Negative Examples Research Articles

Identifying Performance Issues in Cloud Service Systems Based on Relational-Temporal Features

Cloud systems, typically comprised of various components ( e.g. , microservices), are susceptible to performance issues, which may cause service-level agreement violations and financial losses. Identifying performance issues is thus of paramount importance for cloud vendors. In current practice, crucial metrics, i.e. , key performance indicators (KPIs), are monitored periodically to provide insight into the operational status of components. Identifying performance issues is often formulated as an anomaly detection problem, which is tackled by analyzing each metric independently. However, this approach overlooks the complex dependencies existing among cloud components. Some graph neural network-based methods take both temporal and relational information into account, however, the correlation violations in the metrics that serve as indicators of underlying performance issues are difficult for them to identify. Furthermore, a large volume of components in a cloud system results in a vast array of noisy metrics. This complexity renders it impractical for engineers to fully comprehend the correlations, making it challenging to identify performance issues accurately. To address these limitations, we propose Identifying Performance Issues based on Relational-Temporal Features (ISOLATE ), a learning-based approach that leverages both the relational and temporal features of metrics to identify performance issues. In particular, it adopts a graph neural network with attention to characterizing the relations among metrics and extracts long-term and multi-scale temporal patterns using a GRU and a convolution network, respectively. The learned graph attention weights can be further used to localize the correlation-violated metrics. Moreover, to relieve the impact of noisy data, ISOLATE utilizes a positive unlabeled learning strategy that tags pseudo labels based on a small portion of confirmed negative examples. Extensive evaluation on both public and industrial datasets shows that ISOLATE outperforms all baseline models with 0.945 F1-score and 0.920 Hit rate@3. The ablation study also proves the effectiveness of the relational-temporal features and the PU-learning strategy. Furthermore, we share the success stories of leveraging ISOLATE to identify performance issues in Huawei Cloud, which demonstrates its superiority in practice.

SGO: An innovative oversampling approach for imbalanced datasets using SVM and genetic algorithms

Imbalanced datasets present a challenging problem in machine learning and artificial intelligence. Since most models typically assume balanced data distributions, imbalanced positive and negative examples can lead to significant bias in prediction or classification tasks. Current over-sampling methods frequently encounter issues like overfitting and boundary bias. A novel imbalanced data augmentation technique called SVM-GA over-sampling (SGO) is proposed in this paper, which integrates Support Vector Machines (SVM) with Genetic Algorithms (GA). Our approach leverages SVM to identify the decision boundary and uses GA to generate new minority samples along this boundary, effectively addressing both over-fitting and boundary biases. It has been experimentally validated that SGO outperforms the traditional methods on most datasets, providing a novel and effective approach to address imbalanced data problems, with potential application prospects and generalization value.

Effective Hard Negative Mining for Contrastive Learning-based Code Search

Background. Code search aims to find the most relevant code snippet in a large codebase based on a given natural language query. An accurate code search engine can increase code reuse and improve programming efficiency. The focus of code search is how to represent the semantic similarity of code and query. With the development of code pre-trained models, the pattern of using numeric feature vectors (embeddings) to represent code semantics and using vector distance to represent semantic similarity has replaced traditional string matching methods. The quality of semantic representations is critical to the effectiveness of downstream tasks such as code search. Currently, the state-of-the-art (SOTA) learning method uses the contrastive learning paradigm. The objective of contrastive learning is to maximize the similarity between matching code and query (positive samples) and minimize the similarity between mismatched pairs (negative samples). To increase the reusing of negative samples, prior contrastive learning approaches use a large queue (memory bank) to store embeddings. Problem. However, there is still a lot of room for improvement in using negative examples for code search: ① Due to the random selection of negative samples, semantic representations learned by existing models cannot distinguish similar codes well. ② Since semantic vectors in the memory bank are reused from previous inference results and then directly used for loss function calculation without gradient descent, the model cannot effectively learn the negative sample semantic information. Method. To solve the above problems, we propose a contrastive learning code search model with hard negative mining called CoCoHaNeRe: ❶ To enable the model to distinguish similar codes, we introduce hard negative examples into contrastive training, which are negative examples in the codebase that are most similar to positive examples. As a result, hard negative examples are most likely to make the model make mistakes. ❷ To improve the learning efficiency of negative samples during training, we add all hard negative examples to the model's gradient descent process. Result. To verify the effectiveness of CoCoHaNeRe, we conducted experiments on large code search datasets with six programming languages, as well as similar retrieval tasks code clone detection and code question answering. Experimental results show that our model achieves SOTA performance. In the code search task, the average MRR score of CoCoHaNeRe exceeds CodeBERT, GraphCodeBERT, and UniXcoder by 11.25%, 8.13%, and 7.38%, respectively. It has also made great progress in code clone detection and code question answering. In addition, our method performs well in different programming languages and code pre-training models. Furthermore, qualitative analysis shows that our model effectively distinguishes high-order semantic differences between similar codes.

MT-Net: Single image dehazing based on meta learning, knowledge transfer and contrastive learning

Single image dehazing is becoming increasingly important as its results impact the efficiency of subsequent computer vision tasks. While many methods have been proposed to address this challenge, existing dehazing approaches often exhibit limited adaptability to different types of images and lack future learnability. In light of this, we propose a dehazing network based on meta-learning, knowledge transfer, and contrastive learning, abbreviated as MT-Net. In our approach, we combine knowledge transfer with meta-learning to tackle these challenges, thus enhancing the network’s generalization performance. We refine the structure of knowledge transfer by introducing a two-phases approach to facilitate learning under the guidance of teacher networks and learning committee networks. We also optimize the negative examples of contrastive learning to reduce the contrast space. Extensive experiments conducted on synthetic and real datasets demonstrate the remarkable performance of our method in both quantitative and qualitative comparisons. The code has been released on https://github.com/71717171fan/MT-Net.

Tennis teaching assistance model based on double chain shared unsupervised action recognition algorithm

The traditional self-supervised methods based on skeleton data frequently categorize the various enhancements of a given sample as positive examples, while the remaining samples are designated as negative examples. This approach results in a significant imbalance in the ratio of positive to negative samples, which in turn constrains the efficacy of samples with identical semantic information. Therefore, to further improve the training quality in tennis teaching and enhance the accuracy of action recognition, a dual chain sharing unsupervised action recognition algorithm has been proposed. This study first designs a skeleton topology data augmentation method based on the physical connections of human joints to obtain advanced semantic embeddings. Then, an improved positive sample expansion strategy is utilized to enhance the diversity and quality of training data. Next, an unsupervised learning mechanism is employed to autonomously learn and recognize complex patterns of tennis movements. To measure the performance of the model, tests were conducted on its accuracy, F1 score, recognition time, and fitting degree. The experimental results showed that the proposed algorithm could reach its optimal state after 23 iterations of training, and its F1 value reached 0.918, with an average accuracy of 92.9 % and an average recognition time of 7.4 s. The research algorithms were superior to the multi-input branch graph convolutional network action recognition algorithms used for comparison, pull-push contrastive loss action recognition algorithms, and multi-granularity anchor contrastive representation learning action recognition algorithms. Its accuracy was leading by 8.6 %-23.5 %, and the recognition time has been reduced by 2.3s-7.4 s. In addition, in terms of the fitting degree of action recognition, compared with other methods, the proposed method had a fitting degree of up to 95.7 %, which was at least 8.1 % higher. This research method can improve the timeliness of feedback from trainers by accurately and quickly recognizing movements, helping coaches develop more targeted training plans, thereby improving the efficiency and quality of tennis teaching.

Aerial imagery dataset of lost oil wells

Orphaned wells are wells for which the operator is unknown or insolvent. The location of hundreds of thousands of these wells remain unknown in the United States alone. Cost-effective techniques are essential to locate orphaned wells to address environmental problems. In this paper, we present a dataset consisting of 120,948 aerial images of recently documented orphan wells. Each of these 512 × 512 images is paired with segmentation masks that indicate the presence or absence of such well. These images, sourced from the National Agriculture Imagery Program, cover the continental United States with spatial resolutions ranging from 30 centimeters to 1 meter. Additionally, we included negative examples by selecting locations uniformly across the United States. Accompanying metadata includes the IDs and spatial resolution of the original images, which are available for free through the United States Geological Survey, and the pixel coordinates of documented orphaned wells identified in these images. This dataset is intended to support the development of deep-learning models that can help locating undocumented orphan wells from such imagery, thereby blunting the environmental damage they do.

Automated Detection of Hillforts in Remote Sensing Imagery With Deep Multimodal Segmentation

ABSTRACTRecent advancements in remote sensing and artificial intelligence can potentially revolutionize the automated detection of archaeological sites. However, the challenging task of interpreting remote sensing imagery combined with the intricate shapes of archaeological sites can hinder the performance of computer vision systems. This work presents a computer vision system trained for efficient hillfort detection in remote sensing imagery. Equipped with an adapted multimodal semantic segmentation model, the system integrates LiDAR‐derived LRM images and aerial orthoimages for feature fusion, generating a binary mask pinpointing detected hillforts. Post‐processing includes margin and area filters to remove edge inferences and smaller anomalies. The resulting inferences are subjected to hard positive and negative mining, where expert archaeologists classify them to populate the training data with new samples for retraining the segmentation model. As the computer vision system is far more likely to encounter background images during its search, the training data are intentionally biased towards negative examples. This approach aims to reduce the number of false positives, typically seen when applying machine learning solutions to remote sensing imagery. Northwest Iberia experiments witnessed a drastic reduction in false positives, from 5678 to 40 after a single hard positive and negative mining iteration, yielding a 99.3% reduction, with a resulting F1 score of 66%. In England experiments, the system achieved a 59% F1 score when fine‐tuned and deployed countrywide. Its scalability to diverse archaeological sites is demonstrated by successfully detecting hillforts and other types of enclosures despite their typical complex and varied shapes. Future work will explore archaeological predictive modelling to identify regions with higher archaeological potential to focus the search, addressing processing time challenges.

The role of behavioral decision-making in panic buying events during COVID-19: From the perspective of an evolutionary game based on prospect theory

Sudden public health crises such as the COVID-19 pandemic have triggered widespread panic buying incidents globally, causing disruptions in commodity markets characterized by imbalances between supply and demand. This has posed a serious threat to social stability. During sudden public health events, panic buying typically complicates epidemic prevention and control efforts, leading to adverse consequences such as resource wastage, supply-demand imbalances, and market failures, which pose a threat to social stability and economic development. The progression of panic buying is significantly influenced by the actions and choices of both the public and local authorities. Therefore, studying the impact of the behavioral decisions made by the public and local governments plays a pivotal role in the evolution of such events. Building upon this premise, this paper studies the role of behavioral decisions in the COVID-19 panic buying event from the perspective of an evolutionary game. First, the public’s attitude toward the development of the event is divided into two categories: those with a negative attitude and those with a positive attitude. At the same time, the concepts of mental benefits and mental costs are introduced and quantified, serving as key factors in public decision-making. Second, prospect theory is used to elucidate the risk preferences and loss aversion of decision-makers. An evolutionary game model is constructed with the public and local governments as the main entities based on prospect theory. The replication dynamic equation is constructed using the benefit perception matrix, and the equilibrium point and stability conditions of the model are deduced. Third, the impact of participants' behavioral decisions during the panic buying event on the event’s evolution process is discussed through numerical simulation analysis. The results show that (1) in the context of sudden public health emergencies, material demand is the primary cause of public panic buying behavior. Postpanic buying events, such behavior is influenced primarily by the proportion of individuals with negative attitudes within the population. Local governments should release positive information regarding supplies in a timely manner during the nascent stage of panic buying events and actively monitor the proportion of individuals with negative attitudes. (2) There is a positive correlation between risk preference and public panic buying behavior, whereas loss aversion has the opposite effect. The spiritual benefits perceived by the public from engaging in panic buying increase with higher risk preference, whereas perceived costs related to health issues increase with deeper levels of loss aversion. (3) When response costs exceed what local governments are willing to bear, reputation gains and losses play a crucial role in their decision-making. The probability of local governments adopting responsive strategies increases with their emphasis on reputation and aversion to reputation losses. Higher-level governments should enhance the dissemination of positive and negative examples of government handling, enhancing local governments' sense of responsibility and honor to improve their perception of reputation gains and losses, thereby encouraging proactive management of panic buying events.

Impact of different parametric Patlak imaging approaches and comparison with a 2-tissue compartment pharmacokinetic model with a long axial field-of-view (LAFOV) PET/CT in oncological patients.

The recently introduced Long-Axial-Field-of-View (LAFOV) PET-CT scanners allow for the first-time whole-body dynamic- and parametric imaging. Primary aim of this study was the comparison of direct and indirect Patlak imaging as well as the comparison of different time frames for Patlak calculation with the LAFOV PET-CT in oncological patients. Secondary aims of the study were lesion detectability and comparison of Patlak analysis with a two-tissue-compartment model (2TCM). 50 oncological patients with 346 tumor lesions were enrolled in the study. All patients underwent [18F]FDG PET/CT (skull to upper thigh). Here, the Image-Derived-Input-Function) (IDIF) from the descending aorta was used as the exclusive input function. Four sets of images have been reviewed visually and evaluated quantitatively using the target-to-background (TBR) and contrast-to-noise ratio (CNR): short-time (30min)-direct (STD) Patlak Ki, short-time (30min)-indirect (STI) Patlak Ki, long-time (59.25min)-indirect (LTI) Patlak Ki, and 50-60min SUV (sumSUV). VOI-based 2TCM was used for the evaluation of tumor lesions and normal tissues and compared with the results of Patlak model. No significant differences were observed between the four approaches regarding the number of tumor lesions. However, we found three discordant results: a true positive liver lesion in all Patlak Ki images, a false positive liver lesion delineated only in LTI Ki which was a hemangioma according to MRI and a true negative example in a patient with an atelectasis next to a lung tumor. STD, STI and LTI Ki images had superior TBR in comparison with sumSUV images (2.9-, 3.3- and 4.3-fold higher respectively). TBR of LTI Ki were significantly higher than STD Ki. VOI-based k3 showed a 21-fold higher TBR than sumSUV. Parameters of different models vary in their differential capability between tumor lesions and normal tissue like Patlak Ki which was better in normal lung and 2TCM k3 which was better in normal liver. 2TCM Ki revealed the highest correlation (r = 0.95) with the LTI Patlak Ki in tumor lesions group and demonstrated the highest correlation with the STD Patlak Ki in all tissues group and normal tissues group (r = 0.93 and r = 0.74 respectively). Dynamic [18F]-FDG with the new LAFOV PET/CT scanner produces Patlak Ki images with better lesion contrast than SUV images, but does not increase the lesion detection rate. The time window used for Patlak imaging plays a more important role than the direct or indirect method. A combination of different models, like Patlak and 2TCM may be helpful in parametric imaging to obtain the best TBR in the whole body in future.

THDet: A Lightweight and Efficient Traffic Helmet Object Detector based on YOLOv8

Traffic helmet object detection is playing an increasing important role in the smart traffic fields. However, object size variation and small-shaped helmet detection has still been a challenging problem by reason of their poor visual appearance in the image. In this work, we present an efficient traffic helmet detector through feature enhancement and lightweight design based on YOLOv8n called THDet. Specifically, we employ the coordinate attention into C2f blocks combined with softmax activate function to achieve feature channel aggregation and strong non-linear expression of the backbone for further effective feature extraction; Next, Focal_CIoU loss function embedded with Focal Loss method is utilized for the more precise measure of various objects bounding box regression and balance of positive and negative examples during training; Then, a new lightweight detection head style is designed only with two proper position heads (P3 & P4) to perform final classification and localization, through this scheme saving the 33.7% parameters than baseline method. Finally, Attention Refined Features Module (ARFM) is built to calibrate the multi-scale fused features by introducing 3-D weights generated from SimAttention to boost the final detection accuracy. Extensive experiments have demonstrated that our proposed method realizes noticeable performance in terms of detection accuracy and inference speed compared with baseline YOLOv8n and many end-to-end detectors of similar model size. Concretely, THDet achieves 0.447 at the overall evaluation metric of mAP0.5−0.95, accomplishing 3.2% detection accuracy improvement than YOLOv8n. Besides, THDet only holds 2.2M parameters with 295 FPS inference speed, reducing 33.4% parameters compared with YOLOv8n. The experimental results validate the effectiveness of our proposed method, showcasing that THDet outperforms the mainstream real-time detection algorithms in the terms of accuracy, inference speed and lightweight model design for traffic helmet object detection.

Function-Genes and Disease-Genes Prediction Based on Network Embedding and One-Class Classification.

Using genes which have been experimentally-validated for diseases (functions) can develop machine learning methods to predict new disease/function-genes. However, the prediction of both function-genes and disease-genes faces the same problem: there are only certain positive examples, but no negative examples. To solve this problem, we proposed a function/disease-genes prediction algorithm based on network embedding (Variational Graph Auto-Encoders, VGAE) and one-class classification (Fast Minimum Covariance Determinant, Fast-MCD): VGAEMCD. Firstly, we constructed a protein-protein interaction (PPI) network centered on experimentally-validated genes; then VGAE was used to get the embeddings of nodes (genes) in the network; finally, the embeddings were input into the improved deep learning one-class classifier based on Fast-MCD to predict function/disease-genes. VGAEMCD can predict function-gene and disease-gene in a unified way, and only the experimentally-verified genes are needed to provide (no need for expression profile). VGAEMCD outperforms classical one-class classification algorithms in Recall, Precision, F-measure, Specificity, and Accuracy. Further experiments show that seven metrics of VGAEMCD are higher than those of state-of-art function/disease-genes prediction algorithms. The above results indicate that VGAEMCD can well learn the distribution characteristics of positive examples and accurately identify function/disease-genes.

Improving dictionary-based named entity recognition with deep learning.

Dictionary-based named entity recognition (NER) allows terms to be detected in a corpus and normalized to biomedical databases and ontologies. However, adaptation to different entity types requires new high-quality dictionaries and associated lists of blocked names for each type. The latter are so far created by identifying cases that cause many false positives through manual inspection of individual names, a process that scales poorly. In this work, we aim to improve block list s by automatically identifying names to block, based on the context in which they appear. By comparing results of three well-established biomedical NER methods, we generated a dataset of over 12.5 million text spans where the methods agree on the boundaries and type of entity tagged. These were used to generate positive and negative examples of contexts for four entity types (genes, diseases, species, and chemicals), which were used to train a Transformer-based model (BioBERT) to perform entity type classification. Application of the best model (F1-score = 96.7%) allowed us to generate a list of problematic names that should be blocked. Introducing this into our system doubled the size of the previous list of corpus-wide blocked names. In addition, we generated a document-specific list that allows ambiguous names to be blocked in specific documents. These changes boosted text mining precision by ∼5.5% on average, and over 8.5% for chemical and 7.5% for gene names, positively affecting several biological databases utilizing this NER system, like the STRING database, with only a minor drop in recall (0.6%). All resources are available through Zenodo https://doi.org/10.5281/zenodo.11243139 and GitHub https://doi.org/10.5281/zenodo.10289360.

Meta-learning triplet contrast network for few-shot text classification

Few-shot text classification (FSTC) strives to predict classes not involved in the training by learning from a few labeled examples. Currently, most tasks construct meta-tasks in a randomized manner that fails to give more priority to hard-to-identify classes and samples. Besides, some tasks incorporated a contrast strategy, but the sample could only be compared to positive or negative examples individually. In this work, we propose a Meta-learning Triplet Contrast Network (Meta-TCN) with bidirectional contrast capability to solve the above problem. Specifically, Meta-TCN uses external knowledge with labeled information as the class examples, which decouples the embedding of prototypes from the support pool. Meanwhile, the class examples combine the support samples to construct triplet pairs used for learning. Unlike previous studies, the model can learn negative and positive knowledge simultaneously, ensuring that understanding is enriched and enhances learning. Further, we improve the shortcomings of randomness in the meta-task construction process by proposing a Dynamic Rate of Change (DRC) sampling strategy. DRC enhances the model’s focus on difficult-to-classify samples. We conducted extensive experiments on six benchmark datasets such as Huffpost and RCV1. Experiments show that the average accuracy of Meta-TCN can achieve state-of-the-art performance in the vast majority of tasks.

Leisure constraints towards accessible tourism: self-characteristics of people with disability along tourism value chains

ABSTRACT Despite the intensifying legal focus on enhancing accessible infrastructure, people with disability are still often marginalized in traveling. To facilitate the advancement of accessible tourism, this research employs the self-categorization theory, incorporating the experiences of individuals with disabilities within all stages of the travel process. Following a qualitative approach, leisure constraints of people with hearing, seeing, learning, and mobility disability in traveling are examined, deriving potential pathways towards holistic accessible tourism value chains from a multidimensional disability perspective. The findings of this study show a lack of awareness among industry stakeholders regarding the needs of people with disabilities. This is indicated by inaccessibility and weak reliability of information, insufficient sensitisation of staff, or the absence of adequate feedback possibilities. Inherent to this, participants evolve structural and interpersonal constraints, particularly in the pre-travel and on-site phases. While positive travel experiences lead to an increase in the self-confidence of people with disabilities, negative examples trigger intrapersonal constraints, whereby travel plans are discarded due to a self-categorization as individuals with limited opportunities for social inclusion. Theoretical contributions as well as implications for practice are derived to address shortcomings along accessible tourism value chains.

Benchmarking the negatives: Effect of negative data generation on the classification of miRNA-mRNA interactions.

MicroRNAs (miRNAs) are small non-coding RNAs that regulate gene expression post-transcriptionally. In animals, this regulation is achieved via base-pairing with partially complementary sequences on mainly 3' UTR region of messenger RNAs (mRNAs). Computational approaches that predict miRNA target interactions (MTIs) facilitate the process of narrowing down potential targets for experimental validation. The availability of new datasets of high-throughput, direct MTIs has led to the development of machine learning (ML) based methods for MTI prediction. To train an ML algorithm, it is beneficial to provide entries from all class labels (i.e., positive and negative). Currently, no high-throughput assays exist for capturing negative examples. Therefore, current ML approaches must rely on either artificially generated or inferred negative examples deduced from experimentally identified positive miRNA-target datasets. Moreover, the lack of uniform standards for generating such data leads to biased results and hampers comparisons between studies. In this comprehensive study, we collected methods for generating negative data for animal miRNA-target interactions and investigated their impact on the classification of true human MTIs. Our study relies on training ML models on a fixed positive dataset in combination with different negative datasets and evaluating their intra- and cross-dataset performance. As a result, we were able to examine each method independently and evaluate ML models' sensitivity to the methodologies utilized in negative data generation. To achieve a deep understanding of the performance results, we analyzed unique features that distinguish between datasets. In addition, we examined whether one-class classification models that utilize solely positive interactions for training are suitable for the task of MTI classification. We demonstrate the importance of negative data in MTI classification, analyze specific methodological characteristics that differentiate negative datasets, and highlight the challenge of ML models generalizing interaction rules from training to testing sets derived from different approaches. This study provides valuable insights into the computational prediction of MTIs that can be further used to establish standards in the field.

Mohandas Gandhi and the uses of Esperanto

Abstract In the vast literature that deals with Mohandas Gandhi, very little attention has been paid to his opinions on questions of language. This essay studies Gandhi’s use of the international auxiliary language, Esperanto, as a negative example of a way to find a common language. Gandhi was willing and able to collaborate with Esperantists. And yet, Esperanto was repeatedly invoked by Gandhi in order to attack it as an utopian ideal irrelevant in Indian contexts. Gandhi’s arguments against Esperanto established the rejection of an internationalism of which he perceived Esperanto to be a part. His rejection of Esperanto as an artificially constructed language is integral to his failure to acknowledge Hindi’s artificial construction and imposition in South Indian contexts. The Gandhi-led All Indian Education Conference at Wardha (1937) is situated in this essay, therefore, within a wider context of linguistic parochialisms.

Medical Research on Human Subjects – Historical Landmarks

Medical clinical research, otherwise defined as research on human subjects, represents the standardized method unanimously accepted by the medical, scientific community, as well as by civil society, as the only way that allows the approval of the use in medical practice of a new drug, as well as of a new method of investigation or treatment. We will briefly present the history of medical research, with reference to Hippocrates, Galen, Avicenna, Ambroise Paré, James Lind, Edward Jenner, Pasteur, as well as about the negative examples and errors that manifested themselves in the name of medical research, from the perspective of medical ethics.

Positive-Unlabelled learning for identifying new candidate Dietary Restriction-related genes among ageing-related genes

Dietary Restriction (DR) is one of the most popular anti-ageing interventions; recently, Machine Learning (ML) has been explored to identify potential DR-related genes among ageing-related genes, aiming to minimize costly wet lab experiments needed to expand our knowledge on DR. However, to train a model from positive (DR-related) and negative (non-DR-related) examples, the existing ML approach naively labels genes without known DR relation as negative examples, assuming that lack of DR-related annotation for a gene represents evidence of absence of DR-relatedness, rather than absence of evidence. This hinders the reliability of the negative examples (non-DR-related genes) and the method’s ability to identify novel DR-related genes. This work introduces a novel gene prioritization method based on the two-step Positive-Unlabelled (PU) Learning paradigm: using a similarity-based, KNN-inspired approach, our method first selects reliable negative examples among the genes without known DR associations. Then, these reliable negatives and all known positives are used to train a classifier that effectively differentiates DR-related and non-DR-related genes, which is finally employed to generate a more reliable ranking of promising genes for novel DR-relatedness. Our method significantly outperforms (p<0.05) the existing state-of-the-art approach in three predictive accuracy metrics with up to ∼40% lower computational cost in the best case, and we identify 4 new promising DR-related genes (PRKAB1, PRKAB2, IRS2, PRKAG1), all with evidence from the existing literature supporting their potential DR-related role.

Constructing Deterministic Parity Automata from Positive and Negative Examples

We present a polynomial time algorithm that constructs a deterministic parity automaton (DPA) from a given set of positive and negative ultimately periodic example words. We show that this algorithm is complete for the class of $\omega$-regular languages, that is, it can learn a DPA for each regular $\omega$-language. For use in the algorithm, we give a definition of a DPA, that we call the precise DPA of a language, and show that it can be constructed from the syntactic family of right congruences for that language (introduced by Maler and Staiger in 1997). Depending on the structure of the language, the precise DPA can be of exponential size compared to a minimal DPA, but it can also be a minimal DPA. The upper bound that we obtain on the number of examples required for our algorithm to find a DPA for $L$ is therefore exponential in the size of a minimal DPA, in general. However we identify two parameters of regular $\omega$-languages such that fixing these parameters makes the bound polynomial.

BELHD: improving biomedical entity linking with homonym disambiguation.

Biomedical entity linking (BEL) is the task of grounding entity mentions to a given knowledge base (KB). Recently, neural name-based methods, system identifying the most appropriate name in the KB for a given mention using neural network (either via dense retrieval or autoregressive modeling), achieved remarkable results for the task, without requiring manual tuning or definition of domain/entity-specific rules. However, as name-based methods directly return KB names, they cannot cope with homonyms, i.e. different KB entities sharing the exact same name. This significantly affects their performance for KBs where homonyms account for a large amount of entity mentions (e.g. UMLS and NCBI Gene). We present BELHD (Biomedical Entity Linking with Homonym Disambiguation), a new name-based method that copes with this challenge. BELHD builds upon the BioSyn model with two crucial extensions. First, it performs pre-processing of the KB, during which it expands homonyms with a specifically constructed disambiguating string, thus enforcing unique linking decisions. Second, it introduces candidate sharing, a novel strategy that strengthens the overall training signal by including similar mentions from the same document as positive or negative examples, according to their corresponding KB identifier. Experiments with 10 corpora and 5 entity types show that BELHD improves upon current neural state-of-the-art approaches, achieving the best results in 6 out of 10 corpora with an average improvement of 4.55pp recall@1. Furthermore, the KB preprocessing is orthogonal to the prediction model and thus can also improve other neural methods, which we exemplify for GenBioEL, a generative name-based BEL approach. The code to reproduce our experiments can be found at: https://github.com/sg-wbi/belhd.