Entropy Guided Diversification and Preference Elicitation in Agentic Recommendation Systems

Abstract: Users on e-commerce platforms can be uncertain about their preferences early in their search. Queries to recommendation systems are frequently ambiguous, incomplete, or weakly specified. Agentic systems are expected to proactively reason, ask clarifying questions, and act on the user's behalf, which makes handling such ambiguity increasingly important. In existing platforms, ambiguity led to excessive interactions and question fatigue or overconfident recommendations prematurely collapsing the search space. We present an Interactive Decision Support System (IDSS) that addresses ambiguous user queries using entropy as a unifying signal. IDSS maintains a dynamically filtered candidate product set and quantifies uncertainty over item attributes using entropy. This uncertainty guides adaptive preference elicitation by selecting follow-up questions that maximize expected information gain. When preferences remain incomplete, IDSS explicitly incorporates residual uncertainty into downstream recommendations through uncertainty-aware ranking and entropy-based diversification, rather than forcing premature resolution. We evaluate IDSS using review-driven simulated users grounded in real user reviews, enabling a controlled study of diverse shopping behaviors. Our evaluation measures both interaction efficiency and recommendation quality. Results show that entropy-guided elicitation reduces unnecessary follow-up questions, while uncertainty-aware ranking and presentation yield more informative, diverse, and transparent recommendation sets under ambiguous intent. These findings demonstrate that entropy-guided reasoning provides an effective foundation for agentic recommendation systems operating under uncertainty.

• The paper introduces IDSS, a conversational recommendation system that uses candidate-set entropy to select high-value questions, guide ranking, and organize diversified results across an interaction loop.
• Entropy-guided questioning raises question novelty from 0.602 to 0.946 for short queries while preserving near-ceiling relevance, showing its strongest benefit when user intent is underspecified.
• The evaluation finds that MMR substantially increases diversity, while embedding similarity works best for short queries and coverage-risk ranking better exploits detailed preferences in long queries, though validation relies mainly on simulated users.

This paper presents an Interactive Decision Support System (IDSS) for conversational recommendation that uses Shannon entropy over the candidate set as a single uncertainty signal governing three stages of the pipeline: preference elicitation, candidate ranking, and result presentation. The work targets a well-documented failure mode in interactive recommendation: ambiguous or underspecified user queries lead either to excessive clarifying questions (question fatigue) or to premature commitment to a narrow interpretation of intent, which collapses the search space and excludes viable alternatives (2603.11399). Rather than treating elicitation, ranking, and diversification as independently optimized components, IDSS propagates residual uncertainty through the entire loop, asking informative questions when doing so is valuable and explicitly encoding remaining uncertainty into ranking and presentation when it is not.

System architecture

IDSS operates over a structured product database (cars, with a brief qualitative extension to electronics) via a turn-level loop. An LLM-based semantic parser, implemented with GPT-5 under JSON schema enforcement, converts free-form input into explicit filters $\mathcal{F}$, liked/disliked feature sets $\mathcal{P}^+/\mathcal{P}^-$, and a patience signal $s$ used to terminate questioning early when impatience is detected. Retrieval applies these filters to produce a candidate set $\mathcal{C}$.

Entropy-guided question selection. For each unspecified attribute dimension $d$, the system computes normalized Shannon entropy over the value distribution within $\mathcal{C}$, using quantile-based discretization ($k=3$ bins) for continuous attributes such as price, mileage, and year. The next question targets the dimension with maximum entropy among available dimensions, subject to a minimum threshold $\tau_H = 0.3$; if no dimension exceeds this threshold, the system proceeds directly to recommendation rather than asking low-value questions. Dimension selection is data-driven, while natural-language question generation is delegated to an LLM conditioned on distribution statistics and conversation history — a hybrid design that separates statistical decision-making from language generation.

Dual ranking strategies. After elicitation concludes (question budget exhausted, entropy below threshold, or impatience detected), candidates are ranked by one of two methods. The first combines embedding similarity against a query constructed from extracted preferences with Maximal Marginal Relevance (MMR) reranking at $\lambda = 0.85$. The second, Coverage-Risk Optimization, performs greedy set selection maximizing coverage of liked features (via thresholded cosine similarity, $\tau = 0.6$, between user feature embeddings and precomputed pros/cons phrase embeddings from item reviews, using all-mpnet-base-v2) minus a risk penalty on disliked features at $\mathcal{P}^+/\mathcal{P}^-$0. Because the coverage term is submodular, the greedy algorithm carries a $\mathcal{P}^+/\mathcal{P}^-$1 approximation guarantee; grid search showed results are robust across $\mathcal{P}^+/\mathcal{P}^-$2.

Entropy-based presentation. Final recommendations are organized into a labeled grid partitioned along the highest-normalized-entropy unspecified dimension (e.g., fuel type), making trade-offs visible and supporting preference discovery through comparison rather than additional dialogue. Edge cases include progressive filter relaxation for zero-result queries, ordered so cosmetic constraints relax before fundamental ones.

Evaluation methodology

Evaluation uses review-driven simulated users: real car reviews serve as behavioral anchors, augmented by LLM rewriting into stylistically consistent variants across makes/models, then enriched into structured personas with latent attributes (writing style, budget, openness to alternatives). A dataset of 150 personas was evaluated under two query-verbosity regimes — Short (under 10 words, representing underspecified intent) and Long (under 120 words, imposing more specific and potentially conflicting constraints). Each interaction involves the agent, a simulated user, and an LLM judge scoring follow-up questions (Relevance, Newness) and recommendations (satisfaction labels, attribute-level assessments). The protocol reports both raw metrics and confidence-filtered metrics at a deliberately permissive threshold $\mathcal{P}^+/\mathcal{P}^-$3, with majority-vote reassessment triggered when list-level judge confidence is low. This attention to judge noise is appropriate given documented reliability limits of LLM-based user simulators (2603.11399).

Results

The ablation studies yield three principal findings.

MMR dominates diversity but trades off relevance. Removing MMR collapses intra-list diversity (ILD): from 0.779 to 0.279 for Embedding Similarity under Short queries, and from 0.412 to 0.241 under Long queries, with analogous trends for Coverage-Risk. However, removing MMR often improves precision — ES Prec@9 rises from 0.744 to 0.837 under Long queries — confirming a standard relevance–diversity trade-off that the full configuration balances.

Entropy-guided questioning improves novelty where it matters most. In the Short-query setting, question Newness rises sharply from 0.602 without EntropyQ to 0.946 with it, while Relevance remains near-ceiling (0.967–1.00) in all configurations. The primary benefit of entropy-guided selection is therefore eliciting novel information, not topical alignment. Gains diminish for Long queries (Newness 0.976 vs. 0.980), consistent with fewer high-uncertainty dimensions remaining after detailed upfront specification. Notably, removing EntropyQ also degrades downstream recommendation quality even under Long queries (CR Prec@9 drops from 0.801 to 0.753), suggesting clarification value persists beyond what users state initially.

Ranking methods are complementary across regimes. ES performs best under Short queries (Prec@9 = 0.903), where semantic similarity provides a robust inductive bias despite sparse constraints; CR performs best under Long queries (Prec@9 = 0.801), where its explicit constraint-coverage objective exploits richer preference signals. Long queries are uniformly harder than Short ones (e.g., ES Full Prec@9 of 0.903 vs. 0.744), which the authors attribute to more specific and conflicting constraints rather than information poverty.

A small pilot survey ($\mathcal{P}^+/\mathcal{P}^-$4) complements the simulation: participants ranked the two-follow-up policy best in 9/12 cases, the zero-follow-up policy never ranked best and worst in 10/12, with qualitative feedback highlighting that diversified grid presentation made trade-offs easier to compare.

Limitations and open questions

The evaluation rests entirely on review-driven LLM-simulated users and LLM judges. The authors acknowledge this may not capture real-user variability or strategic behavior, and the pilot human survey is too small ($\mathcal{P}^+/\mathcal{P}^-$5) to support strong claims about perceived transparency or satisfaction. Domain generality is asserted only qualitatively: the electronics experiment is described as showing "qualitatively consistent behavior" without quantitative results, and adapting IDSS elsewhere requires re-specifying attributes and diversification objectives. Several design choices are empirical rather than principled — the entropy threshold $\mathcal{P}^+/\mathcal{P}^-$6, MMR's $\mathcal{P}^+/\mathcal{P}^-$7, and the discretization bin count — and no sensitivity analysis is reported for the first two. The paper also leaves open whether entropy over attribute marginals is the right uncertainty measure compared to joint or conditional formulations, and how the relevance–diversity balance should adapt dynamically as user intent evolves during interaction.

Conclusion

IDSS demonstrates that a single information-theoretic quantity — normalized entropy over the feasible candidate set — can coherently drive question selection, uncertainty-aware ranking, and diversified presentation in an agentic recommendation pipeline. The empirical evidence supports the central claim that entropy-guided elicitation substantially improves question novelty under sparse queries while improving final recommendation quality, and that MMR-based diversification is the dominant factor in intra-list diversity at a modest precision cost. The strongest caveat is methodological: conclusions derive from simulated users and judges, and validation with real users at scale remains the key open question before these design principles can be considered established for production agentic recommender systems.