What is an Explanation Structure Model?
An Explanation Structure Model (ESM) is the explainable package around an XNN. It connects the data transformations, predictive model, symbolic identities, relationships and metadata needed to execute a query and organise its explanation.
Why a predictive model needs context
A trained model does not exist in isolation. Its inputs may have been encoded, transformed or decomposed. Its output may need to be grouped into business concepts. Different stakeholders may need different levels of detail. A reviewer may also need to know which dataset, model version and processing steps were involved.
An ESM preserves and organises that context. It provides a structure in which the predictive model and its explanations can be interpreted as parts of an end-to-end pipeline rather than as disconnected artefacts.
The explainable pipeline
The ESM tracks how raw inputs become model-ready features. Depending on the data, the pipeline may include categorical encoding, time decomposition, sequence processing or vector encoding. A categorical feature such as Employment Status may become several encoded parameters, but those parameters remain linked to the original feature and its possible values.
This link is important for explanation. The XNN rule may operate on an encoded value such as Employment Status[Self-employed]; the user-facing explanation can still refer to the original business concept.
The XNN inside the ESM
One or more Explainable Neural Networks (XNNs) provide the predictive component. The ESM can distinguish a trainable XNN from a serving XNN. The trainable form retains the structure and parameters required for later training. The serving form is prepared for inference and embedded or referenced within the production explanation package.
The XNN contributes its modules, partitions, rules, background value and attribution logic. The ESM adds the relationships and metadata needed to navigate those components.
Symbolic identities and the hypergraph
Features, categorical values, bins, partitions, rules and concepts receive unique identifiers. The neuro-symbolic hypergraph records the links between them.
Consider a rule that activates when Age ≤ 29.5 AND Education > 8. The graph can connect the rule to the Age and Education symbols, their relevant bins, the interaction module and the resulting attribution. It can also connect those objects to higher-level groupings used for explanation, such as demographic or qualification factors.
These relationships support filtering, ranking and rollup. They also provide lineage when a reviewer moves from a concise explanation back to the underlying model objects.
Dataset and model metadata
ESMs contain metadata including the data schema, variable statistics, categorical mappings, feature distributions, anomaly information and an optional causal model. This information provides the baseline against which new inputs and model behaviour can be assessed.
Metadata should not be confused with proof that a model is suitable for a particular use. It is evidence that supports validation, monitoring and reproducibility. Suitability still depends on the data, evaluation, controls and deployment context.
Structuring explanations for an audience
The same query can be presented differently without changing the underlying result. A case handler may need the principal factors and the next review step. A data scientist may need module attributions, rule expressions and activation data. An auditor may need the model version, timestamp, identifiers and integrity check.
The ESM supports this by selecting, grouping and summarising explanation elements according to a defined objective. This is a structured transformation of model-native information, not a licence to invent a narrative. Every presented factor should remain traceable to the source explanation objects.
How an ESM is created
The overall process within the framework has three broad stages:
- Dataset analysis and preprocessing: establish the schema, statistics, distributions, transformations and explainable pipeline.
- XNN induction and training: discover feature interactions, learn modules and partitions, and optimise the model.
- Neuro-symbolic processing: assign canonical references, construct the symbolic relationships and package the resulting model and metadata as an ESM.
This separation allows an existing induced structure to be retrained and a new serving XNN to be incorporated into a refreshed ESM when appropriate.
Why it matters
The ESM helps prevent explanation from becoming a disconnected report. It preserves the relationship between original data, transformed features, XNN behaviour, symbolic concepts and the presentation shown to a stakeholder.