Predictive Accuracy vs. Model Complexity

Business Context

More complex machine-learning models can capture nonlinear relationships across traffic, equipment condition, failures, service behavior, and customer patterns. Higher complexity can improve predictive accuracy but increases computing requirements, training effort, data dependency, explainability difficulty, and lifecycle management burden. Simpler models are easier to operate but may miss important patterns.

Telecommunications TRIZ Resolution

Model complexity should vary according to the value and consequence of the prediction. Simple models can manage routine or stable conditions, while more sophisticated models operate on difficult cases, exceptions, or high-value decisions. Knowledge distillation, model cascades, feature selection, and hierarchical prediction can preserve accuracy without forcing every prediction through the most complex model.

Applicable TRIZ Principles

Principle 1 – Segmentation assigns different models to different prediction conditions.

Principle 10 – Prior Action precomputes features and model artifacts before real-time prediction.

Principle 28 – Mechanics Substitution replaces computationally expensive prediction stages with more efficient analytical mechanisms where appropriate.

Expected Outcome

High predictive accuracy

Lower computing and operational burden

Faster inference

Easier model lifecycle management

Decision Indicators

Early indicators include:

Model accuracy improves only through large increases in computational cost.

Complex models are used for routine decisions that simpler models could handle.

Inference latency rises as models become more sophisticated.

Model maintenance consumes disproportionate engineering effort.

Operational teams simplify models primarily to maintain acceptable performance.