A method of weight optimization is introduced when fitting penalized ridge regression models. A penalty term added to a likelihood may be viewed in the light of a hierarchical likelihood. Under this context a method to estimate the variance of a random effect in a mixed model can be employed to obtain an estimate of the penalization weight. We review the theory of ridge penalties from a Bayesian point of view and show how an algorithm for estimating the variance of a random effect can be combined with hierarchical likelihood. The method is compared with other commonly used methods to obtain a penalty weight, such as leave-one-out cross validation, generalized cross validation, penalized quasi-likelihood methods and principal components estimation. Simulation studies are performed to compare the different approaches. For each of the methods we use packages already publicly available in the statistical software R.
An extensive discussion concerning formal, empirical, and ontological status of latent variables in psychological literature concerns the distinction between the realist and anti-realist positions within the classical test theory and item response theory (IRT) psychometric traditions in measurement of latent variables (Measurement 6:25–53, 2008; Salzberger and Koller, J Bus Res 66:1307–1317, 2013). However, this bi-polar view seems to be too distant from the perspectives of schools of thought in the marketing discipline and actual developments of measurement models in specific fields of marketing research. An extensive discussion concerning the reflective–formative latent variables dilemma and relational status of constructs in the contemporary marketing opens space for the redefinition of the nature and role of latent variables in marketing science. The aim of the paper is to outline the interlink between theoretical schools within marketing discipline and contemporary discussion concerning the nature and use of latent variables in marketing.
In recent years, a number of approaches have been developed for the automatic recognition of music genres, but also more specific categories (styles, moods, personal preferences, etc.). Among the different sources for building classification models, features extracted from the audio signal play an important role in the literature. Although such features can be extracted from any digitised music piece independently of the availability of other information sources, their extraction can require considerable computational costs and the audio alone does not always contain enough information for the identification of the distinctive properties of a musical category. In this work we consider playlists that are created and shared by music listeners as another interesting source for feature extraction and music categorisation. The main idea is that the tracks of a playlist are often from the same artist or belong to the same category, e.g. they have the same genre or style, which allows us to exploit their co-occurrences for classification tasks. In the paper, we evaluate strategies for better genre and style classification based on the analysis of larger collections of user-provided playlists and compare them to a recent classification technique from the literature. Our first results indicate that an already comparably simple playlist-based classifiers can in some cases outperform an advanced audio-based classification technique.
In this work we analyse data for 314 participants of a methadone study over 180 days. Dosages in milligram were converted for better interpretability to seven categories in which six categories have an ordinal scale for representing dosages and one category for missing dosages. We develop a dissimilarity measure and cluster the time series using "partitioning around medoids" (PAM). The dissimilarity measure is based on assessing the interpretative dissimilarity between categories. It quantifies the structure of the categories which is partly categorical, partly ordinal and also involves quantitative information. The principle behind the measure can be used for other applications as well, in which there is more information about the meaning of categories than just that they are "ordinal" or "categorical".
The paper describes an attempt to improve the TextRank Textrank algorithm. TextRank Textrank is an algorithm for unsupervised text summarisation. It has two main stages: first stage is representing a text as a weighted directed graph, where nodes stand for single sentences, and edges are weighted with sentence similarity Similarity and connect consequent sentences. The second stage is applying the PageRank algorithm as is to the graph. The nodes that get the highest ranks form the summary of the text. We focus on the first stage, especially on measuring the sentence similarity. Mihalcea and Tarau suggest to employ the common scheme: use the vector space model (VSM), so that every text is a vector in space of words or stems, and compute cosine similarity Similarity between these vectors. Our idea is to replace this scheme by using the annotated suffix trees (AST) model for sentence representation. The AST overcomes several limitations of the VSM model, such as being dependent on the size of vocabulary, the length of sentences and demanding stemming or lemmatisation. This is achieved by taking all fuzzy matches between sentences into account and computing probabilities of matched concurrencies. For testing the method on Russian texts we made our own collection based on newspapers articles with some sentences highlighted as being more important. Using the AST similarity Similarity measure on this collection allows to achieve a slight improvement in comparison with using the cosine similarity Similarity measure.
Concept lattices fulfil one of the aims of classification by providing a description by attributes of each class of objects. We introduce here two new similarity/dissimilarity measures: a similarity measure between concepts (elements) of a lattice and a dissimilarity measure between concept lattices defined on the same set of objects and attributes. Both measures are based on the overhanging relation previously introduced by the author, which are a cryptomorphism of lattices.
Various studies have examined multi-way proximity generalizations of multidimensional scaling (MDS). Some of these have proposed one-mode three-way proximity data analyses to investigate triadic relationships among three objects. However, the results of a triadic distance model are generally similar to those of a one-mode two-way MDS. Moreover, no technique for judging whether a triadic distance model or one-mode two-way MDS is more appropriate has been developed. Thus, it would be valuable to establish a technique for examining the need for a one-mode three-way MDS analysis. Here, we propose a technique to evaluate the need for a triadic distance model using a log-linear model. When the analysis of the log-linear model shows that three objects, i, j, and k, are not independent, the one-mode three-way proximity data should be analyzed with a triadic distance model. However, one-mode three-way proximity data should not be analyzed with a triadic distance model when the analysis of the log-linear model shows that the three objects i, j, and k are independent.
Textual information extraction is a challenging issue in Information Retrieval. Two main approaches are commonly distinguished: texture-based and region-based. In this paper, we propose a method guided by the quadtree decomposition. The principle of the method is to recursively decompose regions of a document image is four equal regions, starting from the image of the whole document. At each step of the decomposition process an OCR engine is used for retrieving a given textual information from the obtained regions. Experiments on real invoice data provide promising results.
Predictive models in analytical CRM (customer relationship management) are closely related to the customer's life cycle. Prediction of binary dependent variable refers to the most common areas such as customer acquisition, customer development (cross-selling and up-selling), and customer retention (churn analysis). While building static predictive models one usually applies decision trees, logistic regression, support vector machines or ensemble methods such as different algorithms of boosted decision trees or random forest. Recently one can observe increasing use of hybrid models in the analytical CRM, i.e. those that combine several different analytical tools, e.g. cluster analysis with decision trees, genetic algorithms with neural networks, or decision trees with logistic regression. The purpose of this paper is to compare the results obtained by using hybrid predictive CART-logit models with single decision tree models and logistic regression models. All analyses have been conducted on the basis of data sets relating to analytical CRM.
There exist several algorithms for tone onset detection, but finding the best one is a challenging task, as there are many categorical and numerical parameters to optimize. The aim of this task is to detect as many true onsets as possible while avoiding false detections. In recent years, model-based optimization (MBO) has been introduced for solving similar problems. The main idea of MBO is modeling the relationship between parameter settings and the response by a socalled surrogate model. After evaluating the points of an initial design-each point represents here one possible algorithm configuration-the main idea is a loop of two steps: firstly, updating a surrogate model, and secondly, proposing a new promising point for evaluation. While originally this technique has been developed mainly for numerical parameters, here, it needs to be adapted for optimizing categorical parameters as well. Unfortunately, optimization steps are very time-consuming, since the evaluation of each new point has to be performed on a large data set of music instances for getting realistic results. Nevertheless, many bad configurations could be rejected much faster, since their expected performance might appear to be very low after evaluating them on just a small partition of instances. Hence, the basic idea is to evaluate each proposed point on a small sample and only evaluate on the whole data set if the results seem to be promising.
The head-up display (HUD) projects a virtual image in the driver’s field of vision. Here, the image quality plays an important role. However, assembly tolerances cause image distortions. The evaluation of these distortions is a current issue, because procedures for the assessment of optical aberrations cannot be applied. Therefore new features and methods are implemented, which evaluate the subjective impression of distortions. The overall objective is to investigate the correlation between subjective labels and objective features. A total of 13 features are required to describe the image quality. Subsequently, the relationship between the labels and the features is adapted to a regression equation. For it, representative images are needed, which are selected by cluster analytical methods.
In this article author presents the results of comparative analysis in reference to scales based on 5-, 7-, 9-, and 11-point response categories. An attempt was made to find the optimum number of responses among these scales but in this regard to the assumptions underlying the Confirmatory Factor Model and MultiTrait-MultiMethod. For this purpose, the data was collected from a sample of young consumers (n = 200) studying at the universities in Poland. The specific aim of the research was focused on their attitudes, which measured different aspects of the companies’ unethical behavior in the context of marketing activities. For the comparison of scales, the author has applied four models derived from the generalized CFA-MTMM model. This model allowed the recommendation of the best scale, and also helped to evaluate the effects associated with the use of particular type of scale on the CFA-MTMM alternative models and extracted, through their agency, factors.
In large discrete data sets which requires classification into signal and noise components, the distribution of the signal is often very bumpy and does not follow a standard distribution. Therefore the signal distribution is further modelled as a mixture of component distributions. However, when the signal component is modelled as a mixture of distributions, we are faced with the challenges of justifying the number of components and the label switching problem caused by multimodality of the likelihood function). To circumvent these challenges, we propose a non-parametric structure for the signal component. This new method is more efficient in terms of precise estimates and better classifications. We demonstrated the efficacy of the methodology using a ChIP-sequencing data set.
In educational and scientific publishing processes scientists and prospective scientists (students) in their different roles (author, editor, reviewer, production editor, lector, reference librarians) invest a large amount of work into the proper handling of scientific literature in the widest sense. In this contribution we introduce the LitObject middleware and its combination with the popular open-source tool Zotero. The LitObject middleware supports the exchange of sets of scientific objects (literature objects) consisting of bibliographic references and documents (e.g. PDF-documents) by scientists. In our contribution we emphasize several process improvements with a special focus on the cogeneration of citation networks.
Variable selection is a problem of increasing interest in many areas of multivariate statistics such as classification, clustering and regression. In contradiction to supervised classification, variable selection in cluster analysis is a much more difficult problem because usually nothing is known about the true class structure. In addition, in clustering, variable selection is highly related to the main problem of the determination of the number of clusters K to be inherent in the data. Here we present a very general bottom-up approach to variable selection in clustering starting with univariate investigations of stability. The hope is that the structure of interest may be contained in only a small subset of variables. Very general means, we make only use of non-parametric resampling techniques for purposes of validation, where we are looking for clusters that can be reproduced to a high degree under resampling schemes. So, our proposed technique can be applied to almost any cluster analysis method.
Over the last years, researchers have focused their attention on a new approach, supervised clustering, that combines the main characteristics of both traditional clustering and supervised classification tasks. Motivated by the importance of pre-processing approaches in the traditional clustering context, this paper explores to what extent supervised pre-processing steps could help traditional clustering to obtain better performance on supervised clustering tasks. This paper reports experiments which show that indeed standard clustering algorithms are competitive compared to existing supervised clustering algorithms when supervised pre-processing steps are carried out.
Interval-valued data arise in practical situations such as recording monthly interval temperatures at meteorological stations, daily interval stock prices, etc. This paper presents a comparison study for clustering efficiency (according to adjusted Rand index) for spectral, ensemble, and spectral-mean shifted clustering methods for symbolic data. Evaluation studies with application of artificial data with known cluster structure (obtained from mlbench and clusterSim packages of R) show the usefulness and stable results of the ensemble clustering compared to spectral and spectral-mean shift method.
Electronic transactions have revolutionized the way that consumers shop, making the small and local retailers, which were being affected by the worldwide crisis, accessible to the entire world. As e-commercemarket expands, the number of commercial transactions supported by credit cards-Card or Customer Not Present also increases. This growing relationship, quite natural and expected, has clear advantages, facilitating e-commerce transactions and attracting new possibilities for trading. However, at the same time a big and serious problem emerges: the occurrence of fraudulent situations in payments. In this work, we used a signature based method to establish the characteristics of user behavior and detect potential fraud cases. A signature is defined by a set of attributes that receive a diverse range of variables-e.g., the average number of orders, time spent per order, number of payment attempts, number of days since last visit, and many others-related to the behavior of a user, referring to an e-commerce application scenario. Based on the analysis of user behavior deviation, detected by comparing the user's recent activity with the user behavior data, which is expressed through the user signature, it is possible to detect potential fraud situations (deviant behavior) in useful time, giving a more robust and accurate decision support system to the fraud analysts on their daily job.
