For a long time, housing price dynamics have been a significant area of economic research, with various models and theories attempting to explain its complexity. However, the impact of market participants’ heterogeneity and isomerism on housing price fluctuations has not been thoroughly explored in these models. In this work, an improved Sugarscape model is used to construct an agent-based model to investigate the dynamics of the real estate market and the mechanisms behind its cyclical fluctuations. The model considers the range of agents’ vision, heterogeneity of expectations, and isomerism of roles to analyze how these factors contribute to the formation of housing price cycles. Simulation results show that expanding agents’ vision within a certain range increases market stability; however, beyond a specific value, it may lead to information overload, increasing market complexity and uncertainty. Furthermore, market volatility significantly increases when agents’ expectations converge. Analysis of isomerism indicates that agents with different roles have distinctly different impacts on housing prices, with varying behaviors and decision-making logics among market participants playing a crucial role in price fluctuations.

Nowadays artificial intelligence (AI) has been applied in many high-stake decision-making tasks. The black box AI models which are lack of explainability can cause serious problems in practice. In the justice, an explainable model becomes more and more important. Since tree-based machine learning models are explainable, we propose an explainable legal judgment prediction model using concept trees with collegiate bench mechanism in this paper. A concept tree is constructed to check the classification labels predicted by the original multi-classifier. A revising process is designed to deal with the scenario when the results of the original multi-classifier and the concept trees are conflicted. Meanwhile, the concept trees grow into concept forest because of the existence of arbitration classifiers. The judicial judgment process is simulated, which not only makes the good classification performance with collegiate bench mechanism, but also has the model explanation from the features in the conceptual levels. The experiments validate the validity of our model with both better explainability and better accuracy.

Multiple-choice questions (MCQs) are widely used in the evaluation of large language models (LLMs) due to their simplicity and efficiency. However, there are concerns about whether MCQs can truly measure LLM's capabilities, particularly in knowledge-intensive scenarios where long-form generation (LFG) answers are required. The misalignment between the task and the evaluation method demands a thoughtful analysis of MCQ's efficacy, which we undertake in this paper by evaluating nine LLMs on four question-answering (QA) datasets in two languages: Chinese and English. We identify a significant issue: LLMs exhibit an order sensitivity in bilingual MCQs, favoring answers located at specific positions, i.e., the first position. We further quantify the gap between MCQs and long-form generation questions (LFGQs) by comparing their direct outputs, token logits, and embeddings. Our results reveal a relatively low correlation between answers from MCQs and LFGQs for identical questions. Additionally, we propose two methods to quantify the consistency and confidence of LLMs' output, which can be generalized to other QA evaluation benchmarks. Notably, our analysis challenges the idea that the higher the consistency, the greater the accuracy. We also find MCQs to be less reliable than LFGQs in terms of expected calibration error. Finally, the misalignment between MCQs and LFGQs is not only reflected in the evaluation performance but also in the embedding space. Our code and models can be accessed at https://github.com/Meetyou-AI-Lab/Can-MC-Evaluate-LLMs.

Explainable Artificial Intelligence (XAI) is a crucial research topic in the intersection of social science and artificial intelligence. In legal provisions recommendation, the existing models insufficiently utilize legal provisions which have the hierarchical structure. The hierarchical structure is based on granular computing. That overlooks the relationship between legal provisions and related case details. Concurrently, category imbalance problem exists in current judicial documents. This can easily lead to model training imbalance, subsequently making it difficult to accurately predict legal provisions which appear less frequently. The inherent hierarchical structure of legal provisions is exploited to learn the semantic expressions of major crime categories and legal provisions. Simultaneously, the collaborative attention mechanism is introduced to enhance the explainability. In addition, the legal cases can be augmented with text enhancement strategies. We propose a pre-trained language model incorporated with prior knowledge. The experiments were conducted on CAIL2018 and compared with six benchmark models and two ablation models. The results show our model's effectiveness and explainability.

Natural language processing is a significant branch of machine learning, and pre-trained models such as BERT have been widely used in it. Previous research has shown that sentence embeddings from pre-trained language models without fine-tune have difficulty in capturing their exact semantics. The ambiguous semantics leads to poor performance on semantic text similarity (STS) tasks. However, fine-tune tends to skew the model toward high-frequency distributions due to the heterogeneous nature of word frequency and word sense distributions. Therefore, fine-tune is not a optimal choice. To address this issue, we propose an unsupervised flow-based contrastive learning model. The model maps sentence embedding distributions to smooth and isotropic Gaussian distributions, thus mitigating the impact caused by irregular word frequency distributions. To evaluate the performance of our model, we use an industry-recognized method that outperforms competing baselines in different sentence-related tasks.

An advanced tomography method based on Bayesian probability theory is presented in this article. In the method, Gaussian Process (GP) prior is adopted as an effective approach to smoothness regularization which can be optimized based on the balance between model complexity and data constraint. In particular, to address the problem of varying smoothness in space, a non-stationary version of the GP has been developed and resolved via Bayesian hierarchical algorithm to implement locally adaptive smoothness regularization such that the accuracy of the reconstruction can be improve significantly. The Bayesian formulism allows the reliability of the reconstruction result to be examined by the confidence interval of a posterior probability. Through a wide range of applications, this tomography method is proved to be a robust tool for the study of magnetohydrodynamics (MHD) activity and impurity transport during HL-2A experimental campaigns. (C) 2022 The Japan Society of Plasma Science and Nuclear Fusion Research

Bayesian rule lists (BRL) model aims to provide interpretability for prediction and decision-making process, where researchers have conducted sufficient work to make such probabilistic model applicable to be integrated into more complex architecture. Deep neural net with prototype layers initially created to deal with dimension reduction, visualization and image data interpretability in computer vision tasks, is now also widely used in other fields due to its pluggable interpretability. However, pure BRL is lack of model expressiveness especially for settings where approximation for highly complex non-linear mapping is requisite. Meanwhile, the prototype deep net, although can output certain kind of prototypes, relies on other interpretable models or problem definition to provide concrete interpretable logics for the generated prototypes. In this paper, we mainly deal with difficulty of providing both interpretability and decent prediction precision for decision process, proposing a deep neural network with prototype layers based on BRL model, known as self-interpretable Bayesian rule lists (SiBRL). This model provides interpretability through Bayesian rules and prototype with decent model performance based on expressive deep net. We first use truncated Poisson distribution and uniform distribution to generate rules and decision points which are put into deep net together with training data, and then compute the likelihood based on deep net architecture with Bayesian rule process. Finally, cutting plane algorithm (CPA) is used to optimize the parameters and hyperparameters in the BRL model with the objective being the likelihood given by the deep net part. Experiment based on a default prediction dataset shows that our model can provide reasonable interpretability while keeping prediction precision superior to non-ensembled machine learning models such as decision tree and BRL.

Knowledge graphs (KGs) automatic construction generally involves fine-grained entity typing, i.e., assigning the types to given entities as (entity, entity type). Since the non-negligible inaccuracy of entity typing systems and lack of sufficient human supervision, KGs inevitably face entity type noises. However, most conventional entity type embedding models unreasonably assume that all entity type instances in existing KGs are completely correct, which ignore noises and could lead to potential errors for down-stream tasks. To address this issue, we propose TrustE to build trustworthiness-aware entity type structured embeddings, which takes possible entity type noises into consideration for learning better representations. Specifically, since entities and entity types are completely distinct objects, we encode them in separate entity space and entity type space with a structural projecting matrix, and learn entity type embeddings with tuple trustworthiness. To make the trustworthiness more universal, we only utilize the internal structural knowledge in existing KGs and build two tuple trustworthiness considering the local tuple and global triple information respectively, which correspondingly makes it more challenging due to the limited knowledge. We evaluate our models on three tasks: entity type noise detection, entity type prediction and classification. Experimental results on real-world datasets (FB15kET and YAGO43kET) show that our models outperform all baselines on all tasks, which verify the capability of TrustE in learning better entity type structural embeddings on noisy KGs. The source code and data of this paper can be obtained from https://github.com/Quan-SWUFE/TrustE

Relation extraction between entity pairs is an increasingly critical area in natural language processing. Recently, the pre-trained bidirectional encoder representation from transformer (BERT) performs excellently on the text classification or sequence labelling tasks. Here, the high-level syntactic features that consider the dependency between each word and the target entities into the pre-trained language models are incorporated. Our model also utilizes the intermediate layers of BERT to acquire different levels of semantic information and designs multi-granularity features for final relation classification. Our model offers a momentous improvement over the published methods for the relation extraction on the widely used data sets.

事件抽取是自然语言处理领域的重要研究方向.传统的事件类型抽取系统采用分类方式,无法解决跨句子的事件角色和事件类型匹配问题.为了解决该问题,提出一种基于序列标注的事件联合抽取模型,结合卷积神经网络(convolutional neural networks,CNN)与长短期记忆网络(long short-term memory,LSTM)提取全局特征和局部特征;并在浅层LSTM层共享参数实现联合抽取,以序列标注方式抽取事件论元并匹配事件类型.实验结果表明,模型能有效提取司法领域的文档事件信息.

Social media popularity estimation refers to predict the post's popularity using multimodal contents. The prediction performance heavily relies on the feature extraction part and fully leveraging multimodal heterogeneous data is of a great challenge in the practical settings. Despite remarkable progress have been made, most of the previous attempts are restrained from the essentially limited property of the employed single modality. Inspired by the recent success of multimodal learning, we propose a novel multimodal deep learning framework for the popularity prediction task, which aims to leverage the complementary knowledge from different modalities. Moreover, an attention mechanism is introduced in our framework, with the goal to assign large weights to specified modalities during the training and inference phases. To empirically investigate the effectiveness and robustness of the proposed approach, we conduct extensive experiments on the 2020 SMP challenge. The obtained results show that the proposed framework outperforms related approaches.

XAI (eXplainable Artificial Intelligence) has been an important cross domain topic between social sciences and artificial intelligence. Especially in the field of Legal Judgment Prediction (LJP), the computer systems aim to predict the judgments based on the facts of legal cases. The features of the subject matters, the subjects' behaviors, and the objective results are highly related to the crimes and punishments. Then the results should be coarsely explainable to people. However, many machine learning algorithms cannot make full use of such information and cannot give people the explaination for the results of LJP. In this paper, an Interpretable Conditional Classification Tree model (ICCT) is proposed to study the multi-class problem in LJP. Our model uses the prior information to recursively generate tree nodes. A feature search method for the feature domain construction, a data clustering algorithm and a grouping algorithm for tree node construction are proposed. The growth processes of the conditional classification tree realize the transition from coarse-grained classification to fine-grained classification which is called multi-granularity. The experimental results show the ICCT which has better interpretability achieves better performances over the baselines on the judgment prediction tasks.

随着机器学习在社会中的广泛使用,带来的歧视问题引发广泛的社会争议,这逐步引起了产业界和学术界对机器学习算法公平性问题的浓厚兴趣.目前对公平性度量和机器学习公平性机制的研究仍然处于初级阶段.本文对公平性机器学习的研究进行了调研,首先从公平性的定义出发,比较了衡量公平性指标的方法,然后调研了公平性数据集,对公平性问题的产生进行了分析,接下来对现有的公平性机器学习算法进行归类和比较,最后总结了当前公平性机器学习研究中存在的问题,并对关键问题和重大挑战进行了讨论.

目前面向分类的差分隐私保护算法中,大部分都是基于决策树或者随机森林等树模型.若数据集中同时存在连续数据和离散数据时,算法往往会选择调用2次指数机制,并且进行隐私预算分配时往往选择平均分配.这都使得隐私预算过小、噪声过大、时间成本增加以及分类准确性降低.如何在保证数据隐私的同时尽可能地保证数据可用性,并提高算法性能,成为目前差分隐私保护技术研究的重点.提出了面向决策树和随机森林的2种差分隐私保护数据挖掘算法,使用Laplace机制来处理离散型特征,使用指数机制处理连续型特征,选择最佳分裂特征和分裂点,并采用最优特征选择策略和等差预算分配加噪策略.对金融数据集的测试结果表明,提出的基于树模型的差分隐私保护算法都能在保护数据隐私的同时,具有较高的分类准确性,并且能够充分利用隐私保护预算,节省了时间成本.

Relation classification is a very important Natural Language Processing (NLP) task to classify the relations from the plain text. It is one of the basic tasks of constructing a knowledge graph. Most existing state-of-the-art methods are primarily based on Convolutional Neural Networks(CNN) or Long Short-Term Memory Networks(LSTM). Recently, many pre-trained Bidirectional Encoder Representation from Transformers (BERT) models have been successfully used in the sequence labeling and many NLP classification tasks. Relation classification is different in that it needs to pay attention to not only the sentence information but also the entity pairs. In this paper, a Siamese BERT model with Adversarial Training (SBERT-AT) is proposed for relation classification. Firstly, the features of the entities and the sentence can be extracted separately to improve the performance of relation classification. Secondly, the adversarial training is applied to the SBERT architecture to improve the robustness. Lastly, the experimental results demonstrate that we achieve significant improvement compared with the other methods on real-world datasets.

The aim of dimensionality reduction is to obtain the faithful low-dimensional representations of high-dimensional data by preserving the data quality. It is beneficial to better visualize the high-dimensional data and improve the classification or clustering performance. Many dimensionality reduction methods based on the framework of stochastic neighbor embedding have been developed. However, most of them use the Euclidean distance to describe the dissimilarity of data points in high-dimensional space, which is not suitable for high-dimensional data with non-linear manifold structure. In addition, they usually use the family of normal distributions as their embedding distributions in low-dimensional space. This will incur that they are only suitable to deal with the spherical data. In order to deal with these issues, we present a novel dimensionality reduction method by integrating the Wasserstein distance and t-copula function into the stochastic neighbor embedding model. We first employ the Gaussian distribution equipped with the Wasserstein distance to describe the pairwise similarity in the high-dimensional space. Then, the t-copula function is used to generate a general heavy-tailed distribution for the description of low-dimensional pairwise similarity, which can process different shapes of data and avoid the crowding problem. Furthermore, Kullback–Leibler divergence is employed to measure the difference between the high-dimensional and low-dimensional similarities. Finally, a gradient descent algorithm with adaptive moment estimation is developed to solve the proposed objective function. Extensive experiments are conducted on eight real-world datasets to demonstrate the effectiveness of the proposed method in terms of the dimensional reduction quality, classification and clustering evaluation metrics.

The performance of the traditional Word2Vec model heavily depends on the quality and quantity of the corpus, which violates the way of the humans learn. To understand word meaning, human beings prefer a two-stage learning process. That is, reading a Linguist-compiled dictionary as well as doing reading comprehension. These two stages complement each other. Traditional Word2Vec is an analogy of reading comprehension. While the first stage, learning the semantic rules from a language dictionary, such as the knowledge of thesaurus and etymology, is usually ignored by existing methods. In this work, we propose a robust word embedding learning framework by imitating the two-stage human learning process. In particular, we construct a semantic manifold based on the thesaurus and etymology to approximate the first stage. Then, we regularize the second stage (Word2Vec model) with this semantic manifold. We train the proposed model on three corpora (Wikipedia, enwik9 and text8). The experimental results demonstrate that the proposed method learns much smoother vector representations. Also, the performance on learning word embedding is robust even when the method is trained with a very simple corpus.

Advanced cyberattacks are often featured by multiple types, layers, and stages, with the goal of cheating the monitors. Existing anomaly detection systems usually search logs or traffics alone for evidence of attacks but ignore further analysis about attack processes. For instance, the traffic detection methods can only detect the attack flows roughly but fail to reconstruct the attack event process and reveal the current network node status. As a result, they cannot fully model the complex multistage attack. To address these problems, we present Traffic-Log Combined Detection (TLCD), which is a multistage intrusion analysis system. Inspired by multiplatform intrusion detection techniques, we integrate traffics with network device logs through association rules. TLCD correlates log data with traffic characteristics to reflect the attack process and construct a federated detection platform. Specifically, TLCD can discover the process steps of a cyberattack attack, reflect the current network status, and reveal the behaviors of normal users. Our experimental results over different cyberattacks demonstrate that TLCD works well with high accuracy and low false positive rate.

Relation classification is an important semantic processing task in natural language processing (NLP). Traditional works on relation classification are primarily based on supervised methods and distant supervision which rely on the large number of labels. However, these existing methods inevitably suffer from wrong labeling problem and may not perform well in resource-poor domains. We thus utilize transfer learning methods on relation classification to enable relation classification system to adapt resource-poor domains along with different relation type. In this paper, we exploit a convolutional neural network to extract lexical and syntactic features and apply transfer learning approaches for transferring the parameters of convolutional layer pre-training on general-domain corpus. The experimental results on real-world datasets demonstrate that our approach is effective and outperforms several competitive baseline methods.

Ion internal transport barriers (iITBs) arc first observed in neutral beam injection (NBI) heated plasmas at the I HL-2A tokamak, The position of the barrier foot, in the stationary state, coincides with the q = 1 surface within its uncertainty of measurement. iITBs can develop more easily at the beginning of NBI heating. Also, iITBs are unstable for the sawtooth plasma, Simulations reveal that the thermal diffusivity of ions (x) inside the barrier can be as low as the neoclassical level. It is observed that the flow shear in the stationary iITB state reaches the level required for suppressing the ion temperature gradient mode instability, which indicates the important role of flow shear in sustaining the iITB,