The machine learning workflow is the end-to-end sequence for defining a problem, building and evaluating a model, deploying it, and maintaining it over time. section 6.0 section 20.1.5
It starts with problem framing and data collection, then moves through preprocessing, baseline modeling, validation, and tuning, before shipping and monitoring the system in production. section 6.0 section 20.1.5
Typical phases are:
- Define the task, choose success measures, and collect or understand the dataset. section 6.0 section 20.1.5
- Develop the model by preprocessing data, selecting an evaluation protocol, beating a baseline, and improving generalization. section 6.0 section 20.1.5
- Deploy the model, present it to stakeholders, and monitor performance in the wild. section 6.0 section 6.3.4
- Keep collecting new data and annotating difficult samples to train the next generation. section 6.3.4 section 20.1.5
Example workflow domains
The workflow is introduced through tasks such as movie review sentiment, news classification, house-price prediction, spam detection, recommendation, fraud detection, click-through prediction, anomaly detection, and satellite-image search. section 4.0 section 6.0
A common emphasis is that model development is only one step, and often not the hardest one; framing the problem and collecting, annotating, and cleaning data can be harder. section 6.2 section 20.1.5
