..For regression, just average out the results..Ensembles are an extensively field-tested and real-world-proven technique to boost prediction accuracy..Give them a try!Scikit learn has a nice ensembling module.Feature EngineeringFeature engineering involves the careful selection and possible manipulation of your data’s features..The purpose of this is to feed your model only the most optimal form of input..If you can consistently give your model only the parts of the data it needs to make accurate predictions, then it doesn’t have to deal with any extra noise that comes from the rest of the data.If you apply Principal Component Analysis and find that one of your features have very low correlation with the output, then you probably don’t need to be processing it..Some features are going to be intuitively not useful, such as the ID or perhaps recording date..Or maybe you only want certain features to be considered in the first place.To give your model the best your data has to offer, do some data exploration to find out what information and features are actually needed for predictions.. More details

Leave a Reply