Search results for: classifiers ensembles

Authors: G. Singer, M. Golan

Abstract:

Providing students with learning disabilities (LD) with extra time to grant them equal access to the exam is a necessary but insufficient condition to compensate for their LD; there should also be a clear indication that the additional time was actually used. For example, if students with LD use more time than students without LD and yet receive lower grades, this may indicate that a different accommodation is required. If they achieve higher grades but use the same amount of time, then the effectiveness of the accommodation has not been demonstrated. The main goal of this study is to evaluate the effect of including parameters related to LD and extended exam time, along with other commonly-used characteristics (e.g., student background and ability measures such as high-school grades), on the ability of ordinal decision-tree algorithms to predict exam performance. We use naturally-occurring data collected from hundreds of undergraduate engineering students. The sub-goals are i) to examine the improvement in prediction accuracy when the indicator of exam performance includes 'actual time used' in addition to the conventional indicator (exam grade) employed in most research; ii) to explore the effectiveness of extended exam time on exam performance for different courses and for LD students with different profiles (i.e., sets of characteristics). This is achieved by using the patterns (i.e., subgroups) generated by the algorithms to identify pairs of subgroups that differ in just one characteristic (e.g., course or type of LD) but have different outcomes in terms of exam performance (grade and time used). Since grade and time used to exhibit an ordering form, we propose a method based on ordinal decision-trees, which applies a weighted information-gain ratio (WIGR) measure for selecting the classifying attributes. Unlike other known ordinal algorithms, our method does not assume monotonicity in the data. The proposed WIGR is an extension of an information-theoretic measure, in the sense that it adjusts to the case of an ordinal target and takes into account the error severity between two different target classes. Specifically, we use ordinal C4.5, random-forest, and AdaBoost algorithms, as well as an ensemble technique composed of ordinal and non-ordinal classifiers. Firstly, we find that the inclusion of LD and extended exam-time parameters improves prediction of exam performance (compared to specifications of the algorithms that do not include these variables). Secondly, when the indicator of exam performance includes 'actual time used' together with grade (as opposed to grade only), the prediction accuracy improves. Thirdly, our subgroup analyses show clear differences in the effect of extended exam time on exam performance among different courses and different student profiles. From a methodological perspective, we find that the ordinal decision-tree based algorithms outperform their conventional, non-ordinal counterparts. Further, we demonstrate that the ensemble-based approach leverages the strengths of each type of classifier (ordinal and non-ordinal) and yields better performance than each classifier individually.

Keywords: actual exam time usage, ensemble learning, learning disabilities, ordinal classification, time extension

Procedia PDF Downloads 68

Authors: Mehrnaz Mostafavi

Abstract:

The assessment and categorization of incidental lung nodules present a considerable challenge in healthcare, often necessitating resource-intensive multiple computed tomography (CT) scans for growth confirmation. This research addresses this issue by introducing a distinct computational approach leveraging radiomics and deep-learning methods. However, understanding local services is essential before implementing these advancements. With diverse tracking methods in place, there is a need for efficient and accurate identification approaches, especially in the context of managing lung nodules alongside pre-existing cancer scenarios. This study explores the integration of text-based algorithms in medical data curation, indicating their efficacy in conjunction with machine learning and deep-learning models for identifying lung nodules. Combining medical images with text data has demonstrated superior data retrieval compared to using each modality independently. While deep learning and text analysis show potential in detecting previously missed nodules, challenges persist, such as increased false positives. The presented research introduces a Structured-Query-Language (SQL) algorithm designed for identifying pulmonary nodules in a tertiary cancer center, externally validated at another hospital. Leveraging natural language processing (NLP) and machine learning, the algorithm categorizes lung nodule reports based on sentence features, aiming to facilitate research and assess clinical pathways. The hypothesis posits that the algorithm can accurately identify lung nodule CT scans and predict concerning nodule features using machine-learning classifiers. Through a retrospective observational study spanning a decade, CT scan reports were collected, and an algorithm was developed to extract and classify data. Results underscore the complexity of lung nodule cohorts in cancer centers, emphasizing the importance of careful evaluation before assuming a metastatic origin. The SQL and NLP algorithms demonstrated high accuracy in identifying lung nodule sentences, indicating potential for local service evaluation and research dataset creation. Machine-learning models exhibited strong accuracy in predicting concerning changes in lung nodule scan reports. While limitations include variability in disease group attribution, the potential for correlation rather than causality in clinical findings, and the need for further external validation, the algorithm's accuracy and potential to support clinical decision-making and healthcare automation represent a significant stride in lung nodule management and research.

Keywords: lung cancer diagnosis, structured-query-language (SQL), natural language processing (NLP), machine learning, CT scans

Procedia PDF Downloads 39

Authors: Julia Ive, Lucia Specia

Abstract:

Understanding and detecting toxicity is an important problem to support safer human interactions online. Our work focuses on the important problem of contextual toxicity detection, where automated classifiers are tasked with determining whether a short textual segment (usually a sentence) is toxic within its conversational context. We use “toxicity” as an umbrella term to denote a number of variants commonly named in the literature, including hate, abuse, offence, among others. Detecting toxicity in context is a non-trivial problem and has been addressed by very few previous studies. These previous studies have analysed the influence of conversational context in human perception of toxicity in controlled experiments and concluded that humans rarely change their judgements in the presence of context. They have also evaluated contextual detection models based on state-of-the-art Deep Learning and Natural Language Processing (NLP) techniques. Counterintuitively, they reached the general conclusion that computational models tend to suffer performance degradation in the presence of context. We challenge these empirical observations by devising better contextual predictive models that also rely on NLP data augmentation techniques to create larger and better data. In our study, we start by further analysing the human perception of toxicity in conversational data (i.e., tweets), in the absence versus presence of context, in this case, previous tweets in the same conversational thread. We observed that the conclusions of previous work on human perception are mainly due to data issues: The contextual data available does not provide sufficient evidence that context is indeed important (even for humans). The data problem is common in current toxicity datasets: cases labelled as toxic are either obviously toxic (i.e., overt toxicity with swear, racist, etc. words), and thus context does is not needed for a decision, or are ambiguous, vague or unclear even in the presence of context; in addition, the data contains labeling inconsistencies. To address this problem, we propose to automatically generate contextual samples where toxicity is not obvious (i.e., covert cases) without context or where different contexts can lead to different toxicity judgements for the same tweet. We generate toxic and non-toxic utterances conditioned on the context or on target tweets using a range of techniques for controlled text generation(e.g., Generative Adversarial Networks and steering techniques). On the contextual detection models, we posit that their poor performance is due to limitations on both of the data they are trained on (same problems stated above) and the architectures they use, which are not able to leverage context in effective ways. To improve on that, we propose text classification architectures that take the hierarchy of conversational utterances into account. In experiments benchmarking ours against previous models on existing and automatically generated data, we show that both data and architectural choices are very important. Our model achieves substantial performance improvements as compared to the baselines that are non-contextual or contextual but agnostic of the conversation structure.

Keywords: contextual toxicity detection, data augmentation, hierarchical text classification models, natural language processing

Procedia PDF Downloads 137

Authors: Mohammad Karimi Moridani, Mohammad Abdi Zadeh, Zahra Shahiazar Mazraeh

Abstract:

In recent years, the use of intelligent systems in biomedical engineering has increased dramatically, especially in the diagnosis of various diseases. Also, due to the relatively simple recording of the electrocardiogram signal (ECG), this signal is a good tool to show the function of the heart and diseases associated with it. The aim of this paper is to design an intelligent system for automatically detecting a normal electrocardiogram signal from abnormal one. Using this diagnostic system, it is possible to identify a person's heart condition in a very short time and with high accuracy. The data used in this article are from the Physionet database, available in 2016 for use by researchers to provide the best method for detecting normal signals from abnormalities. Data is of both genders and the data recording time varies between several seconds to several minutes. All data is also labeled normal or abnormal. Due to the low positional accuracy and ECG signal time limit and the similarity of the signal in some diseases with the normal signal, the heart rate variability (HRV) signal was used. Measuring and analyzing the heart rate variability with time to evaluate the activity of the heart and differentiating different types of heart failure from one another is of interest to the experts. In the preprocessing stage, after noise cancelation by the adaptive Kalman filter and extracting the R wave by the Pan and Tampkinz algorithm, R-R intervals were extracted and the HRV signal was generated. In the process of processing this paper, a new idea was presented that, in addition to using the statistical characteristics of the signal to create a return map and extraction of nonlinear characteristics of the HRV signal due to the nonlinear nature of the signal. Finally, the artificial neural networks widely used in the field of ECG signal processing as well as distinctive features were used to classify the normal signals from abnormal ones. To evaluate the efficiency of proposed classifiers in this paper, the area under curve ROC was used. The results of the simulation in the MATLAB environment showed that the AUC of the MLP and SVM neural network was 0.893 and 0.947, respectively. As well as, the results of the proposed algorithm in this paper indicated that the more use of nonlinear characteristics in normal signal classification of the patient showed better performance. Today, research is aimed at quantitatively analyzing the linear and non-linear or descriptive and random nature of the heart rate variability signal, because it has been shown that the amount of these properties can be used to indicate the health status of the individual's heart. The study of nonlinear behavior and dynamics of the heart's neural control system in the short and long-term provides new information on how the cardiovascular system functions, and has led to the development of research in this field. Given that the ECG signal contains important information and is one of the common tools used by physicians to diagnose heart disease, but due to the limited accuracy of time and the fact that some information about this signal is hidden from the viewpoint of physicians, the design of the intelligent system proposed in this paper can help physicians with greater speed and accuracy in the diagnosis of normal and patient individuals and can be used as a complementary system in the treatment centers.

Keywords: neart rate variability, signal processing, linear and non-linear features, classification methods, ROC Curve

Procedia PDF Downloads 224

Authors: Panagiota Katsikouli, William Hamilton Byrne, Thomas Gammeltoft-Hansen, Tijs Slaats

Abstract:

An individual who demonstrates a well-founded fear of persecution or faces real risk of being subjected to torture is eligible for asylum. In Danish law, the exact legal thresholds reflect those established by international conventions, notably the 1951 Refugee Convention and the 1950 European Convention for Human Rights. These international treaties, however, remain largely silent when it comes to how states should assess asylum claims. As a result, national authorities are typically left to determine an individual’s legal eligibility on a narrow basis consisting of an oral testimony, which may itself be hampered by several factors, including imprecise language interpretation, insecurity or lacking trust towards the authorities among applicants. The leaky ground, on which authorities must assess their subjective perceptions of asylum applicants' credibility, questions whether, in all cases, adjudicators make the correct decision. Moreover, the subjective element in these assessments raises questions on whether individual asylum cases could be afflicted by implicit biases or stereotyping amongst adjudicators. In fact, recent studies have uncovered significant correlations between decision outcomes and the experience and gender of the assigned judge, as well as correlations between asylum outcomes and entirely external events such as weather and political elections. In this study, we analyze a publicly available dataset containing approximately 8,000 summaries of asylum cases, initially rejected, and re-tried by the Refugee Appeals Board (RAB) in Denmark. First, we look for variations in the recognition rates, with regards to a number of applicants’ features: their country of origin/nationality, their identified gender, their identified religion, their ethnicity, whether torture was mentioned in their case and if so, whether it was supported or not, and the year the applicant entered Denmark. In order to extract those features from the text summaries, as well as the final decision of the RAB, we applied natural language processing and regular expressions, adjusting for the Danish language. We observed interesting variations in recognition rates related to the applicants’ country of origin, ethnicity, year of entry and the support or not of torture claims, whenever those were made in the case. The appearance (or not) of significant variations in the recognition rates, does not necessarily imply (or not) bias in the decision-making progress. None of the considered features, with the exception maybe of the torture claims, should be decisive factors for an asylum seeker’s fate. We therefore investigate whether the decision can be predicted on the basis of these features, and consequently, whether biases are likely to exist in the decisionmaking progress. We employed a number of machine learning classifiers, and found that when using the applicant’s country of origin, religion, ethnicity and year of entry with a random forest classifier, or a decision tree, the prediction accuracy is as high as 82% and 85% respectively. tentially predictive properties with regards to the outcome of an asylum case. Our analysis and findings call for further investigation on the predictability of the outcome, on a larger dataset of 17,000 cases, which is undergoing.

Keywords: asylum adjudications, automated decision-making, machine learning, text mining

Procedia PDF Downloads 66

Authors: Jennifer Buz, Alvin Spivey

Abstract:

The most common greenhouse gas emitted through human activities, carbon dioxide (CO2), is naturally consumed by plants during photosynthesis. This process is actively being monetized by companies wishing to offset their carbon dioxide emissions. For example, companies are now able to purchase protections for vegetated land due-to-be clear cut or purchase barren land for reforestation. Therefore, by actively preventing the destruction/decay of plant matter or by introducing more plant matter (reforestation), a company can theoretically offset some of their emissions. One of the biggest issues in the carbon credit market is validating and verifying carbon offsets. There is a need for a system that can accurately and frequently ensure that the areas sold for carbon credits have the vegetation mass (and therefore for carbon offset capability) they claim. Traditional techniques for measuring vegetation mass and determining health are costly and require many person-hours. Orbital Sidekick offers an alternative approach that accurately quantifies carbon mass and assesses vegetation health through satellite hyperspectral imagery, a technique which enables us to remotely identify material composition (including plant species) and condition (e.g., health and growth stage). How much carbon a plant is capable of storing ultimately is tied to many factors, including material density (primarily species-dependent), plant size, and health (trees that are actively decaying are not effectively storing carbon). All of these factors are capable of being observed through satellite hyperspectral imagery. This abstract focuses on speciation. To build a species classification model, we matched pixels in our remote sensing imagery to plants on the ground for which we know the species. To accomplish this, we collaborated with the researchers at the Teakettle Experimental Forest. Our remote sensing data comes from our airborne “Kato” sensor, which flew over the study area and acquired hyperspectral imagery (400-2500 nm, 472 bands) at ~0.5 m/pixel resolution. Coverage of the entire teakettle experimental forest required capturing dozens of individual hyperspectral images. In order to combine these images into a mosaic, we accounted for potential variations of atmospheric conditions throughout the data collection. To do this, we ran an open source atmospheric correction routine called ISOFIT1 (Imaging Spectrometer Optiman FITting), which converted all of our remote sensing data from radiance to reflectance. A database of reflectance spectra for each of the tree species within the study area was acquired using the Teakettle stem map and the geo-referenced hyperspectral images. We found that a wide variety of machine learning classifiers were able to identify the species within our images with high (>95%) accuracy. For the most robust quantification of carbon mass and the best assessment of the health of a vegetated area, speciation is critical. Through the use of high resolution hyperspectral data, ground-truth databases, and complex analytical techniques, we are able to determine the species present within a pixel to a high degree of accuracy. These species identifications will feed directly into our carbon mass model.

Keywords: hyperspectral, satellite, carbon, imagery, python, machine learning, speciation

Procedia PDF Downloads 93