Describe-Predict-Explain Framework

• Describe-Predict-Explain Framework is a tripartite approach that distinctly defines data description, outcome prediction, and causal explanation.
• It integrates methods such as SHAP, Grad-CAM, and surrogate modeling to provide clear insights and actionable interpretations of model behavior.
• The modular pipeline supports transparent decision-making in diverse domains, including healthcare, process monitoring, and deepfake detection.

The Describe-Predict-Explain framework is a tripartite schema for separating three distinct objectives in modeling and decision support: describing patterns in data and model behavior, predicting outcomes for new instances, and explaining outcomes in a causal or mechanistic sense. In the methodological formulation associated with Shmueli, description, prediction, and explanation are distinct scientific tasks, and conflation between explanatory and predictive modeling is common; in the healthcare formulation of Carriero et al., description is association-based, prediction is oriented toward out-of-sample performance and calibration, and explanation is reserved for causal claims and “what if” questions (Shmueli, 2011, Carriero et al., 7 Aug 2025). In domain-specific systems such as predictive process monitoring and deepfake detection, the same triad is operationalized as a pipeline in which a predictive model is accompanied by descriptive representations and explanatory interfaces for human stakeholders (Mehdiyev et al., 2020, Tariq et al., 11 Aug 2025).

1. Conceptual structure and scientific objectives

The framework distinguishes three aims that are often treated as if they were interchangeable. In the healthcare exposition, Describe is defined as summarizing and characterizing patterns in the available data and in the fitted model’s mapping from inputs to outputs; Predict is defined as building models that yield accurate and clinically useful individual-level risk predictions for diagnosis or prognosis; and Explain is defined as understanding causal mechanisms and answering “what if” questions for actionable interventions (Carriero et al., 7 Aug 2025). Shmueli’s methodological analysis places this distinction in a broader statistical context: explanation is aligned with causation, theory, retrospective testing, and bias reduction, whereas prediction is aligned with association, data-driven learning, prospective generalization, and the bias–variance trade-off (Shmueli, 2011).

Component	Aim	Typical forms in the cited literature
Describe	Summarize and characterize patterns	Summary statistics, correlations, SHAP, LIME, Grad-CAM, event-log featurization
Predict	Yield accurate predictions for new cases	Deep neural networks, risk scores, CLIP/XceptionNet classifiers
Explain	Support causal or decision-justifying understanding	Causal estimands, local surrogate rules, narrative refinement

These distinctions can be written formally. For prediction, the healthcare formulation uses empirical risk minimization,

$$\min_{\hat f \in \mathcal{F}} \sum_{i=1}^n L\big(y_i,\hat f(x_i)\big),$$

with performance judged by discrimination and calibration (Carriero et al., 7 Aug 2025). For explanation, it uses causal estimands such as

$$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$

A central implication is that a model can be highly predictive without identifying causes, and a descriptive attribution method can be faithful to associations while remaining non-causal (Carriero et al., 7 Aug 2025).

2. The “Describe” phase as representation of evidence

In predictive process monitoring, the Describe phase begins with process-aware information systems that record event logs comprising sequences of executed activities with timestamps and, when available, actor, resource, and case attributes. A process instance is represented as a trace

$\tau = \langle e_1, e_2, \ldots, e_n \rangle,$

and featurization maps each trace into

$x = \phi(\tau) \in \mathbb{R}^d,$

where $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$ concatenates n-gram vectors, temporal scalars, structural or case statistics, and one-hot categorical vectors (Mehdiyev et al., 2020). The descriptive function here is not causal interpretation but domain-grounded representation: experts can inspect transitions such as “Accepted → In.Progress,” durations such as $t_k - t_1$, and case attributes such as SR.Latest.Impact ∈ {Low, Medium, High}.

In DF-P2E, the Describe phase is visual and semantic rather than tabular. Grad-CAM is used to expose the spatial evidence supporting a classifier’s decision by computing gradient-derived weights

$$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$

and the class-specific saliency map

$$L^c_{Grad\text{-}CAM} = ReLU\Big(\sum_k \alpha_k^c A^k\Big).$$

The resulting heatmap $A^{cam}$ highlights facial regions such as the eyes, nose, mouth, and jawline that most influenced the “fake vs. real” score (Tariq et al., 11 Aug 2025). DF-P2E then translates saliency into semantically grounded text with a captioning function

$\hat{c} = g_\phi(x, A^{cam}),$

fine-tuning BLIP, BLIP2, GIT, OFA, ViT-GPT2, and PaliGemma on MSCOCO plus a custom dataset of Grad-CAM overlays paired with region-level human annotations such as “warped eye edges,” “blurred cheek texture,” and “irregular mouth geometry” (Tariq et al., 11 Aug 2025).

Across these instantiations, description is best understood as making the model’s evidential substrate inspectable. This suggests a common role for the Describe phase: it establishes what the model “sees,” whether in the form of event-log features or saliency-grounded image regions, without yet asserting either predictive adequacy or causal explanation.

3. The “Predict” phase as optimization and generalization

The Predict phase centers on a learned mapping from inputs to outputs. In predictive process monitoring, the classifier is a fully connected deep neural network with rectifier activations and dropout regularization. Hidden layers compute

$$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$0

with dropout rates $$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$1 on input and $$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$2 on hidden layers; the output is a logistic unit

$$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$3

Training minimizes binary cross-entropy,

$$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$4

using SGD with a lock-free parallelization scheme, ADADELTA, and early stopping on AUROC with relative tolerance $$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$5 and patience $$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$6 (Mehdiyev et al., 2020). The reported classifier achieved an Area Under the ROC Curve of $$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$7, and a threshold $$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$8 can be chosen by minimizing the balanced error rate

$$\mathbb{E}[Y \mid do(X=x)], \quad \Delta = \mathbb{E}[Y \mid do(X=1)] - \mathbb{E}[Y \mid do(X=0)].$$9

with the reported use case giving $\tau = \langle e_1, e_2, \ldots, e_n \rangle,$0 (Mehdiyev et al., 2020).

In DF-P2E, the predictive backbone is a binary deepfake detector

$\tau = \langle e_1, e_2, \ldots, e_n \rangle,$1

instantiated with XceptionNet, CLIP-base, and CLIP-large and trained on DF40 using

$\tau = \langle e_1, e_2, \ldots, e_n \rangle,$2

Frame-level AUC on DF40 is reported as approximately $\tau = \langle e_1, e_2, \ldots, e_n \rangle,$3 for XceptionNet, $\tau = \langle e_1, e_2, \ldots, e_n \rangle,$4 for CLIP-base, and $\tau = \langle e_1, e_2, \ldots, e_n \rangle,$5 for CLIP-large, with CLIP-large attaining subset scores of $\tau = \langle e_1, e_2, \ldots, e_n \rangle,$6 on CelebDF (FS), $\tau = \langle e_1, e_2, \ldots, e_n \rangle,$7 on CelebDF (FR), $\tau = \langle e_1, e_2, \ldots, e_n \rangle,$8 on CelebDF (EFS), and $\tau = \langle e_1, e_2, \ldots, e_n \rangle,$9 on DeepFaceLab (Tariq et al., 11 Aug 2025). Because it achieved the best average AUC across subsets, CLIP-large is used as the backbone for saliency generation downstream.

The predictive phase therefore provides the numerical substrate on which the rest of the framework depends. Shmueli’s analysis is directly relevant here: prediction is evaluated prospectively, by its performance on new observations, and a model with strong explanatory fit need not have strong predictive performance, while a biased or underspecified model can sometimes predict better if it reduces variance and therefore lowers expected prediction error (Shmueli, 2011).

4. The “Explain” phase: local justification, narrative refinement, and causal limits

The Explain phase is the most contested part of the framework because different literatures use “explain” differently. In predictive process monitoring, explanation is local and post hoc. The deep network’s last hidden layer defines a latent code

$x = \phi(\tau) \in \mathbb{R}^d,$0

and neighborhoods are formed in latent space using Euclidean distance

$x = \phi(\tau) \in \mathbb{R}^d,$1

K-means partitions the validation codes into clusters, a new instance is assigned by

$x = \phi(\tau) \in \mathbb{R}^d,$2

and, for each cluster, a surrogate decision tree $x = \phi(\tau) \in \mathbb{R}^d,$3 is trained on original features to regress the DNN’s scores (Mehdiyev et al., 2020). Cluster validity is quantified by the ratio $x = \phi(\tau) \in \mathbb{R}^d,$4, interpreted as the proportion of total variance explained by clustering, with the reported value approximately $x = \phi(\tau) \in \mathbb{R}^d,$5, and local fidelity is measured by

$x = \phi(\tau) \in \mathbb{R}^d,$6

The explanation bundle includes the global model score, the surrogate score, the cluster index, local fidelity, the rule path, and rule confidence (Mehdiyev et al., 2020).

In DF-P2E, explanation is narrative refinement. A multimodal LLM,

$x = \phi(\tau) \in \mathbb{R}^d,$7

takes the caption, the image, the Grad-CAM heatmap, and metadata $x = \phi(\tau) \in \mathbb{R}^d,$8 encoding user persona and explanation intent, and outputs a fluent, context-aware explanation. Fine-tuning uses PEFT with QLoRA adapters and the conditional next-token objective

$x = \phi(\tau) \in \mathbb{R}^d,$9

Because the model conditions on the image, heatmap, caption, and metadata, the narratives explicitly reference the same facial regions emphasized in $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$0 and adapt tone, detail level, and domain vocabulary to the target audience (Tariq et al., 11 Aug 2025).

Carriero et al. sharply delimit what such explanations can claim. Their central argument is that explainable AI methods applied to correlation-based prediction models are good descriptive tools, but are limited in their ability to explain why a model works in terms of true underlying biological mechanisms and cause-and-effect relations (Carriero et al., 7 Aug 2025). In this sense, latent-region rules and saliency-grounded narratives may justify or clarify a model decision, yet they do not by themselves deliver causal guarantees.

5. Integrated pipeline formulations

A full Describe-Predict-Explain pipeline is explicit in both process monitoring and deepfake detection. In the process-mining framework, the end-to-end workflow is: ingest event log $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$1 and formalize traces $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$2; engineer features $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$3; build dataset $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$4 and labels $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$5; train a deep feedforward network $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$6 with latent mapping $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$7; compute latent codes on the validation set; cluster them with k-means; train a local surrogate decision tree for each cluster; and, for a query instance, locate the latent cluster and present a rule-based explanation with confidence together with $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$8 and $x = [v_{ngram}, v_{time}, v_{case}, v_{cat}]$9 (Mehdiyev et al., 2020). The rationale given is that grounded feature construction supports expert understanding, a robust classifier establishes baseline reliability, and local surrogate rules help experts justify individual decisions and understand process behaviors leading to outcomes.

DF-P2E makes the same logic multimodal. Its pipeline is defined as: Predict, where $t_k - t_1$0 produces $t_k - t_1$1 and Grad-CAM computes $t_k - t_1$2; Describe, where $t_k - t_1$3 generates a concise manipulation-aware caption $t_k - t_1$4; and Explain, where $t_k - t_1$5 outputs a user-tailored narrative $t_k - t_1$6 (Tariq et al., 11 Aug 2025). The system presents $t_k - t_1$7 in an interactive UI so that users can inspect the heatmap, read a semantic summary, and then consume a fuller narrative in plain language. The design principles are visual grounding, semantic alignment, and narrative refinement.

The evaluation reported for DF-P2E emphasizes both predictive and explanatory outputs. BLIP2-Flan-T5-xxl achieved the strongest automatic captioning metrics, including BLEU-4 approximately $t_k - t_1$8, CIDEr approximately $t_k - t_1$9, and SPICE approximately $$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$0, but was slow; BLIP-large, with CIDEr approximately $$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$1, was chosen as the default for user-facing deployments because it offered the best balance of quality and latency (Tariq et al., 11 Aug 2025). Human evaluation on a 5-point Likert scale gave Usefulness $$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$2, Understandability $$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$3, and Explainability $$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$4, and feedback emphasized the value of layered outputs: “Seeing where the model looked and reading why it mattered helped me understand and trust the decision” (Tariq et al., 11 Aug 2025).

The case studies supplied for DF-P2E illustrate the integrated pipeline directly. In a mouth-region manipulation, CLIP-large yields $$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$5 for fake, Grad-CAM concentrates on the lips and adjacent cheek area, BLIP-large reports “Model focuses on the mouth; slight warping and blur around the lips and lower cheek suggest manipulation,” and the LLaMA-3.2-11B-Vision narrative ties this evidence to geometric inconsistencies and local blurring typical of synthetic blending (Tariq et al., 11 Aug 2025). Additional cases describe eye-edge artifacts with $$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$6 and a real image with $$\alpha_k^c = (1/Z) \sum_i \sum_j \frac{\partial y^c}{\partial A^k_{ij}}$$7 and diffuse attention over the face.

6. Misconceptions, limitations, and domain-specific significance

A recurrent misconception is that interpretability automatically yields causal explanation. The healthcare analysis rejects this equivalence. Post-hoc attributions such as SHAP, LIME, and Grad-CAM describe how a model uses measured features or regions, but they do not convert a non-causal model into a causal one (Carriero et al., 7 Aug 2025). The paper’s kidney-disease example makes the point with three mechanisms. First, mediator dominance: hypertension can dominate attributions even when smoking is causally upstream, because once the mediator is in the model, smoking adds little predictive value. Second, collider bias: in a hospitalized subpopulation, age can appear negatively associated with kidney disease even when the true causal effect is positive. Third, confounding: insulin prescription can show the strongest association with kidney disease because of unmeasured diabetes confounding, so descriptive prominence is not causal explanation (Carriero et al., 7 Aug 2025).

Shmueli’s distinction reinforces the same caution from a statistical perspective. Explanatory power and predictive accuracy are different dimensions, and the paper recommends reporting both explanatory power and predictive accuracy rather than assuming one implies the other (Shmueli, 2011). A plausible implication is that a Describe-Predict-Explain system should be assessed on at least three fronts: whether its descriptive layer is faithful to the data and the fitted model, whether its predictive layer generalizes, and whether its explanatory layer is merely justificatory or genuinely causal.

Within those limits, the framework has clear domain-specific significance. In predictive process monitoring, it supports domain experts in validating data, trusting predictions, and justifying individual outcomes with concise, cognitively plausible rules (Mehdiyev et al., 2020). In deepfake detection, it addresses a different problem: existing detection systems have made significant progress in classification accuracy but typically function as black-box models, offering limited transparency and minimal support for human reasoning. DF-P2E responds by unifying prediction and explanation in a coherent, human-aligned pipeline for non-expert users in adversarial media environments (Tariq et al., 11 Aug 2025).

The Describe-Predict-Explain framework is therefore best understood not as a single algorithm, but as an organizing principle for separating evidential description, predictive inference, and explanatory ambition. In some implementations, especially deepfake detection and process monitoring, the three phases form a modular architecture. In the stricter methodological sense articulated in healthcare and statistical modeling, however, only the predictive component is optimized for forecasting, only the descriptive component characterizes model behavior, and only causal methods can answer explanatory “what if” questions in a fully explanatory sense (Carriero et al., 7 Aug 2025, Shmueli, 2011).