Beyond Additive Decompositions: Interpretability Through Separability

The paper's core claim is that additive methods don't just lose fidelity on interacted features, they actively produce wrong explanations through signal cancellation. TSL avoids this by learning multiplicative structure directly from first-order partials rather than marginalizing interactions away.

This sits in a broader pattern visible across recent work: systems that adapt to real-world complexity rather than forcing data into convenient assumptions. The on-device learning survey from late May exposed how static models fail under distribution shift. TSL makes a similar argument about static decomposition assumptions, proposing instead that practitioners should let the model reveal its actual interaction structure. The financial RAG work also illustrates this principle: frozen components with adaptive layers outperform monolithic retraining. Here, TSL keeps the greedy stagewise framework but makes the orthogonal refitting step adaptive to feature relationships.

If practitioners report that TSL reconstructions match original model predictions on held-out interaction-heavy subsets (e.g., feature pairs with high mutual information) while GAM/SHAP explanations diverge, that confirms the signal cancellation claim. If adoption stalls because the rank-1 tensor constraint proves too restrictive on real datasets, that signals the paper solved a theoretical problem that doesn't match practice.