Overview
f-RecX proves that explanation interface design — not just content — determines whether users actually see and understand recommender explanations.
One in three explanations in commercial recommender systems goes entirely unread. On Netflix, only 5% of users correctly identified the full textual explanation. f-RecX maps four input dimensions — explanation style, goals, domain dynamics, and recommender technique — to a single output: the visual parameters that make explanations consumable.
"Even a perfectly faithful, personalized explanation will fail the user if it is set in a 10-pixel font with low contrast and buried below the recommendation list."— Thesis Chapter 8, f-RecX
The Problem
Textual explanations are the dominant mode of explanation in commercial recommender systems. Despite growing sophistication in generation, a persistent weakness in presentation remains largely unaddressed. In an empathy workshop with 19 non-expert users, only 66% correctly identified the explanation in familiar interfaces — and the Netflix explanation was fully recognized by just 5%.
Existing explainability frameworks focus on algorithmic faithfulness, not visual presentation. There is no principled guidance connecting the type of explanation, the goals behind it, and the domain risk to a concrete set of visual design decisions. Designers and engineers are left without evidence-backed rules for font size, contrast, position, or word count.
The Framework
f-RecX maps four input dimensions — characterizing the style, goal, domain dynamics, and algorithmic foundation of a textual explanation — to one output dimension specifying the visual parameters used to present it. Its central concept is consumability: the ease with which users can locate, understand, engage with, and derive value from a textual explanation.
Input Dimension 1
How the explanation is framed — case-based, collaborative-based, content-based, demographic-based, knowledge-based, or utility-based. Style should align with the recommendation domain: case-based explanations perform best in e-commerce; content-based in entertainment.
Input Dimension 2
The objective behind explaining — transparency, trust, scrutability, effectiveness, efficiency, satisfaction, and persuasiveness, extended with simplicity and timeliness. No single goal configuration is optimal across all domains.
Input Dimension 3
Four contextual factors on a low-to-high scale: impact, risk, urgency, and familiarity. In high-risk, high-impact domains, explanations should be more detailed and prominently positioned. In low-risk, familiar domains, shorter explanations may suffice.
Input Dimension 4
The underlying algorithmic approach — content-based filtering, collaborative filtering, or context-aware systems. Using the wrong explanation style for the technique risks producing misleading explanations that undermine user trust.
Output Dimension
The visual parameters for presenting the explanation — derived from an online study systematically varying font size, font weight, contrast, position, number of recommendation items, and explanation length. Consumability and aesthetic preference are measured separately, as they do not always align.
Design Guidelines
70% rated 32px as most consumable; 24px was most aesthetically preferred (36%). Both exceed the WCAG minimum of 16px. Target 24px when explanation visibility is a priority.
Bold weight was rated most consumable (34%), with Medium close behind (33%). Heavier weights improve findability when explanations must compete visually with item thumbnails and ratings.
White-on-black and black-on-white produced the highest consumability ratings, consistent with WCAG standards. Black text on dark background was strongly preferred aesthetically.
Top-left placement, directly above the recommendation list, was both the most consumable (54%) and most preferred (56%). This aligns with F-pattern reading behavior.
Approximately 20 words was the most preferred explanation length. No participant favored a full paragraph. Users value sufficient context but resist explanations that require sustained reading effort.
A shorter recommendation list focuses attention and makes each explanation more prominent. Longer lists reduce the relative salience of any individual explanation, making it easier to skip.
Visual Study
The study distinguished objective usability from subjective preference — a distinction that proved relevant across several characteristics. What users find most readable is not always what they find most beautiful, and f-RecX captures both.
f-RecX: A Framework for Designing Effective Textual Explanations in Recommender Systems' User Interfaces
IJHCS 2025f-RecX is a conceptual framework that maps four input dimensions — explanation style, explanation goals, domain dynamics, and recommender technique — to a single output dimension specifying the visual presentation parameters that determine whether users actually read and understand textual explanations in recommender system interfaces. Grounded in an empathy workshop (N = 19) and an online study systematically varying six visual characteristics across a realistic movie recommender interface, the framework introduces consumability as its central construct and provides evidence-backed design guidelines for font size, weight, contrast, position, word count, and the number of displayed recommendation items. f-RecX serves as the characterization framework for subsequent work in ScrollyPOI and HUMMUS.
← Back to Research