Chunxiao Xing's research while affiliated with Tsinghua University and other places

... Note that the methods with other enhancing techniques, such as data augmentation [2,17,50,51] or AutoML [18,37,44,47,48] are orthogonal to our approach for comparison. ...

... Such methods only consider the preceding information but ignore the succeeding information, which is insufficient to obtain the optimal result. Different from the greedy strategy, another solution is the context-wise re-ranking strategy [1,6,11,20,25], which uses a context-wise evaluation model to capture the mutual influence among items and re-predict the CTR of each item. Methods like PRM [20], directly take the initial ranking list as input, and generates the optimal permutation based on the predict value given by context-wise model. ...

... Unfortunately, both R-Tree implementations do not disclose their source code. Finally, learned indexes [53] achieve high performance on GPUs, since they heavily depend on linear algebra operations, which have been extensively optimized on GPUs, and some GPUs even offer specialized accelerators for these operations. Regrettably, we were unable to check whether the existing learned index takes advantage of hardware acceleration, as there is no accompanying implementation available. ...

... We employ weak models to capture features useful for stance detection but not useful for intermediate sub-tasks as bias features through adversarial learning (Belinkov et al., 2019). processing, including visual question answering (Niu et al., 2021), natural language inference (Tian et al., 2022), question answering , named entity recognition , and text classification (Wang and Culotta, 2020;Qian et al., 2021;Choi et al., 2022). Different from existing task-agnostic counterfactual frameworks, our model incorporates an adversarial bias learning module that leverages intermediate sub-tasks of stance detection to learn bias more accurately. ...

... Data prefetching or caching are techniques that are highly used mostly in databases [12,13] in order to boost the execution performance by fetching data in the memory before it is needed. These methodologies are widely used in database systems or the prefetching of small-sized data such as websites [14,15] or Internet of Things (IoT) data [16]. ...

... If the distance of u is larger than a(u) + r(u) + l(u) which means the node u cannot be the follower, we set the status of u as closed, and use the Shrink of Algorithm 4 to update its neighbors' distance and status (Lines 12-13). Otherwise, we set the status of u as activated (Line 15) and process each listening node z who is the neighbors of u (Lines [16][17][18][19][20][21]. If node z is not in the k-core and p(z) > p(u) , we push z into Q and update the distance u and z by subtracting 1. ...

... Triple classification task is intended to infer whether the triples of facts in the knowledge base are correct [36]. The classification accuracy rates of three standard data sets WN18, WN18RR and FB15K-237 are shown in Table 7. ...

... There are many applications regarding trajectory similarity computation. Several previous studies [3,8,36] aimed at accelerating the similarity search and join over trajectory data by devising index and pruning techniques following the idea from string similarity search [37][38][39][40]. Specifically, tree-based index structures [22,27] , such as K-D tree or R-tree are employed to organize the trajectories. ...

... First, most studies examined HRV in patients in the subacute (>7 days following stroke onset) or chronic phase of stroke (>6 months), whereas the most severe stroke complications, including autonomic dysfunction, appear in the acute phase of stroke. Second, most studies included only patients with minor stroke with less disability and smaller infarcts with frequent favorable outcomes and a low risk of cardiac complications, which increases proportionally to the severity of ischemic stroke and neurological deficits (Yoshimura et al., 2008;Guan et al., 2019;Zhang et al., 2019;Li et al., 2021). Third, the studies had different follow-ups, and the risk of stroke complications was the highest in the first 3 months after stroke. ...

... The method that might be used is a data clustering technique that has been successfully applied in data mining (Cheng et al., 2014). Data clustering can be done with several techniques or algorithms including single link, average link, minimum spanning tree, k-means (Li et al., 2019;Yang, 2017). Data clustering aims to extract relationships in a dataset and to determine interesting patterns based on sameness of the sample by grouping several data objects into groups or clusters so that objects in a cluster have a high degree of similarity and very different from objects in other clusters. ...