Probabilistic First-Order Theory Revision from Examples

When determining the class of the unknown example by using naive Bayesian classifier, we need to estimate the class conditional probabilities for the continuous attributes. In flexible Bayesian classifier, the Gaussian kernel function is frequently used for classification task under the framework of Parzen window method. In this paper, the other six kernel functions (uniform, triangular, epanechnikov, biweight, triweight and cosine) are introduced in the flexible naive Bayesian. The performances of these seven kernels are compared in 30 UCI datasets. The experimental comparisons are carried out according to the following three aspects: the classification accuracy, ranking performance and the class probability estimation. The latter two are measured by the area under the ROC curve (AUC) and the conditional log likelihood (CLL). The related kernels are compared via two-tailed t-test with a 95 percent confidence level and the Friedman's test using the 0.05 critical level. The experimental results show that the most commonly used Gaussian kernel can not achieve the best classification accuracy and AUC. However, on the CLL, the Gaussian kernel is statistically significantly better than the other six kernels. Finally, the corresponding analyses are given based on the experimental results.

This paper summarizes the research on discriminative language modeling focusing on its application to automatic speech recognition (ASR). A discriminative language model (DLM) is typically a linear or log-linear model consisting of a weight vector associated with a feature vector representation of a sentence. This flexible representation can include linguistically and statistically motivated features that incorporate morphological and syntactic information. At test time, DLMs are used to rerank the output of an ASR system, represented as an N-best list or lattice. During training, both negative and positive examples are used with the aim of directly optimizing the error rate. Various machine learning methods, including the structured perceptron, large margin methods and maximum regularized conditional log-likelihood, have been used for estimating the parameters of DLMs. Typically positive examples for DLM training come from the manual transcriptions of acoustic data while the negative examples are obtained by processing the same acoustic data with an ASR system. Recent research generalizes DLM training by either using automatic transcriptions for the positive examples or simulating the negative examples.KeywordsAutomatic speech recognitionDiscriminative trainingLanguage modeling

Currently, the dominant technology for providing non-technical users with access to Linked Data is keyword-based search. This is problematic because keywords are often inadequate as a means for expressing user intent. In addition, while a structured query language can provide convenient access to the information needed by advanced analytics, unstructured keyword-based search cannot meet this extremely common need. This makes it harder than necessary for non-technical users to generate analytics. We address these difficulties by developing a natural language-based system that allows non-technical users to create well-formed questions. Our system, called TR Discover, maps from a fragment of English into an intermediate First Order Logic representation, which is in turn mapped into SPARQL or SQL. The mapping from natural language to logic makes crucial use of a feature-based grammar with full formal semantics. The fragment of English covered by the natural language grammar is domain specific and tuned to the kinds of questions that the system can handle. Because users will not necessarily know what the coverage of the system is, TR Discover offers a novel auto-suggest mechanism that can help users to construct well-formed and useful natural language questions. TR Discover was developed for future use with Thomson Reuters Cortellis, which is an existing product built on top of a linked data system targeting the pharmaceutical domain. Currently, users access it via a keyword-based query interface. We report results and performance measures for TR Discover on Cortellis, and in addition, to demonstrate the portability of the system, on the QALD-4 dataset, which is associated with a public shared task. We show that the system is usable and portable, and report on the relative performance of queries using SQL and SPARQL back ends.

Let I_1,...,I_n be independent but not necessarily identically distributed Bernoulli random variables, and let X_n=\sum_{j=1}^nI_j. For \nu in a bounded region, a local central limit theorem expansion of P(X_n=EX_n+\nu) is developed to any given degree. By conditioning, this expansion provides information on the high-order correlation structure of dependent, weighted sampling schemes of a population E (a special case of which is simple random sampling), where a set d\subset E is sampled with probability proportional to \prod_{A\in d}x_A, where x_A are positive weights associated with individuals A\in E. These results are used to determine the asymptotic information, and demonstrate the consistency and asymptotic normality of the conditional and unconditional logistic likelihood estimator for unmatched case-control study designs in which sets of controls of the same size are sampled with equal probability.

From first principles, we derive a quadratic nonlinear, first-order dynamical system capable of performing exact Bayes-Markov inferences for a wide class of biologically plausible stimulus-dependent patterns of activity while simultaneously providing an online estimate of model performance. This is accomplished by constructing a dynamical system that has solutions proportional to the probability distribution over the stimulus space, but with a constant of proportionality adjusted to provide a local estimate of the probability of the recent observations of stimulus-dependent activity-given model parameters. Next, we transform this exact equation to generate nonlinear equations for the exact evolution of log likelihood and log-likelihood ratios and show that when the input has low amplitude, linear rate models for both the likelihood and the log-likelihood functions follow naturally from these equations. We use these four explicit representations of the probability distribution to argue that, in contrast to the arguments of previous work, the dynamical system for the exact evolution of the likelihood (as opposed to the log likelihood or log-likelihood ratios) not only can be mapped onto a biologically plausible network but is also more consistent with physiological observations.

The aim of this work is to localize a query photograph by finding other images depicting the same place in a large geotagged image database. This is a challenging task due to changes in viewpoint, imaging conditions and the large size of the image database. The contribution of this work is two-fold. First, we cast the place recognition problem as a classification task and use the available geotags to train a classifier for each location in the database in a similar manner to per-exemplar SVMs in object recognition. Second, as only few positive training examples are available for each location, we propose a new approach to calibrate all the per-location SVM classifiers using only the negative examples. The calibration we propose relies on a significance measure essentially equivalent to the p-values classically used in statistical hypothesis testing. Experiments are performed on a database of 25,000 geotagged street view images of Pittsburgh and demonstrate improved place recognition accuracy of the proposed approach over the previous work.

A partially supervised classification approach to dominant and recessive human disease gene prediction

Previous work has shown that proteins that have the potential to be vaccine candidates can be predicted from features derived from their amino acid sequences. In this work, we make an empirical comparison across various machine learning classifiers on this sequence-based inference problem. Using systematic cross validation on a dataset of 200 known vaccine candidates and 200 negative examples, with a set of 525 features derived from the AA sequences and feature selection applied through a greedy backward elimination approach, we show that simple classification algorithms often perform as well as more complex support vector kernel machines. The work also includes a novel cross validation applied across bacterial species, i.e. the validation proteins all come from a specific species of bacterium not represented in the training set. We termed this type of validation Leave One Bacteria Out Validation (LOBOV).

We address the problem of predicting links and link change in friends networks and introduce a new supervised learning method for both types of prediction. This extends previous based on directed graph features such as the indegree of candidate friends and pair dependent relational features such as common interests. In this new work, we consider how differential user data, such as that produced using regular crawls from a social network site, can be used to produce a time series with which we can identify prediction problems over both links and link change. A key issue we address is the rarity of change between two successive versions of a social network, resulting in severe imbalance between positive and negative examples of change. We compare existing approaches towards coping with this problem, present positive results on new crawls of LiveJournal, and consider how temporal data can enhance the relational link mining process. INTRODUCTION The problem of predicting links between entities such as users and communities in a friends network can be treated as one of supervised inductive learning for classification. In previous work (Hsu et al. 2007), we introduced a system for classifying pairs of users who were known to lie within a radius of 2 of one another as friends or friends of friends. This classification task was defined on a data sets consisting of 1000 and 4000 users from the blog service LiveJournal. Analysis of the graph structure and pair-dependent sets (e.g., mutual friends and common interests) produced a set of 12 features for each candidate pair. From this set of features, an effective predictor for link existence could be learned. However, there were two key limitations to this approach. First, the features made available to machine learning algorithms included certain information that is not always available for prediction tasks. For example, in many social networks such as Facebook and LinkedIn, a user who has been added to the friend set of another is prompted for whether to issue a reciprocal link, whereas realistic prediction may require link existence to be identified before such information is known. Thus, some latency is inherent in realistic prediction task specifications. Second, a key unaddressed problem in this and other related work is that treating link existence as a function of a single snapshot of the friends network fails to take into account the full history of the graph. We show in this paper that data about changes to the link structure over time can provide an effective

In many real-world data mining applications, accurate class probability estimations are often required to make optimal decisions. For example, in direct marketing, we often need to deploy different promotion strategies to customers with different likelihood (probability) of buying some products. When our learning task is to build a model with accurate class probability estimations, C4.4 is the most popular one for achieving this task because of its efficiency and effect. In this paper, we present a locally weighted version of C4.4 to scale up its class probability estimation performance by combining locally weighted learning with C4.4. We call our improved algorithm locally weighted C4.4, simply LWC4.4. We experimentally tested LWC4.4 using the whole 36 UCI data sets selected by Weka, and compared it to other related algorithms: C4.4, NB, KNN, NBTree, and LWNB. The experimental results show that LWC4.4 significantly outperforms all the other algorithms in term of conditional log likelihood, simply CLL. Thus, our work provides an effective algorithm to produce accurate class probability estimation.

Advanced mass spectrometric procedures have been extensively used to provide an accurate structural characterization of aspartate aminotransferase from Sulfolobus solfataricus. The amino acid sequence of this enzyme had previously been deduced from the DNA sequence. The accurate molecular mass of the protein, determined using electrospray mass spectrometry, demonstrated that the amino acid sequence deduced was correct and ruled out the possible presence of large covalent modifications which had been postulated to fit the much higher molecular mass obtained from previous SDS/PAGE experiments. The definition of the entire primary structure of aspartate aminotransferase from S. solfataricus was achieved by exploiting a new mass spectrometric mapping strategy. Initially, the molecular mass of relatively large protein fragments produced by CNBr hydrolysis was accurately determined using electrospray mass spectrometry. The protein regions where structural modifications had occurred were easily identified from their anomalous mass values. The corresponding CNBr fragments were then subdigested with suitable proteases and the resulting peptide mixtures were analysed by fast-atom-bombardment mass spectrometry. This mapping approach led to the detection of two partially modified lysine residues at positions 202 and 384, which had been converted to their N-epsilon-methyl derivatives to a substoichiometric extent.

Molecular Mechanisms of the R61T Mutation in Apolipoprotein E4: A Dynamic Rescue

The performance of gene finding from genome sequences strongly depends on the accuracy of splice site prediction. Recent gene finding programs, however, still do not reach enough levels. To improve the accuracy of splice site prediction, it is required to understand the splicing mechanism and to make a model from clear experimental evidences. For this purpose, genomic full-length precursor mRNA sequences (FL-pre-mRNAs), together with expression information are indispensable. The FL-pre-mRNAs have entire gene structure such as the 5’ and 3’ end of mRNA, initiation codon, splice sites, stop codon, and polyadenylation signals, etc. They also contain all the alternative splice sites except the first or last exons in alternative transcripts. However, databases of FL-pre-mRNAs are still not reported in previous works. Aligning expressed sequence tags (ESTs) to the genomic sequences has been a common method for gene prediction or splice site analysis (1, 3). However, ESTs are not suitable for collecting FL-premRNAs because ESTs are partial sequences and the 5’ ends of mRNAs are unknown in most cases, and even EST contigs clustered in UniGene (2) or RefSeq database (4) are not evident to be full-length. It is because ESTs are single sequencing reads that contain mutations, insertions, or deletions (5). Growing genomic and EST sequence data, computational approach has become one of methods to annotate the sequences as putative genes or ORFs. Whereas, Genbank database has accumulated the entries in which genomic complete protein-coding sequences or full-length mRNA sequences are characterized by experimental evidence. The sequences and the annotation (the positions of gene boundaries and functional signals) with the information more reliable than that determined by in silico prediction are expected to be high quality. Thus, we constructed datasets with experimental annotation from Genbank database for gene structure prediction and splice site analysis. Moreover, the analysis for constitutive and alternative splice sites with the correlation with several biological descriptors will be discussed.

In the pursuit of a lighter composite wing design, fast and effective methodologies for sizing and validating the wing members (e.g., spar, ribs, skins, etc.) are required. In the present paper, the preliminary design methodology of an airliner main composite wing, which has an innovative multispar configuration instead of the conventional two-spar design, is investigated. The investigated aircraft wing is a large-scale composite component, requiring an efficient analysis methodology; for this purpose, the initial wing sizing is mostly based on simplified Finite Element (FE) stress analysis combined to analytically formulated design criteria. The proposed methodology comprises three basic modules, namely, computational stress analysis of the wing structure, comparison of the stress–strain results to specific design allowable and a suitable resizing procedure, until all design requirements are satisfied. The design constraints include strain allowable for the entire wing structure, stability constraints for the upper skin and spar webs, as well as bearing bypass analysis of the riveted/bolted joints of the spar flanges/skins connection. A comparison between a conventional (2-spar) and an innovative 4-spar wing configuration is presented. It arises from the comparison between the conventional and the 4-spar wing arrangement, that under certain conditions the multispar configuration has significant advantages over the conventional design.

This paper presents a method for weakly supervised learning of visual models. The visual model is based on a two-layer image description: a set of “generic” descriptors and their distribution over neighbourhoods. “Generic” descriptors represent sets of similar rotational invariant feature vectors. Statistical spatial constraints describe the neighborhood structure and make our description more discriminant. The joint probability of the frequencies of “generic” descriptors over a neighbourhood is multi-modal and is represented by a set of “neighbourhood-frequency” clusters. Our image description is rotationally invariant, robust to model deformations and characterizes efficiently “appearance-based” visual structure. The selection of distinctive clusters determines model features (common to the positive and rare in the negative examples). Visual models are retrieved and localized using a probabilistic score. Experimental results for “textured” animals and faces show a very good performance for retrieval as well as localization.