H.3.8. Natural Language Processing
Milad Allahgholi; Hossein Rahmani; Parinaz Soltanzadeh
Abstract
Stance detection is the process of identifying and classifying an author's point of view or stance towards a specific target in a given text. Most of previous studies on stance detection neglect the contextual information hidden in the input data and as a result lead to less accurate results. In this ...
Read More
Stance detection is the process of identifying and classifying an author's point of view or stance towards a specific target in a given text. Most of previous studies on stance detection neglect the contextual information hidden in the input data and as a result lead to less accurate results. In this paper, we propose a novel method called ConSPro, which uses decoder-only transformers to consider contextual input data in the process of stance detection. First, ConSPro applies zero-shot prompting of decoder only transformers to extract the context of target in the input data. Second, in addition to target and input text, ConSPro uses the extracted context as the third type of parameter for the ensemble method. We evaluate ConSPro on SemEval2016 and the empirical results indicate that ConSPro outperforms the non-contextual approaches methods, on average 9% with respect to f-measure. The findings of this study show the strong capabilities of zero-shot prompting for extracting the informative contextual information with significantly less effort comparing to previous methods on context extraction.
H.3.8. Natural Language Processing
Milad Allhgholi; Hossein Rahmani; Amirhossein Derakhshan; Saman Mohammadi Raouf
Abstract
Document similarity matching is essential for efficient text retrieval, plagiarism detection, and content analysis. Existing studies in this field can be categorized into three approaches: statistical analysis, deep learning, and hybrid approaches. However, to the best of our knowledge, none have incorporated ...
Read More
Document similarity matching is essential for efficient text retrieval, plagiarism detection, and content analysis. Existing studies in this field can be categorized into three approaches: statistical analysis, deep learning, and hybrid approaches. However, to the best of our knowledge, none have incorporated the importance of named entities into their methodologies. In this paper, we propose DOSTE, a method that first extracts name entities and then utilizes them to enhance document similarity matching through statistical and graph-based analysis. Empirical results indicate that DOSTE achieves better results by emphasizing named entities, resulting in an average improvement of 9% in the average recall metric compared to baseline methods. Also, DOSTE unlike LLM-based approaches, does not require extensive GPU resources. Additionally, non-empirical interpretations of the results indicate that DOSTE is particularly effective in identifying similarity in short documents and complex document comparisons.
H.3. Artificial Intelligence
Mohamad Mahdi Yadegar; Hossein Rahmani
Abstract
In recent years, new technologies have brought new innovations into the financial and commercial world, giving fraudsters many ways to commit fraud and cost companies big time. We can build systems that detect fraudulent patterns and prevent future incidents using advanced technologies. Machine learning ...
Read More
In recent years, new technologies have brought new innovations into the financial and commercial world, giving fraudsters many ways to commit fraud and cost companies big time. We can build systems that detect fraudulent patterns and prevent future incidents using advanced technologies. Machine learning algorithms are being used more for fraud detection in financial data. But the common challenge is the imbalance of the dataset which hinders traditional machine learning methods. Finding the best approach towards these imbalance datasets is the problem many of the researchers are facing when trying to use machine learning methods. In this paper, we propose the method called FinFD-GCN that use Graph Convolutional Networks (GCNs) for fraud detection in credit card transaction datasets. FinFD-GCN represents transactions as graph in which each node represents a transaction and each edge represents similarity between transactions. By using this graph representation FinFD-GCN can capture complex relationships and anomalies that may have been overlooked by traditional methods or were even impossible to detect with conventional approaches, thus enhancing the accuracy and robustness of fraud detection in financial data. We use common evaluation metrics and confusion matrices to evaluate the proposed method. FinFD-GCN achieves significant improvements in recall and AUC compared to traditional methods such as logistic regression, support vector machines, and random forests, making it a robust solution for credit card fraud detection. By using the GCN model for fraud detection in this credit card dataset we outperformed base models 5% and 10%, with respect to F1 and AUC, respectively.
H.3. Artificial Intelligence
Sajjad Alizadeh Fard; Hossein Rahmani
Abstract
Fraud in financial data is a significant concern for both businesses and individuals. Credit card transactions involve numerous features, some of which may lack relevance for classifiers and could lead to overfitting. A pivotal step in the fraud detection process is feature selection, which profoundly ...
Read More
Fraud in financial data is a significant concern for both businesses and individuals. Credit card transactions involve numerous features, some of which may lack relevance for classifiers and could lead to overfitting. A pivotal step in the fraud detection process is feature selection, which profoundly impacts model accuracy and execution time. In this paper, we introduce an ensemble-based, explainable feature selection framework founded on SHAP and LIME algorithms, called "X-SHAoLIM". We applied our framework to diverse combinations of the best models from previous studies, conducting both quantitative and qualitative comparisons with other feature selection methods. The quantitative evaluation of the "X-SHAoLIM" framework across various model combinations revealed consistent accuracy improvements on average, including increases in Precision (+5.6), Recall (+1.5), F1-Score (+3.5), and AUC-PR (+6.75). Beyond enhanced accuracy, our proposed framework, leveraging explainable algorithms like SHAP and LIME, provides a deeper understanding of features' importance in model predictions, delivering effective explanations to system users.
Document and Text Processing
Mina Tabatabaei; Hossein Rahmani; Motahareh Nasiri
Abstract
The search for effective treatments for complex diseases, while minimizing toxicity and side effects, has become crucial. However, identifying synergistic combinations of drugs is often a time-consuming and expensive process, relying on trial and error due to the vast search space involved. Addressing ...
Read More
The search for effective treatments for complex diseases, while minimizing toxicity and side effects, has become crucial. However, identifying synergistic combinations of drugs is often a time-consuming and expensive process, relying on trial and error due to the vast search space involved. Addressing this issue, we present a deep learning framework in this study. Our framework utilizes a diverse set of features, including chemical structure, biomedical literature embedding, and biological network interaction data, to predict potential synergistic combinations. Additionally, we employ autoencoders and principal component analysis (PCA) for dimension reduction in sparse data. Through 10-fold cross-validation, we achieved an impressive 98 percent area under the curve (AUC), surpassing the performance of seven previous state-of-the-art approaches by an average of 8%.
Mohammad Nazari; Hossein Rahmani; Dadfar Momeni; Motahare Nasiri
Abstract
Graph representation of data can better define relationships among data components and thus provide better and richer analysis. So far, movies have been represented in graphs many times using different features for clustering, genre prediction, and even for use in recommender systems. In constructing ...
Read More
Graph representation of data can better define relationships among data components and thus provide better and richer analysis. So far, movies have been represented in graphs many times using different features for clustering, genre prediction, and even for use in recommender systems. In constructing movie graphs, little attention has been paid to their textual features such as subtitles, while they contain the entire content of the movie and there is a lot of hidden information in them. So, in this paper, we propose a method called MoGaL to construct movie graph using LDA on subtitles. In this method, each node is a movie and each edge represents the novel relationship discovered by MoGaL among two associated movies. First, we extracted the important topics of the movies using LDA on their subtitles. Then, we visualized the relationship between the movies in a graph, using the cosine similarity. Finally, we evaluated the proposed method with respect to measures genre homophily and genre entropy. MoGaL succeeded to outperforms the baseline method significantly in these measures. Accordingly, our empirical results indicate that movie subtitles could be considered a rich source of informative information for various movie analysis tasks.
M. Nasiri; H. Rahmani
Abstract
Determining the personality dimensions of individuals is very important in psychological research. The most well-known example of personality dimensions is the Five-Factor Model (FFM). There are two approaches 1- Manual and 2- Automatic for determining the personality dimensions. In a manual approach, ...
Read More
Determining the personality dimensions of individuals is very important in psychological research. The most well-known example of personality dimensions is the Five-Factor Model (FFM). There are two approaches 1- Manual and 2- Automatic for determining the personality dimensions. In a manual approach, Psychologists discover these dimensions through personality questionnaires. As an automatic way, varied personal input types (textual/image/video) of people are gathered and analyzed for this purpose. In this paper, we proposed a method called DENOVA (DEep learning based on the ANOVA), which predicts FFM using deep learning based on the Analysis of variance (ANOVA) of words. For this purpose, DENOVA first applies ANOVA to select the most informative terms. Then, DENOVA employs Word2Vec to extract document embeddings. Finally, DENOVA uses Support Vector Machine (SVM), Logistic Regression, XGBoost, and Multilayer perceptron (MLP) as classifiers to predict FFM. The experimental results show that DENOVA outperforms on average, 6.91%, the state-of-the-art methods in predicting FFM with respect to accuracy.
H. Rahmani; H. Kamali; H. Shah-Hosseini
Abstract
Nowadays, a significant amount of studies are devoted to discovering important nodes in graph data. Social networks as graph data have attracted a lot of attention. There are various purposes for discovering the important nodes in social networks such as finding the leaders in them, i.e. the users who ...
Read More
Nowadays, a significant amount of studies are devoted to discovering important nodes in graph data. Social networks as graph data have attracted a lot of attention. There are various purposes for discovering the important nodes in social networks such as finding the leaders in them, i.e. the users who play an important role in promoting advertising, etc. Different criteria have been proposed in discovering important nodes in graph data. Measuring a node’s importance by a single criterion may be inefficient due to the variety of graph structures. Recently, a combination of criteria has been used in the discovery of important nodes. In this paper, we propose a system for the Discovery of Important Nodes in social networks using Genetic Algorithms (DINGA). In our proposed system, important nodes in social networks are discovered by employing a combination of eight informative criteria and their intelligent weighting. We compare our results with a manually weighted method, that uses random weightings for each criterion, in four real networks. Our method shows an average of 22% improvement in the accuracy of important nodes discovery.
