A Novel Hybrid XAI Solution for Autonomous Vehicles: Real-Time Interpretability Through LIME–SHAP Integration

. 2024 Oct 22;24(21):6776. doi: 10.3390/s24216776

Algorithm 2 Hybrid LIME–SHAP Explanation Generation

Require: Decision

d \in D

, Model M, Feature set

F = {f_{1}, f_{2}, \dots, f_{m}}

Ensure: Hybrid feature importance

ϕ_{i}^{Hybrid}

for all

f_{i} \in F

1:
Input: Decision d, Model M, Feature set F
2:
Output: Hybrid feature importance $ϕ_{i}^{Hybrid}$
3:
Step 1: LIME Approximation
4:
for each feature $f_{i} \in F$ do
5:
Compute LIME importance ${\hat{f}}_{i}^{LIME} = LIME (f_{i}, d)$
6:
end for
7:
Step 2: SHAP Refinement
8:
for each feature $f_{i} \in F$ do
9:
if $f_{i} \in Top - k$ features based on ${\hat{f}}_{i}^{LIME}$ then
10:
Compute SHAP value $ϕ_{i}^{SHAP}$
11:
Compute hybrid importance $ϕ_{i}^{Hybrid} = α \cdot {\hat{f}}_{i}^{LIME} + (1 - α) \cdot ϕ_{i}^{SHAP}$
12:
else
13:
Set $ϕ_{i}^{Hybrid} = {\hat{f}}_{i}^{LIME}$
14:
end if
15:
end for
16:
Step 3: Multi-Model Aggregation
17:
if multiple models $M_{1}, M_{2}, \dots, M_{N}$ are used then
18:
for each feature $f_{i} \in F$ do
19:
Compute aggregated importance $ϕ_{i}^{Hybrid - Combined} = \frac{1}{N} \sum_{j = 1}^{N} ϕ_{i, j}^{Hybrid}$
20:
end for
21:
end if
22:
Return: $ϕ_{i}^{Hybrid}$ or $ϕ_{i}^{Hybrid - Combined}$