Hierarchical Classification Research Articles

Built on Women’s Bodies: An Author’s Response

Across the twentieth century, police were the enforcers of public space, determining which women would be targeted as problems to solve. Sexual policing—the targeting and legal control of people’s bodies and their presumed sexual activities—was considered a marginal and degraded use of public resources, maligned by officers and residents alike. And yet, amid the recurring crises of legitimacy for police during the scandalous enforcement of Prohibition and the social uprisings of the 1960s, the contested practice of sexual policing was also central to the legitimization and legalization of police power. The Streets Belong to Us is a spatial history of the development of urban police regimes as embattled authorities fought to seize and consolidate power. But embedded in that work is an argument that the formation of police power—like the formation of cities themselves—cannot be understood without accounting for gender and the interlocking hierarchies of race, class, and sexuality that make gender meaningful. Focusing on sexual policing, the most historically gendered arena of law enforcement, from Prohibition in segregated Jim Crow to the rise of broken windows policing in the gentrifying 1980s, I argue that police power and the modern city were built on women’s bodies.

Classification of the high‐rank syntaxa of the Central and Eastern Balkan dry grasslands with a new hierarchical expert system approach

AbstractAimsDeveloping a hierarchical classification system for classes, orders and alliances of the diverse dry grasslands of the Central and Eastern Balkan Peninsula and translating this into an electronic expert system (ES) for the automatic assignment of plots.LocationSerbia, Kosovo, North Macedonia, Bulgaria and northern Greece.MethodsWe extracted 5734 plots from the Balkan Dry Grassland Database corresponding to eight classes of dry grasslands reported from the region, using the EuroVegChecklist ES. This data set and later the plots within each derived subunit were subjected to a new numerical approach: starting with an initial partitioning (expert‐interpreted TWINSPAN classification), diagnostic species were determined based on their phi‐values for the target vegetation type and the differences in phi‐values to the next similar types. These diagnostic species were fed into an ES to create a new partitioning, a procedure which was iterated until diagnostic species and species of the ES converged. Then the same approach was applied within each of the derived units to define the units of the next‐lower level.ResultsThe iterative cluster optimisation (ICO) converged in all cases. The resulting hierarchical expert system (HES) classified 95% of all plots to alliances. We distinguished four classes with eight orders and 12 alliances: (1) Tuberarietea guttatae (Romuleion); (2) Stipo‐Brachypodietea distachyi (Clinopodio alpini‐Thymion striati); (3) Festuco‐Brometea with Brachypodietalia pinnati (Chrysopogono‐Danthonion calycinae and Cirsio‐Brachypodion pinnati), Festucetalia valesiacae (Festucion valesiacae), an unnamed order of rocky steppes (with Pimpinello‐Thymion zygioidis) and Koelerietalia splendentis (Centaureo‐Bromion fibrosi, Saturejion montanae and Diantho haematocalycis‐Festucion hirtovaginatae); (four) Koelerio‐Corynephoretea with Sedo acris‐Festucetalia (Festucion vaginatae) and Trifolio arvensis‐Festucetalia ovinae (Armerio rumelicae‐Potentillion and Minuartio montanae‐Poion molinerii all. nov.).ConclusionsWe created a unified hierarchical classification with an electronic ES using diagnostic species defined by phi‐values. Our new approach (ICO‐HES: iterative cluster optimisation for hierarchical expert systems) allows dividing large data sets into meaningful units at several hierarchical levels, and thus has high potential for complex classifications. Importantly, it overcomes the divide between ES species and diagnostic species and re‐unites them into one concept.

Determination, categorization, and hierarchy of content for a pediatric emergency medicine curriculum designed for emergency medicine residents.

Currently, the Accreditation Council of Graduate Medical Education requires time-based pediatric experiences for emergency medicine (EM) residents in both pediatric emergency medicine (PEM) and critical care settings. The American Board of Emergency Medicine has published the Model of the Clinical Practice of Emergency Medicine, which is a list of content an EM resident should learn. However, this list is large and without prioritization and therefore can be difficult to incorporate into time-limited curricula. The primary objective of this study was to develop comprehensive categorization of PEM content using an EM lens. The second objective was to suggest a prioritization for the EM learner of the enumerated PEM elements. We first assembled a comprehensive list of PEM concepts, diagnoses, and procedures that might be taught to EM residents. We then convened focus groups composed of key stakeholders to help formulate content and concept themes important for EM resident training. Once the themes were identified, we divided the list of PEM topics into appropriate themes and then carried out a second round of focus groups expanded to include more diverse expert input for prioritizing the elements of the comprehensive list within each theme. We prioritized 168 important PEM concepts from previous standards and emerging PEM literature among 10 identified themes: the pediatric normal, the bottom-line boil-it-down approach, common presentations, high-acuity pediatric cases and procedures, differences between children and adults, same between children and adults, red flags, infrequency of caring for a child compared with an adult, keep breadth but promote self-directed depth, and triage and disposition. Based on input from stakeholders in EM resident education, we identified key themes within PEM education and created a framework for the hierarchical categorization of PEM content for within an EM residency.

Hierarchical and Non- Hierarchical Cluster Classification of Precipitation Time Series Data in Nigeria

Precipitation is highly unpredictable over space and time, and is an indicator for climate change. The present work classified average yearly precipitation over the 36 states and the Federal Capital Territory (FCT) in Nigeria into various groups. The hierarchical and non- hierarchical (k-means) cluster method was used in the classification of 37- years precipitation data (1984-2020). The results showed that six major clusters were formed. The Precipitation across the 36 states and the FCT in Nigeria can be classified as Light precipitation for Sokoto, Borno, Kastina, Yobe and Kano States; Moderate precipitation for Kebbi, Zamfara, Dutse, Yola, Gombe and Bauchi States; Heavy precipitation for Kaduna, Plateau, Kwara, Taraba, Nassarawa, Benue and Kogi States; Very Heavy precipitation for Osun, Ogun, FCT, Ekiti, Niger and Lagos States; Shower for Edo, Abia, Delta, Anambra, Ebonyi, Enugu, Ondo and Oyo States and lastly, Heavy Shower for Akwaibom, Imo, Rivers, Bayelsa and Cross River States.. The rangs of precipitation based on the six classification are 1.5-1.8mm (Light precipitation), 1.9-2.7mm (Moderate precipitation), 2.8-3.6mm (Heavy precipitation), 3.7-4.3mm (Very Heavy precipitation), 4.4-5.4mm (Shower) and greater than 5.5mm (Heavy Shower). The hierarchical cluster method performed better than the non- hierarchical (k-means) cluster method in structuring the precipitation data and shaping it into groups (clusters) such that observations within the same group are more similar than those in different groups. The results of this study can contribute to a better understanding of the climate condition of the36 States and the FCT in Nigeria.

Mathematical topology and geometry-based classification of tauopathies

Neurodegenerative diseases, like Alzheimer’s, are associated with the presence of neurofibrillary lesions formed by tau protein filaments in the cerebral cortex. While it is known that different morphologies of tau filaments characterize different neurodegenerative diseases, there are few metrics of global and local structure complexity that enable to quantify their structural diversity rigorously. In this manuscript, we employ for the first time mathematical topology and geometry to classify neurodegenerative diseases by using cryo-electron microscopy structures of tau filaments that are available in the Protein Data Bank. By employing mathematical topology metrics (Gauss linking integral, writhe and second Vassiliev measure) we achieve a consistent, but more refined classification of tauopathies, than what was previously observed through visual inspection. Our results reveal a hierarchy of classification from global to local topology and geometry characteristics. In particular, we find that tauopathies can be classified with respect to the handedness of their global conformations and the handedness of the relative orientations of their repeats. Progressive supranuclear palsy is identified as an outlier, with a more complex structure than the rest, reflected by a small, but observable knotoid structure (a diagrammatic structure representing non-trivial topology). This topological characteristic can be attributed to a pattern in the beginning of the R3 repeat that is present in all tauopathies but at different extent. Moreover, by comparing single filament to paired filament structures within tauopathies we find a consistent change in the side-chain orientations with respect to the alpha carbon atoms at the area of interaction.

Comparing Hierarchical Approaches to Enhance Supervised Emotive Text Classification

The performance of emotive text classification using affective hierarchical schemes (e.g., WordNet-Affect) is often evaluated using the same traditional measures used to evaluate the performance of when a finite set of isolated classes are used. However, applying such measures means the full characteristics and structure of the emotive hierarchical scheme are not considered. Thus, the overall performance of emotive text classification using emotion hierarchical schemes is often inaccurately reported and may lead to ineffective information retrieval and decision making. This paper provides a comparative investigation into how methods used in hierarchical classification problems in other domains, which extend traditional evaluation metrics to consider the characteristics of the hierarchical classification scheme, can be applied and subsequently improve the classification of emotive texts. This study investigates the classification performance of three widely used classifiers, Naive Bayes, J48 Decision Tree, and SVM, following the application of the aforementioned methods. The results demonstrated that all the methods improved the emotion classification. However, the most notable improvement was recorded when a depth-based method was applied to both the testing and validation data, where the precision, recall, and F1-score were significantly improved by around 70 percentage points for each classifier.

Multipartite entanglement detection via generalized Wigner–Yanase skew information

The detection of multipartite entanglement in multipartite quantum systems is a fundamental and key issue in quantum information theory. In this paper, we investigate k-nonseparability and k-partite entanglement of N-partite quantum systems from the perspective of the generalized Wigner–Yanase skew information. We deduce the upper bounds on the generalized Wigner–Yanase skew information of k-separable states and k-producible states in arbitrary dimensional multipartite quantum systems. And based on the upper bounds, we develop two different approaches in form of simple inequalities to construct entanglement criteria, which are expressed in terms of the generalized Wigner–Yanase skew information. Any violation of these inequalities by a quantum state reveals its k-nonseparability or k-partite entanglement, so these inequalities present the hierarchic classifications of k-nonseparability or k-partite entanglement for all N-partite quantum states from N-nonseparability to 2-nonseparability or from 2-partite entanglement to N-partite entanglement. Some well-known entanglement criteria are the special cases of our criteria. We illustrate the significance of our criteria by showing that they can reveal k-nonseparability and k-partite entanglement undetectable by well-known criteria through several typical examples, and their advantages over well-known criteria.

Hierarchical Text Classification and Its Foundations: A Review of Current Research

While collections of documents are often annotated with hierarchically structured concepts, the benefits of these structures are rarely taken into account by classification techniques. Within this context, hierarchical text classification methods are devised to take advantage of the labels’ organization to boost classification performance. In this work, we aim to deliver an updated overview of the current research in this domain. We begin by defining the task and framing it within the broader text classification area, examining important shared concepts such as text representation. Then, we dive into details regarding the specific task, providing a high-level description of its traditional approaches. We then summarize recently proposed methods, highlighting their main contributions. We also provide statistics for the most commonly used datasets and describe the benefits of using evaluation metrics tailored to hierarchical settings. Finally, a selection of recent proposals is benchmarked against non-hierarchical baselines on five public domain-specific datasets. These datasets, along with our code, are made available for future research.

Afghan Women’s Resistance - Forty Years of Struggle Against Gender Apartheid

The defeat of the U.S. client regime in Afghanistan and the seizure of power by the Taliban in August 2021 marked a real turning point. These events represented another major setback for the United States in the wake of a failed war in Iraq. Journalists rushed to compare the debacle in Kabul in 2021 to Saigon in 1975, as Afghans fearful of Taliban rule scrambled to get onto US planes. Many were left behind as the United States rushed to get its own forces and those of its allies out. 
 The August 2021 regime collapse in Afghanistan, although sudden in its final manifestation, was a long time coming. The United States realized it had been defeated at least by 2020, as the Trump administration agreed to a total US withdrawal in direct negotiations with the Taliban. The Biden administration continued this policy, which had two basic aspects: the United States would withdraw by the end of August 2021, and the Taliban would not attack US forces during the period of withdrawal. Both sides kept to the bargain; the Afghan people were not consulted at all, nor was the US-backed government of Ashraf Ghani, who was not even included in the negotiations.
 There was an important difference from the situation in Saigon in 1975, however. The forces that defeated the United States in Vietnam included female combatants and officers. Moreover, the regime they installed to replace the US client state espoused a modernist, if authoritarian, ideology that extolled gender equality, land reform, and other forms of social and economic transformation.
 In contrast, the return to power of the Taliban was instead a setback for women’s rights of epochal proportions, and for other social and political rights as well. They set about establishing an ultra-conservative fundamentalist regime of a type not seen since the Islamic State was driven out of Raqqa, Syria, in 2017. The Taliban have again established a theocracy, which openly supports long-standing hierarchies of gender, ethnicity, religion, and class, albeit with a somewhat modern form of organization, including a surveillance apparatus and modern weapons. With its denial of secondary education to girls, the new Taliban regime’s level of gender apartheid far exceeds those of Saudi Arabia and Iran. At this writing, not a single country, not even Saudi Arabia, has formally recognized the Taliban government.

Host-microbiome associations in saliva predict COVID-19 severity.

Established evidence indicates that oral microbiota plays a crucial role in modulating host immune responses to viral infection. Following severe acute respiratory syndrome coronavirus 2, there are coordinated microbiome and inflammatory responses within the mucosal and systemic compartments that are unknown. The specific roles the oral microbiota and inflammatory cytokines play in the pathogenesis of coronavirus disease 2019 (COVID-19) are yet to be explored. Here, we evaluated the relationships between the salivary microbiome and host parameters in different groups of COVID-19 severity based on their oxygen requirement. Saliva and blood samples (n = 80) were collected from COVID-19 and from noninfected individuals. We characterized the oral microbiomes using 16S ribosomal RNA gene sequencing and evaluated saliva and serum cytokines and chemokines using multiplex analysis. Alpha diversity of the salivary microbial community was negatively associated with COVID-19 severity, while diversity increased with health. Integrated cytokine evaluations of saliva and serum showed that the oral host response was distinct from the systemic response. The hierarchical classification of COVID-19 status and respiratory severity using multiple modalities separately (i.e. microbiome, salivary cytokines, and systemic cytokines) and simultaneously (i.e. multimodal perturbation analyses) revealed that the microbiome perturbation analysis was the most informative for predicting COVID-19 status and severity, followed by the multimodal. Our findings suggest that oral microbiome and salivary cytokines may be predictive of COVID-19 status and severity, whereas atypical local mucosal immune suppression and systemic hyperinflammation provide new cues to understand the pathogenesis in immunologically compromised populations.

Self-Paced Unified Representation Learning for Hierarchical Multi-Label Classification

Hierarchical Multi-Label Classification (HMLC) is a well-established problem that aims at assigning data instances to multiple classes stored in a hierarchical structure. Despite its importance, existing approaches often face two key limitations: (i) They employ dense networks to solely explore the class hierarchy as hard criterion for maintaining taxonomic consistency among predicted classes, yet without leveraging rich semantic relationships between instances and classes; (ii) They struggle to generalize in settings with deep class levels, since the mini-batches uniformly sampled from different levels ignore the varying complexities of data and result in a non-smooth model adaptation to sparse data. To mitigate these issues, we present a Self-Paced Unified Representation (SPUR) learning framework, which focuses on the interplay between instance and classes to flexibly organize the training process of HMLC algorithms. Our framework consists of two lightweight encoders designed to capture the semantics of input features and the topological information of the class hierarchy. These encoders generate unified embeddings of instances and class hierarchy, which enable SPUR to exploit semantic dependencies between them and produce predictions in line with taxonomic constraints. Furthermore, we introduce a dynamic hardness measurement strategy that considers both class hierarchy and instance features to estimate the learning difficulty of each instance. This strategy is achieved by incorporating the propagation loss obtained at each hierarchical level, allowing for a more comprehensive assessment of learning complexity. Extensive experiments on several empirical benchmarks demonstrate the effectiveness and efficiency of SPUR compared to state-of-the-art methods, especially in scenarios with missing features.

Scaling Few-Shot Learning for the Open World

Few-shot learning (FSL) aims to enable learning models with the ability to automatically adapt to novel (unseen) domains in open-world scenarios. Nonetheless, there exists a significant disparity between the vast number of new concepts encountered in the open world and the restricted available scale of existing FSL works, which primarily focus on a limited number of novel classes. Such a gap hinders the practical applicability of FSL in realistic scenarios. To bridge this gap, we propose a new problem named Few-Shot Learning with Many Novel Classes (FSL-MNC) by substantially enlarging the number of novel classes, exceeding the count in the traditional FSL setup by over 500-fold. This new problem exhibits two major challenges, including the increased computation overhead during meta-training and the degraded classification performance by the large number of classes during meta-testing. To overcome these challenges, we propose a Simple Hierarchy Pipeline (SHA-Pipeline). Due to the inefficiency of traditional protocols of EML, we re-design a lightweight training strategy to reduce the overhead brought by much more novel classes. To capture discriminative semantics across numerous novel classes, we effectively reconstruct and leverage the class hierarchy information during meta-testing. Experiments show that the proposed SHA-Pipeline significantly outperforms not only the ProtoNet baseline but also the state-of-the-art alternatives across different numbers of novel classes.

ETDPC: A Multimodality Framework for Classifying Pages in Electronic Theses and Dissertations

Electronic theses and dissertations (ETDs) have been proposed, advocated, and generated for more than 25 years. Although ETDs are hosted by commercial or institutional digital library repositories, they are still an understudied type of scholarly big data, partially because they are usually longer than conference and journal papers. Segmenting ETDs will allow researchers to study sectional content. Readers can navigate to particular pages of interest, to discover and explore the content buried in these long documents. Most existing frameworks on document page classification are designed for classifying general documents, and perform poorly on ETDs. In this paper, we propose ETDPC. Its backbone is a two-stream multimodal model with a cross-attention network to classify ETD pages into 13 categories. To overcome the challenge of imbalanced labeled samples, we augmented data for minority categories and employed a hierarchical classifier. ETDPC outperforms the state-of-the-art models in all categories, achieving an F1 of 0.84 -- 0.96 for 9 out of 13 categories. We also demonstrated its data efficiency. The code and data can be found on GitHub (https://github.com/lamps-lab/ETDMiner/tree/master/etd_segmentation).

Abnormal line loss identification and category classification of distribution networks based on semi-supervised learning and hierarchical classification

Line loss refers to the electrical energy that is dissipated as heat during the transmission and distribution of electricity through power lines. However, unusual causes, such as grid topology mismatch and communication failure, can cause abnormal line loss. Efficient abnormal line loss detection contributes not only to minimizing energy wastage and reducing carbon emissions but also to maintaining the stability and reliability of the entire distribution network. In actual situations, the cause of abnormal line loss is not labeled due to the expensive labor cost. This paper proposes a hierarchical abnormal line loss identification and category classification model, considering the unlabeled and unbalanced sample problem. First, an abnormal line loss identification model-based random forest is established to detect whether the line loss is abnormal. Then, an abnormal line loss category classification model is developed with semi-supervised learning for line loss abnormal category classification, considering the unlabeled samples. The real dataset in China is utilized to validate the performance of the proposed model. Its reliability implies the potential to be applied to real-world scenarios to improve the management level and safety of the power grid.

An integrated hierarchical classification and machine learning approach for mapping land use and land cover in complex social-ecological systems

Mapping land use and land cover (LULC) using remote sensing is fundamental to environmental monitoring, spatial planning and characterising drivers of change in landscapes. We develop a new, general and versatile approach for mapping LULC in landscapes with relatively gradual transition between LULC categories such as African savannas. The approach integrates a well-tested hierarchical classification system with the computationally efficient random forest (RF) classifier and produces detailed, accurate and consistent classification of structural vegetation heterogeneity and density and anthropogenic land use. We use Landsat 8 OLI imagery to illustrate this approach for the Extended Greater Masai Mara Ecosystem (EGMME) in southwestern Kenya. We stratified the landscape into eight relatively homogeneous zones, systematically inspected the imagery and randomly allocated 1,697 training sites, 556 of which were ground-truthed, proportionately to the area of each zone. We directly assessed the accuracy of the visually classified image. Accuracy was high and averaged 88.1% (80.5%–91.7%) across all the zones and 89.1% (50%–100%) across all the classes. We applied the RF classifier to randomly selected samples from the original training dataset, separately for each zone and the EGMME. We evaluated the overall and class-specific accuracy and computational efficiency using the Out-of-Bag (OOB) error. Overall accuracy (79.3%–97.4%) varied across zones but was higher whereas the class-specific accuracy (25.4%–98.1%) was lower than that for the EGMME (80.2%). The hierarchical classifier identified 35 LULC classes which we aggregated into 18 intermediate mosaics and further into five more general categories. The open grassed shrubland (21.8%), sparse shrubbed grassland (10.4%) and small-scale cultivation (13.3%) dominated at the detailed level, grassed shrubland (31.9%) and shrubbed grassland (28.9%) at the intermediate level, and grassland (35.7%), shrubland (35.3%) and woodland (12.5%) at the general level. Our granular LULC map for the EGMME is sufficiently accurate for important practical purposes such as land use spatial planning, habitat suitability assessment and temporal change detection. The extensive ground-truthing data, sample site photos and classified maps can contribute to wider validation efforts at regional to global scales.

UP-SDCG: A Method of Sensitive Data Classification for Collaborative Edge Computing in Financial Cloud Environment

The digital transformation of banks has led to a paradigm shift, promoting the open sharing of data and services with third-party providers through APIs, SDKs, and other technological means. While data sharing brings personalized, convenient, and enriched services to users, it also introduces security risks, including sensitive data leakage and misuse, highlighting the importance of data classification and grading as the foundational pillar of security. This paper presents a cloud-edge collaborative banking data open application scenario, focusing on the critical need for an accurate and automated sensitive data classification and categorization method. The regulatory outpost module addresses this requirement, aiming to enhance the precision and efficiency of data classification. Firstly, regulatory policies impose strict requirements concerning data protection. Secondly, the sheer volume of business and the complexity of the work situation make it impractical to rely on manual experts, as they incur high labor costs and are unable to guarantee significant accuracy. Therefore, we propose a scheme UP-SDCG for automatically classifying and grading financially sensitive structured data. We developed a financial data hierarchical classification library. Additionally, we employed library augmentation technology and implemented a synonym discrimination model. We conducted an experimental analysis using simulation datasets, where UP-SDCG achieved precision surpassing 95%, outperforming the other three comparison models. Moreover, we performed real-world testing in financial institutions, achieving good detection results in customer data, supervision, and additional in personally sensitive information, aligning with application goals. Our ongoing work will extend the model’s capabilities to encompass unstructured data classification and grading, broadening the scope of application.

Self-attention-based generative adversarial network optimized with color harmony algorithm for brain tumor classification

ABSTRACT This paper proposes a novel approach, BTC-SAGAN-CHA-MRI, for the classification of brain tumors using a SAGAN optimized with a Color Harmony Algorithm. Brain cancer, with its high fatality rate worldwide, especially in the case of brain tumors, necessitates more accurate and efficient classification methods. While existing deep learning approaches for brain tumor classification have been suggested, they often lack precision and require substantial computational time.The proposed method begins by gathering input brain MR images from the BRATS dataset, followed by a pre-processing step using a Mean Curvature Flow-based approach to eliminate noise. The pre-processed images then undergo the Improved Non-Sub sampled Shearlet Transform (INSST) for extracting radiomic features. These features are fed into the SAGAN, which is optimized with a Color Harmony Algorithm to categorize the brain images into different tumor types, including Gliomas, Meningioma, and Pituitary tumors. This innovative approach shows promise in enhancing the precision and efficiency of brain tumor classification, holding potential for improved diagnostic outcomes in the field of medical imaging. The accuracy acquired for the brain tumor identification from the proposed method is 99.29%. The proposed BTC-SAGAN-CHA-MRI technique achieves 18.29%, 14.09% and 7.34% higher accuracy and 67.92%,54.04%, and 59.08% less Computation Time when analyzed to the existing models, like Brain tumor diagnosis utilizing deep learning convolutional neural network with transfer learning approach (BTC-KNN-SVM-MRI); M3BTCNet: multi model brain tumor categorization under metaheuristic deep neural network features optimization (BTC-CNN-DEMFOA-MRI), and efficient method depending upon hierarchical deep learning neural network classifier for brain tumour categorization (BTC-Hie DNN-MRI) respectively.

Brain tumor grading diagnosis using transfer learning based on optical coherence tomography.

In neurosurgery, accurately identifying brain tumor tissue is vital for reducing recurrence. Current imaging techniques have limitations, prompting the exploration of alternative methods. This study validated a binary hierarchical classification of brain tissues: normal tissue, primary central nervous system lymphoma (PCNSL), high-grade glioma (HGG), and low-grade glioma (LGG) using transfer learning. Tumor specimens were measured with optical coherence tomography (OCT), and a MobileNetV2 pre-trained model was employed for classification. Surgeons could optimize predictions based on experience. The model showed robust classification and promising clinical value. A dynamic t-SNE visualized its performance, offering a new approach to neurosurgical decision-making regarding brain tumors.

CAM based fine-grained spatial feature supervision for hierarchical yoga pose classification using multi-stage transfer learning

In this paper, we propose a technique for hierarchical yoga pose classification (YPC) in a multi-stage multi-tasking framework. We propose a three-stage transfer learning based end-to-end training methodology. Novelty lies in (a) proposed supervised contrastive combined loss function for stage-1 training, (b) proposed Encoder–Decoder network architecture with attention mechanism for stage-3 training, (c) proposed spatial context aware multi-tasking combined loss function for stage-3 training. Firstly, for stage-1 training, we propose the usage of linear combination of three loss functions: cross-entropy, self-supervised contrastive loss and supervised contrastive loss in a multi-tasking manner. We introduce radial and cosine margin into the formulation of self-supervised and supervised contrastive loss to pull feature embeddings of same class closer together compared to feature embeddings of different classes. Weights learned over stage-1 training are subsequently fine-tuned over cross-entropy multi-tasking loss in stage-2. These stage-2 weights are transfer learned and are further fine-tuned in stage-3 training. For stage-3 training, we propose the usage of spatial context aware multi-tasking combined loss function. This loss function leverages on the fine-grained spatial features obtained from HiResCAM. These are processed in parallel with features obtained from XGrad-CAM (sensitivity and conservation axioms satisfying features) to further supervise the learning of hierarchical yoga pose classifier. We exemplify our methodology on the publicly available Yoga-82 large-scale dataset. We report peak Top-1 YPC accuracy of 95.89% over 6 pose classes (Yoga-6), 93.85% over 20 pose classes (Yoga-20) and 90.0% over 82 pose classes (Yoga-82). Our proposed method achieves 6.1% improvement over Top-1 classification accuracy in Yoga-6 hierarchy, 9.3% improvement in Yoga-20 hierarchy and 10.9% improvement in Yoga-82 hierarchy in comparison with state-of-the-art (SOTA) methodology. We achieve the current best Top-1 classification accuracies in all the three YPC hierarchies.

DeepMedFeature: An Accurate Feature Extraction and Drug-Drug Interaction Model for Clinical Text in Medical Informatics

Drug-drug interactions (DDIs) are an important biological phenomenon which can result in medical errors from medical practitioners. Drug interactions can change the molecular structure of interacting agents which may prove to be fatal in the worst case. Finding drug interactions early in diagnosis can be pivotal in side-effect prevention. The growth of big data provides a rich source of information for clinical studies to investigate DDIs. We propose a hierarchical classification model which is double-pass in nature. The first pass predicts the occurrence of an interaction and then the second pass further predicts the type of interaction such as effect, advice, mechanism, and int. We applied different deep learning algorithms with Convolutional Bi-LSTM (ConvBLSTM) proving to be the best. The results show that pre-trained vector embeddings prove to be the most appropriate features. The F1-score of the ConvBLSTM algorithm turned out to be 96.39% and 98.37% in Russian and English language respectively which is greater than the state-of-the-art systems. According to the results, it can be concluded that adding a convolution layer before the bi-directional pass improves model performance in the automatic classification and extraction of drug interactions, using pre-trained vector embeddings such as Fasttext and Bio-Bert.