In recent years,portfolio management problems have been extensively studied in the field of artificial intelligence,but there are some improvements in the existing quantitative trading methods based on deep learning.First of all,the prediction model of stocks is single,usually a model only trains a trading expert,and the decision of trading is only based on the prediction results of the model.Secondly,the data source used in the model is relatively single,only considering the stock's own data,ignoring the impact of the entire market risk on the stock.Aiming at the above problems,a reinforcement learning model based on dynamic se-lection predictor(DSDRL)is proposed.The model is divided into three parts.Firstly,the characteristics of stock data are extrac-ted and introduced into multiple predictors.Multiple prediction models are trained for different investment strategies,and the cur-rent optimal prediction results are obtained by dynamic selector.Secondly,the market environment evaluation module is used to quantify the current market risk and obtain the appropriate proportion of investment amount.Finally,based on the first two mo-dules,a deep reinforcement learning model is established to simulate the real trading environment,and the actual portfolio strate-gy is obtained based on the predicted results and the proportion of investment amount.In this paper,the daily k-line data of China Securities 500 and S & P 500 are used for test verification.The results show that the proposed model is superior to other refe-rence models in Sharpe rate and other indicators.
Time series anomaly detection is one of the critical tasks in time series analysis. There are many existing works using autoencoders for time series anomaly detection; however, autoencoders can not only reconstruct normal sequences but also reconstruct anomalous sequences sometimes, which leads to the poor effect of anomaly detection based on reconstruction error. At the same time, the real-world time series have the characteristic of temporal coupling, and the existing methods often cannot take into account the multiple modes of the time series, which leads to the low accuracy of time series anomaly detection. Aiming at the above problems, we propose a Time Series Anomaly Detection method incorporating Wavelet Decomposition and Temporal Decoupled Autoencoder (WDAE). Specifically, WDAE first reshapes the original 1D time series into 2D data harboring different temporal patterns based on frequency components and independently extracts the different temporal features of the time series data. Then, the original data are reconstructed in the time domain by the autoencoder to obtain the reconstructed time-domain sequence, and in the frequency domain, the 2D data are subjected to the discrete wavelet transform and the inverse wavelet transform after weight adjustment to obtain the reconstructed frequency domain sequence, and the captured normal sequence pattern in the frequency domain restricts the latent feature space in the time domain so that the autoencoder can rebuild the normal sequence to avoid reconstructing the anomalous sequence well. Experiments on five publicly available data show that WDAE significantly improves the F1 value over related state-of-the-art baseline methods.
In order to capture and integrate the structural features and temporal features contained in social graph and diffusion cascade more effectively, an information diffusion prediction model based on the Transformer and Relational Graph Convolutional Network (TRGCN) is proposed. Firstly, a dynamic heterogeneous graph composed of the social network graph and the diffusion cascade graph was constructed, and it was input into the Relational Graph Convolutional Network (RGCN) to extract the structural features of each node. Secondly, the time embedding of each node was reencoded using Bi-directional Long Short-Term Memory (Bi-LSTM). The time decay function was introduced to give different weights to nodes at different time positions, so as to obtain the temporal features of nodes. Finally, structural features and temporal features were input into Transformer and then merged. The spatiotemporal features are obtained for information diffusion prediction. The experimental results on three real datasets of X (formerly known as Twitter), Douban, and Memetracker show that compared with the optimal model in the comparison experiment, the TRGCN model has an average increase of 4.16% in Hits@100 metric and 13.26% in map@100 metric. The validity and rationality of the model are proved.
Most of the existing hypernetwork representation learning methods fail to fully consider the hyperedges, leading to the untapped potential of information contained within the hyperedges. To address this issue, this article proposes a heterogeneous hypernetwork representation learning method with hyperedge fusion abbreviated as HRHF. First, this method incorporates the hyperedges into random walk node sequences by means of incidence graph to enhance tuple relationships, i.e., the hyperedges among the nodes. Second, under the condition of the above random walk node sequences, the cognitive structure model, cognitive set model, and cognitive hyperedge model are jointly optimized to comprehensively consider pairwise relationships and tuple relationships among the nodes to learn high-quality node representation vectors. The experimental results demonstrate that, for the link prediction task, this method outperforms other optimal baseline methods, i.e., hyper-path- based random walks + hyper-gram (HPHG) by 0.99% points on the drug dataset, is comparable to the other optimal method, i.e., Event2vec on the global positioning system (GPS) dataset, is close to the performance of other optimal methods, i.e., hyper-path-based random walks + skip-gram (HPSG) on the MovieLens, and is close to the performance of other optimal methods, i.e., Hyper2vec on the WordNet dataset. For the hypernetwork reconstruction task, this method achieves superior average performance on the drug and GPS datasets compared with other baseline methods.
文本分类任务通常依赖足量的标注数据,针对低资源场景下的分类模型在小样本上的过拟合问题,提出一种基于提示学习的小样本文本分类方法BERT-P-Tuning.首先,利用预训练模型BERT(Bidirectional Encoder Representations from Transformers)在标注样本上学习到最优的提示模板;然后,在每条样本中补充提示模板和空缺,将文本分类任务转化为完形填空任务;最后,通过预测空缺位置概率最高的词并结合它与标签之间的映射关系得到最终的标签.在公开数据集FewCLUE上的短文本分类任务上进行实验,实验结果表明,所提方法相较于基于BERT微调的方法在评价指标上有显著提高.所提方法在二分类任务上的准确率与F1值分别提升了25.2和26.7个百分点,在多分类任务上的准确率与F1值分别提升了6.6和8.0个百分点.相较于手动构建模板的PET(Pattern Exploiting Training)方法,所提方法在两个任务上的准确率分别提升了2.9和2.8个百分点,F1值分别提升了4.4和4.2个百分点,验证了预训练模型应用在小样本任务的有效性.
