Federated learning (FL) enables multiple clients with distributed data sources to collaboratively train a shared model without compromising data privacy. However, existing FL paradigms face challenges due to heterogeneity in client data distributions and system capabilities. Personalized federated learning (pFL) has been proposed to mitigate these problems, but often requires a shared model architecture and a central entity for parameter aggregation, resulting in scalability and communication issues. More recently, model-heterogeneous FL has gained attention due to its ability to support diverse client models, but existing methods are limited by their dependence on a centralized framework, synchronized training, and publicly available datasets. To address these limitations, we introduce Federated Peer-Adaptive Ensemble Learning (FedPAE), a fully decentralized pFL algorithm that supports model heterogeneity and asynchronous learning. Our approach utilizes a peer-to-peer model sharing mechanism and ensemble selection to achieve a more refined balance between local and global information. Experimental results show that FedPAE outperforms existing state-of-the-art pFL algorithms, effectively managing diverse client capabilities and demonstrating robustness against statistical heterogeneity.

Traffic accident forecasting is an important task for intelligent transportation management and emergency response systems. However, this problem is challenging due to the spatial heterogeneity of the environment. Existing data-driven methods mostly focus on studying homogeneous areas with limited size (e.g. a single urban area such as New York City) and fail to handle the heterogeneous accident patterns over space at different scales. Recent advances (e.g. spatial ensemble) utilize pre-defined space partitions and learn multiple models to improve prediction accuracy. However, external knowledge is required to define proper space partitions before training models and predefined partitions may not necessarily reduce the heterogeneity. To address this issue, we propose a novel Learning-Integrated Space Partition Framework (LISA) to simultaneously learn partitions while training models, where the partitioning process and learning process are integrated in a way that partitioning is guided explicitly by prediction accuracy rather than other factors. Experiments using real-world datasets, demonstrate that our work can capture underlying heterogeneous patterns in a self-guided way and substantially improve baseline networks by an average of 13.0%.

Traffic accident forecasting is a significant problem for transportation management and public safety. However, this problem is challenging due to the spatial heterogeneity of the environment and the sparsity of accidents in space and time. The occurrence of traffic accidents is affected by complex dependencies among spatial and temporal features. Recent traffic accident prediction methods have attempted to use deep learning models to improve accuracy. However, most of these methods either focus on small-scale and homogeneous areas such as populous cities or simply use sliding-window-based ensemble methods, which are inadequate to handle heterogeneity in large regions. To address these limitations, this paper proposes a novel Hierarchical Knowledge Transfer Network (HintNet) model to better capture irregular heterogeneity patterns. HintNet performs a multi-level spatial partitioning to separate sub-regions with different risks and learns a deep network model for each level using spatio-temporal and graph convolutions. Through knowledge transfer across levels, HintNet archives both higher accuracy and higher training efficiency. Extensive experiments on a real-world accident dataset from the state of Iowa demonstrate that HintNet outperforms the state-of-the-art methods on spatially heterogeneous and large-scale areas.

Urban dispersal events occur when an unexpectedly large number of people leave an area in a relatively short period of time. It is beneficial for the city authorities, such as law enforcement and city management, to have an advance knowledge of such events, as it can help them mitigate the safety risks and handle important challenges such as managing traffic, and so forth. Predicting dispersal events is also beneficial to Taxi drivers and/or ride-sharing services, as it will help them respond to an unexpected demand and gain competitive advantage. Large urban datasets such as detailed trip records and point of interest ( POI ) data make such predictions achievable. The related literature mainly focused on taxi demand prediction. The pattern of the demand was assumed to be repetitive and proposed methods aimed at capturing those patterns. However, dispersal events are, by definition, violations of those patterns and are, understandably, missed by the methods in the literature. We proposed a different approach in our prior work [32]. We showed that dispersal events can be predicted by learning the complex patterns of arrival and other features that precede them in time. We proposed a survival analysis formulation of this problem and proposed a two-stage framework (DILSA), where a deep learning model predicted the survival function at each point in time in the future. We used that prediction to determine the time of the dispersal event in the future, or its non-occurrence. However, DILSA is subject to a few limitations. First, based on evidence from the data, mobility patterns can vary through time at a given location. DILSA does not distinguish between different mobility patterns through time. Second, mobility patterns are also different for different locations. DILSA does not have the capability to directly distinguish between different locations based on their mobility patterns. In this article, we address these limitations by proposing a method to capture the interaction between POIs and mobility patterns and we create vector representations of locations based on their mobility patterns. We call our new method DILSA+. We conduct extensive case studies and experiments on the NYC Yellow taxi dataset from 2014 to 2016. Results show that DILSA+ can predict events in the next 5 hours with an F1-score of 0.66. It is significantly better than DILSA and the state-of-the-art deep learning approaches for taxi demand prediction.

Inverse classification is the process of perturbing an instance in a meaningful way such that it is more likely to conform to a specific class. Historical methods that address such a problem are often framed to leverage only a single classifier, or specific set of classifiers. These works are often accompanied by naive assumptions. In this work we propose generalized inverse classification (GIC), which avoids restricting the classification model that can be used. We incorporate this formulation into a refined framework in which GIC takes place. Under this framework, GIC operates on features that are immediately actionable. Each change incurs an individual cost, either linear or non-linear. Such changes are subjected to occur within a specified level of cumulative change (budget). Furthermore, our framework incorporates the estimation of features that change as a consequence of direct actions taken (indirectly changeable features). To solve such a problem, we propose three real-valued heuristic-based methods and two sensitivity analysis-based comparison methods, each of which is evaluated on two freely available real-world datasets. Our results demonstrate the validity and benefits of our formulation, framework, and methods.

Along with the rapid advancement of information technology, the traditional hierarchical supply chain has been quickly evolving into a variety of supply networks, which usually incorporate a large number of entities into complex graph topologies. The study of the resilience of supply networks is an important challenge. In this paper, we exploit the resilience embedded in the network topology by investigating in depth the multiple-path reachability of each demand node to other nodes, and propose a novel network resilience metric. We also develop new supply-network growth mechanisms that reflect the heterogeneous roles of different types of nodes in the supply network. We incorporate them into two fundamental network topologies (i.e. random-graph network and scale-free network), and evaluate their resilience against both random disruptions and targeted attacks using the new resilience metric. The experimental results verify the validity of our resilience metric and the effectiveness of our growth model. This research provides a generic framework and important insights into the construction and resilience analysis of complex supply networks.

Community detection and influence analysis are significant notions in social networks. We exploit the implicit knowledge of influence-based connectivity and proximity encoded in the network topology, and propose a novel algorithm for both community detection and influence ranking. Using a new influence cascade model, the algorithm generates an influence vector for each node, which captures in detail how the node's influence is distributed through the network. Similarity in this influence space defines a new, meaningful and refined connectivity measure for the closeness of any pair of nodes. Our approach not only differentiates the influence ranking but also effectively finds communities in both undirected and directed networks, and incorporates these two important tasks into one integrated framework. We demonstrate its superior performance with extensive tests on a set of real-world networks and synthetic benchmarks.

Community detection is an important task for social networks, which helps us understand the functional modules on the whole network. Among different community detection methods based on graph structures, modularity-based methods are very popular recently, but suffer a well-known resolution limit problem. This paper connects modularity-based methods with correlation analysis by subtly reformatting their math formulas and investigates how to fully make use of correlation analysis to change the objective function of modularity-based methods, which provides a more natural and effective way to solve the resolution limit problem. In addition, a novel theoretical analysis on the upper bound of different objective functions helps us understand their bias to different community sizes, and experiments are conducted on both real life and simulated data to validate our findings.

Finding the most interesting correlations among items is essential for problems in many commercial, medical, and scientific domains. Although there are numerous measures available for evaluating correlations, different correlation measures provide drastically different results. Piatetsky-Shapiro provided three mandatory properties for any reasonable correlation measure, and Tan et al. proposed several properties to categorize correlation measures; however, it is still hard for users to choose the desirable correlation measures according to their needs. In order to solve this problem, we explore the effectiveness problem in three ways. First, we propose two desirable properties and two optional properties for correlation measure selection and study the property satisfaction for different correlation measures. Second, we study different techniques to adjust correlation measures and propose two new correlation measures: the Simplified χ2with Continuity Correction and the Simplified χ2with Support. Third, we analyze the upper and lower bounds of different measures and categorize them by the bound differences. Combining these three directions, we provide guidelines for users to choose the proper measure according to their needs.

Click logs from search engines provide a rich opportunity to acquire implicit feedback from users. Patterns derived from the time between a posted query and a click provide information on the ranking quality, reflecting the perceived relevance of a retrieved URL. This paper applies the Kaplan-Meier estimator to study click patterns. The visualization of click curves demonstrates the interaction between the relevance and the rank position of URLs. The observed results demonstrate the potential of using click curves to predict the quality of the top-ranked results.

Recommendation systems suggest items based on user preferences. Collaborative filtering is a popular approach in which recommending is based on the rating history of the system. One of the most accurate and scalable collaborative filtering algorithms is matrix factorization, which is based on a latent factor model. We propose a novel Euclidean embedding method as an alternative latent factor model to implement collaborative filtering. In this method, users and items are embedded in a unified Euclidean space where the distance between a user and an item is inversely proportional to the rating. This model is comparable to matrix factorization in terms of both scalability and accuracy while providing several advantages. First, the result of Euclidean embedding is more intuitively understandable for humans, allowing useful visualizations. Second, the neighborhood structure of the unified Euclidean space allows very efficient recommendation queries. Finally, the method facilitates online implementation requirements such as mapping new users or items in an existing model. Our experimental results confirm these advantages and show that collaborative filtering via Euclidean embedding is a promising approach for online recommender systems.

Collaborative filtering is a popular approach for building recommender systems. Current collaborative filtering algorithms are accurate but also computationally expensive, and so are best in static off-line settings. It is desirable to include the new data in a collaborative filtering model in an online manner, requiring a model that can be incrementally updated efficiently. Incremental collaborative filtering via co-clustering has been shown to be a very scalable approach for this purpose. However, locally optimized co-clustering solutions via current fast iterative algorithms give poor accuracy. We propose an evolutionary co-clustering method that improves predictive performance while maintaining the scalability of co-clustering in the online phase.

Finding the most interesting correlations among items is essential for problems in many commercial, medical, and scientific domains. Much previous research focuses on finding correlated pairs instead of correlated itemsets in which all items are correlated with each other. When designing gift sets, store shelf arrangements, or Website product categories, we are more interested in correlated itemsets than correlated pairs. We solve this problem by finding maximal fully-correlated itemsets (MFCIs), in which all subsets are closely related to all other subsets. Putting the items in an MFCI together can promote sales within this itemset. Though some exsiting methods find high-correlation itemsets, they suffer from both efficiency and effectiveness problems in large datasets. In this paper, we explore high-dimensional correlation in two ways. First, we expand the set of desirable properties for correlation measures and study the advantages and disadvantages of various measures. Second, we propose an MFCI framework to decouple the correlation measure from the need for efficient search. By wrapping the best measure in our MFCI framework, we take advantage of likelihood ratio's superiority in evaluating itemsets, make use of the properties of MFCI to eliminate itemsets with irrelevant items, and still achieve good computational performance.

Ensemble methods have proven to be highly effective in improving the performance of base learners under most circumstances. In this paper, we propose a new algorithm that combines the merits of some existing techniques, namely, bagging, arcing, and stacking. The basic structure of the algorithm resembles bagging. However, the misclassification cost of each training point is repeatedly adjusted according to its observed out-of-bag vote margin. In this way, the method gains the advantage of arcing-building the classifier the ensemble needs - without fixating on potentially noisy points. Computational experiments show that this algorithm performs consistently better than bagging and arcing with linear and nonlinear base classifiers. In view of the characteristics of bacing, a hybrid ensemble learning strategy, which combines bagging and different versions of bacing, is proposed and studied empirically.

The term consideration set is used in marketing to refer to the set of items a customer thought about purchasing before making a choice. While consideration sets are not directly observable, finding common ones is useful for market segmentation and choice prediction. We approach the problem of inducing common consideration sets as a clustering problem on the space of possible item subsets. Our algorithm combines ideas from binary clustering and itemset mining, and differs from other clustering methods by reflecting the inherent structure of subset clusters. Experiments on both real and simulated datasets show that our algorithm clusters effectively and efficiently even for sparse datasets. In addition, a novel evaluation method is developed to compare clusters found by our algorithm with known ones.

Ensemble methods have proved to be highly effective in improving the performance of base learners under most circumstances. In this paper, we propose a new algorithm that combines the merits of some existing techniques, namely bagging, arcing and stacking. The basic structure of the algorithm resembles bagging, using a linear support vector machine (SVM). However, the misclassification cost of each training point is repeatedly adjusted according to its observed out-of-bag vote margin. In this way, the method gains the advantage of arcing - building the classifier the ensemble needs - without fixating on potentially noisy points. Computational experiments show that this algorithm performs consistently better than bagging and arcing.

A classification ensemble is a group of classifiers that all solve the same prediction problem in different ways. It is well-known that combining the predictions of classifiers within the same problem domain using techniques like bagging or boosting often improves the performance. This research shows that sharing classifiers among different but closely related problem domains can also be helpful. In addition, a semi-definite programming based ensemble pruning method is implemented in order to optimize the selection of a subset of classifiers for each problem domain. Computational results on a catalog dataset indicate that the ensembles resulting from sharing classifiers among different product categories generally have larger AUCs than those ensembles trained only on their own categories. The pruning algorithm not only prevents the occasional decrease of effectiveness caused by conflicting concepts among the problem domains, but also provides a better understanding of the problem domains and their relationships.

Ensemble methods have recently garnered a great deal of attention in the machine learning community. Techniques such as Boosting and Bagging have proven to be highly effective but require repeated resampling of the training data, making them inappropriate in a data mining context. The methods presented in this paper take advantage of plentiful data, building separate classifiers on sequential chunks of training points. These classifiers are combined into a fixed-size ensemble using a heuristic replacement strategy. The result is a fast algorithm for large-scale or streaming data that classifies as well as a single decision tree built on all the data, requires approximately constant memory, and adjusts quickly to concept drift.

This paper introduces a personalization method for image retrieval based on the learning of personal preferences. The proposed system indexes objects based on shape and groups them into a set of clusters, or prototypes. Our personalization method refines corresponding prototypes from objects provided by the user in the foreground, and simultaneously adapts the database index in the background.

Feature subset selection is an important problem in knowl- edge discovery, not only for the insight gained from deter- mining relevant modeling variables but also for the improved understandability, scalability, and possibly, accuracy of the resulting models. In this paper we consider the problem of feature selection for unsupervised learning. A number of heuristic criteria can be used to estimate the quality of clusters built from a given featuresubset. Rather than combining such criteria, we use ELSA, an evolutionary lo- cal selection algorithm that maintains a diverse population of solutions that approximate the Pareto front in a multi- dimensional objectiv espace. Each evolved solution repre- sents a feature subset and a number of clusters; a standard K-means algorithm is applied to form the given n umber of clusters based on the selected features. Preliminary results on both real and synthetic data show promise in finding Pareto-optimal solutions through which we can identify the significant features and the correct number of clusters.