Existing question answering (QA) datasets are no longer challenging to most powerful Large Language Models (LLMs). Traditional QA benchmarks like TriviaQA, NaturalQuestions, ELI5 and HotpotQA mainly study ''known unknowns'' with clear indications of both what information is missing, and how to find it to answer the question. A yet unmet need of the NLP community is a bank of non-factoid, multi-perspective questions involving a great deal of unclear information needs, i.e. ''unknown unknowns''. We claim we can find such questions in search engine logs, which is surprising because most question-intent queries are indeed factoid. Furthermore, recent products like Google's DeepResearch (announced a year after this resource was released publicly) specifically address such queries, retrieving hundreds of documents to synthesize report-style responses. We present Researchy Questions, the world's first, only and largest public dataset of ''Deep Research'' questions filtered from real search engine logs to be non-factoid, ''decompositional'' and multi-perspective. We show that users spend substantial ''effort'' on these questions in terms of signals like clicks and session length. We also show that ''slow thinking'' answering techniques, like decomposition into sub-questions shows benefit over answering directly. We release (at https://huggingface.co/datasets/corbyrosset/researchy_questions) about 100k Researchy Questions with a permissive CDLA-2.0 license, along with click histograms on over 350k Clueweb22 URLs that were clicked for each question.

Understanding user intents in information access scenarios can help us provide more relevant and personalized search results and recommendations. However, analyzing user intents is not easy, especially for emerging forms of Web search such as Artificial Intelligence (AI)-driven chat. To understand user intents from retrospective log data, we need a way to label them with meaningful categories that capture their diversity and dynamics. Existing methods rely on manual or Machine-Learned (ML) labeling, which is either expensive or inflexible for large and dynamic datasets. Large Language Models (LLMs) could generate rich and relevant concepts, descriptions, and examples for user intents using log data of user interactions. However, using LLMs to generate a user intent taxonomy and applying it for a given Information Retrieval (IR) application can be problematic for two main reasons: (1) such a taxonomy is not externally validated; and (2) there may be an undesirable feedback loop if an LLM does both these tasks without external validation. To address this, we propose a new methodology with human experts and assessors to verify the quality of the LLM-generated taxonomy. We also present an end-to-end pipeline that uses an LLM with Human-in-the-Loop (HITL) to produce, refine, and apply labels for user intent analysis in log data. We demonstrate its effectiveness by uncovering new insights into user intents from search and chat logs from the Microsoft Bing Web search engine. The novelty in this research stems from the method for generating purpose-driven user intent taxonomies with strong validation. Our approach not only helps remove methodological and practical bottlenecks from intent-focused research, but also provides a new framework for generating, validating, and applying other kinds of taxonomies in a scalable and adaptable way, with reasonable human effort.

Transforming unstructured text into structured and meaningful forms, organized by useful category labels, is a fundamental step in text mining for downstream analysis and application. However, most existing methods for producing label taxonomies and building text-based label classifiers still rely heavily on domain expertise and manual curation, making the process expensive and time-consuming. This is particularly challenging when the label space is under-specified and large-scale data annotations are unavailable. In this paper, we address these challenges with Large Language Models (LLMs), whose prompt-based interface facilitates the induction and use of large-scale pseudo labels. We propose TnT-LLM, a two-phase framework that employs LLMs to automate the process of end-to-end label generation and assignment with minimal human effort for any given use-case. In the first phase, we introduce a zero-shot, multi-stage reasoning approach which enables LLMs to produce and refine a label taxonomy iteratively. In the second phase, LLMs are used as data labelers that yield training samples so that lightweight supervised classifiers can be reliably built, deployed, and served at scale. We apply TnT-LLM to the analysis of user intent and conversational domain for Bing Copilot (formerly Bing Chat), an open-domain chat-based search engine. Extensive experiments using both human and automatic evaluation metrics demonstrate that TnT-LLM generates more accurate and relevant label taxonomies when compared against state-of-the-art baselines, and achieves a favorable balance between accuracy and efficiency for classification at scale.

Recent work has shown that modeling higher-order structures, such as motifs or graphlets, can capture the complex network structure and dynamics in a variety of graph domains (e.g., social sciences, biology, chemistry). However, many dynamic networks are not only rich in structure, but also in content information. For example, an academic citation network has content such as the title and abstracts of the papers. Currently, there is a lack of generative models for dynamic networks that also generate content. To address this gap, in this work we propose DYnamic Attributed Node rolEs (DYANE)-a generative model that (i) captures network structure dynamics through temporal motifs, and (ii) extends the structural roles of nodes in motifs (e.g., a node acting as a hub in a wedge) to roles that generate content embeddings. We evaluate DYANE on real-world networks against other dynamic graph generative model baselines. DYANE outperforms the baselines in graph structure and node behavior, improving the KS score for graph metrics by 21-31% and node metrics by 17-27% on average, and produces content embeddings similar to the observed network. We also derive a methodology to evaluate the content embeddings generated by nodes, taking into account keywords extracted from the content (as topic representations), and using distance metrics.

Email platforms need to generate personalized rankings of emails that satisfy user preferences, which may vary over time. We approach this as a recommendation problem based on three criteria: closeness (how relevant the sender and topic are to the user), timeliness (how recent the email is), and conciseness (how brief the email is). We propose MOSR (Multi-Objective Stationary Recommender), a novel online algorithm that uses an adaptive control model to dynamically balance these criteria and adapt to preference changes. We evaluate MOSR on the Enron Email Dataset, a large collection of real emails, and compare it with other baselines. The results show that MOSR achieves better performance, especially under non-stationary preferences, where users value different criteria more or less over time. We also test MOSR's robustness on a smaller down-sampled dataset that exhibits high variance in email characteristics, and show that it maintains stable rankings across different samples. Our work offers novel insights into how to design email re-ranking systems that account for multiple objectives impacting user satisfaction.

Workplace communication software such as Microsoft Teams, Slack, and Google Workspace have become integral to workplace collaboration, especially due to the rise of remote work. By making it easier to access relevant or useful information, recommender systems for these platforms have the potential to improve efficient cross-team information flow through a company's communication network. While there has been some recent work on recommendation approaches that optimize network objectives, these have focused on static graphs. In this work, we focus on optimizing information flow, which is highly temporal and presents a number of novel algorithmic challenges. To overcome these, we develop tractable measures of temporal information flow and design efficient online recommendation algorithms that jointly optimize for relevance and cross-team information flow. We demonstrate the potential for impact of these approaches on a rich multi-modal dataset capturing one month of communication between 180k Microsoft employees through email, chats and posts on Microsoft Teams, and file sharing on SharePoint. We design an offline model-based evaluation pipeline to estimate the effects of recommendations on the temporal communication network. We show that our recommendation algorithms can significantly improve cross-team information flow with only a small decrease in traditional relevance metrics.

Temporal graph representation learning (T-GRL) aims to learn representations that model how graph edges evolve over time. While recent works on T-GRL have improved link prediction accuracy in temporal settings, their methods optimize a point-wise loss function independently over future links rather than optimize jointly over a candidate set per node. In applications where resources (e.g., attention) are allocated based on ranking links by likelihood, the use of a ranking loss is preferred. However it is not straightforward to develop a T-GRL method to optimize a ranking loss due to a tradeoff between model expressivity and scalability. In this work, we address these issues and propose a Temporal Graph network for Ranking (TGRank), which significantly improves performance for link prediction tasks by (i) optimizing a list-wise loss for improved ranking, and (ii) incorporating a labeling approach designed to allow for efficient inference over the candidate set jointly, while provably boosting expressivity. We extensively evaluate TGRank over six real networks. TGRank outperforms the state-of-the-art baselines on average by 14.21%↑ (transductive) and 16.25% ↑ (inductive) in ranking metrics while being more efficient (up-to 65 × speed-up) to make inference on large networks.

Graph neural networks (GNNs) have achieved tremendous success on multiple graph-based learning tasks by fusing network structure and node features. Modern GNN models are built upon iterative aggregation of neighbor's/proximity features by message passing. Its prediction performance has been shown to be strongly bounded by assortative mixing in the graph, a key property wherein nodes with similar attributes mix/connect with each other. We observe that real world networks exhibit heterogeneous or diverse mixing patterns and the conventional global measurement of assortativity, such as global assortativity coefficient, may not be a representative statistic in quantifying this mixing. We adopt a generalized concept, node-level assortativity, one that is based at the node level to better represent the diverse patterns and accurately quantify the learnability of GNNs. We find that the prediction performance of a wide range of GNN models is highly correlated with the node level assortativity. To break this limit, in this work, we focus on transforming the input graph into a computation graph which contains both proximity and structural information as distinct type of edges. The resulted multi-relational graph has an enhanced level of assortativity and, more importantly, preserves rich information from the original graph. We then propose to run GNNs on this computation graph and show that adaptively choosing between structure and proximity leads to improved performance under diverse mixing. Empirically, we show the benefits of adopting our transformation framework for semi-supervised node classification task on a variety of real world graph learning benchmarks.

Motifs, which have been established as building blocks for network structure, move beyond pair-wise connections to capture longer-range correlations in connections and activity. In spite of this, there are few generative graph models that consider higher-order network structures and even fewer that focus on using motifs in models of dynamic graphs. Most existing generative models for temporal graphs strictly grow the networks via edge addition, and the models are evaluated using static graph structure metrics—which do not adequately capture the temporal behavior of the network. To address these issues, in this work we propose DYnamic MOtif-NoDes (DYMOND)—a generative model that considers (i) the dynamic changes in overall graph structure using temporal motif activity and (ii) the roles nodes play in motifs (e.g., one node plays the hub role in a wedge, while the remaining two act as spokes). We compare DYMOND to three dynamic graph generative model baselines on real-world networks and show that DYMOND performs better at generating graph structure and node behavior similar to the observed network. We also propose a new methodology to adapt graph structure metrics to better evaluate the temporal aspect of the network. These metrics take into account the changes in overall graph structure and the individual nodes’ behavior over time.

Network embedding models aim to learn low-dimensional representations for nodes and/or edges in graphs. For social networks, learning edge representations is especially beneficial as we need to describe or explain the relationships, activities, and interactions between users. Existing approaches that learn stand-alone node embeddings, and represent edges as pairs of node embeddings, are limited in their applicability because nodes participate in multiple relationships, which should be considered. In addition, social networks often contain multiple types of edges, which yields multi-view contexts that need to be considered in the representation. In this paper, we propose a new methodology, MERL, that (1) captures asymmetry in multiple views by learning well-defined edge representations that are responsive to the difference between the source and destination node roles, and (2) incorporates textual communications to identify multiple source of social signals (e.g. strength and affinity) that moderate the impact of different views between users. Our experiments show that MERL outperforms alternative state-of-the-art embedding models on link prediction and multilabel classification tasks across multiple views in social network datasets. We further analyze the learned embeddings of MERL and demonstrate they are more correlated with the existence of view-based edges compared to previous methods.

In many applications, a parsimonious model is often preferred for better interpretability and predictive performance. Online algorithms have been studied extensively for building such models in big data and fast evolving environments, with a prominent example, FTRL-proximal [1]. However, existing methods typically do not provide confidence levels, and with the usage of L1 regularization, the model estimation can be undermined by the uniform shrinkage on both relevant and irrelevant features. To address these issues, we developed OLSS, a Bayesian online sparse learning algorithm based on the spike-and-slab prior. OLSS achieves the same scalability as FTRL-proximal, but realizes appealing selective shrinkage and produces rich uncertainty information, such as posterior inclusion probabilities and feature weight variances. On the tasks of text classification and click-through-rate (CTR) prediction for Yahoo!'s display and search advertisement platforms, OLSS often demonstrates superior predictive performance to the state-of-the-art methods in industry, including Vowpal Wabbit [2] and FTRL-proximal.

The problem of knowledge graph (KG) reasoning has been widely explored by traditional rule-based systems and more recently by knowledge graph embedding methods. While logical rules can capture deterministic behavior in a KG, they are brittle and mining ones that infer facts beyond the known KG is challenging. Probabilistic embedding methods are effective in capturing global soft statistical tendencies and reasoning with them is computationally efficient. While embedding representations learned from rich training data are expressive, incompleteness and sparsity in real-world KGs can impact their effectiveness. We aim to leverage the complementary properties of both methods to develop a hybrid model that learns both high-quality rules and embeddings simultaneously. Our method uses a cross feedback paradigm wherein an embedding model is used to guide the search of a rule mining system to mine rules and infer new facts. These new facts are sampled and further used to refine the embedding model. Experiments on multiple benchmark datasets show the effectiveness of our method over other competitive standalone and hybrid baselines. We also show its efficacy in a sparse KG setting.

In many complex domains, the input data are often not suited for the typical vector representations used in deep learning models. For example, in relational learning and computer vision tasks, the data are often better represented as sets (e.g., the neighborhood of a node, a cloud of points). In these cases, a key challenge is to learn an embedding function that is invariant to permutations of the input. While there has been some recent work on principled methods for learning permutation-invariant representations of sets, these approaches are limited in their applicability to set-of-sets (SoS) tasks, such as subgraph prediction and scene classification. In this work, we develop a deep neural network framework to learn inductive SoS embeddings that are invariant to SoS permutations. Specifically, we propose HATS, a hierarchical sequence model with attention mechanisms for inductive set-of-sets embeddings. We develop stochastic optimization and inference methods for learning HATS, and our experiments demonstrate that HATS achieves superior performance across a wide range of set-of-sets tasks.

We consider hypothesis testing and anomaly detection on datasets where each observation is a weighted network. Current approaches to hypothesis testing for weighted networks typically require thresholding the edge-weights, to transform the data to binary networks. This results in a loss of information, and outcomes are sensitive to choice of threshold levels. Our work avoids this, and we consider weighted-graph observations in two situations, 1) where each graph belongs to one of two populations, and 2) where entities belong to one of two populations, with each entity possessing multiple graphs (indexed e.g. by time). We propose a hierarchical Bayesian hypothesis testing framework that models each population with a mixture of latent space models for weighted networks, and then tests populations of networks for differences in distribution over components.Our framework is capable of population-level, entity-specific, as well as edge-specific hypothesis testing. We apply it to synthetic data and two real-world datasets: a social media dataset involving word co-occurrences from discussions on Twitter of the political unrest in Brazil, and a medical dataset involving fMRI brain-scans of human subjects. The results show that our proposed method has lower Type-I error and higher statistical power compared to previous alternatives that need to threshold the edge weights.

With the availability of vast amounts of user visitation history on location-based social networks (LBSN), the problem of Point-of- Interest (POI) prediction has been extensively studied. However, much of the research has been conducted solely on voluntary check-in datasets collected from social apps such as Foursquare or Yelp. While these data contain rich information about recreational activities (e.g., restaurants, nightlife, and entertainment), information about more prosaic aspects of people's lives is sparse. is not only limits our understanding of users' daily routines, but more importantly the modeling assumptions developed based on characteristics of recreation-based data may not be suitable for richer check-in data. In this work, we present an analysis of education "check-in" data using WiFi access logs collected at Purdue University. We propose a heterogeneous graph-based method to encode the correlations between users, POIs, and activities, and then jointly learn embeddings for the vertices. We evaluate our method compared to previous state-of-the-art POI prediction methods, and show that the assumptions made by previous methods signicantly degrade performance on our data with dense(r) activity signals. We also show how our learned embeddings could be used to identify similar students (e.g., for friend suggestions).

The explosion of interest in KDD and other Data Science/Machine Learning/AI conferences is just one of the many signs that these technologies are no longer confined to the realms of academia and a hand-full of tech companies. As our daily lives seamlessly integrate more and more data-driven applications, people's excitement is tempered by worry about the technologies' potential to disrupt their existence. Having worked for almost 30 years to design and develop these technologies, the KDD community now should examine and debate the impact of Machine Learning & AI on the broader world. Beyond the hype, where do we stand with respect to the dangers? What role can our community play to alleviate concerns around AI taking jobs, or taking over? How can the value derived from data be distributed fairly? Are concerns about inequity well-founded or rather largely problems of perception? What can be done to bring data hunger and data sharing concerns to a level of equilibrium? How do we prepare people to interact with intelligent systems at scale? Can we unleash the incredible responsiveness of the KDD community toward longer-term more impactful projects across sectors that are essential for social good, such as Health, Environmental Sustainability, and Public Welfare.

An important task in network analysis is the detection of anomalous events in a network time series. These events could merely be times of interest in the network timeline or they could be examples of malicious activity or network malfunction. Hypothesis testing using network statistics to summarize the behavior of the network provides a robust framework for the anomaly detection decision process. Unfortunately, choosing network statistics that are dependent on confounding factors like the total number of nodes or edges can lead to incorrect conclusions (e.g., false positives and false negatives). In this article, we describe the challenges that face anomaly detection in dynamic network streams regarding confounding factors. We also provide two solutions to avoiding error due to confounding factors: the first is a randomization testing method that controls for confounding factors, and the second is a set of size-consistent network statistics that avoid confounding due to the most common factors, edge count and node count.

Much of the past work on mining and modeling networks has focused on understanding the observed properties of single example graphs. However, in many real-life applications it is important to characterize the structure of populations of graphs. In this work, we analyze the distributional properties of probabilistic generative graph models (PGGMs) for network populations. PGGMs are statistical methods that model the network distribution and match common characteristics of real-world networks. Specifically, we show that most PGGMs cannot reflect the natural variability in graph properties observed across multiple networks because their edge generation process assumes independence among edges. Then, we propose the mixed Kronecker Product Graph Model (mKPGM), a scalable generalization of KPGMs that uses tied parameters to increase the variability of the sampled networks, while preserving the edge probabilities in expectation. We compare mKPGM to several other graph models. The results show that learned mKPGMs accurately represent the characteristics of real-world networks, while also effectively capturing the natural variability in network structure.

Network sampling is a widely used procedure in social network analysis where a random network is sampled from a generative network model (GNM). Recently proposed GNMs, allow generation of networks with more realistic structural characteristics than earlier ones. This facilitates tasks such as hypothesis testing and sensitivity analysis. However, sampling of networks with correlated vertex attributes remains a challenging problem. While the recent work of \cite{Pfeiffer:14} has provided a promising approach for attributed-network sampling, the approach was developed for use with relatively simple GNMs and does not work well with more complex hierarchical GNMs (which can model the range of characteristics and variation observed in real world networks more accurately). In contrast to simple GNMs where the probability mass is spread throughout the space of edges more evenly, hierarchical GNMs concentrate the mass to smaller regions of the space to reflect dependencies among edges in the network---this produces more realistic network characteristics, but also makes it more difficult to identify candidate networks from the sampling space.

From social science to biology, numerous applications often rely on graphlets for intuitive and meaningful characterization of networks at both the global macro-level as well as the local micro-level. While graphlets have witnessed a tremendous success and impact in a variety of domains, there has yet to be a fast and efficient approach for computing the frequencies of these subgraph patterns. However, existing methods are not scalable to large networks with millions of nodes and edges, which impedes the application of graphlets to new problems that require large-scale network analysis. To address these problems, we propose a fast, efficient, and parallel algorithm for counting graphlets of size k={3,4}-nodes that take only a fraction of the time to compute when compared with the current methods used. The proposed graphlet counting algorithms leverages a number of proven combinatorial arguments for different graphlets. For each edge, we count a few graphlets, and with these counts along with the combinatorial arguments, we obtain the exact counts of others in constant time. On a large collection of 300+ networks from a variety of domains, our graphlet counting strategies are on average 460x faster than current methods. This brings new opportunities to investigate the use of graphlets on much larger networks and newer applications as we show in the experiments. To the best of our knowledge, this paper provides the largest graphlet computations to date as well as the largest systematic investigation on over 300+ networks from a variety of domains.

Many individuals on social networking sites provide traits about themselves, such as interests or demographics. Social networking sites can use this information to provide better content to match their users' interests, such as recommending scheduled events or various relevant products. These tasks require accurate probability estimates to determine the correct answer to return. Relational machine learning (RML) is an excellent framework for these problems as it jointly models the user labels given their attributes and the relational structure. Further, semi-supervised learning methods could enable RML methods to exploit the large amount of unlabeled data in networks.

Many interesting domains in machine learning can be viewed as networks, with relationships (e.g., friendships) connecting items (e.g., individuals). The Active Exploration (AE) task is to identify all items in a network with a desired trait (i.e., positive labels) given only partial information about the network. The AE process iteratively queries for labels or network structure within a limited budget; thus, accurate predictions prior to making each query is critical to maximizing the number of positives gathered. However, the targeted AE query process produces partially observed networks that can create difficulties for predictive modeling. In particular, we demonstrate that these partial networks can exhibit extreme label correlation bias, which makes it difficult for conventional relational learning methods to accurately estimate relational parameters. To overcome this issue, we model the joint distribution of possible edges and labels to improve learning and inference. Our proposed method, Probabilistic Relational Expectation Maximization (PR-EM), is the first AE approach to accurately learn the complex dependencies between attributes, labels, and structure to improve predictions. PR-EM utilizes collective inference over the missing relationships in the partial network to jointly infer unknown item traits. Further, we develop a linear inference algorithm to facilitate efficient use of PR-EM in large networks. We test our approach on four real world networks, showing that AE with PR-EM gathers significantly more positive items compared to state-of-the-art methods.

The recent interest in modeling complex networks has fueled the development of generative graph models, such as Kronecker Product Graph Model (KPGM) and mixed KPGM (mKPGM). The Kronecker family of models are appealing because of their elegant fractal structure, as well as their ability to capture important network characteristics such as degree, diameter, and (in the case of mKPGM) clustering and population variance. In addition, scalable sampling algorithms for KPGMs made the analysis of large-scale, sparse networks feasible for the first time. In this work, we show that the scalable sampling methods, in contrast to prior belief, do not in fact sample from the underlying KPGM distribution and often result in sampling graphs that are very unlikely. To address this issue, we develop a new representation that exploits the structure of Kronecker models and facilitates the development of novel grouped sampling methods that are provably correct. In this paper, we outline efficient algorithms to sample from mKPGMs and KPGMs based on these ideas. Notably, our mKPGM algorithm is the first available scalable sampling method for this model and our KPGM algorithm is both faster and more accurate than previous scalable methods. We conduct both theoretical analysis and empirical evaluation to demonstrate the strengths of our algorithms and show that we can sample a network with 75 million edges in 87 seconds on a single processor.

The prevalence of datasets that can be represented as networks has recently fueled a great deal of work in the area of Relational Machine Learning (RML). Due to the statistical correlations between linked nodes in the network, many RML methods focus on predicting node features (i.e., labels) using the network relationships. However, many domains are comprised of a single, partially-labeled network. Thus, relational versions of Expectation Maximization (i.e., R-EM), which jointly learn parameters and infer the missing labels, can outperform methods that learn parameters from the labeled data and apply them for inference on the unlabeled nodes. Although R-EM methods can significantly improve predictive performance in networks that are densely labeled, they do not achieve the same gains in sparsely labeled networks and can perform worse than RML methods. In this work, we show the fixed-point methods that R-EM uses for approximate learning and inference result in errors that prevent convergence in sparsely labeled networks. We then propose two methods that do not experience this problem. First, we develop a Relational Stochastic EM (R-SEM) method, which uses stochastic parameters that are not as susceptible to approximation errors. Then we develop a Relational Data Augmentation (R-DA) method, which integrates over a range of stochastic parameter values for inference. R-SEM and R-DA can use any collective RML algorithm for learning and inference in partially labeled networks. We analyze their performance with two RML learners over four real world datasets, and show that they outperform independent learning, RML and R-EM -- particularly in sparsely labeled networks.

Sampling is a standard approach in big-graph analytics; the goal is to efficiently estimate the graph properties by consulting a sample of the whole population. A perfect sample is assumed to mirror every property of the whole population. Unfortunately, such a perfect sample is hard to collect in complex populations such as graphs (e.g. web graphs, social networks), where an underlying network connects the units of the population. Therefore, a good sample will be representative in the sense that graph properties of interest can be estimated with a known degree of accuracy.

Online social networks have become ubiquitous to today's society and the study of data from these networks has improved our understanding of the processes by which relationships form. Research in statistical relational learning focuses on methods to exploit correlations among the attributes of linked nodes to predict user characteristics with greater accuracy. Concurrently, research on generative graph models has primarily focused on modeling network structure without attributes, producing several models that are able to replicate structural characteristics of networks such as power law degree distributions or community structure. However, there has been little work on how to generate networks with real-world structural properties and correlated attributes. In this work, we present the Attributed Graph Model (AGM) framework to jointly model network structure and vertex attributes. Our framework learns the attribute correlations in the observed network and exploits a generative graph model, such as the Kronecker Product Graph Model (KPGM) and Chung Lu Graph Model (CL), to compute structural edge probabilities. AGM then combines the attribute correlations with the structural probabilities to sample networks conditioned on attribute values, while keeping the expected edge probabilities and degrees of the input graph model. We outline an efficient method for estimating the parameters of AGM, as well as a sampling method based on Accept-Reject sampling to generate edges with correlated attributes. We demonstrate the efficiency and accuracy of our AGM framework on two large real-world networks, showing that AGM scales to networks with hundreds of thousands of vertices, as well as having high attribute correlation.

The recent interest in networks-social, physical, communication, information, etc.-has fueled a great deal of research on the analysis and modeling of graphs. However, many of the analyses have focused on a single large network (e.g., a sub network sampled from Facebook). Although several studies have compared networks from different domains or samples, they largely focus on empirical exploration of network similarities rather than explicit tests of hypotheses. This is in part due to a lack of statistical methods to determine whether two large networks are likely to have been drawn from the same underlying graph distribution. Research on across-network hypothesis testing methods has been limited by (i) difficulties associated with obtaining a set of networks to reason about the underlying graph distribution, and (ii) limitations of current statistical models of graphs that make it difficult to represent variations across networks. In this paper, we exploit the recent development of mixed-Kronecker Product Graph Models, which accurately capture the natural variation in real world graphs, to develop a model-based approach for hypothesis testing in networks.

There has recently been a great deal of work focused on developing statistical models of graph structure---with the goal of modeling probability distributions over graphs from which new, similar graphs can be generated by sampling from the estimated distributions. Although current graph models can capture several important characteristics of social network graphs (e.g., degree, path lengths), many of them do not generate graphs with sufficient variation to reflect the natural variability in real world graph domains. One exception is the mixed Kronecker Product Graph Model (mKPGM), a generalization of the Kronecker Product Graph Model, which uses parameter tying to capture variance in the underlying distribution [10]. The enhanced representation of mKPGMs enables them to match both the mean graph statistics and their spread as observed in real network populations, but unfortunately to date, the only method to estimate mKPGMs involves an exhaustive search over the parameters.

Given a large time-evolving graph, how can we model and characterize the temporal behaviors of individual nodes (and network states)? How can we model the behavioral transition patterns of nodes? We propose a temporal behavior model that captures the "roles" of nodes in the graph and how they evolve over time. The proposed dynamic behavioral mixed-membership model (DBMM) is scalable, fully automatic (no user-defined parameters), non-parametric/data-driven (no specific functional form or parameterization), interpretable (identifies explainable patterns), and flexible (applicable to dynamic and streaming networks). Moreover, the interpretable behavioral roles are generalizable and computationally efficient. We applied our model for (a) identifying patterns and trends of nodes and network states based on the temporal behavior, (b) predicting future structural changes, and (c) detecting unusual temporal behavior transitions. The experiments demonstrate the scalability, flexibility, and effectiveness of our model for identifying interesting patterns, detecting unusual structural transitions, and predicting the future structural changes of the network and individual nodes.

The popularity of online social networks and social media has increased the amount of linked data available in Web domains. Relational and Gaussian Markov networks have both been applied successfully for classification in these relational settings. However, since Gaussian Markov networks model joint distributions over continuous label space, it is difficult to use them to reason about uncertainty in discrete labels. On the other hand, relational Markov networks model probability distributions over discrete label space, but since they condition on the graph structure, the marginal probability for an instance will vary based on the structure of the subnetwork observed around the instance. This implies that the marginals will not be identical across instances and can sometimes result in poor prediction performance. In this work, we propose a novel latent relational model based on copulas which allows use to make predictions in a discrete label space while ensuring identical marginals and at the same time incorporating some desirable properties of modeling relational dependencies in a continuous space. While copulas have recently been used for descriptive modeling, they have not been used for collective classification in large scale network data and the associated conditional inference problem has not been considered before. We develop an approximate inference algorithm, and demonstrate empirically that our proposed Copula Latent Markov Network models based on approximate inference outperform a number of competing relational classification models over a range of real-world relational classification tasks.

Network sampling is integral to the analysis of social, information, and biological networks. Since many real-world networks are massive in size, continuously evolving, and/or distributed in nature, the network structure is often sampled in order to facilitate study. For these reasons, a more thorough and complete understanding of network sampling is critical to support the field of network science. In this paper, we outline a framework for the general problem of network sampling by highlighting the different objectives, population and units of interest, and classes of network sampling methods. In addition, we propose a spectrum of computational models for network sampling methods, ranging from the traditionally studied model based on the assumption of a static domain to a more challenging model that is appropriate for streaming domains. We design a family of sampling methods based on the concept of graph induction that generalize across the full spectrum of computational models (from static to streaming) while efficiently preserving many of the topological properties of the input graphs. Furthermore, we demonstrate how traditional static sampling algorithms can be modified for graph streams for each of the three main classes of sampling methods: node, edge, and topology-based sampling. Experimental results indicate that our proposed family of sampling methods more accurately preserve the underlying properties of the graph in both static and streaming domains. Finally, we study the impact of network sampling algorithms on the parameter estimation and performance evaluation of relational classification algorithms.

We present a theoretical analysis framework that shows how ensembles of collective classifiers can improve predictions for graph data. We show how collective ensemble classification reduces errors due to variance in learning and more interestingly inference. We also present an empirical framework that includes various ensemble techniques for classifying relational data using collective inference. The methods span single- and multiple-graph network approaches, and are tested on both synthetic and real world classification tasks. Our experimental results, supported by our theoretical justifications, confirm that ensemble algorithms that explicitly focus on both learning and inference processes and aim at reducing errors associated with both, are the best performers.

Recently, a number of modeling techniques have been developed for data mining and machine learning in relational and network domains where the instances are not independent and identically distributed (i.i.d.). These methods specifically exploit the statistical dependencies among instances in order to improve classification accuracy. However, there has been little focus on how these same dependencies affect our ability to draw accurate conclusions about the performance of the models. More specifically, the complex link structure and attribute dependencies in relational data violate the assumptions of many conventional statistical tests and make it difficult to use these tests to assess the models in an unbiased manner. In this work, we examine the task of within-network classification and the question of whether two algorithms will learn models that will result in significantly different levels of performance. We show that the commonly used form of evaluation (paired t -test on overlapping network samples) can result in an unacceptable level of Type I error. Furthermore, we show that Type I error increases as (1) the correlation among instances increases and (2) the size of the evaluation set increases (i.e., the proportion of labeled nodes in the network decreases). We propose a method for network cross-validation that combined with paired t -tests produces more acceptable levels of Type I error while still providing reasonable levels of statistical power (i.e., 1−Type II error).

Recent empirical evaluation has shown that the performance of collective classification models can vary based on the amount of class label information available for use during inference. In this paper, we further demonstrate that the relative performance of statistical relational models learned with different estimation methods changes as the availability of test set labels increases. We reason about the cause of this phenomenon from an information-theoretic perspective and this points to a previously unidentified consideration in the development of relational learning algorithms. In particular, we characterize the high propagation error of collective inference models that are estimated with maximum pseudolikelihood estimation (MPLE), and show how this affects performance across the spectrum of label availability when compared to MLE, which has low propagation error. Our formal study leads to a quantitative characterization that can be used to predict the confidence of local propagation for MPLE models. We use this to propose a mixture model that can learn the best trade-off between high and low propagation models. Empirical evaluation on synthetic and real-world data show that our proposed method achieves comparable, or superior, results to both MPLE and low propagation models across the full spectrum of label availability.

In this paper we present GDR, a Guided Data Repair framework that incorporates user feedback in the cleaning process to enhance and accelerate existing automatic repair techniques while minimizing user involvement. GDR consults the user on the updates that are most likely to be beneficial in improving data quality. GDR also uses machine learning methods to identify and apply the correct updates directly to the database without the actual involvement of the user on these specific updates. To rank potential updates for consultation by the user, we first group these repairs and quantify the utility of each group using the decision-theory concept of value of information (VOI). We then apply active learning to order updates within a group based on their ability to improve the learned model. User feedback is used to repair the database and to adaptively refine the training set for the model. We empirically evaluate GDR on a real-world dataset and show significant improvement in data quality using our user guided repairing process. We also, assess the trade-off between the user efforts and the resulting data quality.

Real-world databases often contain syntactic and semantic errors, in spite of integrity constraints and other safety measures incorporated into modern DBMSs. We present ERACER, an iterative statistical framework for inferring missing information and correcting such errors automatically. Our approach is based on belief propagation and relational dependency networks, and includes an efficient approximate inference algorithm that is easily implemented in standard DBMSs using SQL and user defined functions. The system performs the inference and cleansing tasks in an integrated manner, using shrinkage techniques to infer correct values accurately even in the presence of dirty data. We evaluate the proposed methods empirically on multiple synthetic and real-world data sets. The results show that our framework achieves accuracy comparable to a baseline statistical method using Bayesian networks with exact inference. However, our framework has wider applicability than the Bayesian network baseline, due to its ability to reason with complex, cyclic relational dependencies.

Improving data quality is a time-consuming, labor-intensive and often domain specific operation. Existing data repair approaches are either fully automated or not efficient in interactively involving the users. We present a demo of GDR, a Guided Data Repair system that uses a novel approach to efficiently involve the user alongside automatic data repair techniques to reach better data quality as quickly as possible. Specifically, GDR generates data repairs and acquire feedback on them that would be most beneficial in improving the data quality. GDR quantifies the data quality benefit of generated repairs by combining mechanisms from decision theory and active learning. Based on these benefit scores, groups of repairs are ranked and displayed to the user. User feedback is used to train a machine learning component to eventually replace the user in deciding on the validity of a suggested repair. We describe how the generated repairs are ranked and displayed to the user in a "useful-looking" way and demonstrate how data quality can be effectively improved with minimal feedback from the user.

Relational autocorrelation is ubiquitous in relational domains. This observed correlation between class labels of linked instances in a network (e.g., two friends are more likely to share political beliefs than two randomly selected people) can be due to the effects of two different social processes. If social influence effects are present, instances are likely to change their attributes to conform to their neighbor values. If homophily effects are present, instances are likely to link to other individuals with similar attribute values. Both these effects will result in autocorrelated attribute values. When analyzing static relational networks it is impossible to determine how much of the observed correlation is due each of these factors. However, the recent surge of interest in social networks has increased the availability of dynamic network data. In this paper, we present a randomization technique for temporal network data where the attributes and links change over time. Given data from two time steps, we measure the gain in correlation and assess whether a significant portion of this gain is due to influence and/or homophily. We demonstrate the efficacy of our method on semi-synthetic data and then apply the method to a real-world social networks dataset, showing the impact of both influence and homophily effects.

Previous work analyzing social networks has mainly focused on binary friendship relations. However, in online social networks the low cost of link formation can lead to networks with heterogeneous relationship strengths (e.g., acquaintances and best friends mixed together). In this case, the binary friendship indicator provides only a coarse representation of relationship information. In this work, we develop an unsupervised model to estimate relationship strength from interaction activity (e.g., communication, tagging) and user similarity. More specifically, we formulate a link-based latent variable model, along with a coordinate ascent optimization procedure for the inference. We evaluate our approach on real-world data from Facebook, showing that the estimated link weights result in higher autocorrelation and lead to improved classification accuracy. 1

Recently a number of modeling techniques have been developed for data mining and machine learning in relational and network domains where the instances are not independent and identically distributed (i.i.d.). These methods specifically exploit the statistical dependencies among instances in order to improve classification accuracy. However, there has been little focus on how these same dependencies affect our ability to draw accurate conclusions about the performance of the models. More specifically, the complex link structure and attribute dependencies in network data violate the assumptions of many conventional statistical tests and make it difficult to use these tests to assess the models in an unbiased manner. In this work, we examine the task of within-network classification and the question of whether two algorithms will learn models which will result in significantly different levels of performance. We show that the commonly-used form of evaluation (paired t-test on overlapping network samples) can result in an unacceptable level of Type I error. Furthermore we show that Type I error increases as (1) the correlation among instances increases and (2) the size of the evaluation set increases (i.e., the proportion of labeled nodes in the network decreases). We propose a method for network cross-validation that combined with paired t-tests produces more acceptable levels of Type I error while still providing reasonable levels of statistical power (i.e., Type II error).

The inherent uncertainty of data present in numerous applications such as sensor databases, text annotations, and information retrieval motivate the need to handle imprecise data at the database level. Uncertainty can be at the attribute or tuple level and is present in both continuous and discrete data domains. This paper presents a model for handling arbitrary probabilistic uncertain data (both discrete and continuous) natively at the database level. Our approach leads to a natural and efficient representation for probabilistic data. We develop a model that is consistent with possible worlds semantics and closed under basic relational operators. This is the first model that accurately and efficiently handles both continuous and discrete uncertainty. The model is implemented in a real database system (PostgreSQL) and the effectiveness and efficiency of our approach is validated experimentally.

Recent research has shown that collective classification in relational data often exhibit significant performance gains over conventional approaches that classify instances individually. This is primarily due to the presence of autocorrelation in relational datasets, meaning that the class labels of related entities are correlated and inferences about one instance can be used to improve inferences about linked instances. Statistical relational learning techniques exploit relational autocorrelation by modeling global autocorrelation dependencies under the assumption that the level of autocorrelation is stationary throughout the dataset. To date, there has been no work examining the appropriateness of this stationarity assumption. In this paper, we examine two real-world datasets and show that there is significant variance in the autocorrelation dependencies throughout the relational data graphs. We develop a shrinkage technique for modeling this non-stationary autocorrelation and show that it achieves significant accuracy gains over competing techniques that model either local or global autocorrelation dependencies in isolation.

Many relational domains contain temporal information and dynamics that are important to model (e.g., social networks, protein networks). However, past work in relational learning has focused primarily on modeling static "snapshots" of the data and has largely ignored the temporal dimension of these data. In this work, we extend relational techniques to temporally-evolving domains and outline a representational framework that is capable of modeling both temporal and relational dependencies in the data. We develop efficient learning and inference techniques within the framework by considering a restricted set of temporal-relational dependencies and using parameter-tying methods to generalize across relationships and entities. More specifically, we model dynamic relational data with a two-phase process, first summarizing the temporal-relational information with kernel smoothing, and then moderating attribute dependencies with the summarized relational information. We develop a number of novel temporal-relational models using the framework and then show that the current approaches to modeling static relational data are special cases within the framework. We compare the new models to the competing static relational methods on three real-world datasets and show that the temporal-relational models consistently outperform the relational models that ignore temporal information - achieving significant reductions in error ranging from 15% to 70%.

In this work, we study the problem ofwithin-networkrelational learning and inference, where models are learned on a partially labeled relational dataset and then are applied to predict the classes of unlabeled instances in the same graph. We categorize recent work in statistical relational learning into three alternative approaches for this setting: disjoint learning with disjoint inference, disjoint learning with collective inference, and collective learning with collective inference. Models from each of these categories has been employed previously in different settings, but to our knowledge there has been no systematic comparison of models from all three categories. In this paper, we develop a novel pseudolikelihood EM method that facilitates more generalcollectivelearningandcollectiveinferenceon partially labeled relational networks. We then compare this method to competing methods from the other categories on both synthetic and real-world data. We show that collective learning and inference with the pseudolikelihood EM approach achieves significantly higher accuracy than the other types of models when there are a moderate number of labeled examples in the data graph.

The presence of autocorrelation provides strong motivation for using relational techniques for learning and inference. Autocorrelation is a statistical dependency between the values of the same variable on related entities and is a nearly ubiquitous characteristic of relational data sets. Recent research has explored the use of collective inference techniques to exploit this phenomenon. These techniques achieve significant performance gains by modeling observed correlations among class labels of related instances, but the models fail to capture a frequent cause of autocorrelation---the presence of underlying groups that influence the attributes on a set of entities. We propose a latent group model (LGM) for relational data, which discovers and exploits the hidden structures responsible for the observed autocorrelation among class labels. Modeling the latent group structure improves model performance, increases inference efficiency, and enhances our understanding of the datasets. We evaluate performance on three relational classification tasks and show that LGM outperforms models that ignore latent group structure, particularly when there is little information with which to seed inference.

We describe an application of relational knowledge discovery to a key regulatory mission of the National Association of Securities Dealers (NASD). NASD is the world's largest privatesector securities regulator, with responsibility for preventing and discovering misconduct among securities brokers. Our goal was to help focus NASD’s limited regulatory resources on the brokers who are most likely to engage in securities violations. Using statistical relational learning algorithms, we developed models that rank brokers with respect to the probability that they would commit a serious violation of securities regulations in the near future. Our models incorporate organizational relationships among brokers (e.g., past coworker), which domain experts consider important but cannot easily use otherwise. The learned models were subjected to an extensive evaluation using more than 18 months of data unseen by the model developers and comprising over two person weeks of effort by NASD staff. Model predictions were found to correlate highly with the subjective evaluations of experienced NASD examiners. Furthermore, in all performance measures, our models performed as well as or better than the handcrafted rules that are currently in use at NASD. 1.

Instance independence is a critical assumption of traditional machine learning methods contradicted by many relational datasets. For example, in scientific literature datasets, there are dependencies among the references of a paper. Recent work on graphical models for relational data has demonstrated significant performance gains for models that exploit the dependencies among instances. In this paper, we present relational dependency networks (RDNs), a new form of graphical model capable of reasoning with such dependencies in a relational setting. We describe the details of RDN models and outline their strengths, most notably the ability to learn and reason with cyclic relational dependencies. We present RDN models learned on a number of real-world datasets, and evaluate the models in a classification context, showing significant performance improvements. In addition, we use synthetic data to evaluate the quality of model learning and inference procedures.

Procedures for collective inference make simultaneous statistical judgments about the same variables for a set of related data instances. For example, collective inference could be used to simultaneously classify a set of hyperlinked documents or infer the legitimacy of a set of related financial transactions. Several recent studies indicate that collective inference can significantly reduce classification error when compared with traditional inference techniques. We investigate the underlying mechanisms for this error reduction by reviewing past work on collective inference and characterizing different types of statistical models used for making inference in relational data. We show important differences among these models, and we characterize the necessary and sufficient conditions for reduced classification error based on experiments with real and simulated data.

We present the relational Bayesian classifier (RBC), a modification of the simple Bayesian classifier (SBC) for relational data. There exist several Bayesian classifiers that learn predictive models of relational data, but each uses a different estimation technique for modelling heterogeneous sets of attribute values. The effects of data characteristics on estimation have not been explored. We consider four simple estimation techniques and evaluate them on three real-world data sets. The estimator that assumes each multiset value is independently drawn from the same distribution (INDEPVAL) achieves the best empirical results. We examine bias and variance tradeoffs over a range of data sets and show that INDEPVAL's ability to model more multiset information results in lower bias estimates and contributes to its superior performance.

Classification trees are widely used in the machine learning and data mining communities for modeling propositional data. Recent work has extended this basic paradigm to probability estimation trees. Traditional tree learning algorithms assume that instances in the training data are homogenous and independently distributed. Relational probability trees (RPTs) extend standard probability estimation trees to a relational setting in which data instances are heterogeneous and interdependent. Our algorithm for learning the structure and parameters of an RPT searches over a space of relational features that use aggregation functions (e.g. AVERAGE, MODE, COUNT) to dynamically propositionalize relational data and create binary splits within the RPT. Previous work has identified a number of statistical biases due to characteristics of relational data such as autocorrelation and degree disparity. The RPT algorithm uses a novel form of randomization test to adjust for these biases. On a variety of relational learning tasks, RPTs built using randomization tests are significantly smaller than other models and achieve equivalent, or better, performance.