In short: When a quality measure drifts, the question is always “which variable is causing it?”. Relying on one statistical method means inheriting its blind spot; combining four, and showing where they disagree, gives an answer you can defend.
Once the autocorrelation problem is fixed, there’s still a second question: which single method should you trust to rank feature influence? The honest answer, arrived at after comparing the obvious candidates, was: none of them alone. And that’s the more useful conclusion.
The candidates
For ranking which of several predictor tags most influences a response variable, the natural options were:
- Pearson correlation: fast, simple, captures linear relationships, but blind to anything nonlinear and sensitive to outliers.
- Spearman correlation: a rank-based alternative, more robust to outliers and to monotonic but nonlinear relationships, but still univariate (it looks at one predictor at a time, ignoring interactions).
- PLS regression with VIP scoring: handles multiple correlated predictors well (common in process data, where tags often move together), and gives a variable-importance score that accounts for their joint relationship with the response.
- Ridge regression coefficients: handles multicollinearity via regularization, penalizing large coefficients, and is useful as a cross-check against PLS.
LASSO, Elastic Net and PCA were considered and set aside. LASSO’s tendency to zero out correlated predictors arbitrarily is a poor fit when several tags genuinely move together, and PCA’s components don’t map cleanly back to individual named tags, which matters when the whole point is telling an operator which tag to look at.
Why not just pick the best one
Each of the four remaining methods has a real, distinct blind spot: Pearson misses nonlinearity, Spearman ignores interactions between predictors, and PLS and Ridge both depend on model assumptions that may not perfectly fit a given process. Trusting any single one means trusting its specific blind spot every time the system runs. Combining them doesn’t erase any individual weakness, but it means the ranking isn’t hostage to one method’s failure mode.
Designing the ensemble
Rather than a fixed formula, the ensemble is a configurable, weighted combination of the four normalized scores. Weighting is configurable because different processes may warrant leaning more heavily on the multicollinearity-aware methods (PLS and Ridge) versus the simpler univariate ones, depending on how correlated the tag set actually is.
Normalizing before combining
Pearson and Spearman scores naturally fall in [-1, 1]. PLS VIP scores and Ridge coefficients live on entirely different scales. Combining them requires normalizing each method’s output onto a common scale first. Otherwise the method with the largest raw numbers silently dominates the ensemble regardless of its actual weight setting.
def ensemble_score(pearson, spearman, vip, ridge, weights):
scores = {
"pearson": normalize_abs(pearson),
"spearman": normalize_abs(spearman),
"vip": normalize_abs(vip),
"ridge": normalize_abs(ridge),
}
return sum(scores[k] * weights[k] for k in scores)Making it visible
A ranked bar chart showing each candidate variable’s combined influence score, with the individual method scores available underneath (a drill-down or a grouped view) rather than only the blended number, is what makes this trustworthy in practice. If the four methods agree, the ensemble score is trustworthy and the chart shows a clean consensus. If they disagree, that disagreement is itself useful information: it tells you the ranking is sensitive to how you measure influence, which is exactly the kind of nuance a single blended number would hide. The goal isn’t a prettier chart; it’s not letting the ensemble’s smoothing hide a real disagreement.
The broader lesson
Ensembling isn’t just a machine-learning competition trick. It’s a way to hedge against any one method’s blind spots in a live production setting, where there’s no leaderboard to tell you when a single model has quietly started giving bad answers. Combined with the autocorrelation safeguards from Part 5, this ensemble is what turns “which tag is driving this” from a guess into a defensible, if appropriately humble, answer.
This is Part 6 of a 7-part series on building a real-time process monitoring system. Part 7 covers the final and arguably most important layer: making all of this legible to someone standing on the plant floor.