There has been a lot of research on flight delays. But it is more useful and difficult to estimate the departure delay time especially three hours before the scheduled time of departure, from which passengers can reasonably plan their travel time and the airline and airport staff can schedule flights more reasonably. In this paper, we develop a Spatio-temporal Graph Dual-Attention Neural Network (SGDAN) to learn the departure delay time for each flight with real-time conditions at three hours before the scheduled time of departure. Specifically, it first models the air traffic network as graph sequences, what is, using a heterogeneous graph to model a flight and its adjacent flights with the same departure or arrival airport in a special time interval, and using a sequence to model the flight and its previous flights that share the same aircraft. The main contributions of this paper are using heterogeneous graph-level attention to learn the influence between the flight and its adjacent flight together with sequence-level attention to learn the influence between the flight and its previous flight in the flight sequence. With aggregating features from the learned influence from both graph-level and sequence-level attention, SGDAN can generate node embedding to estimate the departure delay time. Experiments on a real-world large-scale data set show that SGDAN produces better results than state-of-the-art models in the accurate flight delay time estimation task.

Counterfactual reasoning explores what could have happened if the circumstances were different from what actually occurred. As a crucial subtask, counterfactual story generation integrates counterfactual reasoning into the generative narrative chain, which requires the model to preserve minimal edits and ensure narrative consistency. Previous work prioritizes conflict detection as a first step, and then replaces conflicting content with appropriate words. However, these methods mainly face two challenging issues: (a) the causal relationship between story event sequences is not fully utilized in the conflict detection stage, leading to inaccurate conflict detection, and (b) the absence of proper planning in the content rewriting stage results in a lack of narrative consistency in the generated story ending. In this paper, we propose a novel counterfactual generation framework called CLICK based on causal inference in event sequences and commonsense knowledge incorporation. To address the first issue, we utilize the correlation between adjacent events in the story ending to iteratively calculate the contents from the original ending affected by the condition. The content with the original condition is then effectively prevented from carrying over into the new story ending, thereby avoiding causal conflict with the counterfactual conditions. Considering the second issue, we incorporate structural commonsense knowledge about counterfactual conditions, equipping the framework with comprehensive background information on the potential occurrence of counterfactual conditional events. Through leveraging a rich hierarchical data structure, CLICK gains the ability to establish a more coherent and plausible narrative trajectory for subsequent storytelling. Experimental results show that our model outperforms previous unsupervised state-of-the-art methods and achieves gains of 2.65 in BLEU, 4.42 in ENTScore, and 3.84 in HMean on the TIMETRAVEL dataset.

Real-world heterogeneous information networks (HINs) are modeled as heterogeneous graphs, in which features and structures are often incomplete. Existing models employ manual imputation or dynamic adjustment to populate the incomplete data. However, there are some limitations in incomplete heterogeneous graph representation learning: 1) using populated data may lose content and high-level interaction information of HINs, even lead to negative impacts on the performance of downstream tasks; and 2) existing models fail to utilize the high-order heterogeneous structures in original incomplete network data. To resolve the above issues, in this article, we proposed a lattice-based incomplete heterogeneous structural attention network (LIHAN) for learning incomplete heterogeneous node embeddings. LIHAN first constructs characteristic lattice and structure lattice by mining characteristic sets and structure sets according to the partial order relations in between. Then, an improved lattice-based heterogeneous dual-attention mechanism is used to learn the heterogeneous node representations. Extensive node classification experiments are conducted on five open datasets to verify the superior performance of the proposed LIHAN model over the state-of-the-art models. Experimental results illustrate that LIHAN outperforms other methods on the micro-F1 and macro-F1 in node classification tasks. Moreover, experiments on different levels of lattices and the parameter sensitivity analysis shows the great stability during the process of experiments.

It is well known that the social network users' behavior can reflect their personal mental profiles. Therefore, in this paper, we conducted an experiment to study users' gender identity through a mobile social network. Gender identity is the personal sense of one's own gender, which can be the same or different from their sex assigned at birth, that is, one's innermost concept of self as male, female, a blend of both, or neither. In general, a person's psychological gender cannot be observed from the external characteristics of people. And even if some user's attributes such as physical gender and age can be found in social networks, it is difficult to find out the psychological gender of users. Apart from the fact that it will be hidden as the user's privacy, the main reason is that users' behaviors in social networks have no direct relationship with psychological gender. In this paper, we choose a special mobile social network to get around these limitations. Specifically, this paper is based on a real-world large gay social network of more than 10 000 000 users. In this mobile social network, we can indirectly mark the psychological gender of users by their sexual orientation. We collect 1000 users, including their characters, followers, followees and the same data for their followers and followees. Then, we examine an extensive set of features that can be extracted from the data set and train some classifiers to automatically identify the user's psychological gender. In the special mobile social network, a user's psychological gender is related to his characteristics, which include his personal characteristics, dating preference, and so on. This paper mainly analyzes users' personal characteristics, triads, and triadic evolution to infer their psychological gender. The experiments show that all these features are indeed effective in the classification, and even the features considered in previous studies have different implications with gender identity on other social networks. And it turns out that psychological gender as a kind of privacy can be leaked by these unintentional behaviors, and so we give some practical suggestions at the end of this paper to protect users' privacy.

Multi-modal entity alignment (MMEA) identifies identical real-world entities across two multi-modal knowledge graphs. Most existing methods heavily rely on costly manually labeled seed alignments; in response, self-supervised MMEA has emerged to reduce this dependency. However, current self-supervised approaches suffer from two key issues: (1) low-quality pseudo-labels (sparse and noisy) weakening self-supervised signals; and (2) inter-modal semantic inconsistencies (structure, text, vision) hindering unified entity representations. To resolve these issues, we propose OPICE, an Ontology-guided Pseudo-label Generation and Inter-modal Consistency Enhancement for self-supervised MMEA. It adopts a robust pseudo-label generation strategy to produce more initial alignments with less noise, and it uses an inter-modal consistency enhancement module to narrow inter-modal semantic gaps for unified representations. Experiments on FB–DB15K and FB–YAGO15K show that OPICE achieves state-of-the-art performance, improving Hit@1 by 6.8% on average over the strongest self-supervised baseline and being competitive with most supervised baselines under standard reported settings.