Rating Driven Explanations (RDE)

• Rating Driven Explanations (RDE) are paradigms that use ratings as the organizing variable to generate, condition, or evaluate explanations in model predictions.
• RDE integrates deep prediction with interpretability by coupling numerical scores with feature- or time-based rationales in applications like credit scoring, recommendation, and forecasting.
• Diverse methodologies under RDE, from joint rating-review generation to causal attribution metrics, highlight trade-offs between alignment, fluency, and model robustness.

Searching arXiv for the cited papers to ground the article in the current literature. Rating Driven Explanations (RDE) denotes a family of explanation paradigms in which a rating, a rating-derived score, or a rating summary is not treated as an isolated output, but as the organizing variable for explanation. In enterprise credit rating, RDE couples deep prediction with per-feature and per-time rationales (Guo et al., 2021). In recommendation, it appears both as joint rating-and-review generation and as explicit alignment of generated explanations with predicted ratings through mutual information (Xie et al., 2022, Zhao et al., 2024). In automatic evaluation by LLMs, REC generates a rating, then an explanation, then optional citations (Hsu et al., 2024). In time-series forecasting, RDE combines local and global XAI with ordinal ratings derived from causal metrics such as ATE and WRS (Widener et al., 28 Aug 2025). Related work on collaborative explanations studies which summary statistics of rating distributions most affect human choice (Coba et al., 2018). The acronym is also used in a distinct sense for rate–distortion explanation masks in image classification (Kolek et al., 2021).

1. Conceptual scope and terminological range

The literature does not use RDE in a single uniform way. In some papers, the rating is the prediction to be explained; in others, it is a score that conditions the explanation generator; in still others, it is a human-facing summary of prior opinions, or an ordinal assessment of model reliability. Taken together, these usages suggest that RDE is best understood as a design pattern in which explanatory content is explicitly tied to evaluative output, rather than as one fixed architecture or objective.

Context	Rating object	Explanation form
Enterprise credit rating (Guo et al., 2021)	Credit score or rating	Feature-crossing, feature-attention, and temporal-attention rationales
Recommendation generation (Xie et al., 2022)	Predicted numerical rating $\hat r_{u,i}$	Synthetic review $\hat e_{u,i}$
Recommendation alignment (Zhao et al., 2024)	Predicted rating $R$	Explanation optimized for high $I(R;E)$
LLM autoevaluation (Hsu et al., 2024)	Binary or multi-point score $r$	Explanation $e$ and citations $c$ conditioned on $r$
Forecasting reliability (Widener et al., 28 Aug 2025)	Ordinal rating from ATE/WRS	Local/global XAI plus causal stability assessment
Collaborative explanations (Coba et al., 2018)	Aggregate rating summaries	Mean, count, variance, skewness, origin of ratings
Image explanation (Kolek et al., 2021)	Rate in rate–distortion	Sparse mask minimizing expected distortion

A common thread is that explanation is constrained by, generated from, or evaluated against a scalar or ordinal judgment. A plausible implication is that RDE research occupies an intermediate position between prediction, interpretability, and decision support: it does not merely expose model internals, but structures them around the form in which a system is consumed.

2. Formal definitions and objective functions

Several mathematically distinct formulations appear under the RDE label. In joint review-rating recommendation, the model predicts both a rating and a natural-language explanation:

$$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$

Training minimizes a weighted sum of rating-prediction loss and review-generation loss:

$$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$

$\hat e_{u,i}$0

$\hat e_{u,i}$1

Here the explanation is intended to function as the rationale for the rating (Xie et al., 2022).

In REC, the rating is produced first and the explanation is explicitly conditioned on it. For input $\hat e_{u,i}$2 formed by concatenating task prompt and generated response, the chain factorization is

$\hat e_{u,i}$3

with inference proceeding through

$\hat e_{u,i}$4

The effect of this rating-driven factorization is that the explanation network receives $\hat e_{u,i}$5 as an explicit input token, biasing generation toward a confirmatory style for high or “Yes” ratings and a diagnostic style for low or “No” ratings (Hsu et al., 2024).

In forecasting, RDE is formalized as a tuple

$\hat e_{u,i}$6

with $\hat e_{u,i}$7 the absolute error, $\hat e_{u,i}$8 the treatment attribute, $\hat e_{u,i}$9 the protected or confounding attribute, and $R$0 an ordinal score derived from causal metrics. Two central quantities are

$R$1

and

$R$2

Models are ranked by $R$3 and $R$4, or by a composite

$R$5

This formulation makes the rating itself an explanation-oriented reliability score rather than the primary forecast target (Widener et al., 28 Aug 2025).

In explainable enterprise credit rating, the local explanatory object is an importance matrix

$R$6

where feature-attention $R$7, temporal-attention $R$8, and rating probabilities $R$9 are combined so that the top-$I(R;E)$0 entries identify the $I(R;E)$1 pairs that drive a rating decision (Guo et al., 2021).

A separate recommendation line defines rating–explanation alignment through mutual information:

$I(R;E)$2

Because explicit densities are unavailable, the paper uses the Donsker–Varadhan lower bound with a neural statistics network $I(R;E)$3 and optimizes a MINE objective before reinforcement-learning fine-tuning of the explanation generator (Zhao et al., 2024).

3. Architectural realizations

The enterprise credit-rating realization is the Deep Feature Crossing Network, or DeepCross. Numerical and categorical fields are embedded into a common $I(R;E)$4-dimensional space; higher-order feature interactions are then generated by stacked feature-crossing modules that form Cartesian products between first-order and $I(R;E)$5-order features, score candidates with key–query attention, apply a residual combination with $I(R;E)$6, and perform PCA-style selection via $I(R;E)$7-constrained $I(R;E)$8 convolutions. Multi-order features are concatenated, reweighted by feature attention, reweighted again by temporal attention over an $I(R;E)$9-width neighborhood, and finally consumed by a GRU whose last hidden state is mapped to credit logits. Training uses the generalized cross-entropy $r$0 together with $r$1 penalties on the PCA weights (Guo et al., 2021).

Recommendation-oriented RDE systems typically use a two-head architecture with shared user and item embeddings. The rating head predicts

$r$2

while the explanation head generates text either through LSTM-based decoders such as Att2Seq and NRT, Transformer-based models such as PETER, or GPT-2 prompt tuning as in PEPLER. These systems are trained end-to-end so that review generation and rating estimation share representational structure, but the explanation is not guaranteed to be faithful merely because it is jointly trained (Xie et al., 2022).

The MMI framework preserves this backbone structure and then fine-tunes it through reinforcement learning. A MINE estimator is first trained on true $r$3 pairs and randomly paired ratings and explanations. The generator is then viewed as a stochastic policy over full explanations, with reward composed of an MI term for rating alignment, a KL-regularization term that penalizes drift from the pretrained backbone, and an entropy term that preserves diversity:

$r$4

For multi-task backbones that also predict $r$5, the final objective mixes the RL loss with the original backbone loss to avoid destabilizing recommendation accuracy (Zhao et al., 2024).

REC implements rating-driven explanation within a general-purpose LLM evaluator. REC-12B is based on Mistral-Nemo, REC-70B on Llama-3.1-70B, and both are fine-tuned with supervised fine-tuning on the REC-Data mixture using LoRA. The multi-task objective is the sum of cross-entropies over rating, explanation, and citation tokens, with all three loss weights set to $r$6. After the rating is emitted, the model continues decoding an explanation and can then emit citations in four modes: post-fix citation, post-fix with snippet, inline citation, and inline with snippet (Hsu et al., 2024).

Forecasting RDE is organized as an explicit workflow rather than a single network. The procedure splits each series 80/20 into training and test, generates rolling forecasts, computes absolute residuals, derives local explanations by segment-based LIME and global explanations by SHAP, formulates a hypothesis through a choice of $r$7 and $r$8, computes ATE by G-computation and WRS by Student’s $r$9-tests across $e$0 groups, and then assigns ordinal ratings by ranking models on ATE and WRS (Widener et al., 28 Aug 2025).

4. Explanatory objects and user-facing semantics

RDE systems differ substantially in the object they present as an explanation. In enterprise credit rating, one explanatory layer is global and static: sparse non-zero columns in the $e$1-regularized PCA layers reveal a compressed set of meaningful higher-order feature crossings $e$2 that analysts can audit. Another layer is personalized: the importance matrix $e$3 pinpoints which feature combinations at which time points were most persuasive for a particular enterprise. Figure-based examples described in the paper include “net-profit-growth-rate in Q5” driving a positive outlook and “debt-ratio in Q3” signaling a downgrade (Guo et al., 2021).

In collaborative-explanation settings, the explanation is not generated text but a summary of prior ratings. If an item has ratings $e$4, the summary statistics are the number of ratings $e$5, mean

$e$6

variance

$e$7

skewness

$e$8

and the origin of ratings, defined as whether the summary is computed from all users or only from users similar to the current one. In this setting, RDE is a decision aid whose semantics are explicitly statistical rather than generative (Coba et al., 2018).

Recommendation-oriented language explanations aim to verbalize the rationale behind a predicted score. The synthetic review is intended to answer why the model assigned the rating, but the literature shows a distinction between fluency and faithfulness. The MMI framework makes this distinction operational by rewarding explanations that are more aligned with the predicted rating or a pre-defined feature of the recommended item, rather than merely mimicking human review style (Zhao et al., 2024).

REC broadens the explanatory object further by requiring citations in addition to prose explanation. The citation modes expose a latency–granularity trade-off: post-fix citation is fastest but coarse, whereas inline with snippet is most granular because it pins specific claims to supporting evidence, at the cost of greater output length and decoding time (Hsu et al., 2024).

Forecasting RDE explicitly separates “why” from “when” and “for whom.” Local explanations identify influential segments in a context window; global explanations provide feature-level importance ranks; ATE rates consistency across treatment levels such as series identity; WRS rates stability across temporal or subgroup partitions such as month or day-of-week. This decomposition allows a user to inspect feature attributions while also asking whether a model behaves uniformly across subpopulations or calendar periods (Widener et al., 28 Aug 2025).

5. Empirical findings across domains

In enterprise credit rating, DeepCross is reported to achieve state-of-the-art accuracy, with AUC $e$9 on CH-Stocks and $c$0 on US-Stocks, while also surfacing a compact set of higher-order feature crossings and per-sample explanatory heatmaps. Table 3 is described as showing second-order crossings such as $c$1 with large $c$2-weights, matching accounting intuition (Guo et al., 2021).

The conjoint study on collaborative explanations quantifies which rating summaries drive choice. In the overall sample, high mean rating had coefficient $c$3 (SE $c$4, $c$5), large number of ratings had $c$6 (SE $c$7, $c$8), personalized origin had $c$9 ($r$0), high variance had $r$1 ($r$2), and skewness had no significant effect. Users high on the Decision Difficulty subscale were extremely sensitive to mean $r$3, strongly avoided high variance $r$4, and were relatively insensitive to number of ratings when it traded off with mean (Coba et al., 2018).

The joint review-rating literature reports a more skeptical picture of explanatory validity. Across Amazon Movies & TV, TripAdvisor, and Yelp, standard models achieved competitive RMSE of approximately $r$5–$r$6, but AIR hovered around random, MRR-AE was only $r$7–$r$8, entailment ratios were in the single digits except for PETER$r$9 at about $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$0, and PEPLER showed a trade-off in which good AIR could coexist with worse RMSE. The paper’s qualitative examples include “great service” being replaced by “great cookies,” and positive ratings whose sentiment-flipped explanations received lower perplexity than the originals, indicating brittleness (Xie et al., 2022).

The MMI framework addresses alignment rather than raw fluency. On TripAdvisor, Att2Seq improved from NMI $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$1 to $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$2 and sentiment accuracy from $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$3 to $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$4 after MMI fine-tuning, while PETER improved from NMI $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$5 to $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$6 and from $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$7 to $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$8 on five-class sentiment accuracy. On Yelp feature alignment, ApRef2Seq improved Feature Match Ratio from $$\hat r_{u,i}\in\mathbb R,\qquad \hat e_{u,i}=(e_1,\dots,e_T).$$9 to $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$0 and $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$1 from $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$2 to $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$3. Human studies further reported higher agreement rates and higher informativeness, relevance, and satisfaction for MI-enhanced explanations. The same experiments also showed that dropping the original backbone loss in multi-task settings damaged recommendation quality, whereas a mixed objective preserved near-original MAE and RMSE while improving alignment (Zhao et al., 2024).

REC reports strong benchmark performance for rating-conditioned explanation and citation. REC-70B achieved $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$4 AutoAIS and $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$5 FActscore on ExpertQA, scored $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$6 on both rating accuracy and explanation accuracy with citation-F1 $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$7 on ABCD summarization, reached an average of $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$8 on RewardBench, obtained $$L_{\text{rating}} = \frac{1}{|\mathcal D|}\sum_{(u,i)\in\mathcal D} (\hat r_{u,i}-r_{u,i})^2,$$9 average on LLM-AggreFact, and had the lowest average cognitive-bias score, $\hat e_{u,i}$00, on CoBBLEr. The paper describes REC-70B and REC-12B as surpassing or matching state-of-the-art autoevaluator baselines on rating accuracy, explanation quality, and citation fidelity (Hsu et al., 2024).

Forecasting RDE adds a second layer of evaluation beyond point-prediction error. Feature-engineered Gradient Boosting consistently achieved the lowest MASE and sMAPE in volatile or sparse domains such as Power and Car Parts and also had the lowest ATE, but high WRS exposed pronounced seasonal or day-of-week error variability. Chronos attained the lowest sMAPE in the stable, trend-driven Finance domain and the lowest ATE there, though its WRS remained moderate. LLaMA-FT reduced errors relative to base LLaMA but remained inferior to specialized models; ARIMA performed well in short, near-stationary finance series but had high ATE in sparse domains. In the Car Parts example, GBoost had $\hat e_{u,i}$01, $\hat e_{u,i}$02, $\hat e_{u,i}$03, and $\hat e_{u,i}$04, illustrating how a model can be highly series-consistent yet temporally unstable (Widener et al., 28 Aug 2025).

6. Limitations, design tensions, and acronym ambiguity

A recurrent tension in RDE research is the gap between alignment, faithfulness, and usability. The recommendation literature shows that fluent synthetic reviews can correlate with ratings without faithfully exposing the model’s rationale. This motivates the recommendation that generated reviews be treated as probabilistic hints rather than deterministic rationales, and that adversarial and semantic coherence evaluations such as AIR, MRR-AE, and entailment accompany any claim of trustworthiness (Xie et al., 2022).

A second tension concerns optimization pathology. In MMI-based recommendation, optimizing with MI reward alone produced degenerate “reward-hacking” behavior with repetitive or unreadable outputs; adding KL regularization prevented total collapse but made outputs overly short; adding entropy restored diversity and length. Likewise, when the original MSE objective was removed from a multi-task backbone, rating quality collapsed even as alignment improved. These results indicate that rating-driven objectives often need explicit regularization against linguistic drift and predictive degradation (Zhao et al., 2024).

In REC, improved citation and explanation quality does not remove all bias or deployment concerns. The paper states that no formal fairness constraint was imposed during supervised fine-tuning, multilingual performance was untested, and latency overhead increases with more granular citation modes. Its bias mitigation strategy relies on diverse training data, exact-match checking of synthetic citation snippets, and human cross-verification rather than explicit re-weighting for demographic or topical bias (Hsu et al., 2024).

Forecasting RDE introduces its own caution: raw SHAP or LIME attributions do not reveal whether a model is stable across subgroups or time partitions, so local and global explanations should be complemented by ATE and WRS if the deployment question concerns reliability, robustness, or fairness under interventions. A plausible implication is that RDE in forecasting is less about replacing attribution methods than about wrapping them in a model-selection and audit protocol (Widener et al., 28 Aug 2025).

The acronym also remains genuinely ambiguous. In the image-explanation literature, RDE refers to the rate–distortion explanation framework rather than rating-driven explanation. There the objective is to find a sparse mask $\hat e_{u,i}$05 minimizing expected distortion under obfuscation:

$\hat e_{u,i}$06

or, in relaxed form,

$\hat e_{u,i}$07

CartoonX instantiates this in the wavelet domain, using the inverse discrete wavelet transform as $\hat e_{u,i}$08, masking wavelet coefficients rather than pixels, and exploiting the fact that natural images are approximately cartoon-like and sparse in a wavelet basis. Its reported rate–distortion curve lies strictly below Pixel RDE, Integrated Gradients, SmoothGrad, LIME, and Grad-CAM, so that for a given target rate it achieves lower distortion with fewer coefficients. This is conceptually separate from rating-conditioned explanation, but important because the same acronym appears in arXiv-adjacent interpretability work (Kolek et al., 2021).