Collider bias occurs when conditioning on a common effect (collider) of two variables $X, Y$. In this manuscript, we quantify the collider bias in the estimated association between exposure $X$ and outcome $Y$ induced by selecting on one value of a binary collider $S$ of the exposure and the outcome. In the case of logistic regression, it is known that the magnitude of the collider bias in the exposure-outcome regression coefficient is proportional to the strength of interaction $\delta_3$ between $X$ and $Y$ in a log-additive model for the collider: $\mathbb{P} (S = 1 | X, Y) = \exp \left\{ \delta_0 + \delta_1 X + \delta_2 Y + \delta_3 X Y \right\}$. We show that this result also holds under a linear or Poisson regression model for the exposure-outcome association. We then illustrate by simulation that even if a log-additive model with interactions is not the true model for the collider, the interaction term in such a model is still informative about the magnitude of collider bias. Finally, we discuss the implications of these findings for methods that attempt to adjust for collider bias, such as inverse probability weighting which is often implemented without including interactions between variables in the weighting model.
相關內容
We investigate the internal behavior of Transformer-based Large Language Models (LLMs) when they generate factually incorrect text. We propose modeling factual queries as Constraint Satisfaction Problems and use this framework to investigate how the model interacts internally with factual constraints. Specifically, we discover a strong positive relation between the model's attention to constraint tokens and the factual accuracy of its responses. In our curated suite of 11 datasets with over 40,000 prompts, we study the task of predicting factual errors with the Llama-2 family across all scales (7B, 13B, 70B). We propose SAT Probe, a method probing self-attention patterns, that can predict constraint satisfaction and factual errors, and allows early error identification. The approach and findings demonstrate how using the mechanistic understanding of factuality in LLMs can enhance reliability.
Multi-behavioral sequential recommendation has recently attracted increasing attention. However, existing methods suffer from two major limitations. Firstly, user preferences and intents can be described in fine-grained detail from multiple perspectives; yet, these methods fail to capture their multi-aspect nature. Secondly, user behaviors may contain noises, and most existing methods could not effectively deal with noises. In this paper, we present an attentive recurrent model with multiple projections to capture Multi-Aspect preferences and INTents (MAINT in short). To extract multi-aspect preferences from target behaviors, we propose a multi-aspect projection mechanism for generating multiple preference representations from multiple aspects. To extract multi-aspect intents from multi-typed behaviors, we propose a behavior-enhanced LSTM and a multi-aspect refinement attention mechanism. The attention mechanism can filter out noises and generate multiple intent representations from different aspects. To adaptively fuse user preferences and intents, we propose a multi-aspect gated fusion mechanism. Extensive experiments conducted on real-world datasets have demonstrated the effectiveness of our model.
TCP and QUIC can both leverage ECN to avoid congestion loss and its retransmission overhead. However, both protocols require support of their remote endpoints and it took two decades since the initial standardization of ECN for TCP to reach 80% ECN support and more in the wild. In contrast, the QUIC standard mandates ECN support, but there are notable ambiguities that make it unclear if and how ECN can actually be used with QUIC on the Internet. Hence, in this paper, we analyze ECN support with QUIC in the wild: We conduct repeated measurements on more than 180M domains to identify HTTP/3 websites and analyze the underlying QUIC connections w.r.t. ECN support. We only find 20% of QUIC hosts, providing 6% of HTTP/3 websites, to mirror client ECN codepoints. Yet, mirroring ECN is only half of what is required for ECN with QUIC, as QUIC validates mirrored ECN codepoints to detect network impairments: We observe that less than 2% of QUIC hosts, providing less than 0.3% of HTTP/3 websites, pass this validation. We identify possible root causes in content providers not supporting ECN via QUIC and network impairments hindering ECN. We thus also characterize ECN with QUIC distributedly to traverse other paths and discuss our results w.r.t. QUIC and ECN innovations beyond QUIC.
Cognitive biases exert a significant influence on human thinking and decision-making. In order to identify how they influence the occurrence of architectural technical debt, a series of semi-structured interviews with software architects was performed. The results show which classes of architectural technical debt originate from cognitive biases, and reveal the antecedents of technical debt items (classes) through biases. This way, we analysed how and when cognitive biases lead to the creation of technical debt. We also identified a set of debiasing techniques that can be used in order to prevent the negative influence of cognitive biases. The observations of the role of organisational culture in the avoidance of inadvertent technical debt throw a new light on that issue.
We analyze the running time of Hartigan's method, an old algorithm for the $k$-means clustering problem. First, we construct an instance on the line on which the method can take $2^{\Omega(n)}$ steps to converge, demonstrating that Hartigan's method has exponential worst-case running time even when $k$-means is easy to solve. As this is in contrast to the empirical performance of the algorithm, we also analyze the running time in the framework of smoothed analysis. In particular, given an instance of $n$ points in $d$ dimensions, we prove that the expected number of iterations needed for Hartigan's method to terminate is bounded by $k^{12kd}\cdot \poly(n, k, d, 1/\sigma)$ when the points in the instance are perturbed by independent $d$-dimensional Gaussian random variables of mean $0$ and standard deviation $\sigma$.
Crafting an effective Automatic Speech Recognition (ASR) solution for dialects demands innovative approaches that not only address the data scarcity issue but also navigate the intricacies of linguistic diversity. In this paper, we address the aforementioned ASR challenge, focusing on the Tunisian dialect. First, textual and audio data is collected and in some cases annotated. Second, we explore self-supervision, semi-supervision and few-shot code-switching approaches to push the state-of-the-art on different Tunisian test sets; covering different acoustic, linguistic and prosodic conditions. Finally, and given the absence of conventional spelling, we produce a human evaluation of our transcripts to avoid the noise coming from spelling inadequacies in our testing references. Our models, allowing to transcribe audio samples in a linguistic mix involving Tunisian Arabic, English and French, and all the data used during training and testing are released for public use and further improvements.
We consider signal source localization from range-difference measurements. First, we give some readily-checked conditions on measurement noises and sensor deployment to guarantee the asymptotic identifiability of the model and show the consistency and asymptotic normality of the maximum likelihood (ML) estimator. Then, we devise an estimator that owns the same asymptotic property as the ML one. Specifically, we prove that the negative log-likelihood function converges to a function, which has a unique minimum and positive definite Hessian at the true source's position. Hence, it is promising to execute local iterations, e.g., the Gauss-Newton (GN) algorithm, following a consistent estimate. The main issue involved is obtaining a preliminary consistent estimate. To this aim, we construct a linear least-squares problem via algebraic operation and constraint relaxation and obtain a closed-form solution. We then focus on deriving and eliminating the bias of the linear least-squares estimator, which yields an asymptotically unbiased (thus consistent) estimate. Noting that the bias is a function of the noise variance, we further devise a consistent noise variance estimator that involves $3$-order polynomial rooting. Based on the preliminary consistent location estimate, a one-step GN iteration suffices to achieve the same asymptotic property as the ML estimator. Simulation results demonstrate the superiority of our proposed algorithm in the large sample case.
In this paper, we investigate how the initial models and the final models for the polynomial functors can be uniformly specified in matching logic.
Graph Neural Networks (GNNs) have achieved promising performance in a variety of graph-focused tasks. Despite their success, existing GNNs suffer from two significant limitations: a lack of interpretability in results due to their black-box nature, and an inability to learn representations of varying orders. To tackle these issues, we propose a novel Model-agnostic Graph Neural Network (MaGNet) framework, which is able to sequentially integrate information of various orders, extract knowledge from high-order neighbors, and provide meaningful and interpretable results by identifying influential compact graph structures. In particular, MaGNet consists of two components: an estimation model for the latent representation of complex relationships under graph topology, and an interpretation model that identifies influential nodes, edges, and important node features. Theoretically, we establish the generalization error bound for MaGNet via empirical Rademacher complexity, and showcase its power to represent layer-wise neighborhood mixing. We conduct comprehensive numerical studies using simulated data to demonstrate the superior performance of MaGNet in comparison to several state-of-the-art alternatives. Furthermore, we apply MaGNet to a real-world case study aimed at extracting task-critical information from brain activity data, thereby highlighting its effectiveness in advancing scientific research.
Graph Neural Networks (GNNs) are state-of-the-art models for performing prediction tasks on graphs. While existing GNNs have shown great performance on various tasks related to graphs, little attention has been paid to the scenario where out-of-distribution (OOD) nodes exist in the graph during training and inference. Borrowing the concept from CV and NLP, we define OOD nodes as nodes with labels unseen from the training set. Since a lot of networks are automatically constructed by programs, real-world graphs are often noisy and may contain nodes from unknown distributions. In this work, we define the problem of graph learning with out-of-distribution nodes. Specifically, we aim to accomplish two tasks: 1) detect nodes which do not belong to the known distribution and 2) classify the remaining nodes to be one of the known classes. We demonstrate that the connection patterns in graphs are informative for outlier detection, and propose Out-of-Distribution Graph Attention Network (OODGAT), a novel GNN model which explicitly models the interaction between different kinds of nodes and separate inliers from outliers during feature propagation. Extensive experiments show that OODGAT outperforms existing outlier detection methods by a large margin, while being better or comparable in terms of in-distribution classification.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191