Secure aggregation protocols ensure the privacy of users’ data in federated learning by preventing the disclosure of local gradients. Many existing protocols impose significant communication and computational burdens on participants and may not efficiently handle the large update vectors typical of machine learning models. Correspondingly, we present e-SeaFL, an efficient verifiable secure aggregation protocol taking only one communication round during the aggregation phase. e-SeaFL allows the aggregation server to generate proof of honest aggregation to participants via authenticated homomorphic vector commitments. Our core idea is the use of assisting nodes to help the aggregation server, under similar trust assumptions existing works place upon the participating users. Our experiments show that the user enjoys an order of magnitude efficiency improvement over the state-of-the-art (IEEE S&P 2023) for large gradient vectors with thousands of parameters. Our open-source implementation is available at https://github.com/vt-asaplab/e-SeaFL.

OBJECTIVE: Public Health Announcements (PHAs) on television are a means of raising awareness about risk behaviors and chronic conditions. PHAs' scarce airtime puts stress on their target audience reach. We seek to help health campaigns select television shows for their PHAs about smoking, binge drinking, drug overdose, obesity, diabetes, STDs, and other conditions using available statistics. MATERIALS AND METHODS: Using Nielsen's TV viewership database for the entire US panel, we presented a novel show discovery methodology for PHAs that combined (i) pattern discovery from high-dimensional data (ii) nonparametric tests for validation, and (iii) online experiments on Facebook. RESULTS: The nonparametric tests verified the robustness of the discovered associations between the popularity of certain shows and health conditions. Findings from fifty (independent) online experiments (where our awareness messages were seen by nearly 1.5 million American adults) empirically demonstrated the value of the methodology. DISCUSSION: For 2016, the methodology identified several shows whose popularities were genuinely associated with certain health conditions, opening up the possibility of health agencies embracing both big data and large-scale experimentation to address an old problem in a new way. CONCLUSION: Policy makers can repeatedly apply the methodology as new data streams in, with perhaps different feature sets, pattern discovery techniques, and online experiments running over longer periods. The comparatively lower initial investment in the methodology can pay off by identifying several shows for a potentially national television campaign. As simply a by-product, the initial investment also results in awareness messages that might reach millions of individuals.

Abstract Mass media sentiment of financial news significantly influences investment decisions of investors. Hence, studying how this sentiment emerges is important. In years past, this was straightforward, often dictated by journalists who cover financial news, but this has become more complex now. In this paper, we focus on how social media sentiment affects mass media sentiment. Using data from Sina Weibo and Sina Finance (around 60 million weibos and 6.2 million news articles), we show that social media does influence mass media sentiment emergence for financial news. The sentiment consistency between social media reaction and prior news articles amplifies the persistence of mass media sentiment over time. By contrast, we found limited evidence of social media reducing the persistence of mass media sentiment over time. The results have significant implications for understanding how 2 types of media, treated separately in the literature, may be connected.

Engagement on online learning platforms is essential for user retention, learning, and performance. However, there is a paucity of research addressing latent engagement measurement using user activities. In this work in progress paper, we present a novel engagement score consisting of three sub-dimensions - cognitive engagement, emotional engagement, and behavioral engagement using a comprehensive set of user activities. We plan to evaluate our score on a large scale online learning platform and compare our score with measurements from a user survey-based engagement scale from the literature.

Objective: Develop an approach, One-class-at-a-time, for triaging psychiatric patients using machine learning on textual patient records. Our approach aims to automate the triaging process and reduce expert effort while providing high classification reliability. Materials and Methods: The One-class-at-a-time approach is a multistage cascading classification technique that achieves higher triage classification accuracy compared to traditional multiclass classifiers through 1) classifying one class at a time (or stage), and 2) identification and application of the highest accuracy classifier at each stage. The approach was evaluated using a unique dataset of 433 psychiatric patient records with a triage class label provided by "I2B2 challenge," a recent competition in the medical informatics community. Results: The One-class-at-a-time cascading classifier outperformed state-of-the-art classification techniques with overall classification accuracy of 77% among 4 classes, exceeding accuracies of existing multiclass classifiers. The approach also enabled highly accurate classification of individual classes-the severe and mild with 85% accuracy, moderate with 64% accuracy, and absent with 60% accuracy. Discussion: The triaging of psychiatric cases is a challenging problem due to the lack of clear guidelines and protocols. Our work presents a machine learning approach using psychiatric records for triaging patients based on their severity condition. Conclusion: The One-class-at-a-time cascading classifier can be used as a decision aid to reduce triaging effort of physicians and nurses, while providing a unique opportunity to involve experts at each stage to reduce false positive and further improve the system's accuracy.

In prior work we addressed a major problem faced by media sites with popularity based recommender systems such as the top-10 list of most liked or most clicked posts. We showed that the hard cutoff used in these systems to generate the "Top N"lists is prone to unduly penalizing good articles that may have just missed the cutoff. A solution to this was to generate recommendations probabilistically, which is an approach that has been shown to be robust against some manipulation techniques as well. The aim of this research is to introduce a class of probabilistic news recommender systems that incorporates widely practiced recommendation techniques as a special case. We establish our results in a special case of two articles using the urn models with feedback mechanism from probability theory.

We address the discovery of periodic patterns in sequence data. Building on prior work in this area, we present definitions and new methods for characterizing and identifying four types of periodic patterns. A unifying concept across the different types of periodic patterns we consider is the use of statistical variance to define periodicity. This lends itself to efficient variance-reduction algorithms for identifying periodic patterns. We motivate and test our approach using both extensive simulated sequences and real sequence data from online clickstream data.

The broad motivation for our research is to build manipulation resistant news recommender systems. However, there can be several different algorithms that are used to generate news recommendations, and the strategies for manipulation resistance are likely specific to the algorithm (or class of algorithm) employed. In this paper we will focus on a common method used by many media sites of recommending the N most read (or popular) articles (e.g. New York Times, BBC, Wall Street Journal all prominently use this). Through simulation results we show that whereas recommendation of the N most read articles is easily susceptible to manipulation, a simple probabilistic variant is more robust to common manipulation strategies. Further, for the "N most read" recommender, probabilistic selection has other desirable properties. Specifically, the (N+1)th article, which may have "just" missed making the cutoff, is unduly penalized under common user models. Small differences initially are easily amplified -- an observation that can be used by manipulators. Probabilistic selection, on the other hand, creates no such artificial penalty. We also use classical results from urn models to derive theoretical results for special cases.

Recommending news articles has become a promising research direction as the Internet provides fast access to real-time information from multiple sources around the world. Traditional news recommendation systems strive to adapt their services to individual users by virtue of both user and news content information. However, the latent relationships among different news items, and the special properties of new articles, such as short shelf lives and value of immediacy, render the previous approaches inefficient.

Using an ensemble of models often results in better performance than using a single “best” model. We present a new approach based on data envelopment analysis (DEA) for model combination. We prove that for two-class classification problems, DEA models identify the same convex hull as does the popular receiver operating characteristics (ROC) analysis used for model combination. We further develop two DEA-based methods to combine classifiers for the more general k-class classification problems. Our results demonstrate that the two methods outperform other benchmark methods and suggest that DEA can be a powerful tool for model combination.

A drawback of traditional data-mining methods is that they do not leverage prior knowledge of users. In prior work, we proposed a method that could discover unexpected patterns in data by using domain knowledge in a systematic manner. In this paper, we present new methods for discovering a minimal set of unexpected patterns by combining the two, independent concepts of minimality and unexpectedness, both of which have been well-studied in the KDD literature. We demonstrate the strengths of this approach experimentally using a case study in a marketing domain.

Grouping customer transactions into segments may help understand customers better. The marketing literature has concentrated on identifying important segmentation variables (e.g., customer loyalty) and on using cluster analysis and mixture models for segmentation. The data mining literature has provided various clustering algorithms for segmentation without focusing specifically on clustering customer transactions. Building on the notion that observable customer transactions are generated by latent behavioral traits, in this paper, we investigate using a pattern-based clustering approach to grouping customer transactions. We define an objective function that we maximize in order to achieve a good clustering of customer transactions and present an algorithm, GHIC, that groups customer transactions such that itemsets generated from each cluster, while similar to each other, are different from ones generated from others. We present experimental results from user-centric Web usage data that demonstrates that GHIC generates a highly effective clustering of transactions.

In this paper we study market share rules, rules that have a certain market share statistic associated with them. Such rules are particularly relevant for decision making from a business perspective. Motivated by market share rules, in this paper we consider statistical quantitative rules (SQ rules) that are quantitative rules in which the RHS can be any statistic that is computed for the segment satisfying the LHS of the rule. Building on prior work, we present a statistical approach for learning all significant SQ rules, i.e., SQ rules for which a desired statistic lies outside a confidence interval computed for this rule. In particular we show how resampling techniques can be effectively used to learn significant rules. Since our method considers the significance of a large number of rules in parallel, it is susceptible to learning a certain number of "false" rules. To address this, we present a technique that can determine the number of significant SQ rules that can be expected by chance alone, and suggest that this number can be used to determine a "false discovery rate" for the learning procedure. We apply our methods to online consumer purchase data and report the results.

This paper proposes a novel Data Envelopment Analysis (DEA) based approach for model combination. We first prove that for the 2-class classification problems DEA models identify the same convex hull as the popular ROC analysis used for model combination. For general k-class classifiers, we then develop a DEA-based method to combine multiple classifiers. Experiments show that the method outperforms other benchmark methods and suggest that DEA can be a promising tool for model combination.

This position paper advocates a vision in the development of automated legal reasoning and presents evidence supporting the plausibility of that vision. The paper observes that original creation of documents of legal import in either fully formal or semistructured form offers the prospect of greatly reducing the cost and expanding the scope of knowledge engineering for legal reasoning. This, it is claimed, is most likely to be achieved via formalization of various sublanguages of legal discourse. SeaSpeak is an example of such a sublanguage and it appears to be amenable to full formalization. Short of that, much can be done with partial formalization and semistructured documents. The paper presents a tabular format for message expression, motivated by a formal agent communication language.

Grouping customer transactions into categories helps understand customers better. The marketing literature has concentrated on identifying important segmentation variables (e.g. customer loyalty) and on using clustering and mixture models for segmentation. The data mining literature has provided various clustering algorithms for segmentation. We investigate using "pattern-based" clustering approaches to grouping customer transactions. We argue that there are clusters in transaction data based on natural behavioral patterns, and present a new technique, YACA, that groups transactions such that itemsets generated from each cluster, while similar to each other, are different from ones generated from others. We present experimental results from user-centric Web usage data that demonstrates that YACA generates a highly effective clustering of transactions.

The literature on web-usage mining is replete with data preprocessing techniques, which correspond to many closely related problem formulations. We survey data preprocessing techniques for session-level pattern discovery and compare three of these techniques in the context of understanding session-level purchase behavior on the web. Using real data collected from 20,000 users' browsing behavior over a period of six months, four different models (linear regressions, logistic regressions, neural networks, and classification trees) are built based on data preprocessed using three different techniques. The results demonstrate that the three approaches result in radically different conclusions and provide initial evidence that a data preprocessing bias exists, the effect of which can be significant.

Many applications are characterized by having naturally incomplete data on customers - where data on only some fixed set of local variables is gathered However, having a more complete picture can help build better models. The naive solution to this problem - acquiring complete data for all customers s often impractical due to the costs of doing so. A possible alternative is to acquire complete data for "some" customers and to use this to improve the models built. The data acquisition problem is determining how many, and which, customers to acquire additional data from. In this paper we suggest using active learning based approaches for the data acquisition problem. In particular, we present initial methods for data acquisition and evaluate these methods experimentally on web usage data and UCI datasets. Results show that the methods perform well and indicate that active learning based methods for data acquisition can be a promising area for data mining research.

Clickstream data collected at any web site (site-centric data) is inherently incomplete, since it does not capture users' browsing behavior across sites (user-centric data). Hence, models learned from such data may be subject to limitations, the nature of which has not been well studied. Understanding the limitations is particularly important since most current personalization techniques are based on site-centric data only. In this paper, we empirically examine the implications of learning from incomplete data in the context of two specific problems: (a) predicting if the remainder of any given session will result in a purchase and (b) predicting if a given user will make a purchase at any future session. For each of these problems we present new algorithms for fast and accurate data preprocessing of clickstream data. Based on a comprehensive experiment on user-level clickstream data gathered from 20,000 users' browsing behavior, we demonstrate that models built on user-centric data outperform models built on site-centric data for both prediction tasks.

Article Small is beautiful: discovering the minimal set of unexpected patterns Share on Authors: Balaji Padmanabhan The Wharton School, University of Pennsylvania, 1310 Steinberg-Dietrich Hall, Philadelphia, PA The Wharton School, University of Pennsylvania, 1310 Steinberg-Dietrich Hall, Philadelphia, PAView Profile , Alexander Tuzhilin Stern School of Business, New York University, 44 West 4lt;supgt;thlt;/supgt; Street, New York, NY Stern School of Business, New York University, 44 West 4lt;supgt;thlt;/supgt; Street, New York, NYView Profile Authors Info & Claims KDD '00: Proceedings of the sixth ACM SIGKDD international conference on Knowledge discovery and data miningAugust 2000 Pages 54–63https://doi.org/10.1145/347090.347103Online:01 August 2000Publication History 100citation873DownloadsMetricsTotal Citations100Total Downloads873Last 12 Months6Last 6 weeks1 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access

Organizations are taking advantage of “data-mining” techniques to leverage the vast amounts of data captured as they process routine transactions. Data mining is the process of discovering hidden structure or patterns in data. However, several of the pattern discovery methods in data-mining systems have the drawbacks that they discover too many obvious or irrelevant patterns and that they do not leverage to a full extent valuable prior domain knowledge that managers have. This research addresses these drawbacks by developing ways to generate interesting patterns by incorporating managers' prior knowledge in the process of searching for patterns in data. Specifically, we focus on providing methods that generate unexpected patterns with respect to managerial intuition by eliciting managers' beliefs about the domain and using these beliefs to seed the search for unexpected patterns in data. Our approach should lead to the development of decision-support systems that provide managers with more relevant patterns from data and aid in effective decision making.