Sifting Truth from Coincidences: A Two-Stage Positive and Unlabeled Learning Model for Coincidental Correctness Detection

Tax evasion detection plays a crucial role in reducing tax revenue loss and many efforts have been made to develop detection models based on machine learning techniques. To train an effective model to detect tax evaders, a large amount of data is required, especially sufficient labeled data. However, the expensive and time-consuming annotation process results in small amount of labeled data being available, which makes the development of detection models difficult. To address this issue, we propose a tax evasion detection method based on positive and unlabeled learning (TEDM-PU), to identify tax evasion by utilizing limited annotated tax evasion taxpayers and a large amount of unlabeled data. The TEDM-PU framework consists of three stages: a preprocessing stage extracting taxpayer features based on random forest, a pseudo labeling stage assigning pseudo labels to unlabeled samples based on PUAdapter, and a model training stage based on LightGBM method. To evaluate the effectiveness of our proposed TEDM-PU, we conduct experimental tests on real-world tax data. The results demonstrate that TEDM-PU method can detect tax evaders with higher accuracy and better interpretability than state-of-the-art methods.

Positive and Unlabeled (PU) learning aims to train a suitable classifier simply based on a set of positive data and unlabeled data. Existing PU methods usually follow a discriminative framework and yield limited classification performance, because the lack of explicit negative labels poses a great barrier in training a discriminative PU model. To address the challenge of limited supervisory information faced by discriminative PU methods, this paper introduces generative operation to PU learning in addition to the conventional discriminative operation, and proposes a novel algorithm dubbed "Discriminative-Generative Positive and Unlabeled Learning" (DGPU). Specifically, our proposed DGPU consists of a data generation stage and a discriminative annotation stage, which can benefit from each other in an iterative manner. In data generation stage, we employ a tailored diffusion model to generate high-quality negative examples and positive examples to efficiently enrich the supervisory information. In discriminative annotation stage, the classifier is further refined on the initial and generated training data. To the best of our knowledge, this study represents the first attempt to integrate diffusion models into PU learning to make generative model and discriminative model benefit from each other in a collaborative way. Thanks to this, our proposed DGPU significantly outperforms existing PU methods across a wide range of synthetic and real-world benchmark datasets. In particular, our DGPU is almost comparable to the fully supervised counterpart, and improves the test accuracy of existing state-of-the-art methods by 3.89% and 2.56% on CIFAR-10 and CelebA datasets, respectively.

To identify the feasibility of deep learning-based diagnostic models for detecting and assessing lower-extremity fatigue fracture severity on plain radiographs. This retrospective study enrolled 1151 X-ray images (tibiofibula/foot: 682/469) of fatigue fractures and 2842 X-ray images (tibiofibula/foot: 2000/842) without abnormal presentations from two clinical centers. After labeling the lesions, images in a center (tibiofibula/foot: 2539/1180) were allocated at 7:1:2 for model construction, and the remaining images from another center (tibiofibula/foot: 143/131) for external validation. A ResNet-50 and a triplet branch network were adopted to construct diagnostic models for detecting and grading. The performances of detection models were evaluated with sensitivity, specificity, and area under the receiver operating characteristic curve (AUC), while grading models were evaluated with accuracy by confusion matrix. Visual estimations by radiologists were performed for comparisons with models. For the detection model on tibiofibula, a sensitivity of 95.4%/85.5%, a specificity of 80.1%/77.0%, and an AUC of 0.965/0.877 were achieved in the internal testing/external validation set. The detection model on foot reached a sensitivity of 96.4%/90.8%, a specificity of 76.0%/66.7%, and an AUC of 0.947/0.911. The detection models showed superior performance to the junior radiologist, comparable to the intermediate or senior radiologist. The overall accuracy of the diagnostic model was 78.5%/62.9% for tibiofibula and 74.7%/61.1% for foot in the internal testing/external validation set. The deep learning-based models could be applied to the radiological diagnosis of plain radiographs for assisting in the detection and grading of fatigue fractures on tibiofibula and foot. • Fatigue fractures on radiographs are relatively difficult to detect, and apt to be misdiagnosed. • Detection and grading models based on deep learning were constructed on a large cohort of radiographs with lower-extremity fatigue fractures. • The detection model with high sensitivity would help to reduce the misdiagnosis of lower-extremity fatigue fractures.

Many online system fraud detection techniques employ deep learning models for identifying malicious user activity sessions. In many real-world scenarios, few labeled malicious and many unlabeled sessions exist. In such scenarios, the fraud detection problem can be effectively addressed through the Positive Unlabeled (PU) learning technique. Despite this fact, possible malicious sessions can be extremely diverse, which makes learning a good decision boundary challenging. In this paper, we present a novel contrastive positive unlabeled learning (ConPU) model for fraud detection and in particular, propose a contrastive loss function for PU learning. ConPU indirectly approximates the cluster center of normal sessions in the representation space by using distributions of unlabeled sessions and malicious sessions, predicts labels of sessions in the unlabeled set by analyzing their proximity between normal and malicious session cluster centers in the representation space, and then incorporates both positive pairs and negative pairs into the contrastive loss function. As a result, ConPU can derive separable representations as well as accurate cluster centers of normal and malicious sessions in the representation space. We theoretically demonstrate the efficacy o f o ur d eveloped l oss f unction i n C onPU. Additionally, we empirically evaluate ConPU on benchmark datasets, in which, ConPU demonstrates substantial performance improvement over state-of-the-art baselines.

Positive and Unlabeled (PU) learning aims to learn a binary classifier from only positive and unlabeled training data. The state-of-the-art methods usually formulate PU learning as a cost-sensitive learning problem, in which every unlabeled example is simultaneously treated as positive and negative with different class weights. However, the ground-truth label of an unlabeled example should be unique, so the existing models inadvertently introduce the label noise which may lead to the biased classifier and deteriorated performance. To solve this problem, this paper proposes a novel algorithm dubbed as "Positive and Unlabeled learning with Label Disambiguation'' (PULD). We first regard all the unlabeled examples in PU learning as ambiguously labeled as positive and negative, and then employ the margin-based label disambiguation strategy, which enlarges the margin of classifier response between the most likely label and the less likely one, to find the unique ground-truth label of each unlabeled example. Theoretically, we derive the generalization error bound of the proposed method by analyzing its Rademacher complexity. Experimentally, we conduct intensive experiments on both benchmark and real-world datasets, and the results clearly demonstrate the superiority of the proposed PULD to the existing PU learning approaches.

Positive and unlabeled (PU) learning aims to train a suitable classifier simply based on a set of positive data and unlabeled data. The state-of-the-art methods usually formulate PU learning as a cost-sensitive learning problem, in which every unlabeled example is treated as negative with modified class weights. However, existing methods fail to generate high-quality data representations, which brings about negative-prediction preference and performance decline. To overcome this problem, this article proposes a novel algorithm dubbed weighted contrastive learning with hard negative mining for positive and unlabeled learning (termed WConPU), which specifically designs a new prototypical contrastive strategy for gaining discriminative representations for PU learning. Specifically, our proposed WConPU consists of a contrastive learning (CL) module and a classifier training module, which can benefit from each other in an iterative manner. Moreover, a novel weighted contrastive objective function equipped with a prototype-based hard negative mining module is proposed to further enhance the representation quality. Theoretically, we show that our WConPU can be justified from the perspective of the expectation-maximization (EM) algorithm. Empirically, we compare our method with state-of-the-art PU algorithms on a wide range of real-world benchmark datasets, and the experimental results firmly demonstrate the advantage of our proposed method over the existing PU learning approaches.

As the first decentralized peer-to-peer (P2P) cryptocurrency system allowing people to trade with pseudonymous addresses, Bitcoin has become increasingly popular in recent years. However, the P2P and pseudonymous nature of Bitcoin make transactions on this platform very difficult to track, thus triggering the emergence of various illegal activities in the Bitcoin ecosystem. Particularly, mixing services in Bitcoin, originally designed to enhance transaction anonymity, have been widely employed for money laundry to complicate trailing illicit fund. In this paper, we focus on the detection of the addresses belonging to mixing services, which is an important task for anti-money laundering in Bitcoin. Specifically, we provide a feature-based network analysis framework to identify statistical properties of mixing services from three levels, namely, network level, account level and transaction level. To better characterize the transaction patterns of different types of addresses, we propose the concept of Attributed Temporal Heterogeneous motifs (ATH motifs). Moreover, to deal with the issue of imperfect labeling, we tackle the mixing detection task as a Positive and Unlabeled learning (PU learning) problem and build a detection model by leveraging the considered features. Experiments on real Bitcoin datasets demonstrate the effectiveness of our detection model and the importance of hybrid motifs including ATH motifs in mixing detection.

Positive and unlabeled (PU) learning targets a binary classifier on labeled positive data and unlabeled data containing data samples of positive and unknown negative classes, whereas multi-class positive and unlabeled (MPU) learning aims to learn a multi-class classifier assuming labeled data from multiple positive classes. In this paper, we propose a two-step approach for MPU learning on high dimensional data. In the first step, negative samples are selected from unlabeled data using an ensemble of k-nearest neighbors-based outlier detection models in a low dimensional space which is embedded by a linear discriminant function. We present an approach for binary prediction which determines whether a data sample is a negative data sample. In the second step, the linear discriminant function is optimized on the labeled positive data and negative samples selected in the first step. It alternates between updating the parameters of the linear discriminant function and selecting reliable negative samples by detecting outliers in a low-dimensional space. Experimental results using high dimensional text data demonstrate the high performance of the proposed MPU learning method.

Discrete State Models of Information Processing

Suicide and self-harm remain major public health concerns in the United States. Early identification is critical for effective intervention, yet underdiagnosis and undercoding are common across mental health conditions, and only positive cases are typically labeled in healthcare data. As a result, reliable negative examples are missing. Positive and Unlabeled (PU) learning is well-suited to such data, enabling estimation of phenotype prevalence and identification of undiagnosed individuals at elevated risk for self-harm as well as other mental illnesses. To identify U.S. Veterans whose self-harm events were not explicitly captured through diagnostic codes in electronic health records (EHRs) and estimate the prevalence of ever self-harm cases among Veterans using a novel PU learning algorithm applicable to undetected mental health diagnoses. We performed a retrospective observational study using Veterans Health Administration EHRs (from October 1, 1999 to August 31, 2019), selecting a random 25% sample of 1,329,120 Veterans out of 5,316,480 (1,193,563 males and 135,557 females) with at least 2 years of observation. The study cohort comprised 24,625 veterans with coded self-harm and 1,304,495 uncoded for self-harm, with the mean age for coded individuals 38.39 (SD 12.17) and uncoded individuals 48.76 (SD 15.04). We applied our PULSNAR (Positive Unlabeled Learning Selected Not At Random) algorithm to estimate the proportion of individuals with uncoded self-harm. The selected covariates included age at enrollment and the presence or absence of recorded medical conditions, procedures, and clinical observations throughout the observation period. Four experts (raters) independently reviewed charts of 97 uncoded Veterans, each selected from 1% intervals of calibrated PULSNAR probabilities from 0.01 to 0.97. Agreement was assessed among raters, PULSNAR classifications, and consensus review decisions. Post hoc calibration was used to refine prevalence estimates. Of the 159,049 covariates in the dataset, the XGBoost model within our PULSNAR framework identified 1,302 (0.82%) as informative for classification. Only 1.85% (24,625/1,329,120) of Veterans had diagnostic codes indicating self-harm events, while PULSNAR estimated an overall prevalence of 10.46% (139,026/1,329,120) by identifying an additional ?=8.77% (114,404/1,304,495) of self-harm cases among the uncoded population. Of the 97 chart-reviewed patients, 39 had documented but uncoded self-harm. PULSNAR estimates were post hoc calibrated such that their sum over the 97 cases equaled 39, which resulted in PULSNAR adjusted coded and imputed estimation of 7.91% (105,133/1,329,120). When applied to the 1.3M Veterans, PULSNAR suggests that coded self-harm represents only 23.4% (95% CI: 17.76% to 31.51%) of all documented (coded + notes) self-harm. Under the Selected Not At Random assumption, PULSNAR provides an innovative and scalable framework for estimating the clinically documented prevalence of mental health conditions and identifying the uncoded individuals with calibrated prediction, without requiring confirmed negative labels. This method offers an alternative to time-consuming chart reviews for detecting likely cases missing structured coding capture. By addressing diagnostic undercoding of mental health conditions in EHRs, this approach has the potential to enhance the estimation of mental health prevalence and support screening, activation of automated clinical decision support, targeted intervention, better resource allocation, and research to improve outcomes in real-world settings.

e18023 Background: Accurate detection of lymph node metastases (LNM) is crucial in the management of head and neck cancers. Artificial intelligence (AI) techniques, including deep learning (DL) and hand-crafted radiomics (HCR) models, have shown potential in improving diagnostic accuracy. This study aims to compare the performance of DL and HCR models for detecting LNM. Methods: Studies employing DL and HCR models for LNM detection were systematically analyzed. Internal validation datasets were utilized due to limited external validation in the literature. Diagnostic performance metrics, including sensitivity, specificity, and area under the curve (AUC), were evaluated using summary receiver operating characteristic (SROC) curves and paired forest plots. Heterogeneity was assessed using the I² statistic, and leave-one-out sensitivity analyses were performed to identify outliers. Data analysis was conducted using R software environment (version 4.2.1, R Foundation for Statistical Computing, Vienna, Austria). Results: The pooled AUCs were 92.1% (95% CI: 84.9–94.7%) for DL models and 90.5% (95% CI: 82.9–91.8%) for HCR models, with no statistically significant difference in diagnostic accuracy (p=0.978). Sensitivities and specificities for DL models were 83.9% (95% CI: 77.6–88.7%) and 87.0% (95% CI: 81.6–91.1%), while HCR models achieved 82.7% (95% CI: 76.9–87.2%) and 86.2% (95% CI: 81.1–90.5%), respectively. Substantial heterogeneity was observed (DL: I² = 72.5–93.1%; HCR: I² = 31.3–65.9%). However, the exclusion of outliers did not significantly alter the results (p=0.981). Conclusions: DL and HCR models exhibit comparable diagnostic accuracy for detecting LNM in head and neck cancers. While both approaches show promise, significant heterogeneity highlighted the need for external validation to confirm their reliability across diverse clinical settings.

Access to information in mass and social media was once thought of as a universal “leveler” for all. Unfortunately learning opportunities have been corrupted by biased, false, deliberately inaccurate, and unverified information. Negative learning (NL) occurs when Internet users are unable to recognize this. Positive learning (PL), the acquisition of new, warranted and morally justified knowledge, is PLATO’s focus. PLATO should capitalize on high fidelity simulations to: (a) study NL and PL, (b) teach PL skills, and (c) assess PL skills. A model of teaching and learning developed for “Pluraliteracies” provides a framework for PLATO to develop, teach and assess PL competencies. Two warnings as PLATO progresses: (1) education alone cannot overcome NL opportunities; technologies (e.g., artificial intelligence) are needed to assist citizens in identifying NL environments; and (2) research must go beyond “person” to “person-in-context” recognizing that environments can foster either NL or PL.

Applications in which researchers aim to extract a single land type from remotely sensed data are quite common in practical scenarios: extract the urban footprint to make connections with socio-economic factors; map the forest extent to subsequently retrieve biophysical variables and detect a particular crop type to successively calibrate and deploy yield prediction models. In this scenario, the (positive) targeted class is well defined, while the negative class is difficult to describe. This one-class classification setting is also referred to as positive unlabelled learning (PUL) in the general field of machine learning. To deal with this challenging setting, when satellite image time series data are available, we propose a new framework named positive and unlabelled learning of satellite image time series (PUL-SITS). PUL-SITS involves two different stages: In the first one, a recurrent neural network autoencoder is trained to reconstruct only positive samples with the aim to higight reliable negative ones. In the second stage, both labelled and unlabelled samples are exploited in a semi-supervised manner to build the final binary classification model. To assess the quality of our approach, experiments were carried out on a real-world benchmark, namely Haute-Garonne, located in the southwest area of France. From this study site, we considered two different scenarios: a first one in which the process has the objective to map Cereals/Oilseeds cover versus the rest of the land cover classes and a second one in which the class of interest is the Forest land cover. The evaluation was carried out by comparing the proposed approach with recent competitors to deal with the considered positive and unlabelled learning scenarios.

Epileptic seizures significantly impact patients' lives due to their unpredictability, making early and accurate detection crucial for effective treatment. Machine learning (ML) models based on electroencephalogram (EEG) signals have been explored for automated seizure detection. This meta-analysis reviews the performance of ML models in seizure detection and analyzes factors such as the model type (deep learning vs. traditional ML), data preprocessing methods, and dataset types. This study aims to provide an evidence-based foundation for the future development of intelligent tools by evaluating the performance of ML models in detecting epileptic seizures through a meta-analysis. A systematic search of multiple databases up to April 2025 identified 60 studies and 93 datasets. The pooled sensitivity, specificity, and area under the curve (AUC) were calculated using Stata 17.0. Subgroup analyses were performed to identify sources of heterogeneity. Publication bias was assessed using Deek's test and funnel plots. The pooled sensitivity, specificity, and AUC were 0.96 (95% CI 0.95-0.97), 0.97 (95% CI 0.96-0.98), and 0.99 (95% CI 0.98-1.00), respectively, indicating a good performance of ML in seizure detection. Subgroup analyses revealed that the model type, data preprocessing methods, and dataset type contributed to heterogeneity. ML shows a strong potential for EEG-based seizure detection. Imaging devices integrating ML may serve as effective tools for early epilepsy diagnosis. However, larger, multicenter clinical studies are needed to validate these algorithms and enhance their interpretability, safety, and applicability in real-world clinical settings.

IntroductionInaccurate identification of implants on X-rays may lead to prolonged surgical duration as well as increased complexity and costs during implant removal. Deep learning models may help to address this problem, although they typically require large datasets to effectively train models in detecting and classifying objects, e.g. implants. This can limit applicability for instances when only smaller datasets are available. Transfer learning can be used to overcome this limitation by leveraging large, publicly available datasets to pre-train detection and classification models. The aim of this study was to assess the effectiveness of deep learning models in implant localisation and classification on a lower limb X-ray dataset.MethodFirstly, detection models were evaluated on their ability to localise four categories of implants, e.g. plates, screws, pins, and intramedullary nails. Detection models (Faster R-CNN, YOLOv5, EfficientDet) were pre-trained on the large, freely available COCO dataset (330000 images). Secondly, classification models (DenseNet121, Inception V3, ResNet18, ResNet101) were evaluated on their ability to classify five types of intramedullary nails. Localisation and classification accuracy were evaluated on a smaller image dataset (204 images).ResultThe YOLOv5s model showed the best capacity to detect and distinguish between different types of implants (accuracy: plate=82.1%, screw=72.3%, intramedullary nail=86.9%, pin=79.9%). Screw implants were the most difficult implant to detect, likely due to overlapping screw implants visible in the image dataset. The DenseNet121 classification model showed the best performance in classifying different types of intramedullary nails (accuracy=73.7%). Therefore, a deep learning model pipeline with the YOLOv5s and DenseNet121 was proposed for the most optimal performance of automating implants localisation and classification for a relatively small dataset.ConclusionThese findings support the potential of deep learning techniques in enhancing implant detection accuracy. With further development, AI-based implant identification may benefit patients, surgeons and hospitals through improved surgical planning and efficient use of theatre time.