The task of converting monolithic programs to mi-croservices architecture is complex, hindered by the intertwined nature of program control and data flows. Conventional methods for microservice extraction often fall short in capturing the essen-tial connections within a monolithic structure and in propagating properties to distant neighbors for effective clustering. To address these issues, we introduce an innovative graph-based deep clustering technique that utilizes both control flow and data flow graphs. This approach offers a thorough analysis of class interactions within monolithic applications, facilitating accurate identification and extraction of microservices. Furthermore, we present the Microservice Extraction Graph Neural Network (MEGNN), an advanced graph attention network designed to enhance message transmission depth and enable nodes to assimilate features from k-hop neighbors. This method extends the reach of message distribution across node chains and mitigates the issue of feature homogenization, leading to more cohesive clustering of related nodes and improving the quality of microservices extraction. Experimental evaluations on data from three publicly accessible Java monolithic programs confirm that our proposed method surpasses existing techniques in microservices extraction efficacy.

The utilization of text classification techniques has demonstrated great promise in the field of detecting consumer fraud based on consumer reviews. However, persistent challenges remain in handling large samples at the borders and identifying unforeseen fraud behaviors. To address these challenges, we propose a novel approach that combines a channel biattention convolutional neural network (CNN) with a pretrained language model. Specifically, we propose a similarity computation module for implicitly learning a metric matrix to characterize the similarity between prior knowledge and consumer reviews in vector space. Through this process, the model is able to learn and understand the relationship between prior knowledge and corresponding samples during training, thereby improving its ability to identify unforeseen fraudulent behaviors. Additionally, we propose a channel biattention CNN module to adaptively emphasize the importance of relevant prior knowledge to enhance the model’s ability to accurately classify boundary samples. To ensure effective model training, we expand and organize a real-world dataset, reducing noise and increasing the number of fraud samples available for analysis. Experimental results demonstrate that our approach achieves state-of-the-art performance in fraud detection. Notably, our model is capable of detecting unforeseen fraud cases without the need for retraining or fine-tuning, making it highly adaptable and efficient in practical applications.

Matching APP reviews with bug reports can help APP developers to quickly identify new bugs from the users’ feedback. Existing solutions represent the semantics of APP reviews and bug reports via carefully designed features and models, the performance of which however depends heavily on the manually designed model and the training data set. Large-scale pretrained models can well capture the semantics of text and have demonstrated their success in many NLP tasks. Inspired by this, we explore the effect of various pretrained models on the matching accuracy of app review and bug report. We conduct a systematic study to analyze the factors of four major pretrained models (including T5, Sentence T5, Sentence MiniLM, Sentence BERT and so on) on the matching accuracy. We find that the accuracy of Sentence T5 and Sentence MiniLM in four open source applications is significantly greater than that of the state-of-the-art approach DeepMatcher. Based on the findings, we design a novel approach to match the APP reviews with bug reports based on the pretrained model Sentence T5 and Sentence MiniLM to calculate the sentence similarity. We test it on four open source applications and the results show that our method outperforms the existing solution. On average, the precision of Sentence T5 and Sentence MiniLM are increased by 17% and 13%, respectively, and the hit ratio are increased by 15% and 14%, respectively.

Mobile application markets usually enact policies to describe in detail the minimum requirements that an application should comply with. User comments on mobile applications contain a large amount of information that can be used to find out APP's violations of market policies in a cost-effective way. Existing state-of-the-art methods match user comments with the violations of market policies based on well-designed syntax rules, which however cannot well capture the semantics of user comments and cannot be generalized to the scenarios not covered by the rules. To address this issue, we propose an innovative method, UBC-BERT, to detect undesired behavior from user comments based on their semantics. By incorporating sentence embeddings with attention, we train a classification model for 21 groups of undesirable behaviors based on the fine-tuning of a pre-trained model BERT-BASE. The experimental results show that our solution outperforms the baseline solutions in terms of a higher precision(up to 60.5% more).

The rapid development of tourism economy has brought new challenges such as the prevention of consumer fraud for smart city applications. Traditional fraud detection approaches such as telecommunication fraud and credit card fraud detection need a data set containing both the normal behavior and abnormal behavior. Therefore, they are incompetent to address such challenges in the tourism market where the data set is open and very few records of fraud behaviors are available. To address this issue, we propose a P-feature conversion algorithm to construct third-party features to expose the outliers. These features reveal the internal characteristics of different businesses and their useful internal connections. Then, we build a fraud detection model for tourism based on the Local Outlier Factor anomaly detection algorithm. Experimental results show that our model can effectively identify fraudulent merchants in the tourism market.

Different news events have different effects on stock price changes. If they are simply fed to the neural network for prediction, the accuracy will be affected. We propose a method to predict stock price trend based on time series news information. First, we extract events from news text and represent them as dense vectors by event embedding technique. Further-more, we employ attention mechanism to figure out event is the main cause of the price fluctuation. Then, we use a Gated Recurrent Unit to model the influence of events on stock market. Experimental results show that our model achieve a certain improvement on S&P500 index compared to baseline methods.

Stock trend forecasting plays a great role in maximizing the profit of stock investment. However, due to the high volatility and non-stationarity of the stock market, accurate trend prediction is very difficult. With the development of the Internet and deep learning technology, people can use deep learning methods to reveal market trends and volatility from the explosive information on the Internet. Unfortunately, there is a large amount of content related to the stock market, and a large part of it is useless information. As a result, how to extract the effective information and combine this information as different characteristics to effectively predict stock trends has become the biggest challenge. In order to cope with these challenges, we use TextCNN as the news text feature extractor for feature extraction of news information, and propose a prediction method based on multi-feature fusion: Bi-LSTNAA, to predict the Chinese stock market. Extensive experiments on actual stock market data show that the our method has a greater improvement in the accuracy of stock trend prediction.