Outputs from models (trained on different data) as inputs to another model?

Question

Let's say I want to detect new species of fish.

I have several models, each trained to recognize a different characteristic, e.g., the speed of known fish, the size of known fish, their known shapes, etc. Naturally, each model will be trained on a distinct set of data: dataset_1 used to train model_speed, dataset_2 used to train model_size, dataset_3 used to train model_shape.

• Is it reasonable to combine the outputs from each of those models and use that combined output as an input to another model, such that together, my 'final' model can better understand how likely it is that it is observing a new type of fish?
• What is the name for such an approach, if it exists?
• How else would I tackle such a problem?

Answer 1

The characteristics you mention are typically called features. My interpretation of your description is that each of the models performs feature engineering. You'll still need to put all the features into a single model to predict species of fish. Predicting species of fish is a multi-class classification problem, assign a single categorical label to each data instance.

If you think there may be novel species of fish that do not appear in the training set, you can handle that by modeling the uncertainty in your model's predictions. In other words, your model will have category predictions and associated probabilities. If the largest probability is below a threshold, the system should predict that the features of the current fish are not likely associated with any species of fish that appeared in the training data set.