Lain Su

Research

Publications

Evolutionary Policy Optimization

ABSTRACT: A key challenge in reinforcement learning (RL) is managing the exploration-exploitation trade-off without sacrificing sample efficiency. Policy gradient (PG) methods excel in exploitation through fine-grained, gradient-based optimization but often struggle with exploration due to their focus on local search. In contrast, evolutionary computation (EC) methods excel in global exploration, but lack mechanisms for exploitation. To address these limitations, this paper proposes Evolutionary Policy Optimization (EPO), a hybrid algorithm that integrates neuroevolution with policy gradient methods for policy optimization. EPO leverages the exploration capabilities of EC and the exploitation strengths of PG, offering an efficient solution to the exploration-exploitation dilemma in RL. EPO is evaluated on the Atari Pong and Breakout benchmarks. Experimental results show that EPO improves both policy quality and sample efficiency compared to standard PG and EC methods, making it effective for tasks that require both exploration and local optimization.

GECCO 2025

Read More

Spam Detection Using Bidirectional Transformers and Machine Learning Classifier Algorithms

ABSTRACT: Spam email has accounted for a high percentage of email traffic and has created problems worldwide. The deep learning transformer model is an efficient tool in natural language processing. This study proposed an efficient spam detection approach using a pre-trained Bidirectional Encoder Representation from Transformer (BERT) and machine learning algorithms to classify ham or spam emails. Email texts were fed into the BERT and features obtained from the BERT outputs were used to represent the texts. Four classifier algorithms in machine learning were employed to classify the features of the text into ham or spam categories. The proposed model were tested using two public datasets in the experiments. The results of the evaluation metrics demonstrate the logistic regression algorithm achieved the best classification performance in both datasets. They also justified the efficient ability of the proposed model in detecting spam emails.

April 24, 2022

Read More

LSTM-Based Analysis of Company Sentiments Regarding Cryptocurrencies

ABSTRACT: This study proposes an LSTM-based model to analyze company sentiments toward cryptocurrencies. The model was trained on a dataset that was collected by downloading 10-K files of 277 companies from the SEC Edgar Database and extracting sentences containing keywords related to cryptocurrencies. It achieved 99% accuracy on the training set and 83.59% accuracy on the testing set. By analyzing the relationship between company sentiment and price, the proposed model can be used to predict the price of cryptocurrencies. This study can also be extended to determine companies' intentions in disclosing the use of cryptocurrencies in their financial statements.

May 5, 2022

Read More

Conferences

Sigma Xi Scientific Honor Society Student Research Showcase

Spam Detection Using Bidirectional Transformers and Machine Learning Classifier Algorithms

ABSTRACT: Spam emails account for 85% of all e-mail traffic and harm both individuals and institutions. A novel approach to spam detection involves transformers, deep learning models that use the self-attention mechanism to perform natural language processing (NLP) tasks. This study proposes an efficient spam email detection model based on a pre-trained Bidirectional Encoder Representation from Transformer (BERT) and machine learning algorithms to classify ham or spam emails. Email texts are fed into the BERT and features obtained from the BERT outputs are used to represent the texts. Four classifier algorithms in machine learning: support vector machine, logistic regression, random forest, and k-nearest neighbors, are employed to classify the features of the text into ham or spam categories. In the experiments, two public datasets were used to train and test the proposed model. The performance of the model was evaluated using three metrics: precision, recall, and F1 score. The experimental results demonstrate that the logistic regression algorithm achieved the best classification performance in both datasets, with 97.86% precision, 97.83% recall, and an F1 score of 97.84% on the first dataset and 95.95% precision, 96% recall, and an F1 score of 95.92% on the second dataset. The results also proved the feasibility and effectiveness of the proposed model in detecting spam emails.

April 25-May 9, 2022

UIS Student Technology, Arts & Research Symposium (STARS)

Spam Detection Using Bidirectional Transformers and Machine Learning Classifier Algorithms

ABSTRACT: Spam emails account for a high percentage of e-mail traffic and create worldwide problems. The deep learning transformer model is an efficient tool in natural language processing. This study proposed an efficient spam detection model based on a pre-trained Bidirectional Encoder Representation from Transformer (BERT) and machine learning algorithms to classify ham or spam emails. Email texts were fed into the BERT and features obtained from the BERT outputs were used to represent the texts. Four classifier algorithms in machine learning: support vector machine, logistic regression, random forest, k-nearest neighbors, were employed to classify the features of the text into ham or spam categories. In the experiments, two public datasets were used to train and test the proposed model. The results of the evaluation metrics demonstrate that the logistic regression algorithm achieved the best classification performance in both datasets. They also justified the efficient ability of the proposed model in detecting spam emails.

April 7-8, 2022

Honors Council of the Illinois Region Student Research Symposium

Textual Analysis of Companies’ Performance During the COVID-19 Pandemic

ABSTRACT: This research explores the use of text analysis to evaluate the reporting of specific topics in company financial statements. Our objective is to determine the aspects in which the COVID-19 pandemic affected the performance of companies, and companies’ intentions in reporting COVID-19 in their financial statements. First, we use a Python script to automatically download selected companies’ 10-Ks which include certain keywords relating to COVID-19. Then, we split these 10-Ks into sentences and use KH-Coder to build co-occurrence graphs, multi-dimensional scaling, and hierarchical cluster analysis to verify the association among keywords. Finally, the K-means clustering algorithm is employed to identify the ways in which COVID-19 impacted companies’ performance and their intentions in disclosing this impact.

February 16, 2022

Eaton Corporation Tech Poster Competition 2021

Application of Text Analysis to Company Financial Statements

My abstract was one of the 15 selected for the 2021 Tech Poster Competition held by Eaton Corporation.
ABSTRACT: This research explores how to use text analysis in evaluating reporting specific matters in the notes of the financial statements. Our objective is to determine the objectives of self-disclose companies of cryptocurrencies in their 10-Ks. Text data from different sources (e.g., Edgar database, and SEC) are extracted and cleaned. After identifying 10-Ks that include any aspects of cryptocurrencies, these reports are split into sentences, and preprocessing is applied into the sentences, such as removing invalid characters, short sentences. Then a pre-trained natural language processing model, FinBERT, is used to encode each word/token in sentences and to extract features on each sentence. Principal Component Analysis (PCA) is employed to select the significant features. Different clustering analysis algorithms such as K-means, fuzzy c-means are utilized to answer the question that companies self-disclose the use of cryptocurrencies in the 10-Ks. Also, we use KH-Coder to build co-occurrence graphs to verify the association among keywords. Our preliminary analysis results indicate that startup companies self-disclose cryptocurrencies in the business section and a few in the Risk Factor section in the 10-Ks.

November 19, 2021

Midwest/West Regional Undergraduate Research, Scholarly, and Creative Activity (URSCA) Virtual Conference

Application of Text Analysis to Company Financial Statements

ABSTRACT: This research explores how to use text analysis in evaluating reporting specific matters in the notes of the financial statements. Our objective is to determine the objectives of self-disclose companies of cryptocurrencies in their 10-Ks. Text data from different sources (e.g., Edgar database, and SEC) are extracted and cleaned. After identifying 10-Ks that include any aspects of cryptocurrencies, these reports are split into sentences, and preprocessing is applied into the sentences, such as removing invalid characters, short sentences. Then a pre-trained natural language processing model, FinBERT, is used to encode each word/token in sentences and to extract features on each sentence. Principal Component Analysis (PCA) is employed to select the significant features. Different clustering analysis algorithms such as K-means, fuzzy c-means are utilized to answer the question that companies self-disclose the use of cryptocurrencies in the 10-Ks. Also, we use KH-Coder to build co-occurrence graphs to verify the association among keywords. Our preliminary analysis results indicate that startup companies self-disclose cryptocurrencies in the business section and a few in the Risk Factor section in the 10-Ks.

Improved Convolutional Neural Network Model for Binary Malicious Data Flow Detection

ABSTRACT: Network intrusion detection systems can protect the traffic data in a network from the threat of network attacks. However, current research mostly uses the statistical features of traffic data to design algorithms, which heavily depend on expert experience and have low accuracy. This research proposes an improved convolutional neural network (ICNN) model to extract content features from traffic packets to detect network intrusions. The steps include binary malicious payload data generation, sample construction, ICNN model construction and training, and model evaluation. 100 samples each from 9 categories of binary malicious payloads are generated through Metasploit and preprocessed into 2-dimensional (2D) format. The model is trained and tested using 2D samples. The performance of the model is evaluated using a confusion matrix and other metrics. The experimental results indicate that the proposed model has an outstanding ability to accurately detect binary malicious data flow with high performance.

November 13, 2021

Louis Stokes Midwest Regional Center of Excellence Annual Conference

Application of Text Analysis to Company Financial Statements

ABSTRACT: This research explores how to use text analysis in evaluating reporting specific matters in the notes of the financial statements. Our objective is to determine the objectives of self-disclose companies of cryptocurrencies in their 10-Ks. Text data from different sources (e.g., Edgar database, and SEC) are extracted and cleaned. After identifying 10-Ks that include any aspects of cryptocurrencies, these reports are split into sentences, and preprocessing is applied into the sentences, such as removing invalid characters, short sentences. Then a pre-trained natural language processing model, FinBERT, is used to encode each word/token in sentences and to extract features on each sentence. Principal Component Analysis (PCA) is employed to select the significant features. Different clustering analysis algorithms such as K-means, fuzzy c-means are utilized to answer the question that companies self-disclose the use of cryptocurrencies in the 10-Ks. Also, we use KH-Coder to build co-occurrence graphs to verify the association among keywords. Our preliminary analysis results indicate that startup companies self-disclose cryptocurrencies in the business section and a few in the Risk Factor section in the 10-Ks.

Convolutional Neural Network Model for Malicious Data Flow Detection

ABSTRACT: In recent years, there has been a dramatic increase in the number of network attacks by intruders who aim to exploit the operating system or software vulnerabilities and then transmit malicious data. Network intrusion detection systems can protect the data in a network from the threat of network attacks. Nowadays, network intrusion detection research works mostly use the statistical features of traffic data to design algorithms, which heavily depend on expert experience and have low accuracy. This research proposes a convolutional neural network (CNN) model to extract the content features from traffic packets to detect network intrusions. The model includes three stages: generation of binary malicious payload data sets, data preprocessing, construction and training of the CNN model, and model evaluation. Firstly, 9 different categories of binary malicious payloads are generated through Metasploit software and each category contains 100 samples. The generated binary malicious payloads are transformed into the format required by the CNN model and preprocessed as inputs for the model. Then, the features are extracted and used to train the proposed CNN model. Finally, the malicious payload samples in the testing set are classified by the trained model, and the performance of the model is quantitatively evaluated using a confusion matrix and other metrics. The results indicate that the proposed CNN model has an outstanding ability to accurately detect network intrusions with high performance.

October 22-24, 2021