Source-linked AI summary
Who's Blocking Whom? Candidate Generation and Block Prediction on Bluesky
Cecilia Galbiati, Carlo Bono, Carlo Piccardi, Francesco Pierri
TL;DR
The paper asks who will block whom on Bluesky and how candidate construction affects directed block prediction. It evaluates alternative candidate and negative-example formulations using large-scale temporal interaction data, finding that candidate coverage and conditional ranking performance depend strongly on experimental design. The findings inform moderation-tool evaluation while remaining subject to observational, platform-specific, and dual-use limitations.
Problem
Existing predictive work links observable behavior to whether accounts are blocked generally, but does not identify which user will block which other account.
Method
The study compares directed block-prediction formulations that vary candidate generation, negative-example construction, evaluation scope, models, and feature groups using Bluesky block and interaction data.
Results
5.2% of future blocks are covered by recent direct interaction, rising to approximately 22.0% with common-neighbor candidates after excluding highly connected intermediaries; conditional ranking substantially exceeds random.
Takeaways & Limitations
Candidate coverage should be evaluated separately from conditional classification or ranking because candidate generation determines which block events a framework can capture.
Takeaways & Limitations
Results are specific to observed Bluesky data and experimental settings, and the observational data do not establish why users block or whether predictive suggestions improve safety.
Abstract
from arXiv · showhide
Blocking is a widely used tool that helps people manage unwanted interactions on social platforms. We study the problem of predicting block events on Bluesky: whether a given user will block a particular account, given recent interaction, network, activity, and content signals. Using more than three million block events and over 260 million user interactions, we examine several formulations of the directed block prediction problem, differing in which possible targets are considered and how negative examples are selected. We observe that only about 5% of user blocks are preceded by a recent direct interaction. Expanding the set of possible targets to include accounts connected through a common neighbor raises the share to about 22%, once high-degree accounts are excluded as intermediaries. Conditional on candidate inclusion, Hits@1 ranges from 61.3% to 72.4%, compared with a random baseline of 16.7%. However, predictive performance and the signals used by the models depend strongly on how candidates and comparison examples are constructed. Although our goal is to empirically analyze user behavior rather than to propose a deployable system, these findings are relevant to user-facing moderation tools that might help users identify accounts they may wish to avoid or block. In such tools, candidate generation would be a substantive design and evaluation choice rather than merely a preprocessing step.
Introduction
The paper studies directed, pair-specific block prediction on Bluesky and argues that candidate generation is part of the prediction problem, not merely preprocessing. It evaluates how candidate coverage, evaluation design, and feature relevance shape measured performance and the implications for moderation tools.
- Blocking research has mostly predicted whether accounts are generally likely to be blocked, leaving unclear who is likely to block whom.
- Directed block prediction requires restricting the enormous space of ordered user pairs to plausible candidate targets.Considering every possible ordered pair would be computationally impractical.
- Candidate-generation rules determine which future blocks can be captured, so accurate ranking cannot compensate for excluding the true target.
- The study uses more than three million block events and over 260 million interactions to compare candidate generation, negative-example construction, evaluation scope, models, and feature groups.
- Recent direct interaction covers 5.2% of future blocks, whereas adding common-neighbor candidates raises estimated coverage to approximately 22.0% after excluding highly connected intermediaries.
- The work is intended to study blocking behavior rather than propose a deployable system, but it bears on moderation tools that surface accounts users may wish to avoid or block.
Related Work
Related work frames directed block prediction as temporal negative-link prediction while emphasizing that negative selection and candidate-set construction materially affect evaluation. The paper connects this technical problem to blocking’s role in user control, safety, governance, and exclusion.
- Temporal and negative-link prediction: Directed block prediction extends temporal link prediction to future, negative, ordered edges using interaction histories that include follows, likes, reposts, replies, and prior blocks.
- Temporal and negative-link prediction: Random negatives can inflate performance, motivating historical and future-aware alternatives such as observed-future negatives.
- Temporal and negative-link prediction: Grouped evaluations use multiple negatives, Hits@K, and MRR, while this study extends the evaluation concern to deciding which ordered pairs are eligible for ranking.
- Blocking and platform governance: Blocking differs from positive-tie prediction because it is deliberate, strongly directional, and associated with structural and ideological patterns in negative relations.
- Blocking and platform governance: Blocking functions both as a safety mechanism and as platform governance, shifting some authority toward communities and end users in decentralized systems.
- Blocking and platform governance: Personalized moderation can increase user control while shifting configuration and moderation labor onto users, and collective blocklists can create false positives and perceived injustice.
- Blocking and platform governance: Unlike prior Bluesky work on network structure or platform-wide blocking, this study predicts future blocks between specific user pairs and examines candidate-generation effects.
Problem Formulation and Data
The paper defines directed block prediction over time using ordered blocker-target pairs, recent interaction histories, and explicit negative-labeling rules. Its two-stage setup first admits candidates and then scores them with structured behavioral, network, and content features from Bluesky data.
- Pair definition: For each prediction day, an ordered pair (u, v) identifies u as potential blocker and v as candidate target; reversing the order defines a different event.
- Two-stage prediction task: Candidate generation determines which user pairs enter the task, after which a supervised model scores admitted candidates using pair-, source-, target-, graph-, and content-level features.
- Pair definition: The analysis retains only the first observed block for each ordered pair and defines y_u,v,t = 1 on the day that block occurs.
- Feature construction: Features use activity from the seven days before prediction day t, including interaction direction, type, frequency, recency, user activity, graph position, visibility, blocking history, and content statistics.
- Negative examples: Same-day negatives have no observed block on day t, whereas observed-future negatives remain unblocked through the observation period.The final seven prediction days are excluded to ensure at least seven observable future days.
- Dataset: The dataset contains 3 071 079 first observed user-to-user block events, 2 561 333 users, and 263 416 428 timestamped follows, likes, reposts, and replies.
- Dataset: Because Bluesky activity is publicly observable through its Firehose, the study uses public event data but withholds raw block and interaction records to reduce privacy risks.
Experimental Design
The study evaluates seven directed block-prediction settings that vary candidate eligibility, negative-example construction, and whether candidates are ranked globally or within source-day groups. It combines chronological evaluation, model screening, classification and ranking metrics, uncertainty estimation, and feature-importance analyses.
- Prediction settings: Seven settings vary candidate targets, negative definitions, and whether evaluation is global or grouped by potential blocker and prediction day.Settings A–C use ungrouped classification, while D–G use source-day groups; Setting G expands eligibility to direct-or-common-neighbor pairs.
- Prediction settings: Setting F selects five matched same-source negatives using interaction-pattern restrictions and standardized historical-activity similarity.Negative targets are chosen from the five nearest eligible cases by Euclidean distance, making discrimination more challenging.
- Prediction settings: Setting G admits pairs connected directly or through one common neighbor in a seven-day undirected interaction graph, excluding the top 1% of nodes as intermediaries.Interaction direction and type are ignored when constructing the candidate graph but remain model features.
- Evaluation protocol: The design uses chronological training, validation, and test intervals, with shortened periods for settings requiring seven days of observable follow-up.Settings A–B and C–G use different date ranges, while C, D, F, E, and G impose the follow-up constraint described in the passages.
- Models and metrics: LightGBM is selected for the remaining experiments after screening because it achieves the highest ROC-AUC with nearly the highest AP and lower fitting cost than the custom GNN.The screening compares proximity heuristics, embeddings, a graph neural network, and tabular classifiers; the paper uses separate LightGBM instances for each setting.
- Models and metrics: Classification uses average precision as the main threshold-independent metric, while grouped settings report tie-aware Hits@1, Hits@3, and mean reciprocal rank against five negatives.The random Hits@1 baseline for grouped settings is 1/6 ≈ 0.167; uncertainty is estimated with 1000 cluster-bootstrap resamples.
Results
Results show that candidate coverage is limited by recent direct interaction, while predictive performance and feature importance vary substantially with candidate construction and comparison design.
- Pre-block connectivity: 5.2% of 2,878,871 first observed blocks involved recent direct interaction, leaving 94.8% without such interaction.Among direct-interaction cases, 64.0% involved only target-to-source interaction, 20.1% only source-to-target interaction, and 15.9% reciprocal interaction.
- Pre-block connectivity: 85.7% of no-direct-interaction block pairs were connected within the historical interaction graph, versus 76.4% of same-source, degree-matched non-block pairs.The broader connectivity pattern supports examining local interaction structure beyond direct pairwise contact.
- Candidate coverage: 22.0% of observed block events were admitted by the filtered common-neighbor candidate rule, compared with 5.2% for direct-interaction-only candidates.Without hub filtering, common-neighbor expansion reached approximately 41.0% coverage; common-neighbor-only pairs added 16.9 percentage points under the filtered rule.
- Predictive performance: Hits@1 fell from 0.724 in Setting D and 0.723 in Setting E to 0.657 in Setting F and 0.613 in Setting G, above the 0.167 random baseline.These ranking values are conditional on the true target being admitted to the candidate set; Setting G’s rule admits approximately 22% of observed block events.
- Feature effects: Feature importance depended on task design: source-level features mattered most for ungrouped AP, whereas target-level features mattered most for same-source Hits@1.Removing target-level features reduced Hits@1 by 0.093, 0.115, and 0.178 in Settings D, F, and G, respectively; removing source-level features reduced AP by 0.136 and 0.140 in Settings B and C.
- Feature effects: Recent received-block count was the strongest individual predictor, while approximately 8–9% of features recovered 97–99% of achievable AP.Used alone, recent received-block count achieved ROC-AUC values of 0.722–0.805 and AP values of 0.354–0.510, but its underlying association remains unclear.
- Feature effects: Directional interaction associations weakened under behaviorally matched negatives, while Adamic–Adar became informative when common-neighbor candidates were included.In Setting D, target-to-source interaction had a 54.0% positive rate versus 33.8% in Setting F; in Setting G, Adamic–Adar alone achieved ROC-AUC 0.622 and AP 0.279.
Discussion and Conclusion
The paper shows that candidate generation and evaluation choices shape both directed block-prediction performance and the features models appear to rely on. It also emphasizes that predictive moderation tools must account for ambiguous behavioral signals, user control, and dual-use risks.
- Candidate selection substantially improves coverage, but predictive performance remains conditional on which targets enter the candidate set.The paper argues that coverage and ranking performance should be evaluated separately because they use different denominators.
- Feature importance changes across candidate-generation, grouping, and negative-sampling procedures.This means model interpretations are tied to the evaluation design rather than solely to underlying user behavior.
- Recent blocks received by a target are strongly associated with model performance, but the association may reflect harmful behavior, visibility, controversy, polarization, or coordinated blocking.The paper cautions against treating prior blocks as direct evidence of harmful conduct.
- The results are specific to observed Bluesky data and experimental settings, and predictive ranking should not be interpreted as platform-wide recall.The observational design also cannot establish why users block or whether predictive suggestions improve safety.
- Future work should improve candidate coverage without creating excessively large or irrelevant candidate sets, while testing robustness, user effects, and causal consequences.The paper calls for evaluation across longer periods and user groups and for studying effects on safety, exposure, and behavior.
- AI-assisted moderation should treat candidate generation, user control, and unfair-exclusion risks as core evaluation concerns because they shape social consequences and behavioral signals.These concerns extend beyond model performance alone.