Considerable research has focused on the challenge of estimating individual treatment effects (ITE) from observational data, primarily due to the presence of treatment assignment bias. To address this, practitioners often adjust for relevant covariates to correct potential biases. However, indiscriminately adjusting for all observed covariates risks including 'bad controls'-variables that can introduce bias when conditioned upon, thereby compromising ITE estimation accuracy. To tackle this issue, we propose Invariant Treatment Effect Estimation via Consistent Constraints and Information Bottleneck (CIBITE). This method mitigates the impact of bad controls by leveraging diverse environments and adjusting for confounding factors in observational data, enabling robust ITE estimation. We introduce a novel invariant causal prediction framework to eliminate bad controls while retaining sufficient information for confounding adjustment. This is achieved by imposing consistent constraints on both the representation and output layers of the neural network. Additionally, the Information Bottleneck is employed to reduce the influence of pseudo-invariant features. To further address confounding, we propose a balanced representation learning framework using adversarial training. Extensive experiments on synthetic, semi-simulated, and real-world datasets demonstrate the effectiveness of our approach. The proposed method significantly outperforms state-of-the-art ITE estimation techniques and existing Invariant Risk Minimization (IRM)-based methods.

Estimating Individual Treatment Effects (ITE) from observational data has been widely applied across various domains. The challenge lies in the fact that observational data only includes outcomes under one treatment, while potential outcomes under different treatments need to be inferred. We propose GLCN, a novel causal effect estimation neural network that fuses global-local modeling. The Global branch, trained on the entire dataset, captures overall causal relationships to ensure estimation stability, while the Local branch adopts a matching approach to borrow outcomes from neighboring instances, thereby capturing local heterogeneity. A gating network dynamically integrates predictions from both branches through an adaptive weighting mechanism to enhance adaptability to local variations. A key difficulty in the Local branch lies in defining a reasonable distance metric for neighboring instances. To address this, our method employs a match-and-reconstruct strategy, where the reconstruction error serves as a supervisory signal to guide learning. To mitigate confounding bias, we introduce a confusion loss based on adversarial training. Extensive experiments on public benchmarks and real-world industrial datasets demonstrate that our method outperforms state-of-the-art approaches.

Causal inference in time series data is a challenging yet crucial task in real-world applications such as healthcare and economics, where time-varying confounders often complicate causal effect estimation. In this work, we propose LLMCE, a novel approach that leverages frozen LLMs for counterfactual estimation in time series settings. First, we adapt LLMs to time series by encoding time-varying covariates, past outcomes, and past and future treatments using a reprogramming mechanism. This aligns time series with textual modalities, enabling LLMs to generalize efficiently for time series analysis. Second, to address time-varying confounders, we introduce an adversarial debiasing strategy. This ensures that learned representations predict outcomes without incorporating incremental information about future treatments beyond what can be inferred from past treatments. This dual objective enhances causal estimate reliability while maintaining predictive accuracy. We evaluate LLMCE on synthetic and real-world datasets, demonstrating superior performance compared to state-of-the-art baselines.

In recent years, semi-supervised graph learning with data augmentation (DA) has been the most commonly used and best-performing method to improve model robustness in sparse scenarios with few labeled samples. However, most existing DA methods are based on the homogeneous graph, but none are specific for the heterogeneous graph. Differing from the homogeneous graph, DA in the heterogeneous graph faces greater challenges: heterogeneity of information requires DA strategies to effectively handle heterogeneous relations, which considers the information contribution of different types of neighbors and edges to the target nodes. Furthermore, over-squashing of information is caused by the negative curvature formed by the non-uniformity distribution and the strong clustering in a complex graph. To address these challenges, this article presents a novel method named HG-MDA (Semi-Supervised Heterogeneous Graph Learning with Multi-Level Data Augmentation). For the problem of heterogeneity of information in DA, node and topology augmentation strategies are proposed for the characteristics of the heterogeneous graph. Additionally, meta-relation-based attention is applied as one of the indexes for selecting augmented nodes and edges. For the problem of over-squashing of information, triangle-based edge adding and removing are designed to alleviate the negative curvature and bring the gain of topology. Finally, the loss function consists of the cross-entropy loss for labeled data and the consistency regularization for unlabeled data. To effectively fuse the prediction results of various DA strategies, sharpening is used. Existing experiments on public datasets (i.e., ACM, DBLP, and OGB) and the industry dataset MB show that HG-MDA outperforms current SOTA models. Additionally, HG-MDA is applied to user identification in internet finance scenarios, helping the business to add 30% key users, and increase loans and balances by 3.6%, 11.1%, and 9.8%.

Traffic characteristics over space and time constitute an important aspect of cellular networks in consideration of resource provision, traffic engineering and system optimization. Despite recent progress in revealing temporal dynamics and spatial inhomogeneity of cellular traffic, limited knowledge about traffic dependence is gained. One of challenges comes from the absence of sustained observations at a network-wide scale. In this paper, we make an analysis on the week-long traffic generated by a large population of users in a city of China, and model traffic dependence along both space and time dimensions. The evaluation results suggest connections between spatio-temporal dependence of cellular traffic and the organization of human lives. Region differences are observed to impact traffic dependence to a great extent. Additionally, interactive knowledge between space and time enhances traffic prediction with a decrease in root-mean-square error of 2.8%~25.2%. We believe that these achievements will benefit multiple research and development areas such as network deploying and simulation researches.