Recent work studying the generalization of diffusion models with locally linear UNet-based denoisers reveals inductive biases that can be expressed via geometry-adaptive harmonic bases. For such locally linear UNets, these geometry-adaptive harmonic bases can be conveniently visualized through the eigen-decomposition of a UNet’s Jacobian matrix. In practice, however, more recent denoising networks are often transformer-based, e.g., the diffusion transformer (DiT). Due to the presence of nonlinear operations, similar eigen-decomposition analyses cannot be used to reveal the inductive biases of transformer-based denoisers. This motivates our search for alternative ways to explain the strong generalization ability observed in DiT models. Investigating a DiT’s pivotal attention modules, we find that locality of attention maps in a DiT’s early layers are closely associated with generalization. To verify this finding, we modify the generalization of a DiT by restricting its attention windows and observe an improvement in generalization. Furthermore, we empirically find that both the placement and the effective attention size of these local attention windows are crucial factors. Experimental results on the CelebA, ImageNet, MSCOCO, and LSUN data show that strengthening the inductive bias of a DiT can improve both generalization and generation quality when less training data is available. Source code is available at https://github.com/DiT-Generalization/DiT-Generalization.

Identifying urban employment centers (UECs) is crucial for understanding urban spatial structures. Conventional identification methods use census-type aggregated data. The increasing pervasiveness of mobile phone data provides new possibilities to analyze UECs at finer spatial resolution but requires novel identification methods. This study proposes a new approach, the Locally Decaying Model (LDM), to fit the employment density locally and identify UECs based on statistically significant local peaks. We compared the proposed LDM with conventional methods by Monte-Carlo simulation and real data experiments on employment distributions generated from Shanghai mobile phone data. The Monte-Carlo simulation showed that the LDM performed significantly better. The Shanghai case study demonstrated greater stability of the LDM in identifying UEC numbers and locations on aggregated data than the conventional methods. Furthermore, UECs extracted by LDM from the employment density raster were more consistent with the existing local plan. This research contributed a new subcenter identification method that can be applied to recently available urban big data and a comprehensive analytical framework for evaluating such methods.

With billions of active social media accounts and millions of live video cameras, live new big data offer many opportunities for smart applications. However, the main consumers of the new big data have been humans. We envision the research on live knowledge , to automatically acquire real-time, validated, and actionable information. Live knowledge presents two significant and diverging technical challenges: big noise and concept drift. We describe the EBKA (evidence-based knowledge acquisition) approach, illustrated by the LITMUS landslide information system. LITMUS achieves both high accuracy and wide coverage, demonstrating the feasibility and promise of EBKA approach to achieve live knowledge.

We present Argoverse, a dataset designed to support autonomous vehicle perception tasks including 3D tracking and motion forecasting. Argoverse includes sensor data collected by a fleet of autonomous vehicles in Pittsburgh and Miami as well as 3D tracking annotations, 300k extracted interesting vehicle trajectories, and rich semantic maps. The sensor data consists of 360 degree images from 7 cameras with overlapping fields of view, forward-facing stereo imagery, 3D point clouds from long range LiDAR, and 6-DOF pose. Our 290km of mapped lanes contain rich geometric and semantic metadata which are not currently available in any public dataset. All data is released under a Creative Commons license at Argoverse.org. In baseline experiments, we use map information such as lane direction, driveable area, and ground height to improve the accuracy of 3D object tracking. We use 3D object tracking to mine for more than 300k interesting vehicle trajectories to create a trajectory forecasting benchmark. Motion forecasting experiments ranging in complexity from classical methods (k-NN) to LSTMs demonstrate that using detailed vector maps with lane-level information substantially reduces prediction error. Our tracking and forecasting experiments represent only a superficial exploration of the potential of rich maps in robotic perception. We hope that Argoverse will enable the research community to explore these problems in greater depth.

In machine learning research, the proximal gradient methods are popular for solving various optimization problems with non-smooth regularization. Inexact proximal gradient methods are extremely important when exactly solving the proximal operator is time-consuming, or the proximal operator does not have an analytic solution. However, existing inexact proximal gradient methods only consider convex problems. The knowledge of inexact proximal gradient methods in the non-convex setting is very limited. To address this challenge, in this paper, we first propose three inexact proximal gradient algorithms, including the basic version and Nesterov’s accelerated version. After that, we provide the theoretical analysis to the basic and Nesterov’s accelerated versions. The theoretical results show that our inexact proximal gradient algorithms can have the same convergence rates as the ones of exact proximal gradient algorithms in the non-convex setting. Finally, we show the applications of our inexact proximal gradient algorithms on three representative non-convex learning problems. Empirical results confirm the superiority of our new inexact proximal gradient algorithms.

We study the problem of using Social Media to detect natural disasters, of which we are interested in a special kind, namely landslides. Employing information from Social Media presents unique research challenges, as there exists a considerable amount of noise due to multiple meanings of the search keywords, such as "landslide" and "mudslide". To tackle these challenges, we propose REX, a rapid ensemble classification system which can filter out noisy information by implementing two key ideas: (I) a new method for constructing independent classifiers that can be used for rapid ensemble classification of Social Media texts, where each classifier is built using randomized Explicit Semantic Analysis; and (II) a self-correction approach which takes advantage of the observation that the majority label assigned to Social Media texts belonging to a large event is highly accurate. We perform experiments using real data from Twitter over 1.5 years to show that REX classification achieves 0.98 in F-measure, which outperforms the standard Bag-of-Words algorithm by an average of 0.14 and the state-of-the-art Word2Vec algorithm by 0.04. We also release the annotated datasets used in the experiments as a contribution to the research community containing 282k labeled items.

In multi-task learning, it is paramount to discover the relational structure of tasks and utilize the learned task structure. Previous works have been using the low-rank latent feature subspace to capture the task relations, and some of them aim to learn the group based relational structure of tasks. However, in many cases, the low-rank subspace may not exist for the specific group of tasks, thus using this paradigm would not work. To discover the task relational structures, we propose a novel multi-task learning method using the structured sparsity-inducing norms to automatically uncover the relations of tasks. Instead of imposing the low-rank constraint, our new model uses a more meaningful assumption, in which the tasks from the same relational group should share the common feature subspace. We can discover the group relational structure of tasks and learn the shared feature subspace for each task group, which help to improve the predictive performance. Our proposed algorithm avoids the high computational complexity of integer programming, thus it converges very fast. Empirical studies conducted on both synthetic and real-world data show that our method consistently outperforms related multi-task learning methods.

The use of Social Media for event detection, such as detection of natural disasters, has gained a booming interest from research community as Social Media has become an immensely important source of real-time information. However, it poses a number of challenges with respect to high volume, noisy information and lack of geo-tagged data. Extraction of high quality information (e.g., Accurate locations of events) while maintaining good performance (e.g., Low latency) are the major problems. In this paper, we propose two approaches for tackling these issues: an augmented Explicit Semantic Analysis approach for rapid classification and a composition of clustering algorithms for location estimation. Our experiments demonstrate over 98% in precision, recall and F-measure when classifying Social Media data while producing a 20% improvement in location estimation due to clustering composition approach. We implement these approaches as part of the landslide detection service LITMUS, which is live and openly accessible for continued evaluation and use.

Feature selection has been an important research topic in data mining, because the real data sets often have high-dimensional features, such as the bioinformatics and text mining applications. Many existing filter feature selection methods rank features by optimizing certain feature ranking criterions, such that correlated features often have similar rankings. These correlated features are redundant and don't provide large mutual information to help data mining. Thus, when we select a limited number of features, we hope to select the top non-redundant features such that the useful mutual information can be maximized. In previous research, Ding et al. recognized this important issue and proposed the minimum Redundancy Maximum Relevance Feature Selection (mRMR) model to minimize the redundancy between sequentially selected features. However, this method used the greedy search, thus the global feature redundancy wasn't considered and the results are not optimal. In this paper, we propose a new feature selection framework to globally minimize the feature redundancy with maximizing the given feature ranking scores, which can come from any supervised or unsupervised methods. Our new model has no parameter so that it is especially suitable for practical data mining application. Experimental results on benchmark data sets show that the proposed method consistently improves the feature selection results compared to the original methods. Meanwhile, we introduce a new unsupervised global and local discriminative feature selection method which can be unified with the global feature redundancy minimization framework and shows superior performance.

Landslides are an illustrative example of multi-hazards, which can be caused by earthquakes, rainfalls and human activity among other reasons. Detection of landslides presents a significant challenge, since there are no physical sensors that would detect landslides directly. A more recent approach in detection of natural hazards, such as earthquakes, involves the use of social media. We propose a multi-service composition approach and describe LITMUS, which is a landslide detection service that combines data from both physical and social information services by filtering and then joining the information flow from those services based on their spatiotemporal features. Our results show that with such approach LITMUS detects 25 out of 27 landslides reported by USGS in December 2013 and 40 more landslide locations unreported by USGS during this period. LITMUS is a prototype tool that is used to investigate and implement research ideas in the area of disaster detection. We list some of the current work being done on refining the system that allows us to identify 137 landslide locations unreported by USGS during a more recent period of September 2014. Finally, we describe a live demonstration that displays landslide detection results on a web map in real-time.

Social media have been used in the detection and management of natural hazards such as earthquakes. However, disasters often lead to other kinds of disasters, forming multi-hazards. Landslide is an illustrative example of a multi-hazard, which may be caused by earthquakes, rainfalls, water erosion, among other reasons. Detecting such multi-hazards is a significant challenge, since physical sensors designed for specific disasters are insufficient for multi-hazards. We describe LITMUS -- a landslide detection service based on a multi-service composition approach that combines data from both physical and social information services by filtering and then joining the information flow from those services based on their spatiotemporal features. Our results show that with such approach LITMUS detects 25 out of 27 landslides reported by USGS in December and 40 more landslides unreported by USGS. Also, LITMUS provides a live demonstration that displays results on a web map.

Most semi-supervised learning models propagate the labels over the Laplacian graph, where the graph should be built beforehand. However, the computational cost of constructing the Laplacian graph matrix is very high. On the other hand, when we do classification, data points lying around the decision boundary (boundary points) are noisy for learning the correct classifier and deteriorate the classification performance. To address these two challenges, in this paper, we propose an adaptive semi-supervised learning model. Different from previous semi-supervised learning approaches, our new model needn't construct the graph Laplacian matrix. Thus, our method avoids the huge computational cost required by previous methods, and achieves a computational complexity linear to the number of data points. Therefore, our method is scalable to large-scale data. Moreover, the proposed model adaptively suppresses the weights of boundary points, such that our new model is robust to the boundary points. An efficient algorithm is derived to alternatively optimize the model parameter and class probability distribution of the unlabeled data, such that the induction of classifier and the transduction of labels are adaptively unified into one framework. Extensive experimental results on six real-world data sets show that the proposed semi-supervised learning model outperforms other related methods in most cases.

Identifying and detecting web spam is an ongoing battle between spam-researchers and spammers which has been going on since search engines allowed searching of web pages to the modern sharing of web links via social networks. A common challenge faced by spam-researchers is the fact that new techniques depend on requiring a corpus of legitimate and spam web pages. Although large corpora of legitimate web pages are available to researchers, the same cannot be said about web spam or spam web pages. In this paper, we introduce the Webb Spam Corpus 2011 — a corpus of approximately 330,000 spam web pages — which we make available to researchers in the fight against spam. By having a standard corpus available, researchers can collaborate better on developing and reporting results of spam filtering techniques. The corpus contains web pages crawled from links found in over 6.3 million spam emails. We analyze multiple aspects of this corpus including redirection, HTTP headers, web page content, and classification evaluation. We also provide insights into changes in web spam since the last Webb Spam Corpus was released in 2006. These insights include: (1) spammers manipulate social media in spreading spam; (2) HTTP headers and content also change over time; (3) spammers have evolved and adopted new techniques to avoid the detection based on HTTP header information.

Email spam is a persistent problem, especially today, with the increasing dedication and sophistication of spammers. Even popular social media sites such as Facebook, Twitter, and Google Plus are not exempt from email spam as they all interface with email systems. With an "arms-race'' between spamme

With an average of 80% length reduction, the URL shorteners have become the norm for sharing URLs on Twitter, mainly due to the 140-character limit per message. Unfortunately, spammers have also adopted the URL shorteners to camouflage and improve the user click-through of their spam URLs. In this p

With over 2.5 hours a day spent browsing websites online[1] and with over a billion pages[2], identifying and detecting web spam is an important problem. Although large corpora of legitimate web pages are available to researchers, the same cannot be said about web spam or spam web pages. We int