Xueying Ding, Xingyue Huang, Mingxuan Ju, Liam Collins, Yozen Liu, Leman Akoglu, Neil Shah, Tong Zhao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Softmax attention struggles with long contexts due to structural limitations: the strict sum-toone constraint forces attention sinks on irrelevant tokens, and probability mass disperses as sequence lengths increase.We tackle these problems with Threshold Differential Attention (TDA), a sink-free attention mechanism that achieves ultra-sparsity and improved robustness at longer sequence lengths without the computational overhead of projection methods or the performance degradation caused by noise accumulation of standard rectified attention.TDA applies row-wise extreme-value thresholding with a length-dependent gate, retaining only exceedances.Inspired by the differential transformer, TDA also subtracts an inhibitory view to enhance expressivity.Theoretically, we prove that TDA controls the expected number of spurious survivors per row to O(1) and that consensus spurious matches across independent views vanish as context grows.Empirically, TDA produces > 99% exact zeros and eliminates attention sinks while maintaining competitive performance on standard and long-context benchmarks.

Generative retrieval (GR) has revolutionized recommendation systems by integrating the content of the items into semantic identifiers. However, existing frameworks predominantly isolate modalities (e.g., relying solely on text), overlooking the inherently multimodal nature of real-world items. This work addresses the underexplored challenge of Multimodal Generative Retrieval (MGR). Through a systematic analysis of Early and Late Fusion strategies, we reveal that naive integration fails due to two critical limitations: modality sensitivity, where one modality dominates the representation, and modality correspondence, where the model fails to align distinct semantic IDs across modalities. To overcome these challenges, we introduce MGR-LF++, an enhanced late fusion framework. MGR-LF++ employs contrastive modality alignment to enforce cross-modal consistency and incorporates special tokens to preserve autoregressive integrity. Extensive experiments demonstrate that MGR-LF++ achieves performance improvements of over 20% compared to unimodal and naive multimodal alternatives.

Recent advances in graph machine learning (ML) with the introduction of Graph Neural Networks (GNNs) have led to a widespread interest in applying these approaches to business applications at scale. GNNs enable differentiable end-to-end (E2E) learning of model parameters given graph structure which enables optimization towards popular node, edge (link) and graph-level tasks. While the research innovation in new GNN layers and training strategies has been rapid, industrial adoption and utility of GNNs has lagged considerably due to the unique scale challenges that large-scale graph ML problems create. In this work, we share our approach to training, inference, and utilization of GNNs at Snapchat. To this end, we present GiGL (Gigantic Graph Learning), an open-source library to enable large-scale distributed graph ML to the benefit of researchers, ML engineers, and practitioners. We use GiGL internally at Snapchat to manage the heavy lifting of GNN workflows, including graph data preprocessing from relational DBs, subgraph sampling, distributed training, inference, and orchestration. GiGL is designed to interface cleanly with open-source GNN modeling libraries prominent in academia like PyTorch Geometric (PyG), while handling scaling and productionization challenges that make it easier for internal practitioners to focus on modeling. GiGL is used in multiple production settings, and has powered over 35 launches across multiple business domains in the last 2 years in the contexts of friend recommendation, content recommendation and advertising. This work details high-level design and tools the library provides, scaling properties, case studies in diverse business settings with large-scale graphs up to hundreds of millions of nodes, tens of billions of edges, and hundreds of node and edge features, and several key lessons learned in employing graph ML at scale on large social data. GiGL is open-sourced at https://github.com/Snapchat/GiGL.

Recent advances in graph machine learning (GML) and Graph Neu- ral Networks (GNNs) have sparked significant practical interest given the ability to model complex relationships between entities. Despite rapid progress in GNN designs, scalability remains a major challenge. Industry applications require solutions that can handle graphs with billions of nodes and edges efficiently. GiGL (Gigantic Graph Learning) is an open-source library from Snapchat, designed for large-scale distributed training and inference with GNNs. It seamlessly integrates with popular open-source GNN libraries like PyTorch Geometric (PyG). GiGL provides simplified configurable interfaces with minimal modeling code requirements, providing in- dustrial practitioners a straightforward way to apply GNNs to large- scale applications and enabling academics to conduct large-scale experiments. At the same time, it enables complex modeling capabil- ities desirable for modeling iteration. In this hands-on tutorial, we will demonstrate how GiGL addresses the scalability challenge in GNNs and provide a step-by-step guide for attendees to complete end-to-end training and inference with GiGL on industry-scale graphs. By the end of our tutorial, participants will have hands-on experience in training GNNs on graphs with billions of nodes and edges - capabilities not easily achievable with open-source graph learning libraries like PyG alone. We anticipate strong interest and participation from both industrial practitioners working on GNN applications and academics conducting large-scale experiments.

Deep recommender systems rely heavily on large embedding tables to handle high-cardinality categorical features such as user/item identifiers, and face significant memory constraints at scale.To tackle this challenge, hashing techniques are often employed to map multiple entities to the same embedding and thus reduce the size of the embedding tables.Concurrently, graph-based collaborative signals have emerged as powerful tools in recommender systems, yet their potential for optimizing embedding table reduction remains unexplored.This paper introduces GraphHash, the first graph-based approach that leverages modularity-based bipartite graph clustering on user-item interaction graphs to reduce embedding table sizes.We demonstrate that the modularity objective has a theoretical connection to message-passing, which provides a foundation for our method.By employing fast clustering algorithms, GraphHash serves as a computationally efficient proxy for message-passing during preprocessing and a plug-andplay graph-based alternative to traditional ID hashing.Extensive experiments show that GraphHash substantially outperforms diverse hashing baselines on both retrieval and click-through-rate prediction tasks.In particular, GraphHash achieves on average a 101.52% improvement in recall when reducing the embedding table size by more than 75%, highlighting the value of graph-based collaborative information for model reduction.

Collaborative filtering (CF) has exhibited prominent results for recommender systems and been broadly utilized for real-world applications. A branch of research enhances CF methods by message passing (MP) used in graph neural networks, due to its strong capabilities of extracting knowledge from graph-structured data, like user-item bipartite graphs that naturally exist in CF. They assume that MP helps CF methods in a manner akin to its benefits for graph-based learning tasks in general (e.g., node classification). However, even though MP empirically improves CF, whether or not this assumption is correct still needs verification. To address this gap, we formally investigate why MP helps CF from multiple perspectives and show that many assumptions made by previous works are not entirely accurate. With our curated ablation studies and theoretical analyses, we discover that (i) MP improves the CF performance primarily by additional representations passed from neighbors during the forward pass instead of additional gradient updates to neighbor representations during the model back-propagation and (ii) MP usually helps low-degree nodes more than high-degree nodes.}Utilizing these novel findings, we present Test-time Aggregation for Collaborative Filtering, namely TAG-CF, a test-time augmentation framework that only conducts MP once at inference time. The key novelty of TAG-CF is that it effectively utilizes graph knowledge while circumventing most of notorious computational overheads of MP. Besides, TAG-CF is extremely versatile can be used as a plug-and-play module to enhance representations trained by different CF supervision signals. Evaluated on six datasets (i.e., five academic benchmarks and one real-world industrial dataset), TAG-CF consistently improves the recommendation performance of CF methods without graph by up to 39.2% on cold users and 31.7% on all users, with little to no extra computational overheads. Furthermore, compared with trending graph-enhanced CF methods, TAG-CF delivers comparable or even better performance with less than 1% of their total training times. Our code is publicly available at https://github.com/snap-research/Test-time-Aggregation-for-CF.

Learning general-purpose user representations based on user behavioral logs is an increasingly popular user modeling approach. It benefits from easily available, privacy-friendly yet expressive data, and does not require extensive re-tuning of the upstream user model for different downstream tasks. While this approach has shown promise in search engines and e-commerce applications, its fit for instant messaging platforms, a cornerstone of modern digital communication, remains largely uncharted. We explore this research gap using Snapchat data as a case study. Specifically, we implement a Transformer-based user model with customized training objectives and show that the model can produce high-quality user representations across a broad range of evaluation tasks, among which we introduce three new downstream tasks that concern pivotal topics in user research: user safety, engagement and churn. We also tackle the challenge of efficient extrapolation of long sequences at inference time, by applying a novel positional encoding method.

Embedding-based retrieval in graph-based recommendation has shown great improvements over traditional graph walk retrieval methods, and has been adopted in large-scale industry applications such as friend recommendations [16]. However, it is not without its challenges: retraining graph embeddings frequently due to changing data is slow and costly, and producing high recall of approximate nearest neighbor search (ANN) on such embeddings is challenging due to the power law distribution of the indexed users. In this work, we address theses issues by introducing a simple query expansion method in ANN, called FriendSeedSelection, where for each node query, we construct a set of 1-hop embeddings and run ANN search. We highlight our approach does not require any model-level tuning, and is inferred from the data at test-time. This design choice effectively enables our recommendation system to adapt to the changing graph distribution without frequent heavy model retraining. We also discuss how we design our system to efficiently construct such queries online to support 10k+ QPS. For friend recommendation, our method shows improvements of recall, and 11% relative friend reciprocated communication metric gains, now serving over 800 million monthly active users at Snapchat.

Pictorial emojis and stickers are commonly used in online social communications. We analyzed social communications using Bitmoji stickers, which are expressive pictorial stickers made from avatars resembling actual users. We collect a large-scale dataset of 3 billion Bitmoji stickers’ metadata, shared among 300 million Snapchat users. We find that individual Bitmoji sticker usage patterns can be characterized jointly on dimensions of reciprocity and selectivity. Generally speaking, users are either both reciprocal and selective about whom they use Bitmoji stickers with or neither reciprocal nor selective. We additionally demonstrate network homophily by showing that friends use Bitmoji stickers at similar rates. Finally, using a quasi-experimental approach, we show that receiving Bitmoji stickers from a friend encourages future Bitmoji sticker usage and overall Snapchat engagement. Our work carries implications for a better understanding of online pictorial communication behaviors.

Culture shapes people’s behavior, both online and offline. Surprisingly, there is sparse research on how cultural context affects network formation and content consumption on social media. We analyzed the friendship networks and dyadic relations between content producers and consumers across 73 countries through a cultural lens in a closed-network setting. Closed networks allow for intimate bonds and self-expression, providing a natural setting to study cultural differences in behavior. We studied three theoretical frameworks of culture - individualism, relational mobility, and tightness. We found that friendship networks formed across different cultures differ in egocentricity, meaning the connectedness between a user’s friends. Individualism, mobility, and looseness also significantly negatively impact how tie strength affects content consumption. Our findings show how culture affects social media behavior, and we outline how researchers can incorporate this in their work. Our work has implications for content recommendations and can improve content engagement.

Graph Neural Networks (GNNs) have shown exceptional performance in the task of link prediction. Despite their effectiveness, the high latency brought by non-trivial neighborhood data dependency limits GNNs in practical deployments. Conversely, the known efficient MLPs are much less effective than GNNs due to the lack of relational knowledge. In this work, to combine the advantages of GNNs and MLPs, we start with exploring direct knowledge distillation (KD) methods for link prediction, i.e., predicted logit-based matching and node representation-based matching. Upon observing direct KD analogs do not perform well for link prediction, we propose a relational KD framework, Linkless Link Prediction (LLP), to distill knowledge for link prediction with MLPs. Unlike simple KD methods that match independent link logits or node representations, LLP distills relational knowledge that is centered around each (anchor) node to the student MLP. Specifically, we propose rank-based matching and distribution-based matching strategies that complement each other. Extensive experiments demonstrate that LLP boosts the link prediction performance of MLPs with significant margins and even outperforms the teacher GNNs on 7 out of 8 benchmarks. LLP also achieves a 70.68x speedup in link prediction inference compared to GNNs on the large-scale OGB dataset.

Understanding the users' patterns of visiting various location categories can help online platforms improve content personalization and user experiences. Current literature on predicting future location categories of a user typically employs features that can be traced back to the user, such as spatial geo-coordinates and demographic identities. Moreover, existing approaches commonly suffer from cold-start and generalization problems, and often cannot specify when the user will visit the predicted location category. In a large social platform, it is desirable for prediction models to avoid using user-identifiable data, generalize to unseen and new users, and be able to make predictions for specific times in the future. In this work, we construct a neural model, LocHabits, using data from Snapchat. The model omits user-identifiable inputs, leverages temporal and sequential regularities in the location category histories of Snapchat users and their friends, and predicts the users' next-hour location categories. We evaluate our model on several real-life, large-scale datasets from Snapchat and FourSquare, and find that the model can outperform baselines by 14.94% accuracy. We confirm that the model can (1) generalize to unseen users from different areas and times, and (2) fall back on collective trends in the cold-start scenario. We also study the relative contributions of various factors in making the predictions and find that the users' visitation preferences and most-recent visitation sequences play more important roles than time contexts, same-hour sequences, and social influence features.

Self-supervised learning on graphs has made large strides in achieving great performance in various downstream tasks. However, many state-of-the-art methods suffer from a number of impediments, which prevent them from realizing their full potential. For instance, contrastive methods typically require negative sampling, which is often computationally costly. While non-contrastive methods avoid this expensive step, most existing methods either rely on overly complex architectures or dataset-specific augmentations. In this paper, we ask: Can we borrow from classical unsupervised machine learning literature in order to overcome those obstacles? Guided by our key insight that the goal of distance-based clustering closely resembles that of contrastive learning: both attempt to pull representations of similar items together and dissimilar items apart. As a result, we propose CARL-G - a novel clustering-based framework for graph representation learning that uses a loss inspired by Cluster Validation Indices (CVIs), i.e., internal measures of cluster quality (no ground truth required). CARL-G is adaptable to different clustering methods and CVIs, and we show that with the right choice of clustering method and CVI, CARL-G outperforms node classification baselines on 4/5 datasets with up to a 79× training speedup compared to the best-performing baseline. CARL-G also performs at par or better than baselines in node clustering and similarity search tasks, training up to 1,500× faster than the best-performing baseline. Finally, we also provide theoretical foundations for the use of CVI-inspired losses in graph representation learning.

Friend recommendation systems in online social and professional networks such as Snapchat helps users find friends and build connections, leading to better user engagement and retention. Traditional friend recommendation systems take advantage of the principle of locality and use graph traversal to retrieve friend candidates, e.g. Friends-of-Friends (FoF). While this approach has been adopted and shown efficacy in companies with large online networks such as Linkedin and Facebook, it suffers several challenges: (i) discrete graph traversal offers limited reach in cold-start settings, (ii) it is expensive and infeasible in realtime settings beyond 1 or 2 hop requests owing to latency constraints, and (iii) it cannot well-capture the complexity of graph topology or connection strengths, forcing one to resort to other mechanisms to rank and find top-K candidates. In this paper, we proposed a new Embedding Based Retrieval (EBR) system for retrieving friend candidates, which complements the traditional FoF retrieval by retrieving candidates beyond 2-hop, and providing a natural way to rank FoF candidates. Through online A/B test, we observe statistically significant improvements in the number of friendships made with EBR as an additional retrieval source in both low- and high-density network markets. Our contributions in this work include deploying a novel retrieval system to a large-scale friend recommendation system at Snapchat, generating embeddings for billions of users using Graph Neural Networks, and building EBR infrastructure in production to support Snapchat scale.

As the state-of-the-art graph learning models, the message passing based neural networks (MPNNs) implicitly use the graph topology as the "pathways" to propagate node features. This implicit use of graph topology induces the MPNNs' over-reliance on (node) features and high inference latency, which hinders their large-scale applications in industrial contexts. To mitigate these weaknesses, we propose the Graph Explicit Neural Network (GENN) framework. GENN can be flexibly applied to various MPNNs and improves them by providing more efficient and accurate inference that is robust in feature-constrained settings. Specifically, we carefully incorporate recent developments in network embedding methods to efficiently prioritize the graph topology for inference. From this vantage, GENN explicitly encodes the topology as an important source of information to mitigate the reliance on node features. Moreover, by adopting knowledge distillation (KD) techniques, GENN takes an MPNN as the teacher to supervise the training for better effectiveness while avoiding the teacher's high inference latency. Empirical results show that our GENN infers dramatically faster than its MPNN teacher by 40x-78x. In terms of accuracy, GENN yields significant gains (more than 40%) for its MPNN teacher when the node features are limited based on our explicit encoding. Moreover, GENN outperforms the MPNN teacher even in feature-rich settings thanks to our KD design.

Explaining machine learning models is an important and increasingly popular area of research interest. The Shapley value from game theory has been proposed as a prime approach to compute feature importance towards model predictions on images, text, tabular data, and recently graph neural networks (GNNs) on graphs. In this work, we revisit the appropriateness of the Shapley value for GNN explanation, where the task is to identify the most important subgraph and constituent nodes for GNN predictions. We claim that the Shapley value is a non-ideal choice for graph data because it is by definition not structure-aware. We propose a Graph Structure-aware eXplanation (GStarX) method to leverage the critical graph structure information to improve the explanation. Specifically, we define a scoring function based on a new structure-aware value from the cooperative game theory proposed by Hamiache and Navarro (HN). When used to score node importance, the HN value utilizes graph structures to attribute cooperation surplus between neighbor nodes, resembling message passing in GNNs, so that node importance scores reflect not only the node feature importance, but also the node structural roles. We demonstrate that GStarX produces qualitatively more intuitive explanations, and quantitatively improves explanation fidelity over strong baselines on chemical graph property prediction and text graph sentiment classification. Code: https://github.com/ShichangZh/GStarX

Social platforms have paved the way in creating new, modern ways for users to communicate with each other. In recent years, multiple platforms have introduced ''Stories'' features, which enable broadcasting of ephemeral multimedia content. Specifically, ''Friend Stories,'' or Stories meant to be consumed by one's close friends, are a popular feature, promoting significant user-user interactions by allowing people to see (visually) what their friends and family are up to. A key challenge in surfacing Friend Stories for a given user, is in ranking over each viewing user's friends to efficiently prioritize and route limited user attention. In this work, we explore the novel problem of Friend Story Ranking from a graph representation learning perspective. More generally, our problem is a link ranking task, where inferences are made over existing links (relations), unlike common node or graph-based tasks, or link prediction tasks, where the goal is to make inferences about non-existing links. We propose ELR, an edge-contextual approach which carefully considers local graph structure, differences between local edge types and directionality, and rich edge attributes, building on the backbone of graph convolutions. ELR handles social sparsity challenges by considering and attending over neighboring nodes, and incorporating multiple edge types in local surrounding egonet structures. We validate ELR on two large country-level datasets with millions of users and tens of millions of links from Snapchat. ELR shows superior performance over alternatives by 8% and 5% error reduction measured by MSE and MAE correspondingly. Further generality, data efficiency and ablation experiments confirm the advantages of ELR.

Data augmentation has been widely used to improve generalizability of machine learning models. However, comparatively little work studies data augmentation for graphs. This is largely due to the complex, non-Euclidean structure of graphs, which limits possible manipulation operations. Augmentation operations commonly used in vision and language have no analogs for graphs. Our work studies graph data augmentation for graph neural networks (GNNs) in the context of improving semi-supervised node-classification. We discuss practical and theoretical motivations, considerations and strategies for graph data augmentation. Our work shows that neural edge predictors can effectively encode class-homophilic structure to promote intra-class edges and demote inter-class edges in given graph structure, and our main contribution introduces the GAug graph data augmentation framework, which leverages these insights to improve performance in GNN-based node classification via edge prediction. Extensive experiments on multiple benchmarks show that augmentation via GAug improves performance across GNN architectures and datasets.

Showing ads delivers revenue for online content distributors, but ad exposure can compromise user experience and cause user fatigue and frustration. Correctly balancing ads with other content is imperative. Currently, ad allocation relies primarily on demographics and inferred user interests, which are treated as static features and can be privacy-intrusive. This paper uses person-centric and momentary context features to understand optimal ad-timing. In a quasi-experimental study on a three-month longitudinal dataset of 100K Snapchat users, we find ad timing influences ad effectiveness. We draw insights on the relationship between ad effectiveness and momentary behaviors such as duration, interactivity, and interaction diversity. We simulate ad reallocation, finding that our study-driven insights lead to greater value for the platform. This work advances our understanding of ad consumption and bears implications for designing responsible ad allocation systems, improving both user and platform outcomes. We discuss privacy-preserving components and ethical implications of our work.

Graph Neural Networks (GNNs) have risen to prominence in learning representations for graph structured data. A single GNN layer typically consists of a feature transformation and a feature aggregation operation. The former normally uses feed-forward networks to transform features, while the latter aggregates the transformed features over the graph. Numerous recent works have proposed GNN models with different designs in the aggregation operation. In this work, we establish mathematically that the aggregation processes in a group of representative GNN models including GCN, GAT, PPNP, and APPNP can be regarded as (approximately) solving a graph denoising problem with a smoothness assumption. Such a unified view across GNNs not only provides a new perspective to understand a variety of aggregation operations but also enables us to develop a unified graph neural network framework UGNN. To demonstrate its promising potential, we instantiate a novel GNN model, ADA-UGNN, derived from UGNN, to handle graphs with adaptive smoothness across nodes. Comprehensive experiments show the effectiveness of ADA-UGNN.

Graph Neural Networks (GNNs) have recently enabled substantial advances in graph learning. Despite their rich representational capacity, GNNs remain under-explored for large-scale social modeling applications. One such industrially ubiquitous application is friend suggestion: recommending users other candidate users to befriend, to improve user connectivity, retention and engagement. However, modeling such user-user interactions on large-scale social platforms poses unique challenges: such graphs often have heavy-tailed degree distributions, where a significant fraction of users are inactive and have limited structural and engagement information. Moreover, users interact with different functionalities, communicate with diverse groups, and have multifaceted interaction patterns.

With the rapid growth and prevalence of social network applications (Apps) in recent years, understanding user engagement has become increasingly important, to provide useful insights for future App design and development. While several promising neural modeling approaches were recently pioneered for accurate user engagement prediction, their black-box designs are unfortunately limited in model explainability. In this paper, we study a novel problem of explainable user engagement prediction for social network Apps. First, we propose a flexible definition of user engagement for various business scenarios, based on future metric expectations. Next, we design an end-to-end neural framework, FATE, which incorporates three key factors that we identify to influence user engagement, namely friendships, user actions, and temporal dynamics to achieve explainable engagement predictions. FATE is based on a tensor-based graph neural network (GNN), LSTM and a mixture attention mechanism, which allows for (a) predictive explanations based on learned weights across different feature categories, (b) reduced network complexity, and (c) improved performance in both prediction accuracy and training/inference time. We conduct extensive experiments on two large-scale datasets from Snapchat, where FATE outperforms state-of-the-art approaches by 10% error and 20% runtime reduction. We also evaluate explanations from FATE, showing strong quantitative and qualitative performance.

While mobile social apps have become increasingly important in people's daily life, we have limited understanding on what motivates users to engage with these apps. In this paper, we answer the question whether users' in-app activity patterns help inform their future app engagement (e.g., active days in a future time window)? Previous studies on predicting user app engagement mainly focus on various macroscopic features (e.g., time-series of activity frequency), while ignoring fine-grained inter-dependencies between different in-app actions at the microscopic level. Here we propose to formalize individual user's in-app action transition patterns as a temporally evolving action graph, and analyze its characteristics in terms of informing future user engagement. Our analysis suggested that action graphs are able to characterize user behavior patterns and inform future engagement. We derive a number of high-order graph features to capture in-app usage patterns and construct interpretable models for predicting trends of engagement changes and active rates. To further enhance predictive power, we design an end-to-end, multi-channel neural model to encode both temporal action graphs, activity sequences, and other macroscopic features. Experiments on predicting user engagement for 150k Snapchat new users over a 28-day period demonstrate the effectiveness of the proposed prediction models. The analysis and prediction framework is also deployed at Snapchat to deliver real world business insights. Our proposed framework is also general and can be applied to any online platform.