Inferring humans' private valuations for goods from their observed market behavior is essential for evaluating market efficiency and improving trading mechanism design. A core challenge lies in uncovering the human decision function that maps private valuations and observed market states to actions. In complex market settings where humans make sequential decisions in stochastic environments, neural networks offer the flexibility to model this decision function. However, training them without access to private valuations or environment dynamics remains challenging. We tackle this challenge and study how to infer heterogeneous human valuations from offline decision data in continuous double auctions. We propose learning the decision function via risk‑sensitive utility maximization. First, we train a generative model on offline bid and ask data to simulate individual trading behavior. Using this generative model, we instantiate simulated markets composed of randomly generated buyers and sellers. Second, we introduce an agent into these simulated markets and use reinforcement learning to learn a risk-sensitive utility-maximizing decision function for the agent. Third, we formulate a bilevel optimization to jointly recover private valuations and risk preference parameters. Our extensive experiments on a large‑scale continuous double auction dataset demonstrate that our framework significantly reduces errors in inferring real human valuations.

Yu Sun, Xingyu Qian, Weiwen Xu, Hao Zhang, Chenghao Xiao, Long Li, Deli Zhao, Wenbing Huang, Tingyang Xu, Qifeng Bai, Yu Rong. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Deep learning models for time series anomaly detection are mainly divided into supervised learning, semi-supervised learning, and unsupervised learning. Unsupervised learning has grown in importance for identifying anomaly features in time series without labels. Unsupervised learning of anomaly detection needs to be improved in two aspects. One is the calculation time, the other is the anomaly score. Some excellent unsupervised learning algorithms utilize clustering algorithms to filter out normal features, which takes too much time. Moreover, these algorithms use a fixed-length window to calculate anomaly scores, leading to inaccuracies and fluctuations. This paper proposes a Temporal Autoencoder Dynamic Series (TADS) framework. Here are the original contributions of the paper. Firstly, in the process of network compression coding, an enhanced Channel Attention Mechanism is used to identify critical features. Secondly, it uses Dynamic Series to replace the fixed window which improves the inaccuracy and fluctuation and also avoids the random error behavior caused by parameter adjustment. In addition, it is unnecessary to use clustering algorithms to obtain normal features. Finally, on the public MIT-BIH dataset, the experimental results show that TADS can improve the F1 score from 79% to 91% and accuracy from 89% to 93%. Meanwhile, the calculation time and fluctuation decreased by 42%and 19.8%, respectively. The experimental results demonstrate the performance of TADS for unsupervised anomaly detection on the MIT-BIH time series of ECG dataset through qualitative and quantitative methods.