#High Predictive Power
- How much does it influence the prediction? (Predictive Power)
- e.g. is marital status a good predictor of cancer? is it of the likelihood of crashing a car?
#Fast Computability
e.g. learning/inference time
- Training
- the sparser, the less information, the more difficult to train
- Inference
- trade-off between time to inference and quality of inference. The better, the longer.
#Reliability
Can we trust that we will get the value of a feature (e.g. API answer)? Can we trust that value?
#Un-correlatedness
Highly correlated features affect model performance. Small changes in input data may affect many variables and, thus, the model’s overall performance.
#Unitary, easy to understand and maintain
The feature represents a certain, simple to understand and to explain quantity. e.g. weight, length, width, and color.
Properties of a Successful Model