Estimating individual treatment effects (ITE) from networked observational data is a fundamental but challenging task. Network connections not only introduce complex confounding bias but also generate spillover effects, making causal inference in such settings particularly difficult. To tackle these challenges, existing methods often attempt to learn confounders from all observed variables to predict potential outcomes. However, this approach overlooks heterogeneous latent factors that differentially influence treatment assignment and outcomes, which may prevent accurate identification of the confounders and reduce the effectiveness of bias correction. To overcome these limitations, we propose the Network Disentangled Identifiable Variational Autoencoder (NDiVAE), a structured framework that learns three distinct latent factors for each unit, including instrumental, confounding, and adjustment factors. NDiVAE further incorporates graph-based neighborhood information to learn aggregated factor representations, applies sample reweighting to mitigate confounding bias, and enforces causal regularization to ensure precise disentanglement of latent factors. We establish theoretical guarantees through the identifiability of these latent factors and derive a generalization error bound for ITE estimation. Extensive experiments on synthetic and semi-synthetic datasets demonstrate that NDiVAE consistently outperforms state-of-the-art methods in estimating treatment effects.

Disentangled representation learning aims to learn low-dimensional representations where each dimension corresponds to an underlying generative factor. While the Variational Auto-Encoder (VAE) is widely used for this purpose, most existing methods assume independence among factors, a simplification that does not hold in many real-world scenarios where factors are often interdependent and exhibit causal relationships. To overcome this limitation, we propose the Disentangled Causal Variational Auto-Encoder (DCVAE), a novel supervised VAE framework that integrates causal flows into the representation learning process, enabling the learning of more meaningful and interpretable disentangled representations. We evaluate DCVAE on both synthetic and real-world datasets, demonstrating its superior ability in causal disentanglement and intervention experiments. Furthermore, DCVAE outperforms state-of-the-art methods in various downstream tasks, highlighting its potential for learning true causal structures among factors.

Click-through rate prediction aims to predict the ratio of clicks to impressions of a specific link, which is challenging due to (1) extremely high-dimensional categorical features; (2) both important original features and their interactions; and (3) reliance on different features and interactions in different time periods. To overcome these difficulties, we propose a new feature interaction detection method based on the idea of frequent itemset mining, named Online Random Intersection Chains (ORIC), which detects informative feature interactions with high interpretability. ORIC can be updated by controlling the importance of the historical and latest data with a tuning parameter, which saves computational burden and makes full use of historical information. Further, Streaming Integrated Model (SIM) is developed to feed the time-varying feature interactions into CTR prediction models. Empirical results on three benchmark datasets show that SIM achieves better performance than many CTR prediction models, as well as the efficiency, consistency, and interpretability of ORIC.

Predicting the probability that a user clicks a specific item is fundamental in online advertising and recommendation. Further, it is crucial to use the latest and historical data appropriately in online scenarios to train CTR models. Online Random Interaction Chains (ORIC) was proposed to detect informative and interpretable feature interactions without retraining on historical data in online scenario, and the Streaming Integrated Model (SIM) framework was designed to integrate these time-varying feature interactions into CTR prediction models. Unfortunately, ORIC exhibits latency when provides the feature interactions used to evaluate SIM, and ORIC is not applicable for numerical features. For these reasons, we propose ORIC-V2 that uses time series models to predict the confidence of candidate evaluating feature interactions and selects reasonable feature interactions, and combines numerical features with ORIC-V2 through a discretization model to obtain DORIC-V2. Feeding the feature interactions found by ORIC-V2 and DORIC-V2 into SIM obtains significant experimental results on three datasets, demonstrating the effectiveness and interpretability of ORIC-V2 and DORIC-V2.

The lack of interpretability often makes black-box models challenging to be applied in many practical domains. For this reason, the current work, from the black-box model input port, proposes to incorporate data-mined knowledge into the black-box soft-margin SVM model to enhance accuracy and interpretability. The concept and incorporation mechanism of data-mined knowledge are successively developed, based on which a partially interpretable soft-margin SVM ( pTsm -SVM) optimization model is designed and then solved through reformulating the optimization problem as standard quadratic programming. An algorithm for mining linear positive (negative) class knowledge from general data sets is also proposed, which generates a linear two-dimensional discriminative rule with specificity (sensitivity) equal to 1 and the highest possible sensitivity (specificity) among all two-dimensional feature spaces. The knowledge-integrated pTsm -SVM works by achieving a good trade-off among the “large margin”, “high specificity”, and “high sensitivity”. Our experimental results on eight UCI datasets demonstrate the superiority of the proposed pTsm -SVM over the standard soft-margin SVM both in terms of accuracy and interpretability.

With the growing size of datasets, feature selection becomes increasingly important. Taking interactions of original features into consideration will lead to extremely high dimension, especially when the features are discrete and one-hot encoding is applied. This makes it more worthwhile mining useful features as well as their interactions. Association rule mining aims to extract interesting correlations between items, but it is difficult to use rules as a qualified classifier themselves. Drawing inspiration from association rule mining, we come up with a method that uses association rules to select features and their interactions, then modify the algorithm for several practical concerns. We analyze the computation complexity of the proposed algorithm to show its efficiency. And the results of a series of experiments verify the effectiveness of the algorithm.