Hybrid LRR-TED Framework

• Hybrid LRR-TED Framework is an explainable AI system that combines automated rule discovery with a Pareto-selected set of expert rules to balance scalability and stability.
• It employs a three-phase pipeline—automated discovery via LRR, expert augmentation via Pareto selection, and hybrid classification with TED-SVC—to optimize explanation accuracy and reduce manual effort.
• In customer churn prediction, the framework achieved 94% accuracy while halving annotation workload compared to an 8-rule manual baseline, demonstrating its practical efficiency.

Hybrid LRR-TED Framework is an explainable AI framework introduced in "Augmenting Intelligence: A Hybrid Framework for Scalable and Stable Explanations" for resolving a stated "Scalability-Stability Dilemma" in XAI (Krukrubo et al., 22 Dec 2025). It combines automated rule discovery with supervised explanation learning: automated rule learners identify broad retention patterns, while human experts contribute a small, Pareto-selected set of high-value churn-trigger rules. In the reported customer churn setting, the framework initializes the explanation matrix with automated safety rules, augments it with four human-defined risk rules, and trains a TED-SVC model on an explanation-augmented label space. The reported result is 94.00% predictive accuracy with a 50% reduction in human annotation effort relative to an 8-rule manual TED baseline (Krukrubo et al., 22 Dec 2025).

1. Problem formulation and the scalability-stability dilemma

The framework is motivated by the claim that explainable AI systems must jointly satisfy two objectives: they must provide stable explanations, and they must scale without prohibitive human annotation cost. The underlying contrast is drawn between post-hoc methods such as LIME and SHAP, which may scale easily but suffer from instability, and supervised explanation frameworks such as TED, which offer stability but require extensive human labeling effort (Krukrubo et al., 22 Dec 2025, Ribeiro et al., 2016, Lundberg et al., 2017).

The formalization uses an explainer $E(x;\theta)$ that produces an explanation for instance $x$ under parameters $\theta$. Stability is defined as

$$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$

that is, one minus the expected perturbation in explanations under small input noise $\delta$. Scalability, represented as a cost-oriented objective, is defined as

$C(\theta) = \frac{1}{1 + H(\theta)},$

where $H(\theta)$ denotes expected human annotation effort, such as manually authored rules or manual labels per instance. Higher $C$ therefore corresponds to lower human cost (Krukrubo et al., 22 Dec 2025).

The framework is situated within a multi-objective optimization view:

$\max_{\theta} [S(\theta), C(\theta)]$

or, in scalarized form,

$$\max_{\theta} \alpha \cdot S(\theta) + (1-\alpha)\cdot C(\theta), \quad \alpha \in [0,1].$$

Its specific design emphasis is a constrained version:

$x$0

for a desired stability threshold $x$1. In this formulation, the framework does not discard the stability objective in favor of efficiency; rather, it attempts to preserve supervised-explanation stability while reducing manual burden (Krukrubo et al., 22 Dec 2025).

2. Asymmetry of discovery and the rule typology

A central claim is the "Asymmetry of Discovery": automated rule learners, specifically GLRM + LRR, are reported to uncover broad "Safety Nets" but to miss fine-grained "Risk Traps" (Krukrubo et al., 22 Dec 2025). Safety Nets are retention rules, while Risk Traps are specific churn triggers. The distinction is formalized by partitioning the rule set into automated safety-net rules

$x$2

where each $x$3 has negative coefficient $x$4 in the LRR objective, and expert-defined risk-trap rules

$x$5

The paper gives rule-level examples in symbolic notation. An automated Safety Net takes the form

$x$6

while a manual Risk Trap takes the form

$x$7

This division is interpreted through the "Anna Karenina Principle of Churn," stated as: "all safe customers behave similarly (dense, homogeneous clusters), but each churner defects for a unique, specific reason (sparse, heterogeneous patterns)" (Krukrubo et al., 22 Dec 2025). Under this view, retention behavior is structurally easier for automated rule discovery because it occupies denser and more homogeneous regions of feature space, whereas churn behavior is harder because it is sparse and heterogeneous.

A common misunderstanding would be to treat the framework as claiming that automated methods are generally insufficient for explanation. The narrower claim is more specific: automated methods tend to recover $x$8, but human experts are needed to supply $x$9. This suggests that the hybridization is asymmetric by design rather than a generic manual correction layer.

3. Hybrid LRR-TED construction

The pipeline is organized into three phases: Automated Discovery via Linear Rule Regression, Domain Augmentation via a Risk Filter and Pareto Selection, and Hybrid Classification via TED-SVC (Krukrubo et al., 22 Dec 2025).

Phase 1: Automated Discovery. Original features $\theta$0 are binarized into $\theta$1 using atomic propositions such as $\theta$2 and $\theta$3. Rule weights are then estimated by solving

$\theta$4

From the resulting rule set, Safety Nets are extracted as

$\theta$5

and these automated safety rules are aggregated into three confidence tiers, Codes 1-3, by sorting according to $\theta$6.

Phase 2: Domain Augmentation. The full expert rule set $\theta$7 consists of eight candidate rules, Codes 4-11. Each candidate rule is assessed by coverage over true churn cases:

$\theta$8

Rule orthogonality is measured with the Jaccard index,

$\theta$9

A Pareto-front selection then chooses a subset $$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$0 of size $$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$1 that maximizes combined Coverage$$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$2 while minimizing $$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$3. The selected subset, denoted the "Golden Quartet," has $$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$4 rules spanning Financial, Structural, Interaction, and Engagement risks.

Phase 3: Hybrid Classification. The method constructs an explanation-augmented label space $$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$5, where $$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$6 encodes Codes 1-12 and Code 12 is the "Drift" default. A Support Vector Classifier is trained by minimizing the joint loss

$$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$7

Here, $$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$8 is initialized from the hybrid explanation matrix that combines automated safety codes with the selected risk codes. Operationally, the pipeline binarizes the data, fits the LRR objective, extracts automated Safety Nets, computes coverage and pairwise overlap for expert rules, selects the Pareto-optimal subset, forms explanation codes for each instance, and then trains the TED-SVC model on the joint target space.

4. Empirical evaluation on customer churn

The reported evaluation uses the IBM AIX360 Churn dataset with $$S(\theta) = 1 - \mathbb{E}_{\delta \sim D_\delta, x \sim D_x}\left[\|E(x;\theta) - E(x+\delta;\theta)\|_1\right],$$9 instances: 860 "Stay" and 1,140 "Churn." The feature set includes demographic variables such as Age and Tenure, billing variables such as MonthlyCharges and PaymentDelay, and usage and support-call metrics (Krukrubo et al., 22 Dec 2025).

Two evaluation measures are explicitly defined. Predictive accuracy is

$\delta$0

and the annotation effort ratio is

$\delta$1

Because manual TED uses 8 rules and the Hybrid configuration uses 4, the reported value is $\delta$2.

Configuration	Rule count	Reported result
Fully automated LRR	0	Acc = 75.15%
Manual TED	8	Acc = 92.90%
Hybrid ("Golden Quartet")	4	Acc = 94.00%

The reported human effort reduction is $\delta$3 (Krukrubo et al., 22 Dec 2025). Detailed class-level performance for the Golden 4 configuration is also reported. For the Churn class, Precision = 0.99, Recall = 0.93, and F1 = 0.96. For the Stay class, Precision = 0.92, Recall = 0.98, and F1 = 0.95. Confidence intervals were not explicitly reported in the paper.

These results support the paper’s central empirical claim: the hybrid configuration exceeds both the fully automated baseline and the full 8-rule manual expert baseline on the stated dataset while using fewer expert-authored risk rules.

5. Pareto efficiency and the changing role of the expert

The framework interprets the reduction from 8 expert rules to 4 as a Pareto-optimal trade-off. The "Golden Quartet" is described as lying on the efficiency frontier: no further reduction in rule count can improve accuracy, and no increase in rules can improve accuracy without adding redundancy (Krukrubo et al., 22 Dec 2025). Within the framework’s own vocabulary, the key optimization criterion is not merely sparse rule selection, but selection of the highest-coverage, lowest-overlap expert rules.

This design leads to a redefinition of human participation in Human-in-the-Loop AI. Instead of functioning as exhaustive "Rule Writers," domain experts are positioned as "Exception Handlers," supplying only the most strategic Risk Traps while automated learners capture broad Safety Nets. The claimed shift is therefore organizational as well as algorithmic: labor is concentrated on exceptions that automated discovery systematically misses.

A possible misconception is that the framework simply halves rule count and therefore trivially lowers cost. The more specific claim is stronger and narrower: the reduction is achieved through Pareto filtering under simultaneous coverage and overlap criteria, and the resulting 4-rule system is reported to outperform the 8-rule manual TED baseline. This suggests that expert effort is being reallocated rather than uniformly reduced.

6. Limitations and future directions

Two limitations are explicitly identified. First, under real-world churn rates below 5%, LRR may default to trivial "always-stay" rules, thereby forcing more manual intervention. This is presented as a class-imbalance failure mode rather than a generic limitation of rule learning (Krukrubo et al., 22 Dec 2025). Second, the framework inherits a linearity assumption: LRR builds axis-aligned hyper-rectangles, so complex perceptual data such as images and text may not binarize cleanly.

The future research directions are correspondingly targeted. One direction is to evaluate the Hybrid LRR-TED framework across a spectrum of churn-rate imbalances, from 1:1 to 1:100, in order to identify the "break-even" point at which automated discovery degrades. Another is to integrate causal discovery, rather than purely correlational rules, so as to distinguish genuine retention drivers from spurious safety nets (Krukrubo et al., 22 Dec 2025).

These limitations clarify the scope of the reported results. The framework is presented as a method for stable, interpretable, and more scalable explanation in a structured-data churn setting, not as a universal solution for all modalities or class regimes. A plausible implication is that its strongest use case lies where binarizable structured features coexist with sparse, domain-salient failure patterns that are difficult to recover automatically.