Temporal graph learning (TGL) has emerged as an important machine learning problem and is widely used in a number of real-world applications, such as traffic prediction [Yuan and Li, 2021, Zhang et al., 2021], knowledge graphs [Cai et al., 2022, Leblay and Chekol, 2018], and recommender systems [Kumar et al., 2019, Rossi et al., 2020, Xu et al., 2020a]. A typical downstream task of temporal graph learning is link prediction, which focuses on predicting future interactions among nodes. For example in an online video recommender system, the user-video clicks can be modeled as a temporal graph whose nodes represent users and videos, and links are associated with timestamps indicating when users click videos. Link prediction between nodes can be used to predict if and when a user is interested in a video. Therefore, designing graph learning models that can capture node evolutionary patterns and accurately predict future links is important. TGL is generally more challenging than static graph learning, thereby requiring more sophisticated algorithms to model the temporal evolutionary patterns [Huang et al., 2023]. In recent years, many TGL algorithms [Kumar et al., 2019, Xu et al., 2020a, Rossi et al., 2020, Sankar et al., 2020, Wang et al., 2021e] have been proposed that leverage memory blocks, self-attention, time-encoding function, recurrent neural networks (RNNs), temporal walks, and message passing to better capture the meaningful structural or temporal patterns. For instance, JODIE [Kumar et al., 2019] maintains a memory block for each node and utilizes an RNN to update the memory blocks upon the occurance of each interaction; TGAT [Xu et al., 2020a] utilizes self-attention message passing to aggregate neighbor information on the temporal graph; TGN [Rossi et al., 2020] combines memory blocks with message passing to allow each node in the temporal graph to have a receptive field that is not limited by the number of message-passing layers; DySAT [Sankar et al., 2020] uses self-attention to capture structural information and uses RNN to capture temporal dependencies; CAW [Wang et al., 2021e] captures temporal dependencies between nodes by performing multiple temporal walks from the root

Speech quality estimation has recently undergone a paradigm shift from human-hearing expert designs to machine-learning models. However, current models rely mainly on supervised learning, which is time-consuming and expensive for label collection. To solve this problem, we propose VQScore, a self-supervised metric for evaluating speech based on the quantization error of a vector-quantized-variational autoencoder (VQ-VAE). The training of VQ-VAE relies on clean speech; hence, large quantization errors can be expected when the speech is distorted. To further improve correlation with real quality scores, domain knowledge of speech processing is incorporated into the model design. We found that the vector quantization mechanism could also be used for self-supervised speech enhancement (SE) model training. To improve the robustness of the encoder for SE, a novel self-distillation mechanism combined with adversarial training is introduced. In summary, the proposed speech quality estimation method and enhancement models require only clean speech for training without any label requirements. Experimental results show that the proposed VQScore and enhancement model are competitive with supervised baselines. The code will be released after publication.

Jointly learning multiple tasks with a unified model can improve accuracy and data efficiency, but it faces the challenge of task interference, where optimizing one task objective may inadvertently compromise the performance of another. A solution to mitigate this issue is to allocate task-specific parameters, free from interference, on top of shared features. However, manually designing such architectures is cumbersome, as practitioners need to balance between the overall performance across all tasks and the higher computational cost induced by the newly added parameters. In this work, we propose \textit{InterroGate}, a novel multi-task learning (MTL) architecture designed to mitigate task interference while optimizing inference computational efficiency. We employ a learnable gating mechanism to automatically balance the shared and task-specific representations while preserving the performance of all tasks. Crucially, the patterns of parameter sharing and specialization dynamically learned during training, become fixed at inference, resulting in a static, optimized MTL architecture. Through extensive empirical evaluations, we demonstrate SoTA results on three MTL benchmarks using convolutional as well as transformer-based backbones on CelebA, NYUD-v2, and PASCAL-Context.

One of the key challenges in fairness for machine learning is to train models that minimize the disparity across various sensitive groups such as race or gender [Caton and Haas, 2020, Ustun et al., 2019, Celis et al., 2019]. This often comes at the cost of reduced model accuracy, a phenomenon termed accuracyfairness trade-off in literature [Valdivia et al., 2021, Martinez et al., 2020]. This trade-off can differ significantly across datasets in practice, depending on factors such as dataset biases, imbalances etc. [Agarwal et al., 2018, Bendekgey and Sudderth, 2021, Celis et al., 2021]. To demonstrate how these trade-offs are inherently dataset-dependent, let's consider a simple example involving two distinct crime datasets. Dataset A has records from a community where crime rates are uniformly distributed across all racial groups, whereas Dataset B comes from a community where historical factors have resulted in a disproportionate crime rate among a specific racial group. Intuitively, training models which are racially agnostic is more challenging for Dataset B, due to the unequal distribution of crime rates across racial groups, and will result in a greater loss in model accuracy as compared to Dataset A. This example underscores that setting a uniform fairness requirement across diverse datasets (such as requiring the fairness violation metric to be below 10% for both datasets), while also adhering to essential accuracy benchmarks is impractical.

Data assimilation (DA) methods use priors arising from differential equations to robustly interpolate and extrapolate data. Popular techniques such as ensemble methods that handle high-dimensional, nonlinear PDE priors focus mostly on state estimation, however can have difficulty learning the parameters accurately. On the other hand, machine learning based approaches can naturally learn the state and parameters, but their applicability can be limited, or produce uncertainties that are hard to interpret. Inspired by the Integrated Nested Laplace Approximation (INLA) method in spatial statistics, we propose an alternative approach to DA based on iteratively linearising the dynamical model. This produces a Gaussian Markov random field at each iteration, enabling one to use INLA to infer the state and parameters. Our approach can be used for arbitrary nonlinear systems, while retaining interpretability, and is furthermore demonstrated to outperform existing methods on the DA task. By providing a more nuanced approach to handling nonlinear PDE priors, our methodology offers improved accuracy and robustness in predictions, especially where data sparsity is prevalent.

We analyze the statistical properties of generalized cross-validation (GCV) and leave-one-out cross-validation (LOOCV) applied to early-stopped gradient descent (GD) in high-dimensional least squares regression. We prove that GCV is generically inconsistent as an estimator of the prediction risk of early-stopped GD, even for a well-specified linear model with isotropic features. In contrast, we show that LOOCV converges uniformly along the GD trajectory to the prediction risk. Our theory requires only mild assumptions on the data distribution and does not require the underlying regression function to be linear. Furthermore, by leveraging the individual LOOCV errors, we construct consistent estimators for the entire prediction error distribution along the GD trajectory and consistent estimators for a wide class of error functionals. This in particular enables the construction of pathwise prediction intervals based on GD iterates that have asymptotically correct nominal coverage conditional on the training data.

In various real-world scenarios like recommender systems and political surveys, pairwise rankings are commonly collected and utilized for rank aggregation to obtain an overall ranking of items. However, preference rankings can reveal individuals' personal preferences, underscoring the need to protect them before releasing for downstream analysis. In this paper, we address the challenge of preserving privacy while ensuring the utility of rank aggregation based on pairwise rankings generated from the Bradley-Terry-Luce (BTL) model. Using the randomized response mechanism to perturb raw pairwise rankings is a common privacy protection strategy used in practice, but a critical challenge arises because the privatized rankings no longer adhere to the BTL model, resulting in significant bias in downstream rank aggregation tasks. Motivated from this, we propose a debiased randomized response mechanism to protect the raw pairwise rankings, ensuring consistent estimation of true preferences and rankings in downstream rank aggregation. Theoretically, we offer insights into the relationship between overall privacy guarantees and estimation errors from private ranking data, and establish minimax rates for estimation errors. This enables the determination of optimal privacy guarantees that balance consistency in rank aggregation with robust privacy protection. We also investigate convergence rates of expected ranking errors for partial and full ranking recovery, quantifying how privacy protection influences the specification of top-$K$ item sets and complete rankings. Our findings are validated through extensive simulations and a real application.

Noise-contrastive estimation (NCE) is a popular method for estimating unnormalised probabilistic models, such as energy-based models, which are effective for modelling complex data distributions. Unlike classical maximum likelihood (ML) estimation that relies on importance sampling (resulting in ML-IS) or MCMC (resulting in contrastive divergence, CD), NCE uses a proxy criterion to avoid the need for evaluating an often intractable normalisation constant. Despite apparent conceptual differences, we show that two NCE criteria, ranking NCE (RNCE) and conditional NCE (CNCE), can be viewed as ML estimation methods. Specifically, RNCE is equivalent to ML estimation combined with conditional importance sampling, and both RNCE and CNCE are special cases of CD. These findings bridge the gap between the two method classes and allow us to apply techniques from the ML-IS and CD literature to NCE, offering several advantageous extensions.

Deep networks are increasingly applied to a wide variety of data, including data with high-dimensional predictors. In such analysis, variable selection can be needed along with estimation/model building. Many of the existing deep network studies that incorporate variable selection have been limited to methodological and numerical developments. In this study, we consider modeling/estimation using the conditional Wasserstein Generative Adversarial networks. Group Lasso penalization is applied for variable selection, which may improve model estimation/prediction, interpretability, stability, etc. Significantly advancing from the existing literature, the analysis of censored survival data is also considered. We establish the convergence rate for variable selection while considering the approximation error, and obtain a more efficient distribution estimation. Simulations and the analysis of real experimental data demonstrate satisfactory practical utility of the proposed analysis.

Given i.i.d. observations uniformly distributed on a closed manifold $\mathcal{M}\subseteq \mathbb{R}^p$, we study the spectral properties of the associated empirical graph Laplacian based on a Gaussian kernel. Our main results are non-asymptotic error bounds, showing that the eigenvalues and eigenspaces of the empirical graph Laplacian are close to the eigenvalues and eigenspaces of the Laplace-Beltrami operator of $\mathcal{M}$. In our analysis, we connect the empirical graph Laplacian to kernel principal component analysis, and consider the heat kernel of $\mathcal{M}$ as reproducing kernel feature map. This leads to novel points of view and allows to leverage results for empirical covariance operators in infinite dimensions.