Recommendation-as-Experience: A framework for context-sensitive adaptation in conversational recommender systems

TL;DR

The paper tackles the gap in conversational recommender systems where emphasis on ranking accuracy neglects nuanced, context-sensitive interaction behaviors. It introduces the Recommendation-as-Experience (RAE) framework, which treats educative, explorative, and affective aims as adaptive state variables that are modulated by domain context, item value, and user traits, including autonomy preferences. Through a multi-domain vignette study with $N=168$, the authors demonstrate robust domain and value effects and show that prior CRS experience is a strong global predictor of aim importance, while demographics have more limited, domain-specific effects. RAE provides an architecture-agnostic state–policy mapping that can drive adaptive policies via heuristic rules, learned controllers, or LLM-based orchestration, enabling context-sensitive dialogue that balances experiential quality with predictive performance. The work offers practical design guidance and safeguards for responsible adaptation, highlighting that dynamic, signal-driven adaptation can improve user alignment and satisfaction in CRS across diverse domains.

Abstract

While Conversational Recommender Systems (CRS) have matured technically, they frequently lack principled methods for encoding latent experiential aims as adaptive state variables. Consequently, contemporary architectures often prioritise ranking accuracy at the expense of nuanced, context-sensitive interaction behaviours. This paper addresses this gap through a comprehensive multi-domain study ($N = 168$) that quantifies the joint prioritisation of three critical interaction aims: educative (to inform and justify), explorative (to diversify and inspire), and affective (to align emotionally and socially). Utilising Bayesian hierarchical ordinal regression, we establish domain profiles and perceived item value as systematic modulators of these priorities. Furthermore, we identify stable user-level preferences for autonomy that persist across distinct interactional goals, suggesting that agency is a fundamental requirement of the conversational experience. Drawing on these empirical foundations, we formalise the Recommendation-as-Experience (RAE) adaptation framework. RAE systematically encodes contextual and individual signals into structured state representations, mapping them to experience-aligned dialogue policies realised through retrieval diversification, heuristic logic, or Large Language Model based controllable generation. As an architecture-agnostic blueprint, RAE facilitates the design of context-sensitive CRS that effectively balance experiential quality with predictive performance.

Figures (7)

• Figure 1: Stacked-bar distribution (%) of user ratings for educative, explorative, and affective interaction aims, comparing low-value and high-value item scenarios. Bars on the left represent low-value items; bars on the right represent high-value items. Colour gradients from red (rating = 1) to dark blue (rating = 5) indicate increasing perceived importance.
• Figure 2: The three-stage methodological workflow for operationalising the Recommendation-as-Experience (RAE) framework. Stage 1 elicits user preferences through high-stakes vignettes; Stage 2 utilises Bayesian modeling to derive robust empirical weights and effect sizes ($r$); and Stage 3 translates these weights into a formal computational policy $\pi$ and state vector $\mathbf{s}_t$ for adaptive system behaviour. Dashed boxes represent the specific sub-components validated at each phase of the pipeline.
• Figure A.1: Descriptive distribution of ratings across domains. Medians (markers), modes (labels), and interquartile ranges (error bars) illustrate central tendency and variability for educative, explorative, and affective interaction. These statistics contextualise the domain-level effects reported in Section \ref{['sec:results']}.
• Figure B.1: Posterior predictive check for the Educative model. Predicted distributions closely align with observed responses across all ordinal categories, indicating robust fit.
• Figure B.2: Posterior predictive check for the Explorative model. Predicted and observed distributions exhibit strong agreement, supporting model adequacy.