The governance of recommender systems (RSs) in very large online platforms (VLOPs) is expected to change significantly under the Digital Services Act (DSA), which imposes specific obligations on algorithmic transparency and user control. However, beyond legal compliance, a critical question remains: How can recommender systems be redesigned to genuinely empower users and foster meaningful personalization? This paper addresses this question by analyzing how three major short-video platforms — Instagram, TikTok, and YouTube — have implemented the DSA requirements for RSs. By reviewing their audit reports, systemic risk assessments, and compliance strategies, we evaluate the extent to which current approaches enhance user autonomy and control over content exposure. Building on this analysis, we outline a perspective for the future of VLOPs’ RSs grounded in speculative design. We argue that meaningful personalization should integrate algorithmic choice, balancing proportionality and granularity in RS customization, and content curation, ensuring diversity and authoritativeness to mitigate systemic risks. Subsequently, we test this approach through a small-scale user study run on a purpose-built short-video recommendation interface based on YouTube Shorts. By bridging legal analysis, platform governance, and user-centred design, this paper outlines actionable pathways for aligning technical developments with regulatory objectives for transparent and controllable recommender systems.
Cross-domain recommendation (CDR) leverages source domain knowledge to alleviate data sparsity in target domains. While existing methods have achieved significant success by utilizing overlapping users as bridges for knowledge transfer, they inadvertently introduce a critical fairness issue we term CDR Overlapping Bias : recommendation quality significantly improves for overlapping users but often stagnates or degrades for non-overlapping users. Furthermore, we observe that this disparity exacerbates as the ratio of overlapping users decreases, a common scenario in real-world applications. To address this inequity, we propose a novel model-agnostic framework named S parsity- C alibrated V irtual U ser G eneration (SC-VUG). SC-VUG synthesizes virtual source profiles for non-overlapping target users, effectively transforming them into virtual overlapping users to enable equitable knowledge transfer. Specifically, we design a Sparsity-Calibrated Generator that employs a dual-view attention mechanism to model virtual users as probabilistic distributions, where the variance explicitly captures generation uncertainty related to the overlapping ratio. To ensure the reliability of these synthesized profiles, we introduce a Structure-Preserving Limiter incorporating Anchor Loss and Isometry Loss, which aligns generated users with the source distribution while preserving the topological structure of user preferences. Comprehensive experiments conducted on three public datasets with five CDR baselines demonstrate that our method effectively mitigates the CDR non-overlapping user bias, while maintaining comparable overall accuracy in most settings. Our code is publicly available at https://github.com/WeixinChen98/VUG/tree/sc-vug.
The expanding online job market and the impact of recommender systems highlight the need for practical job recommendations. While diversity is crucial, job recommendations uniquely require high-quality diversity to accelerate career exploration rather than just maximizing quantitative metrics. Existing systems often fall short, as high diversity scores do not always align with user preferences, resulting in user dissatisfaction. To address this, we propose J ob R ecommendation for C areer E xploration ( JR4CE ), which uses user-job interaction data, explicit preferences, and current user information, aiming to achieve high-quality diversity. JR4CE has three main modules: (1) Knowledge Graph Learning Module represents explicit preferences and current information as graphs and learns latent representations of users and jobs. (2) Collaborative Filtering Module uses interaction data to refine the latent representations from (1) and predicts the likelihood of a target user’s job application. (3) Diversity Data Augmentation Module performs data augmentation using the interaction data of role model users for a target user to enhance diversity, thereby aiming to support career exploration. Based on offline metrics and qualitative case studies, our experiments using datasets constructed from an actual job search website show that JR4CE outperforms several state-of-the-arts in accuracy and the accuracy-diversity balance. Specifically, JR4CE effectively works for users in the early stages of career exploration activities.
User-generated reviews contain rich semantics that can reveal users’ fine-grained preferences beyond what interaction data alone can capture. However, existing review-based recommender systems often fail to achieve semantic alignment between review content and user–item interactions, as they either treat reviews as coarse textual signals or rely on sentiment-oriented heuristics. Achieving such fine-grained alignment is challenging due to the inherent complexity and context dependence of review texts, as well as the semantic gap between explicit review information and implicit interaction preferences. To bridge this gap, we propose LLM-ASAR , a L arge L anguage M odel-driven framework for A spect-based S emantic A lignment in Review-based R ecommendation that explicitly aligns aspect-level review semantics with user–item interaction patterns. Specifically, LLM-ASAR leverages LLMs with Chain-of-Thought (CoT) prompting to perform sentence-level reasoning on user reviews, extracting aspect-specific semantics that are structurally integrated into aspect-specific interaction graphs. A multi-aspect contrastive alignment mechanism is further introduced to align user and item embeddings with corresponding aspect-level review features, ensuring consistency across modalities. Extensive experiments on three real-world datasets demonstrate that LLM-ASAR not only achieves significant accuracy improvements over state-of-the-art baselines but also yields interpretable insights into user preferences through its aspect-aware design. Our data and code are available at https://github.com/HuilinChenJN/LLM-ASAR.
Explainability in recommender systems (RS) remains a pivotal challenge. Counterfactual explanations have emerged as a particularly actionable paradigm, offering intuitive “what-if” reasoning. However, their evaluation lacks principled standards. Current metrics primarily assess whether explanations change the top-ranked recommendation, overlooking two fundamental aspects of explanation quality. First, evaluation results are frequently inconsistent , as metrics are tightly coupled to the underlying recommender’s performance. Second, explanations are rarely assessed for compactness —whether changes are sufficiently small to remain interpretable. Oversized counterfactuals may technically succeed but fail to provide practical insight. In this work, we advocate for a holistic evaluation perspective centered on consistency and compactness . We systematically analyze how extending evaluation beyond top-1 to top-k recommendations improves metric stability and reduces dependence on recommender fluctuations. In parallel, we introduce compactness-aware evaluation criteria that quantify the minimality of counterfactual modifications. Through extensive experiments across multiple datasets and models, we demonstrate that jointly considering these dimensions yields more reliable assessments. Our findings expose key factors driving metric instability, highlight the trade-off between effectiveness and explanation size, and provide practical guidelines toward standardized, faithful, and compact evaluation of counterfactual explanations in recommender systems.
