User retention is a critical objective for online platforms like Pinterest, as it strengthens user loyalty and drives growth through repeated engagement. A key indicator of retention is revisitation, i.e., when users return to view previously saved content, a behavior often sparked by personalized recommendations and user satisfaction. However, modeling and optimizing revisitation poses significant challenges. One core difficulty is accurate attribution: it is often unclear which specific user actions or content exposures trigger a revisit, since many confounding factors (e.g., content quality, user interface, notifications, or even changing user intent) can influence return behavior. Additionally, the scale and timing of revisitations introduce further complexity; users may revisit content days or even weeks after their initial interaction, requiring the system to maintain and associate extensive historical records across millions of users and sessions. These complexities render existing methods insufficient for robustly capturing and optimizing long-term revisitation. To address these gaps, we introduce a novel, lightweight, and interpretable framework for modeling revisitation behavior and optimizing long-term user retention in Pinterest’s search-based recommendation context. By defining a surrogate attribution process that links saves to subsequent revisitations, we reduce noise in the causal relationship between user actions and return visits. Our scalable event aggregation pipeline enables large-scale analysis of user revisitation patterns and enhances the ranking system’s ability to surface items with high retention value. Deployed on Pinterest’s Related Pins surface to serve 500+ million users, the framework led to a significant lift of 0.1% in active users without additional computational costs. Our data analysis reveals novel insights, such as the impact of content topics on revisitation rates; for example, users are more likely to revisit aesthetically pleasing topics.

Semantic IDs (SIDs) are hierarchical item identifiers learned via residual quantization, offering a promising solution to cold-start recommendation, yet their integration into discriminative recommendations remains challenging and lacks systematic investigation. We present Pin-SCALE, a framework for optimized end-to-end integration of SID into dense retrieval in cascading recommender systems. Pin-SCALE instantiates the principles through three technical contributions: (1) cascading pooling preserving residual quantization hierarchy that outperforms common pooling schema and attention;(2) engagement-aware tokenization that encodes collaborative signals to fully utilize both content and user engagement pattern; and (3) multi-view contrastive learning that harnesses query-item and item-item co-occurrence to align SID representations with engagement prediction. Through extensive offline analyses, we collected systematic guidance and proved the effectiveness of Pin-SCALE. Furthermore, the framework has been tested online on multiple surfaces at Pinterest including Closeup, Home Feed and Search, with different content types like organic and e-commerce. Pin-SCALE achieves substantial improvements in fresh engagements (+3.67% repins) as well as platform-level retention (+0.05% DAU). Our work has been shipped into production to serve billions of requests daily and drive discovery journey of our users, as well as provides principled guidelines for SID adoption in discriminative recommendation models at production scale.

Representation learning, a task of learning latent vectors to represent entities, is a key task in improving search and recommender systems in web applications. Various representation learning methods have been developed, including graph-based approaches for relationships among entities, sequence-based methods for capturing the temporal evolution of user activities, and content-based models for leveraging text and visual content. However, the development of a unifying framework that integrates these diverse techniques to support multiple applications remains a significant challenge. This paper presents OmniSage, a large-scale representation framework that learns universal representations for a variety of applications at Pinterest. OmniSage integrates graph neural networks with content-based models and user sequence models by employing multiple contrastive learning tasks to effectively process graph data, user sequence data, and content signals. To support the training and inference of OmniSage, we developed an efficient infrastructure capable of supporting Pinterest graphs with billions of nodes. The universal representations generated by OmniSage have significantly enhanced user experiences on Pinterest, leading to an approximate 2.5% increase in sitewide repins (saves) across five applications. This paper highlights the impact of unifying representation learning methods, and we make the model code publicly available at https://github.com/pinterest/atg-research/tree/main/omnisage.

User activity sequences have emerged as one of the most important signals in recommender systems.We present a foundational model, PinFM, for understanding user activity sequences across multiple applications at a billion-scale visual discovery platform.We pretrain a transformer model with 20B+ parameters using extensive user activity data, then fine-tune it for specific applications, efficiently coupling it with existing models.While this pretrainingand-fine-tuning approach has been popular in other domains, such as Vision and NLP, its application in industrial recommender systems presents numerous challenges.The foundational model must be scalable enough to score millions of items every second while meeting tight cost and latency constraints imposed by these systems,.Additionally, it should capture the interactions between user activities and other features and handle new items that were not present during the pretraining stage.We developed innovative techniques to address these challenges.Our infrastructure and algorithmic optimizations, such as the Deduplicated Cross-Attention Transformer (DCAT), improved our throughput by 600% on Pinterest internal data.We demonstrate that PinFM can learn interactions between user sequences and candidate items * Work done at Pinterest.

User-to-item retrieval has been an active research area in recommendation system, and two tower models are widely adopted due to model simplicity and serving efficiency. In this work, we focus on a variant called conditional retrieval, where we expect retrieved items to be relevant to a condition (e.g. topic). We propose a method that uses the same training data as standard two tower models but incorporates item-side information as conditions in query. This allows us to bootstrap new conditional retrieval use cases and encourages feature interactions between user and condition. Experiments show that our method can retrieve highly relevant items and outperforms standard two tower models with filters on engagement metrics. The proposed model is deployed to power a topic-based notification feed at Pinterest and led to +0.26% weekly active users.

Fairness in Graph Convolutional Neural Networks (GCNs) becomes a more and more important concern as GCNs are adopted in many crucial applications. Societal biases against sensitive groups may exist in many real world graphs. GCNs trained on those graphs may be vulnerable to being affected by such biases. In this paper, we adopt the well-known fairness notion of demographic parity and tackle the challenge of training fair and accurate GCNs efficiently. We present an in-depth analysis on how graph structure bias, node attribute bias, and model parameters may affect the demographic parity of GCNs. Our insights lead to FairSample, a framework that jointly mitigates the three types of biases. We employ two intuitive strategies to rectify graph structures. First, we inject edges across nodes that are in different sensitive groups but similar in node features. Second, to enhance model fairness and retain model quality, we develop a learnable neighbor sampling policy using reinforcement learning. To address the bias in node features and model parameters, FairSample is complemented by a regularization objective to optimize fairness.

Taxonomies describe the definitions of entities, entities' attributes and the relations among the entities, and thus play an important role in building a knowledge graph. In this paper, we tackle the task of taxonomy entity translation, which is to translate the names of taxonomy entities in a source language to a target language. The translations then can be utilized to build a knowledge graph in the target language. Despite its importance, taxonomy entity translation remains a hard problem for AI models due to two major challenges. One challenge is understanding the semantic context in very short entity names. Another challenge is having deep understanding for the domain where the knowledge graph is built.

Graph Neural Networks (GNNs) are AI models that learn embeddings for the nodes in a graph and use the embeddings to perform prediction tasks. In this talk, we present how we developed GNNs for the LinkedIn economic graph. LinkedIn economic graph is a digital representation of the global economy with 1B nodes and 200B edges, consisting of social graphs about members' connections, activity graphs between members and other economic entities, and knowledge graphs about members', companies', job postings' attributes. By applying GNN to this graph, we can utilize the full potential of the economic graph in many search and recommendation products across LinkedIn.

The ability to infer an individual's expertise for a given skill has proven to be crucial in creating economic opportunity for every talent of the global workforce. Applications ranging from recommending relevant job opportunities to talents to providing better candidate suggestions to recruiters, all benefit from deep understanding of the skill "proficiency" of the talent pool.

The main challenge of adapting Graph convolutional networks (GCNs) to large-scale graphs is the scalability issue due to the uncontrollable neighborhood expansion in the aggregation stage. Several sampling algorithms have been proposed to limit the neighborhood expansion. However, these algorithms focus on minimizing the variance in sampling to approximate the original aggregation. This leads to two critical problems: 1) low accuracy because the sampling policy is agnostic to the performance of the target task, and 2) vulnerability to noise or adversarial attacks on the graph.

At LinkedIn, we want to create economic opportunity for everyone in the global workforce. To make this happen, LinkedIn offers a reactive Job Search system, and a proactive Jobs You May Be Interested In (JYMBII) system to match the best candidates with their dream jobs. One of the most challenging tasks for developing these systems is to properly extract important skill entities from job postings and then target members with matched attributes. In this work, we show that the commonly used text-based salience and market-agnostic skill extraction approach is sub-optimal because it only considers skill mention and ignores the salient level of a skill and its market dynamics, i.e., the market supply and demand influence on the importance of skills. To address the above drawbacks, we present Job2Skills, our deployed salience and market-aware skill extraction system. The proposed Job2Skills shows promising results in improving the online performance of job recommendation (JYMBII) (+1.92% job apply) and skill suggestions for job posters (-37% suggestion rejection rate). Lastly, we present case studies to show interesting insights that contrast traditional skill recognition method and the proposed Job2Skills from occupation, industry, country, and individual skill levels. Based on the above promising results, we deployed the Job2Skills online to extract job targeting skills for all 20M job postings served at LinkedIn.

As the world's largest professional network, LinkedIn wants to create economic opportunity for everyone in the global workforce. One of its most critical missions is matching jobs with processionals. Improving job targeting accuracy and hire efficiency align with LinkedIn's Member First Motto. To achieve those goals, we need to understand unstructured job postings with noisy information. We applied deep transfer learning to create domain-specific job understanding models. After this, jobs are represented by professional entities, including titles, skills, companies, and assessment questions. To continuously improve LinkedIn's job understanding ability, we designed an expert feedback loop where we integrated job understanding models into LinkedIn's products to collect job posters' feedback. In this demonstration, we present LinkedIn's job posting flow and demonstrate how the integrated deep job understanding work improves job posters' satisfaction and provides significant metric lifts in LinkedIn's job recommendation system.

At LinkedIn, we want to create economic opportunity for everyone in the global workforce. A critical aspect of this goal is matching jobs with qualified applicants. To improve hiring efficiency and reduce the need to manually screening each applicant, we develop a new product where recruiters can ask screening questions online so that they can filter qualified candidates easily. To add screening questions to all 20M active jobs at Linked In, we propose a new task that aims to automatically generate screening questions for a given job posting. To solve the task of generating screening questions, we develop a two-stage deep learning model called Job2Questions, where we apply a deep learning model to detect intent from the text description, and then rank the detected intents by their importance based on other contextual features. Since this is a new product with no historical data, we employ deep transfer learning to train complex models with limited training data. We launched the screening question product and our AI models to LinkedIn users and observed significant impact in the job marketplace. During our online A/B test, we observed +53.10% screening question suggestion acceptance rate, +22.17% job coverage, +190% recruiter-applicant interaction, and +11 Net Promoter Score. In sum, the deployed Job2Questions model helps recruiters to find qualified applicants and job seekers to find jobs they are qualified for.

Network representations have been shown to improve performance within a variety of tasks, including classification, clustering, and link prediction. However, most models either focus on moderate-sized, homogeneous networks or require a significant amount of auxiliary input to be provided by the user. Moreover, few works have studied network representations in real-world heterogeneous social networks with ambiguous social connections and are often incomplete. In the present work, we investigate the problem of learning low-dimensional node representations in heterogeneous professional social networks (HPSNs), which are incomplete and have ambiguous social connections. We present a general heterogeneous network representation learning model called Star2Vec that learns entity and person embeddings jointly using a social connection strength-aware biased random walk combined with a node-structure expansion function. Experiments on LinkedIn's Economic Graph and publicly available snapshots of Facebook's network show that Star2Vec outperforms existing methods on members' industry and social circle classification, skill and title clustering, and member-entity link predictions. We also conducted large-scale case studies to demonstrate practical applications of the Star2Vec embeddings trained on LinkedIn's Economic Graph such as next career move, alternative career suggestions, and general entity similarity searches.

The main mission of LinkedIn is to connect 610M+ members to the right opportunities. To find the right opportunities, LinkedIn needs to understand each member's skill set and their expertise levels accurately. However, estimating members' skill expertise is challenging due to lack of ground-truth. So far, the industry relied on either hand-created small scale data, or large scale social gestures containing a lot of social bias (e.g., endorsements).

PET image reconstruction is challenging due to the ill-poseness of the inverse problem and limited number of detected photons. Recently, the deep neural networks have been widely and successfully used in computer vision tasks and attracted growing interests in medical imaging. In this paper, we trained a deep residual convolutional neural network to improve PET image quality by using the existing inter-patient information. An innovative feature of the proposed method is that we embed the neural network in the iterative reconstruction framework for image representation, rather than using it as a post-processing tool. We formulate the objective function as a constrained optimization problem and solve it using the alternating direction method of multipliers algorithm. Both simulation data and hybrid real data are used to evaluate the proposed method. Quantification results show that our proposed iterative neural network method can outperform the neural network denoising and conventional penalized maximum likelihood methods.

LinkedIn dynamically delivers update activities from a user's interpersonal network to more than 300 million members in the personalized feed that ranks activities according their "relevance" to the user. This paper discloses the implementation details behind this personalized feed system at LinkedIn which can not be found from related work, and addresses the scalability and data sparsity challenges for deploying the system online. More specifically, we focus on the personalization models by generating three kinds of affinity scores: Viewer-ActivityType Affinity, Viewer-Actor Affinity, and Viewer-Actor-ActivityType Affinity. Extensive experiments based on online bucket tests (A/B experiments) and offline evaluation illustrate the effect of our personalization models in LinkedIn feed.

Networks are a general language for representing relational information among objects. An effective way to model, reason about, and summarize networks, is to discover sets of nodes with common connectivity patterns. Such sets are commonly referred to as network communities. Research on network community detection has predominantly focused on identifying communities of densely connected nodes in undirected networks.

Event sequences, such as patients' medical histories or users' sequences of product reviews, trace how individuals progress over time. Identifying common patterns, or progression stages, in such event sequences is a challenging task because not every individual follows the same evolutionary pattern, stages may have very different lengths, and individuals may progress at different rates. In this paper, we develop a model-based method for discovering common progression stages in general event sequences. We develop a generative model in which each sequence belongs to a class, and sequences from a given class pass through a common set of stages, where each sequence evolves at its own rate. We then develop a scalable algorithm to infer classes of sequences, while also segmenting each sequence into a set of stages. We evaluate our method on event sequences, ranging from patients' medical histories to online news and navigational traces from the Web. The evaluation shows that our methodology can predict future events in a sequence, while also accurately inferring meaningful progression stages, and effectively grouping sequences based on common progression patterns. More generally, our methodology allows us to reason about how event sequences progress over time, by discovering patterns and categories of temporal evolution in large-scale datasets of events.

Networks provide a powerful way to study complex systems of interacting objects. Detecting network communities-groups of objects that often correspond to functional modules-is crucial to understanding social, technological, and biological systems. Revealing communities allows for analysis of system properties that are invisible when considering only individual objects or the entire system, such as the identification of module boundaries and relationships or the classification of objects according to their functional roles. However, in networks where objects can simultaneously belong to multiple modules at once, the decomposition of a network into overlapping communities remains a challenge. Here we present a new paradigm for uncovering the modular structure of complex networks, based on a decomposition of a network into any combination of overlapping, nonoverlapping, and hierarchically organized communities. We demonstrate on a diverse set of networks coming from a wide range of domains that our approach leads to more accurate communities and improved identification of community boundaries. We also unify two fundamental organizing principles of complex networks: the modularity of communities and the commonly observed core-periphery structure. We show that dense network cores form as an intersection of many overlapping communities. We discover that communities in social, information, and food web networks have a single central dominant core while communities in protein-protein interaction (PPI) as well as product copurchasing networks have small overlaps and form many local cores.

One of the main organizing principles in real-world networks is that of network communities , where sets of nodes organize into densely linked clusters. Even though detection of such communities is of great interest, understanding the structure communities in large networks remains relatively limited. In particular, due to the unavailability of labeled ground-truth data, it was traditionally very hard to develop accurate models of network community structure. Here we use six large social, collaboration, and information networks where nodes explicitly state their ground-truth community memberships. For example, nodes in social networks join into explicitly defined interest based groups, and we use such groups as explicitly labeled ground-truth communities. We use such ground-truth communities to study their structural signatures by analyzing how ground-truth communities emerge in networks and how they overlap. We observe some surprising phenomena. First, ground-truth communities contain high-degree hub nodes that reside in community overlaps and link to most of the members of the community. Second, the overlaps of communities are more densely connected than the non-overlapping parts of communities. We show that this in contrast to the conventional wisdom that community overlaps are more sparsely connected than the non-overlapping parts themselves. We then show that many existing models of network communities do not capture dense community overlaps. This in turn means that most present models and community detection methods confuse overlaps as separate communities. In contrast, we present the community-affiliation graph model (AGM), a conceptual model of network community structure. We demonstrate that AGM reliably captures the overall structure of networks as well as the overlapping and hierarchical nature of network communities.

Community detection algorithms are fundamental tools that allow us to uncover organizational principles in networks. When detecting communities, there are two possible sources of information one can use: the network structure, and the features and attributes of nodes. Even though communities form around nodes that have common edges and common attributes, typically, algorithms have only focused on one of these two data modalities: community detection algorithms traditionally focus only on the network structure, while clustering algorithms mostly consider only node attributes. In this paper, we develop Communities from Edge Structure and Node Attributes (CESNA), an accurate and scalable algorithm for detecting overlapping communities in networks with node attributes. CESNA statistically models the interaction between the network structure and the node attributes, which leads to more accurate community detection as well as improved robustness in the presence of noise in the network structure. CESNA has a linear runtime in the network size and is able to process networks an order of magnitude larger than comparable approaches. Last, CESNA also helps with the interpretation of detected communities by finding relevant node attributes for each community.

In an era of information overload, many people struggle to make sense of complex stories, such as presidential elections or economic reforms. We propose a methodology for creating structured summaries of information, which we call zoomable metro maps. Just as cartographic maps have been relied upon for centuries to help us understand our surroundings, metro maps can help us understand the information landscape.

Online social networks have become important channels for users to share content with their connections and diffuse information. Although much work has been done to identify socially influential users, the problem of finding "reputable" sharers, who share good content, has received relatively little attention. Availability of such reputation scores can be useful or various applications like recommending people to follow, procuring high quality content in a scalable way, creating a content reputation economy to incentivize high quality sharing, and many more. To estimate sharer reputation, it is intuitive to leverage data that records how recipients respond (through clicking, liking, etc.) to content items shared by a sharer. However, such data is usually biased --- it has a selection bias since the shared items can only be seen and responded to by users connected to the sharer in most social networks, and it has a response bias since the response is usually influenced by the relationship between the sharer and the recipient (which may not indicate whether the shared content is good). To correct for such biases, we propose to utilize an additional data source that provides unbiased goodness estimates for a small set of shared items, and calibrate biased social data through a novel multi-level hierarchical model that describes how the unbiased data and biased data are jointly generated according to sharer reputation scores. The unbiased data also provides the ground truth for quantitative evaluation of different methods. Experiments based on such ground-truth data show that our proposed model significantly outperforms existing methods that estimate social influence using biased social data.

Network communities represent basic structures for understanding the organization of real-world networks. A community (also referred to as a module or a cluster) is typically thought of as a group of nodes with more connections amongst its members than between its members and the remainder of the network. Communities in networks also overlap as nodes belong to multiple clusters at once. Due to the difficulties in evaluating the detected communities and the lack of scalable algorithms, the task of overlapping community detection in large networks largely remains an open problem.

Nodes in real-world networks organize into densely linked communities where edges appear with high concentration among the members of the community. Identifying such communities of nodes has proven to be a challenging task mainly due to a plethora of definitions of a community, intractability of algorithms, issues with evaluation and the lack of a reliable gold-standard ground-truth. In this paper we study a set of 230 large real-world social, collaboration and information networks where nodes explicitly state their group memberships. For example, in social networks nodes explicitly join various interest based social groups. We use such groups to define a reliable and robust notion of ground-truth communities. We then propose a methodology which allows us to compare and quantitatively evaluate how different structural definitions of network communities correspond to ground-truth communities. We choose 13 commonly used structural definitions of network communities and examine their sensitivity, robustness and performance in identifying the ground-truth. We show that the 13 structural definitions are heavily correlated and naturally group into four classes. We find that two of these definitions, Conductance and Triad-participation-ratio, consistently give the best performance in identifying ground-truth communities. We also investigate a task of detecting communities given a single seed node. We extend the local spectral clustering algorithm into a heuristic parameter-free community detection method that easily scales to networks with more than hundred million nodes. The proposed method achieves 30% relative improvement over current local clustering methods.

One of the main organizing principles in real-world networks is that of network communities, where sets of nodes organize into densely linked clusters. Communities in networks often overlap as nodes can belong to multiple communities at once. Identifying such overlapping communities is crucial for the understanding the structure as well as the function of real-world networks. Even though community structure in networks has been widely studied in the past, practically all research makes an implicit assumption that overlaps between communities are less densely connected than the non-overlapping parts themselves. Here we validate this assumption on 6 large scale social, collaboration and information networks where nodes explicitly state their community memberships. By examining such ground-truth communities we find that the community overlaps are more densely connected than the non-overlapping parts, which is in sharp contrast to the conventional wisdom that community overlaps are more sparsely connected than the communities themselves. Practically all existing community detection methods fail to detect communities with dense overlaps. We propose Community-Affiliation Graph Model, a model-based community detection method that builds on bipartite node-community affiliation networks. Our method successfully captures overlapping, non-overlapping as well as hierarchically nested communities, and identifies relevant communities more accurately than the state-of-the-art methods in networks ranging from biological to social and information networks.

Online content exhibits rich temporal dynamics, and diverse realtime user generated content further intensifies this process. However, temporal patterns by which online content grows and fades over time, and by which different pieces of content compete for attention remain largely unexplored. We study temporal patterns associated with online content and how the content’s popularity grows and fades over time. The attention that content receives on the Web varies depending on many factors and occurs on very different time scales and at different resolutions. In order to uncover the temporal dynamics of online content we formulate a time series clustering problem using a similarity metric that is invariant to scaling and shifting. We develop the K-Spectral Centroid (K-SC) clustering algorithm that effectively finds cluster centroids with our similarity measure. By applying an adaptive wavelet-based incremental approach to clustering, we scale K-SC to large data sets. We demonstrate our approach on two massive datasets: a set of 580 million Tweets, and a set of 170 million blog posts and news media articles. We find that K-SC outperforms the K-means clustering algorithm in finding distinct shapes of time series. Our analysis shows that there are six main temporal shapes of attention of online content. We also present a simple model that reliably predicts the shape of attention by using information about only a small number of participants. Our analyses offer insight into common temporal patterns of the content on the Web and broaden the understanding of the dynamics of human attention.

Social media forms a central domain for the production and dissemination of real-time information. Even though such flows of information have traditionally been thought of as diffusion processes over social networks, the underlying phenomena are the result of a complex web of interactions among numerous participants. Here we develop the Linear Influence Model where rather than requiring the knowledge of the social network and then modeling the diffusion by predicting which node will influence which other nodes in the network, we focus on modeling the global influence of a node on the rate of diffusion through the (implicit) network. We model the number of newly infected nodes as a function of which other nodes got infected in the past. For each node we estimate an influence function that quantifies how many subsequent infections can be attributed to the influence of that node over time. A nonparametric formulation of the model leads to a simple least squares problem that can be solved on large datasets. We validate our model on a set of 500 million tweets and a set of 170 million news articles and blog posts. We show that the Linear Influence Model accurately models influences of nodes and reliably predicts the temporal dynamics of information diffusion. We find that patterns of influence of individual participants differ significantly depending on the type of the node and the topic of the information.