I keep hearing experts say that "data is king" and that feature engineering is the most important part of the pipeline. Is it true that a simple model with great features will outperform a complex neural network with raw data? What are some "must-know" feature engineering tricks for tabular data?
3 answers
For most e-commerce sites, "Batch Learning" is the standard. You collect data throughout the week and retrain your model on a schedule (e.g., every Sunday night). It’s simpler to monitor and less prone to "catastrophic forgetting." "Online Learning" (or Incremental Learning) updates the model with every new data point. It’s great for things like stock market prediction or real-time bidding, but it's much harder to maintain. If a stream of "bad" data comes in, an online model can degrade in minutes. I’d recommend starting with Batch until you absolutely need the real-time speed.
Deborah, for a Batch system, how do you decide the right "retraining frequency"? Is once a week enough, or does it depend on "Data Drift"?
Online learning is cool in theory, but the infrastructure cost for the "pipelines" is usually way higher than just doing a daily batch.
Completely agree, Kelly. Unless you are a giant like Netflix or Amazon, the complexity of Online learning often outweighs the benefits.
Lawrence, it’s all about monitoring "Data Drift." You should set up a system that tracks the statistical distribution of your input features. If the distribution changes significantly—say, because of a new holiday sale or a change in user behavior—you trigger a retrain immediately. Don't just stick to a calendar; let the data tell you when the model is getting "stale." This is a core part of MLOps that ensures your recommendations stay relevant to what users are doing right now.