Relevance-driven ranking on e-commerce category pages often produces repetitive recommendation lists that constrain exploration and concentrate exposure on a narrow slice of the catalog. Personalized diversification mechanisms can enable user-dependent variety by tuning diversity pressure to signals of whether a user behaves like a generalist or a specialist, or by surfacing novelty through exploration-driven contextual bandits, as instantiated in this eyewear recommendation study.
Diversity in recommendation is tightly coupled to how users browse, compare, and revise preferences, especially in shopping environments where exploration is part of the decision process. If ranked lists repeatedly show near-duplicates, the system may inadvertently narrow what a user learns about the catalog, reduce perceived novelty, and reinforce a feedback loop in which the same item types become even more dominant.
At the same time, diversity is not uniformly desirable. Some users behave like specialists, repeatedly engaging with a tight set of product characteristics; others behave like generalists, sampling broadly across styles and attributes. A single, global diversification strategy can therefore be misaligned: it may distract specialists with unwanted variety, while still not giving generalists enough breadth to support discovery. The challenge is to treat diversity as a personalized objective, thus producing a recommendation list that injects the right amount of variety for a given user.
In a study, in cooperation with Lalit Kishore Vyas, and published in the Proceedings of ACM UMAP 2025, we introduce an industrial case study at Lenskart that targets a specific gap: production systems can deliver strong relevance while still offering diversity that is either limited or not personalized, for instance, when injecting the same globally popular items for everyone. We focus on mechanisms that retain relevance while adapting diversification pressure to individual users’ observed breadth of interest.
High-level solution overview
We explore two complementary ways to achieve personalized diversity on top of a relevance model.
First, we treat diversity as a post-processing problem: start from a strong relevance scorer, generate a candidate list, and then re-rank it to increase variety. Crucially, we do not apply the same diversification strength to everyone; instead, we modulate it using a user-level signal that estimates whether the user behaves more like a generalist or a specialist.
Second, we consider an in-processing alternative: a contextual bandit that selects recommendations by balancing exploitation (what seems best) and exploration (what is uncertain). In this view, diversity can emerge naturally from exploration, rather than being injected explicitly after the fact.
Our solution
A user-level signal for “diversity willingness”
The core personalization lever is a Generalist–Specialist signal. Conceptually, we ask: Are this user’s historical interactions tightly clustered around a narrow set of items, or spread across different regions of the catalog? We operationalize this idea by using the latent representation learned by the relevance model and measuring how concentrated the user’s interacted items are in that representation space.
This allows us to convert “diversity preference” from an abstract notion into a usable control signal. Instead of assuming all users want the same amount of novelty, we infer a coarse but actionable characterization of their browsing breadth, which then determines how aggressively we should diversify their ranked lists.
Post-processing diversity as a controlled trade-off
Given a relevance-ranked candidate set, we apply diversity-aware re-ranking in two different conceptual ways.
One approach emphasizes marginal novelty: as we build the recommendation list, we prefer items that are both predicted to be relevant and not overly redundant with items already selected. The intent is to reduce near-duplicates early in the list and ensure the set covers multiple “angles” of the catalog that remain compatible with user taste.
A second approach emphasizes set-level coverage: rather than only controlling redundancy item-by-item, we select a set that jointly balances item quality and mutual dissimilarity, aiming for a list that spreads attention across different item characteristics.
The Generalist–Specialist signal controls the strength of this trade-off. For generalist-like users, we allow stronger diversification so the final list spans more distinct styles. For specialist-like users, we keep diversification milder, preserving a tighter focus while still reducing monotony.
In-processing diversity via exploration in a contextual bandit
We also study a contextual bandit approach (LinUCB) as an alternative route to personalized diversity. Here the key mechanism is uncertainty-aware choice: the model does not only pick items with high predicted reward, but also items where there is meaningful uncertainty, because reducing that uncertainty is valuable for future decisions.
This matters for diversity because uncertainty is often higher for items outside a user’s dominant historical pattern. As a result, exploration can surface items with different attributes, widening the displayed set without imposing an explicit “diversity constraint” during re-ranking. In the paper’s qualitative example, Figure 3 (page 5) visually contrasts how a relevance-only approach can concentrate on similar-looking frames, while the bandit-driven recommendations span a wider mix of styles and attributes, illustrating how exploration pressure can reshape what appears in the top results.
Findings and insights
Across offline experiments on Lenskart interaction data from a short time window on a single category page, we see a consistent pattern: methods that explicitly or implicitly encourage variety increase the breadth of recommended product attributes, but the way diversity is introduced strongly affects the relevance–diversity balance.
The production-inspired baseline that mixes relevance ranking with a small number of globally popular items achieves strong relevance behavior, but its diversity is not personalized: the same popular inserts are shown to everyone, so the additional variety does not adapt to user-specific browsing breadth. In contrast, diversity-aware re-ranking increases the variety of recommended frame characteristics by actively discouraging redundancy within the list. When we personalize the re-ranking strength using the Generalist–Specialist signal, we retain most of the diversity gains while reducing the extent of the relevance sacrifice compared with a one-size-fits-all diversification setting.
The contextual bandit approach produces the strongest qualitative shift toward variety across multiple product facets. The example in the following figure highlights how diversity manifests: rather than simply adding “different” items somewhere in the list, the bandit-style ranking can surface a more varied mix among top-ranked items, consistent with the idea that exploration reshapes early exposure.
An important methodological insight is about evaluation realism. Because the bandit approach is designed to adapt online through interaction, offline tests can only approximate its advantage; historical data reflects what users were shown under past policies, and both exposure and position effects can mask the benefits of a changed ranking strategy. This makes the case that, for diversity mechanisms that alter exploration dynamics, online validation is not just desirable but structurally important for trustworthy conclusions.
Conclusions
This work contributes a practical framing of personalized diversity for e-commerce category pages: rather than choosing between “relevance” and “diversity” globally, we use a user-level signal to decide how much diversity to inject, and we evaluate both post-processing (re-ranking) and in-processing (contextual bandit) mechanisms in a real industrial setting.
Several extensions emerge naturally from the study. One direction is to move beyond binary generalist/specialist segmentation and treat the generalist–specialist signal as continuous, enabling smoother control over diversification strength. Another direction is to validate these approaches in online experiments that can capture behavioral responses, including how users adapt when early-ranked items become more varied. Finally, the contrast between re-ranking and exploration-based learning suggests a hybrid opportunity: combining explicit redundancy control with uncertainty-aware exploration to achieve diversity that is both intentional and adaptive, while remaining aligned with relevance and user satisfaction.