Efforts to improve diversity in recommendations typically come at a cost of accuracy. While several works have proposed methods to overcome this trade-off, they neglect the unique challenges of short-video platforms: extreme long-tail distributions and hierarchical category structures with multiple granularity levels. Meanwhile, graph neural networks have become the state-of-the-art for recommendation because of their ability to achieve collaborative filtering and learn higher-order relationships between users and items. We therefore aim to unlock the capabilities of GNNs for personalization to adapt diversity to individual user preferences and overcome the accuracy-diversity dilemma in short-video recommendation. In this work, we identify two dimensions of diversity preferences—validated by psychological theory and readily inferable from interaction data and modeled using graph neighbor sampling: user preferences for diverse content across category levels, and preferences for fine-grained diversity within broad-level interests. We integrate these dimensions directly into the graph propagation process through neighbor sampling, enabling the model to dynamically balance exploration of diverse categories with exploitation of user preferences during training. We also employ a curriculum training strategy that schedules exposure to items based on popularity-based hardness to determine how this scheduling improves the latent representations of long-tail items and mitigates the accuracy-diversity trade-off. Through extensive experiments against both standard graph and state-of-the-art graph-based diversified recommendation baselines, we show that our solution achieves superior balance between diversity and accuracy on datasets with extreme long-tail distributions while outperforming state-of-the-art methods at deeper levels of the category hierarchy.
Search query logs are invaluable for scientific research, particularly in the medical domain, yet they pose significant privacy risks through the potential exposure of Personally Identifiable Information (PII). This work addresses the dual challenge of quantifying PII vulnerability and evaluating effective anonymization strategies to prevent these vulnerabilities. We collected search query logs and PII information of 11 attributes from 19 consenting participants. Human analysts and an offline Large Language Model (LLM) attempted to identify the PII from the query logs. Our findings show that some PII values are easily identified (e.g., age group, which was identified in 67% of cases) but others, such as participant name, are much more difficult, though the risk is non-negligible (8%). Notably, human analysts exhibit significant inconsistency in the PII they were able to infer. The LLM reached identification results that were statistically indistinguishable from those of human analysts, and both were statistically significant better than classic Named-Entity Recognition methods. We tested the ability to hinder identification of PII through two techniques: Sanitization (data removal) and Obfuscation (noise addition). While the former did not provide statistically significant reductions in identification of PII, the latter did. Specifically, the Dissimilar Addition method, where queries from a user with dissimilar PII values were injected, achieved a statistically significant reduction in identification accuracy (absolute reduction of 18%) while requiring significantly fewer data modifications than the best sanitization techniques. These findings indicate that effectively anonymizing query logs requires masking the signal rather than simple data suppression. Our work shows that query logs reveal some PII values, but that these can be masked to a certain extent using appropriate techniques. This opens the way to protecting privacy in query logs with minimal effect on the utility of the logs and, more broadly, to the development and testing of new deidentification methods.
This special issue brings together six contributions that explore the Web as a space where meaning is produced, negotiated, and preserved. Spanning topics such as digital archiving, generative AI, accessibility, and interactive narrative, the articles reflect a diversity of approaches while converging on a shared concern with how digital infrastructures shape expression and interpretation. Rather than offering a unified framework, the issue highlights key tensions—between scale and situated practices, human authorship and algorithmic generation, and inclusion and exclusion in digital memory—inviting a reconsideration of the Web as a sociotechnical system of meaning-making.
In recent years, causal disentanglement, which seeks to separate user interest from social conformity, has attracted growing attention for its potential to mitigate popularity bias in recommender systems. The key to completing this task is accurately modeling the causal drivers of user interactions with items. Due to the rapid development of causal inference, most state-of-the-art (SOTA) approaches aim to learn representations of interest and conformity from interaction data. Despite their accomplishments, two limitations remain. First, most existing approaches rely on global random sampling, which overlooks the time-varying nature of popularity and leads to temporal confounding bias. Second, most ID-based approaches suffer from semantic sparsity, neglecting the rich textual metadata crucial to understanding the intrinsic content of mobile apps (short for applications). To this end, we propose a temporal causal semantic enhancement (TCSE) framework to robustly disentangle user interest from conformity by integrating time-aware sampling and semantic reasoning. More specifically, we introduce a semantic knowledge enhancement module to augment mobile app representations with generalized semantics extracted from large language models. Then, we design a temporal causal sampling strategy to construct locally unbiased interaction pairs within the same temporal window, effectively capturing dynamic popularity shifts. Extensive experiments on two real-world large-scale datasets indicate that TCSE can outperform SOTA models in both recommendation accuracy and debiasing effectiveness. An ablation study also demonstrates the necessity of TCSE’s semantic and temporal components.
Recommender systems (RS) are widely deployed in domains such as e-commerce and online media, yet often intensify the filter bubble problem by repeatedly exposing similar items. While prior studies generally regard such homogeneity as detrimental, its nuanced effects in interactive recommendation systems (IRS) remain underexplored. In particular, the dynamic feedback loop between user preferences and system responses complicates the relationship between homogeneity, diversity, and engagement. This paper presents the first fine-grained analysis of homogeneity in IRS, revealing that homogeneous recommendations can be beneficial in the early repetition stages before becoming detrimental. Building on this insight, we propose Diversity- and Debias-aware Interactive Recommendation (DDIR), an offline reinforcement learning framework that dynamically balances homogeneity and diversity over user sessions. DDIR employs a Transformer-based state encoder to capture time-aware diversity preferences, a homogeneity-friendly diversity model to estimate evolving tolerance to repetition, and a debiased interest model that disentangles intrinsic preferences from conformity bias while considering item quality. These components jointly guide a policy learner to sustain long-term engagement. Experiments on two realistic RL environments, KuaiEnv and KuaiRand, constructed from large-scale short-video logs, demonstrate that DDIR effectively mitigates filter bubbles, increasing average session length by 4.35% and long-term engagement by 25.65% over state-of-the-art baselines in KuaiRand simulated environment. Codes are available at https://github.com/16061025/DDIR.
