Home Artificial Intelligence Data Drift Vs Concept Drift in Machine learning

Data Drift Vs Concept Drift in Machine learning

5
Data Drift Vs Concept Drift in Machine learning

. Why does model decay occur? Why does the model which was doing good for previous few days/months starts behaving in another way? Lets attempt to deep dive and understand the explanations for this model decay .

The perpetrator is none other but data itself. As we all know Data is king in machine world ,it may possibly make or break your models. is one among major reasons for failures in production.

On this dynamic world,. Machine learning models get affected by this alteration. We are able to define

Lets understand reasons for this model drift:

Data drift occurs when the information a model is trained on changes. The . Because the model trained on old data grow to be useless and preforms badly on latest data. Data drift, feature drift, population, or covariate shift mean the identical thing.

Lets understand it with example. Consider a ML model was trained to predict the likelihood whether customer will buy product based on his income. If distribution of income changes, then model is not going to perform accurately in future.

Data drift can occur for quite a few reasons: (for instance, columns is perhaps added/deleted above in data pipeline), or the even when the structure/schema hasn’t (for instance, whether a salary is taken into account “above average” may change over time).

Concept drift occurs every time the . Consider example of bank card fraud detection. The way in which people use bank cards has modified over time and thus the common characteristics of bank card fraud have also modified. So when “Chip and Pin” technology got here fraudulent transactions began to maneuver more online in comparison with offline.

Concept Drift will be further divided into 4 categories-Sudden ,Gradual,Incremental, Reoccuring

Essentially the most direct strategy to discover model deterioration is to , and assess that performance with the identical evaluation metrics used during model training. Continuous evaluation provides a trigger for when to. So, does that mean we must always retrain model as soon as performance starts to dip? No, It depends. Retraining will be expensive. We should always consider trade off between what amount of degradation of performance is suitable in relation to this cost. We are able to . For instance trigger model retraining when model accuracy falls below 95%. We should always also that’s received during serving and .

Pleased learning!

5 COMMENTS

LEAVE A REPLY

Please enter your comment!
Please enter your name here