Long Tail Research Articles

RepSGG: Novel Representations of Entities and Relationships for Scene Graph Generation.

Scene Graph Generation (SGG) has achieved significant progress recently. However, most previous works rely heavily on fixed-size entity representations based on bounding box proposals, anchors, or learnable queries. As each representation's cardinality has different trade-offs between performance and computation overhead, extracting highly representative features efficiently and dynamically is both challenging and crucial for SGG. In this work, a novel architecture called RepSGG is proposed to address the aforementioned challenges, formulating a subject as queries, an object as keys, and their relationship as the maximum attention weight between pairwise queries and keys. With more fine-grained and flexible representation power for entities and relationships, RepSGG learns to sample semantically discriminative and representative points for relationship inference. Moreover, the long-tailed distribution also poses a significant challenge for generalization of SGG. A run-time performance-guided logit adjustment (PGLA) strategy is proposed such that the relationship logits are modified via affine transformations based on run-time performance during training. This strategy encourages a more balanced performance between dominant and rare classes. Experimental results show that RepSGG achieves the state-of-the-art or comparable performance on the Visual Genome and Open Images V6 datasets with fast inference speed, demonstrating the efficacy and efficiency of the proposed methods.

Mine Your Own Anatomy: Revisiting Medical Image Segmentation With Extremely Limited Labels.

Recent studies on contrastive learning have achieved remarkable performance solely by leveraging few labels in the context of medical image segmentation. Existing methods mainly focus on instance discrimination and invariant mapping (i.e., pulling positive samples closer and negative samples apart in the feature space). However, they face three common pitfalls: (1) tailness: medical image data usually follows an implicit long-tail class distribution. Blindly leveraging all pixels in training hence can lead to the data imbalance issues, and cause deteriorated performance; (2) consistency: it remains unclear whether a segmentation model has learned meaningful and yet consistent anatomical features due to the intra-class variations between different anatomical features; and (3) diversity: the intra-slice correlations within the entire dataset have received significantly less attention. This motivates us to seek a principled approach for strategically making use of the dataset itself to discover similar yet distinct samples from different anatomical views. In this paper, we introduce a novel semi-supervised 2D medical image segmentation framework termed Mine yOur owNAnatomy (MONA), and make three contributions. First, prior work argues that every pixel equally matters to the model training; we observe empirically that this alone is unlikely to define meaningful anatomical features, mainly due to lacking the supervision signal. We show two simple solutions towards learning invariances-through the use of stronger data augmentations and nearest neighbors. Second, we construct a set of objectives that encourage the model to be capable of decomposing medical images into a collection of anatomical features in an unsupervised manner. Lastly, we both empirically and theoretically, demonstrate the efficacy of our MONA on three benchmark datasets with different labeled settings, achieving new state-of-the-art under different labeled semi-supervised settings. MONA makes minimal assumptions on domain expertise, and hence constitutes a practical and versatile solution in medical image analysis. We provide the PyTorch-like pseudo-code in supplementary.

Homogeneous turbophoresis of heavy inertial particles in turbulent flow

Heavy particles suspended in turbulent flow possess inertia and are ejected from violent vortical structures by centrifugal forces. Once piled up along particle paths, this small-scale mechanism leads to an effective large-scale drift. This phenomenon, known as ‘turbophoresis’, causes particles to leave highly turbulent regions and migrate towards calmer regions, explaining why particles transported by non-homogeneous flows tend to concentrate near the minima of turbulent kinetic energy. It is demonstrated here that turbophoretic effects are just as crucial in statistically homogeneous flows. Although the average turbulent activity is uniform, instantaneous spatial fluctuations are responsible for inertial-range inhomogeneities in the particle distribution. Direct numerical simulations are used to probe particle accelerations, specifically how they correlate to local turbulent activity, yielding an effective coarse-grained dynamics that accounts for particle detachment from the fluid and ejection from excited regions through a space- and time-dependent non-Fickian diffusion. This leads to cast fluctuations in particle distributions in terms of a scale-dependent Péclet number ${\textit {Pe}}_\ell$ , which measures the importance of turbulent advection compared with inertial turbophoresis at a given scale $\ell$ . Multifractal statistics of energy dissipation indicate that $ {\textit {Pe}}_\ell \sim \ell ^\delta /\tau _{p}$ with $\delta \approx 0.84$ . Numerical simulations support this behaviour and emphasise the relevance of the turbophoretic Péclet number in characterising how particle distributions, including their radial distribution function, depends on $\ell$ . This approach also explains the presence of voids with inertial-range sizes, and the fact that their volumes have a non-trivial distribution with a power-law tail $p(\mathcal {V}) \propto \mathcal {V}^{-\alpha }$ , with an exponent $\alpha$ that tends to 2 as ${\textit {Pe}}_\ell \to 0$ .

Measuring DNS-over-HTTPS Downgrades: Prevalence, Techniques, and Bypass Strategies

DNS-over-HTTPS (DoH) is a privacy-enhancing protocol that encrypts plaintext query data in DNS resolution. However, DoH often faces accessibility challenges due to phenomena known as DoH downgrades, where DoH queries are reverted to plaintext DNS queries. Unlike downgrades in other security protocols, which are undoubtedly malicious, the act of downgrading DoH queries can be both desirable and undesirable depending on the context; e.g., enterprise networks are officially advised to avoid or downgrade DoH for security reasons. Recent research has drawn attention to the deeper examination of the phenomena of DoH downgrades, focusing on the prevalence, techniques, and potential bypass strategies. However, existing studies on DoH downgrades have several limitations, notably that they severely overestimate the severity of DoH downgrades across the globe as they lack any distinction between desirable and undesirable downgrades of DoH. In this work, we conduct a large-scale measurement study to provide a more accurate depiction of the DoH downgrade landscape. By minimizing the influence of desirable downgrades of DoH in our measurement probes, we show a skewed long-tail distribution of DoH downgrades across the globe. Our stateful probing techniques also reveal hidden DoH filtering mechanisms that were previously undetected. Furthermore, we design near perfect bypass strategies against existing DoH downgrades. Our study expands our limited understanding of DoH downgrades, offering a more accurate, fine-grained, and comprehensive view of the phenomena.

The Boltzmann-Grad Limit of the Lorentz Gas in a Union of Lattices

The Lorentz gas describes an ensemble of noninteracting point particles in an infinite array of spherical scatterers. In the present paper we consider the case when the scatterer configuration P\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${{\\mathcal {P}}}$$\\end{document} is a fixed union of (translated) lattices in Rd\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${\\mathbb {R}}^d$$\\end{document}, and prove that in the limit of low scatterer density, the particle dynamics converges to a random flight process. In the special case when the lattices in P\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${{\\mathcal {P}}}$$\\end{document} are pairwise incommensurable, this settles a conjecture from Marklof and Strömbergsson (J Stat Phys 155:1072–1086, 2014). The proof is carried out by applying a framework developed in recent work by Marklof and Strömbergsson (Mem AMS 294, 2024), and central parts of our proof are the construction of an admissible marking of the point set P\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${{\\mathcal {P}}}$$\\end{document}, and the verification of the uniform spherical equidistribution condition required in Marklof and Strömbergsson (Mem AMS 294, 2024). Regarding the random flight process obtained in the low density limit of the Lorentz gas, we prove that it can be reconstructed from the corresponding limiting flight processes arising from the individual commensurability classes of lattices in P\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${{\\mathcal {P}}}$$\\end{document}. We furthermore prove that the free path lengths of the limit flight process have a distribution with a power law tail, whose exponent depends on the number of commensurability classes in P\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${{\\mathcal {P}}}$$\\end{document}.

A Pareto Tail Plot Without Moment Restrictions

We propose a mean functional that exists for arbitrary probability distributions and characterizes the Pareto distribution within the set of distributions with finite left endpoint. This is in sharp contrast to the mean excess plot, which is meaningless for distributions without an existing mean and has nonstandard behavior when the mean is finite, but the second moment does not exist. The construction of the plot is based on the principle of a single huge jump, which differentiates between distributions with moderately heavy and super heavy tails. We present an estimator of the tail function based on U-statistics and study its large sample properties. Several loss datasets illustrate the use of the new plot.

Multimodal Framework for Long-Tailed Recognition

Long-tailed data distribution (i.e., minority classes occupy most of the data, while most classes have very few samples) is a common problem in image classification. In this paper, we propose a novel multimodal framework for long-tailed data recognition. In the first stage, long-tailed data are used for visual-semantic contrastive learning to obtain good features, while in the second stage, class-balanced data are used for classifier training. The proposed framework leverages the advantages of multimodal models and mitigates the problem of class imbalance in long-tailed data recognition. Experimental results demonstrate that the proposed framework achieves competitive performance on the CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist2018 datasets for image classification.

Relieving popularity bias in recommendation via debiasing representation enhancement

The interaction data used for training recommender systems often exhibit a long-tail distribution. Such highly imbalanced data distribution results in an unfair learning process among items. Contrastive learning alleviates the above issue by data augmentation. However, it lacks consideration of the significant disparity in popularity between items and may even introduce false negatives during the data augmentation, misleading user preference prediction. To address this issue, we combine contrastive learning with a weighted model for negative validation. By penalizing identified false negatives during training, we limit their potential harm within the training process. Meanwhile, to tackle the scarcity of supervision signals for unpopular items, we design Popularity Associated Modeling to mine the correlation among items. Then we guide unpopular items to learn hidden features favored by specific users from their associated popular items, which provides effective supplementary information for their representation modeling. Extensive experiments on three real-world datasets demonstrate that our proposed model outperforms state-of-the-art baselines in recommendation performance, with Recall@20 improvements of 4.2%, 2.4% and 3.6% across the datasets, but also shows significant effectiveness in relieving popularity bias.

Assessing the Recreational Resource Value of National Park Based on Visitor Perception—A Case of Three-River-Source National Park in China

National parks serve as critical practical sites for advancing the concept of “harmonious coexistence between humans and nature” and hold a strategic role in establishing global ecological security barriers. Scholars and decision-makers have expressed significant interest in rigorous assessments of the recreational resource value in national parks. This paper focuses on the Three-River-Source National Park, examining the characteristics and components of its recreational resource value through the lens of human–environment relationship theory. Analysis spans dimensions of geological and geomorphological value, ecological service value, historical and cultural value, and aesthetic landscape value. By extracting visitor comments rich in vocabulary related to their perceptions, this study compares variations in resource values and the “resource value–visitor perception” synergy within Three-River-Source National Park, employing text analysis, semantic network analysis, and coordination analysis methods. The findings reveal that (1) Visitor perceptions of recreational resource value display a clear hierarchy, with aesthetic landscape value (43.6%) ranking highest, followed by geological and geomorphological value (26.7%), historical and cultural value (19.3%), and ecological service value (10.4%), showing significant variation among categories; the vocabulary across these value types exhibits a pronounced long-tail distribution. (2) The recreational resource value in the park forms a distinct core centered on prominent attractions, accompanied by patterns of vocabulary aggregation and dispersion. (3) Visitors demonstrate strong synergy in their perception of geological and aesthetic value, weaker perception regarding historical and cultural value, and a relatively narrow understanding of ecological service value. This research enhances public comprehension of the recreational resource value of national parks and provides a scientific foundation for the conservation and sustainable use of recreational resources in national parks, advancing the realization of their recreational functions.

NIMG-55. DEVELOPING A COMPUTATIONAL PIPELINE TO FORECAST AND SPATIALLY MAP HIGH-GRADE GLIOMA RESPONSE TO RADIOTHERAPY USING BIOPHYSICAL MODELS AND DATA-DRIVEN STATISTICAL ANALYSIS

Abstract High-grade glioma (HGG) presents significant treatment challenges, particularly in identifying less responsive tumor areas to target with adaptive radiotherapy (RT). This study introduces a patient-specific computational pipeline designed to predict and spatially map HGG response to RT, focusing on regions demonstrating resilience to RT. Our computational pipeline is centered around a family of 3D biophysical models describing tumor cell invasion, proliferation, and response to RT which can be personalized for individual patients using longitudinal multiparametric MRI data. Utilizing longitudinal multiparametric MRI data reporting on cell density (diffusion weighted MRI) and the extent of disease (T2-FLAIR and contrast-enhanced T1-weighted MRIs) from 21 patients, ranging from baseline to three weeks into RT, we calibrated a family of biophysical models to forecast tumor cell density up to the final week of RT and one-month follow-up. By employing pareto tail analysis on empirical cumulative distribution function fits of voxel-wise changes in apparent diffusion coefficient maps (ΔADC) between the future images (end of RT, one-month follow-up) and pre-treatment images, we identified areas of increased tumor cellularity on both the ground truth and model predicted cellularity maps. The analysis differentiated tumor voxels into “non-responding” and “responding” categories, based on cutoffs determined through ROC analysis across various conditions—including assessments of both enhancing and non-enhancing tumor regions at different time points. Optimal ΔADC cutoffs for predicted and measured data were established, leading to high area under curve (AUC; ranging from 0.803 to 0.934) values and significant Youden’s index scores (ranging from 0.526 to 0.765) across all conditions. Our findings demonstrate the feasibility of this novel pipeline in delineating biologically relevant RT targets, offering a potential tool for enhancing the efficacy of adaptive RT strategies in treating HGG.

Conditional Variational Autoencoder and generative adversarial network-based approach for long-tailed fault diagnosis for the motor system

This study addresses the challenge of diagnosing motor faults in long-tailed data distributions, characterized by dominant healthy states and rare fault types. We propose the LT-CVAE-GAN model, which integrates a Conditional Variational Autoencoder (CVAE) with a Conditional Generative Adversarial Network (CGAN) to enhance long-tailed fault diagnosis. Initially, we train the CVAE-GAN model using traditional CVAE and CGAN losses such as Kullback–Leibler (KL) divergence loss, reconstruction loss, and adversarial loss. Additionally, we introduce mean feature matching loss and pairwise feature matching loss to mitigate mode collapse and improve model stability, thereby enhancing the generation ability of less frequent fault samples under long-tail conditions. Subsequently, the pre-trained Generator is used to produce infrequent fault mode data to rebalance the dataset. Classifier parameters are fine-tuned in this step to improve fault diagnosis accuracy. Experimental results demonstrate that our LT-CVAE-GAN surpasses state-of-the-art models in diverse long-tailed conditions.

Separating Noisy Samples From Tail Classes for Long-Tailed Image Classification With Label Noise.

Most existing methods that cope with noisy labels usually assume that the classwise data distributions are well balanced. They are difficult to deal with the practical scenarios where training samples have imbalanced distributions, since they are not able to differentiate noisy samples from tail classes' clean samples. This article makes an early effort to tackle the image classification task in which the provided labels are noisy and have a long-tailed distribution. To deal with this problem, we propose a new learning paradigm which can screen out noisy samples by matching between inferences on weak and strong data augmentations. A leave-noise-out regularization (LNOR) is further introduced to eliminate the effect of the recognized noisy samples. Besides, we propose a prediction penalty based on the online classwise confidence levels to avoid the bias toward easy classes which are dominated by head classes. Extensive experiments on five datasets including CIFAR-10, CIFAR-100, MNIST, FashionMNIST, and Clothing1M demonstrate that the proposed method outperforms the existing algorithms for learning with long-tailed distribution and label noise.

An investigation of escape and scaling properties of a billiard system.

We investigate some statistical properties of escaping particles in a billiard system whose boundary is described by two control parameters with a hole on its boundary. Initially, we analyze the survival probability for different hole positions and sizes. We notice that the survival probability follows an exponential decay with a characteristic power-law tail when the hole is positioned partially or entirely over large stability islands in phase space. We find that the survival probability exhibits scaling invariance with respect to the hole size. In contrast, the survival probability for holes placed in predominantly chaotic regions deviates from the exponential decay. We introduce two holes simultaneously and investigate the complexity of the escape basins for different hole sizes and control parameters by means of the basin entropy and the basin boundary entropy. We find a non-trivial relation between these entropies and the system's parameters and show that the basin entropy exhibits scaling invariance for a specific control parameter interval.

Innovation Concentration and Patent Market Structure

Two empirical facts concerning the increase in innovation concentration levels of U.S. industries over the last two decades have been established: innovating firms with larger market shares invest more in R&D and enter into more industries. This paper presents a model of an imperfectly competitive patent market with heterogeneous firms to show that R&D itself increases industry concentration. In the context of an imperfectly competitive patent market structure, firms generate endogenous variable markups. In particular, the price of firms' newly invented knowledge, the profit and survival rate of the innovating firms depend on the market share of the knowledge stock. Firms pay a random fixed cost in each period which together with market share determine the decision on whether to innovate in different sectors. Firms enter different sectors sequentially: they start developing new varieties in a sector with minimal competition, building up private knowledge, and then venturing into other sectors using R&D investment and accumulated knowledge. Finally, we prove that the stationary firm innovation size distribution exhibits a Pareto tail in the steady state.

Cloud droplet size distributions governed by condensation, collision–coalescence, and sedimentation. Part 2: Analytical expressions for size distribution shape and asymptotic behavior for large-droplet tails

Abstract Droplet growth via condensation and collision-coalescence has been investigated in the limit of very weak coalescence growth: a collector-mode approximation emerges for coalescence growth that consists solely of droplets in the tail collecting droplets near the main mode of the cloud droplet size distribution, and that does not deplete the mode. In this second part, the approximation is formalized, leading to a simplified form of the collision-coalescence term in the equation governing the evolution of the droplet size distribution. This form allows for analytical solutions for transient growth and for steady-state distributions representing a balance between condensation, collision-coalescence growth, and removal by sedimentation. The solutions lead to an expression for the drizzle growth time, formally, the finite time in which a droplet approaches infinite size. The drizzle growth time is related to a transition radius at which droplet growth switches from condensation-dominated to coalescence-dominated. The solutions also provide a dimensionless group that governs the steady-state distribution shape, which is referred to as the drizzle number. The solutions are checked by comparing to a Monte Carlo model developed in Part 1, and it is found that the analytical solution is excellent when the drizzle number is much greater than unity, which is the typical situation for the microphysical conditions achievable in a convection-cloud chamber. The steady-state size distribution has the form of a modified gamma distribution, with a transition to power-law tail at sufficiently large radius. The solutions are formally valid for a convection-cloud chamber, but can also be extended to shallow, mixed-layer clouds such as mixing fogs.

MinoritySalMix and adaptive semantic weight compensation for long-tailed classification

In real-world datasets, the widespread presence of a long-tailed distribution often leads models to become overly biased towards majority class samples while ignoring minority class samples. We propose a strategy called MASW (MinoritySalMix and adaptive semantic weight compensation) to improve this problem. First, we propose a data augmentation method called MinoritySalMix (minority-saliency-mixing), which uses significance detection techniques to select significant regions from minority class samples as cropping regions and paste them into the same regions of majority class samples to generate brand new samples, thereby amplifying images containing important regions of minority class samples. Second, in order to make the label value information of the newly generated samples more consistent with the image content of the newly generated samples, we propose an adaptive semantic compensation factor. This factor provides more label value compensation for minority samples based on the different cropping areas, thereby making the new label values closer to the content of the newly generated samples. Improve model performance by generating more accurate new label value information. Finally, considering that some current re-sampling strategies generally lack flexibility in handling class sampling weight allocation and frequently require manual adjustment. We designed an adaptive weight function and incorporated it into the re-sampling strategy to achieve better sampling. The experimental results on three long-tailed datasets show that our method can effectively improve the performance of the model and is superior to most advanced long-tailed methods. Furthermore, we extended MinoritySalMix’s strategy to three balanced datasets for experimentation, and the results indicated that our method surpassed several advanced data augmentation techniques.

Dual-branch network with hypergraph feature augmentation and adaptive logits adjustment for long-tailed visual recognition

The long-tailed distribution presents issues of data scarcity and significant class imbalance, which causes the model to increase the prediction tendency for the head classes and reduce performance on tail classes. The decoupling learning strategy separates the learning process into representation learning and classifier learning to enhance model performance. However, this approach does not delve into the complex semantic relationships within images and overlooks the bias of the model confidence. To address this pervasive issue, we propose a novel algorithm called Hypergraph Feature Augmentation and Adaptive Logits Adjustment for Long-tailed Visual Recognition (HALR) based on a decoupling learning framework. For the representation learning task, we extract hypergraph features from the mixed input samples to capture the global spatial contextual semantic information of the images. For the classifier learning task, we propose an adaptive logits adjustment function that automatically corrects prediction score biases, thereby yielding robust decision boundaries. Extensive experiments on widely-used benchmark datasets, including CIFAR10/100-LT, ImageNet-LT, and iNaturalist 2018, validate the efficacy of HALR. For example, the test accuracy of HALR on CIFAR10-LT with an imbalance factor of 100 is 88.25 %, which is a 17.85 % improvement over the baseline model accuracy and 0.68 % higher than the state-of-the-art model. The results demonstrate that HALR effectively optimizes representation and classifier learning tasks for long-tailed learning, mitigating the negative impact of class imbalance and data scarcity.

Strong Linkage Between Observed Daily Precipitation Extremes and Anthropogenic Emissions Across the Contiguous United States

AbstractThe results of probabilistic event attribution studies depend on the choice of the extreme value statistics used in the analysis, particularly with the arbitrariness in the selection of appropriate thresholds to define extremes. We bypass this issue by using the Extended Generalized Pareto Distribution (ExtGPD), which jointly models low precipitation with a generalized Pareto distribution and extremes with a different Pareto tail, to conduct daily precipitation attribution across the contiguous United States (CONUS). We apply the ExtGPD to 12 general circulation models from the Coupled Model Intercomparison Project Phase 6 and compare counterfactual scenarios with and without anthropogenic emissions. Observed precipitation by the Climate Prediction Center is used for evaluating the GCMs. We find that greenhouse gases rather than natural variability can explain the observed magnitude of extreme daily precipitation, especially in the temperate regions. Our results highlight an unambiguous linkage of anthropogenic emissions to daily precipitation extremes across CONUS.

Learning personalized drug features and differentiated drug-pair interaction information for drug–drug interaction prediction

Multi-drug combination therapies are increasingly used for complex diseases but carry risks of harmful drug interactions. Effective drug–drug interaction prediction (DDIP) is essential for assessing risks among numerous drug pairs. Most DDIP methods involve two main steps: drug representation and drug pair interaction extraction, respectively challenged by the loss of personalized drug information and the need for differentiated interaction data, and are rarely studied. Specifically, personalized drug information refers to the distinct features of each drug. These properties can be easily confused by neighboring information during graph propagation. This issue is especially prominent in drug interaction graphs with long-tail distributions, which poses challenges for personalized drug information learning. Furthermore, it is crucial to learn interactions with differentiation in order to identify diverse drug relationships. Some methods simply concatenate drug features, often ignoring the differences of different drug relationships, while other methods based on substructures rely on professional pharmacological knowledge and are computationally complex. To address these issues, we propose a novel method, learning personalized Drug Features and differentiated Drug-Pair interaction information for drug–drug interaction prediction (DFPDDI). This approach employs a contrastive learning network with edge-aware augmentations and mutual information estimators to capture personalized drug features across various graph distributions. Furthermore, it applies a mutual information constraint to drug-pair representations, enhancing the accuracy of interaction predictions by better distinguishing between different types of drug relationships. The results evaluated on three public datasets demonstrate competitive performance compared to baselines. It also shows potential for accurate predictions, particularly in imbalanced-distribution graphs.

ChatDiff: A ChatGPT-based diffusion model for long-tailed classification

Long-tailed data distributions have been a major challenge for the practical application of deep learning. Information augmentation intends to expand the long-tailed data into uniform distribution, which provides a feasible way to mitigate the data starvation of underrepresented classes. However, most existing augmentation methods face two significant challenges: (1) limited diversity in generated samples, and (2) the adverse effect of generated negative samples on downstream classification performance. In this paper, we propose a novel information augmentation method, named ChatDiff, to provide diverse positive samples for underrepresented classes, and eliminate generated negative samples. Specifically, we start with a prompt template to extract textual prior knowledge from the ChatGPT-3.5 model, enhancing the feature space for underrepresented classes. Then using this prior knowledge, a conditional diffusion model generates semantic-rich image samples for tail classes. Moreover, the proposed ChatDiff leverages a CLIP-based discriminator to screen and remove generated negative samples. This process avoids neural network learning the invalid or erroneous features, and further, improves long-tailed classification performance. Comprehensive experiments conducted on long-tailed benchmarks such as CIFAR10-LT, CIFAR100-LT, ImageNet-LT, and iNaturalist 2018, validate the effectiveness of our ChatDiff method.