The question posed by this Reddit user, retrain from scratch or fine-tune daily, is exactly the kind of practical, decision-level thinking that separates a working prototype from a production system. Our take is plain: incremental fine-tuning is the smarter path for this use case, but only if you structure it to combat drift, not convenience.
The user's proposed sampling strategy is already a strong foundation. Taking 100% of the last 30 days, 50% from days 30 to 90, and 10% from days 90 to 180 shows they understand that user behavior in e-commerce is seasonal, trend-sensitive, and often short-lived. A full retrain from scratch on this weighted window would discard nothing useful and reintroduce computational overhead. Fine-tuning, by contrast, lets the model adjust to subtle shifts in clickstream patterns without forgetting the broader behavioral structures it has already learned. For XGBoost models predicting intent and price sensitivity, this matters. A customer who browsed winter coats in October may behave differently in December, but their underlying sensitivity to discounts probably hasn't changed overnight. Fine-tuning preserves that stability.
Where we see room for improvement is in the monitoring layer. The user mentions daily data ingestion but does not describe how they detect when the model's predictions start degrading. Fine-tuning on stale or noisy data can reinforce bad patterns. A simple performance dashboard tracking prediction confidence against actual conversions, updated weekly, would flag when a retrain from scratch is actually needed. That is the concrete safeguard. Without it, the user risks optimizing for recency while losing the signal from older, still-valid behavior.
As for resources, the user should look into online learning frameworks and incremental learning libraries. Scikit-learn's partial_fit methods, Vowpal Wabbit, or even a lightweight streaming pipeline with Apache Kafka and River (formerly Creme) would provide patterns for this exact scenario. The academic literature on concept drift detection, specifically the drift detection method (DDM) and adaptive windowing (ADWIN), is also directly applicable. These are not theoretical exercises; they are production tools that answer "when to retrain" better than any calendar schedule.
The user has already done the hard part: defining the problem and building a working pipeline. The next step is not a bigger model or more data. It is a repeatable decision rule for when to fine-tune and when to reset. That is the difference between a system that learns and one that just accumulates.