Corrective classification: Learning from data imperfections with aggressive and diverse classifier ensembling

Learning from imperfect (noisy) information sources is a challenging and reality issue for many data mining applications. Common practices include data quality enhancement by applying data preprocessing techniques or employing robust learning algorithms to avoid developing overly complicated structures that overfit the noise. The essential goal is to reduce noise impact and eventually enhance the learners built from noise-corrupted data. In this paper, we propose a novel corrective classification (C2) design, which incorporates data cleansing, error correction, Bootstrap sampling and classifier ensembling for effective learning from noisy data sources. C2 differs from existing classifier ensembling or robust learning algorithms in two aspects. On one hand, a set of diverse base learners of C2 constituting the ensemble are constructed via a Bootstrap sampling process; on the other hand, C2 further improves each base learner by unifying error detection, correction and data cleansing to reduce noise impact. Being corrective, the classifier ensemble is built from data preprocessed/corrected by the data cleansing and correcting modules. Experimental comparisons demonstrate that C2 is not only more accurate than the learner built from original noisy sources, but also more reliable than Bagging [4] or aggressive classifier ensemble (ACE) [56], which are two degenerated components/variants of C2. The comparisons also indicate that C2 is more stable than Boosting and DECORATE, which are two state-of-the-art ensembling methods. For real-world imperfect information sources (i.e. noisy training and/or test data), C2 is able to deliver more accurate and reliable prediction models than its other peers can offer.     

Process and learning outcomes from remotely-operated,simulated, and hands-on student laboratories

A large-scale, multi-year, randomized study compared learning activities and outcomes for hands-on, remotely-operated, and simulation-based educational laboratories in an undergraduate engineering course. Students (N = 458) worked in small-group lab teams to perform two experiments involving stress on a cantilever beam. Each team conducted the experiments in one of three lab formats (hands-on, remotely-operated, or simulation-based), collecting data either individually or as a team. Lab format and data-collection mode showed an interaction, such that for the hands-on lab format learning outcomes were higher when the lab team collected data sets working as a group rather than individually collecting data sets to be combined later, while for remotely-operated labs individual data collection was best. The pattern of time spent on various lab-related activities suggests that working with real instead of simulated data may induce higher levels of motivation. The results also suggest that learning with computer-mediated technologies can be improved by careful design and coordination of group and individual activities.     

Insights from a simple expression for linear fisher information in a recurrently connected population of spiking neurons

A simple expression for a lower bound of Fisher information is derived for a network of recurrently connected spiking neurons that have been driven to a noise-perturbed steady state. We call this lower bound linear Fisher information, as it corresponds to the Fisher information that can be recovered by a locally optimal linear estimator. Unlike recent similar calculations, the approach used here includes the effects of nonlinear gain functions and correlated input noise and yields a surprisingly simple and intuitive expression that offers substantial insight into the sources of information degradation across successive layers of a neural network. Here, this expression is used to (1) compute the optimal (i.e., information-maximizing) firing rate of a neuron, (2) demonstrate why sharpening tuning curves by either thresholding or the action of recurrent connectivity is generally a bad idea, (3) show how a single cortical expansion is sufficient to instantiate a redundant population code that can propagate across multiple cortical layers with minimal information loss, and (4) show that optimal recurrent connectivity strongly depends on the covariance structure of the inputs to the network.     

On NFAs where all states are final,initial, or both

We examine questions involving nondeterministic finite automata where all states are final, initial, or both initial and final. First, we prove hardness results for the nonuniversality and inequivalence problems for these NFAs. Next, we characterize the languages accepted. Finally, we discuss some state complexity problems involving such automata.     

Using LiDAR and quickbird data to model plant production and quantify uncertainties associated with wetland detection and land cover generalizations

Spatiotemporal data from satellite remote sensing and surface meteorology networks have made it possible to continuously monitor global plant production, and to identify global trends associated with land cover/use and climate change. Gross primary production (GPP) and net primary production (NPP) are routinely derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard satellites Terra and Aqua, and estimates generally agree with independent measurements at validation sites across the globe. However, the accuracy of GPP and NPP estimates in some regions may be limited by the quality of model input variables and heterogeneity at fine spatial scales. We developed new methods for deriving model inputs (i.e., land cover, leaf area, and photosynthetically active radiation absorbed by plant canopies) from airborne laser altimetry (LiDAR) and Quickbird multispectral data at resolutions ranging from about 30 m to 1 km. In addition, LiDAR-derived biomass was used as a means for computing carbon-use efficiency. Spatial variables were used with temporal data from ground-based monitoring stations to compute a six-year GPP and NPP time series for a 3600 ha study site in the Great Lakes region of North America. Model results compared favorably with independent observations from a 400 m flux tower and a process-based ecosystem model (BIOME-BGC), but only after removing vapor pressure deficit as a constraint on photosynthesis from the MODIS global algorithm. Fine-resolution inputs captured more of the spatial variability, but estimates were similar to coarse-resolution data when integrated across the entire landscape. Failure to account for wetlands had little impact on landscape-scale estimates, because vegetation structure, composition, and conversion efficiencies were similar to upland plant communities. Plant productivity estimates were noticeably improved using LiDAR-derived variables, while uncertainties associated with land cover generalizations and wetlands in this largely forested landscape were considered less important.     

A new data fusion model for high spatial- and temporal-resolution mapping of forest disturbance based on Landsat and MODIS

Investigating the temporal and spatial pattern of landscape disturbances is an important requirement for modeling ecosystem characteristics, including understanding changes in the terrestrial carbon cycle or mapping the quality and abundance of wildlife habitats. Data from the Landsat series of satellites have been successfully applied to map a range of biophysical vegetation parameters at a 30 m spatial resolution; the Landsat 16 day revisit cycle, however, which is often extended due to cloud cover, can be a major obstacle for monitoring short term disturbances and changes in vegetation characteristics through time.The development of data fusion techniques has helped to improve the temporal resolution of fine spatial resolution data by blending observations from sensors with differing spatial and temporal characteristics. This study introduces a new data fusion model for producing synthetic imagery and the detection of changes termed Spatial Temporal Adaptive Algorithm for mapping Reflectance Change (STAARCH). The algorithm is designed to detect changes in reflectance, denoting disturbance, using Tasseled Cap transformations of both Landsat TM/ETM and MODIS reflectance data. The algorithm has been tested over a 185 × 185 km study area in west-central Alberta, Canada. Results show that STAARCH was able to identify spatial and temporal changes in the landscape with a high level of detail. The spatial accuracy of the disturbed area was 93% when compared to the validation data set, while temporal changes in the landscape were correctly estimated for 87% to 89% of instances for the total disturbed area. The change sequence derived from STAARCH was also used to produce synthetic Landsat images for the study period for each available date of MODIS imagery. Comparison to existing Landsat observations showed that the change sequence derived from STAARCH helped to improve the prediction results when compared to previously published data fusion techniques.     

New global optimization algorithms for model-based clustering

The Expectation-Maximization (EM) algorithm is a very popular optimization tool for mixture problems and in particular for model-based clustering problems. However, while the algorithm is convenient to implement and numerically very stable, it only produces local solutions. Thus, it may not achieve the globally optimal solution in problems that have a large number of local optima. This paper introduces several new algorithms designed to produce global solutions in model-based clustering. The building blocks for these algorithms are methods from the operations research literature, namely the Cross-Entropy (CE) method and Model Reference Adaptive Search (MRAS). One problem with applying these methods directly is the efficient simulation of positive definite covariance matrices. We propose several new solutions to this problem. One solution is to apply the principles of Expectation-Maximization updating, which leads to two new algorithms, CE-EM and MRAS-EM. We also propose two additional algorithms, CE-CD and MRAS-CD, which rely on the Cholesky decomposition. We conduct numerical experiments of varying complexity to evaluate the effectiveness of the proposed algorithms in comparison to classical EM. We find that although a single run of the new algorithms is slower than a single run of EM, all have the potential for producing significantly better solutions. We also find that although repeat application of EM may achieve similar results, our algorithms provide automated, data-driven decision rules which may significantly reduce the burden of searching for the global optimum.     

Detecting palindromes,patterns and borders in regular languages

Given a language L and a non-deterministic finite automaton M, we consider whether we can determine efficiently (in the size of M) if M accepts at least one word in L, or infinitely many words. Given that M accepts at least one word in L, we consider how long a shortest word can be. The languages L that we examine include the palindromes, the non-palindromes, the k-powers, the non-k-powers, the powers, the non-powers (also called primitive words), the words matching a general pattern, the bordered words, and the unbordered words.     

Improving peer-to-peer search performance through intelligent social search

As a large amount of information is added onto the Internet on a daily basis, the efficiency of peer-to-peer (P2P) search has become increasingly important. However, how to quickly discover the right resource in a large-scale P2P network without generating too much network traffic remains highly challenging. In this paper, we propose a novel P2P search method, by applying the concept of social grouping and intelligent social search; we derive peers into social groups in a P2P network to improve search performance. Through a super-peer-based architecture, we establish and maintain virtual social groups on top of a P2P network. The interactions between the peers in the P2P network are used to incrementally build the social relationships between the peers in the associated social groups. In such a P2P network, a search query is propagated along the social groups in the overlay social network. Our preliminary experiments have demonstrated that our method can significantly shorten search routes and result in a higher peer search performance. In addition, our method also enhances the trustworthiness of search results because searches go through trusted peers.