Search

In many mobile applications, user-generated data are presented as set-valued data. To tackle potential privacy threats in analyzing these valuable data, local differential privacy has been attracting substantial attention. However, existing approaches only provide sub-optimal utility and are expensive in computation and communication for set-valued data distribution estimation and heavy-hitter identification. In this paper, we propose a utility-optimal and efficient set-valued data publication method (i.e., Wheel mechanism). On the user side, the computational complexity is only $O(\min \lbrace m\log m, m e^\epsilon \rbrace )$O(min{mlogm,meε}) and communication costs are $O(\epsilon +\log m)$O(ε+logm) bits, where $m$m is the number of items, $d$d is the domain size and $\epsilon$ε is the privacy budget, while existing approaches usually depend on $O(d)$O(d) or $O(\log d)$O(logd) ($d \gg m$d≫m). Our theoretical analyses reveal the estimation errors have been reduced from the previously known $O(\frac{m^{2}d}{n\epsilon ^{2}})$O(m2dnε2) to the optimal rate $O(\frac{m d}{n\epsilon ^{2}})$O(mdnε2). Additionally, for heavy-hitter identification, we present a variant of the Wheel mechanism as an efficient frequency oracle, entailing only $O(\sqrt{n})$O(n) computational complexity. This heavy-hitter protocol achieves an identification bar of $\tilde{O}(\frac{1}{\epsilon }\sqrt{\frac{m}{n} \log d})$O˜(1εmnlogd), reducing by a factor of $\sqrt{m}$m relative to existing protocols. Extensive experiments demonstrate our methods are 3-100x faster than existing approaches and have optimized statistical efficiency.

Rapid identification of numerous coseismic landslides following a major earthquake is essential for emergency response and postdisaster recovery. While convolutional neural networks (CNNs) and vision transformers (ViTs) have shown excellent capability in automatic recognition of coseismic landslides, these models require large amounts of training data and computational resources. In addition, there is a lack of additional constraints to restrict the high false alarms of coseismic landslides. To address these challenges, this study proposes an unsupervised deep clustering method coupling remote sensing and implicit knowledge (UC-RSIK) for the recognition of coseismic landslides. The UC-RSIK model integrates a remote sensing feature extraction branch for image reconstruction and an implicit knowledge branch that derives insights from normalized difference vegetation index (NDVI), seismological, geological, topographic, and geographic data. These extracted features are fused and then processed by a self-training adaptive clustering optimization module. This module incorporates hierarchical clustering and autonomously determines the optimal number of image categories, enhancing the effectiveness and robustness of the clustering process. Tested on datasets from the Iburi, Haiti, and Luding earthquakes, UC-RSIK outperformed the state-of-the-art methods such as K-means, Mini-batch K-means, FCM, FSECSGL, and CAE in key metrics like mean Intersection over Union (mIoU) and $F1$ values. Furthermore, t-SNE visualization confirmed that UC-RSIK effectively separates landslides and no-landslides in the feature space, demonstrating superior clustering and category discriminating abilities. The results highlight UC-RSIK’s potential as a highly accurate, robust, and adaptable tool for coseismic landslide identification in diverse terrain conditions.

Designing machine learning models followed by their deployment in a real-world environment has been an area of recent pursuits, resulting in a large number of successful applications. In particular, these applications target environments that call for a great deal of autonomy and criticality of the developed constructs and ensuing decision processes. An efficient design, carefully structured advanced architecture, high performance, and efficient learning methods are of paramount importance. Equally desired is the confidence of any result produced by the numeric model. In this study, we advocate that the associated information granularity of the numeric models and their results inherently link with the notion of specificity of information granularity. The confidence of results can be quantified in the form of an information granule where the two associated criteria of granular outcomes, such as coverage and specificity, are crucial to the holistic evaluation of the granularity of the results. It is shown that these two characteristics are conflicting and their quality becomes evaluated and optimized. Two main approaches are studied in depth. The first one concerns a granular embedding of numeric models. In the second one, we consider a synergistic environment of Gaussian process models whose results come as probabilistic information granules and can be transformed into interval information granules. An interesting architecture of a rule-based model constructed with the use of innovative clustering takes into account the generative-discriminative aspect of the process of structure discovery, which is accomplished through the optimization of some augmented objective functions. This model is investigated with regard to the two approaches to the design of the mechanism of granular assessment of results. Some illustrative examples are covered to show the essentials of the design process.