IAES International Journal of Artificial Intelligence (IJ-AI)
Not a member yet
1769 research outputs found
Sort by
Revolutionizing recommendations a survey: a comprehensive exploration of modern recommender systems
Get PDFThe rapid proliferation of digital information and online services has fundamentally reshaped user interactions with websites, necessitating the evolution of recommender systems. These systems, crucial in domains such as e-commerce, education, and scientific research, serve to enhance user engagement and satisfaction through personalized recommendations. However, it comes up with new challenges, including information overload, prompting the development of recommender systems that can efficiently navigate this vast group to offer more personalized and relevant suggestions. This survey paper explores the dynamic opinion of recommendation systems, addressing the limitations of traditional approaches, the emergence of deep learning models, and the extended potential for additional data. By investigating various recommendation systems and the evolving role of deep learning, this paper illuminates the path toward more accurate, personalized, and effective recommender systems, considering challenges like sparse data and improved context-based recommendations. The study encompasses three primary recommendation approaches: collaborative filtering, content-based filtering, and hybrid systems. It further investigates into the transformation brought about by deep learning models, showcasing how these models intricate user-item interactions. This survey offers a comprehensive exploration of recommendation systems and their advancements in the digital era, providing insights into the future of personalized content delivery
Interpretable machine learning for academic risk analysis in university students
Get PDFHigher education institutions often grapple with issues related to academic risk among their students. These academic risks encompass low academic performance, study delays, and dropouts. One approach to address these challenges is to predict students’ academic performance as accurately as possible by leveraging advanced computational techniques and utilizing academic and non-academic student data. This research aims to develop a model that accurately identifies students with high potential for academic risk while explaining the contributing factors to this phenomenon in the Faculty of Vocational Studies, Institut Teknologi Sepuluh Nopember (ITS). The prediction model is constructed using the light gradient boosting machine (LightGBM) method and is subsequently interpreted using the Shapley additive explanations (SHAP) value. Additionally, an oversampling method, based on synthetic minority oversampling technique (SMOTE), is implemented to address imbalances in the dataset. The proposed approach achieves 96% and 97% accuracy and specificity rates, respectively. Analysis based on SHAP values reveals that extracurricular activities, choice of major, smoking habit, gender, and friendship circle are among the top five factors impacting students’ academic risk
FaceSynth: text-to-face generation using CLIP and its variants with generative adversarial networks
Get PDFIn recent years, there have been massive developments in the field of generative AI, especially in generative adversarial networks (GANs). GANs generate original images that haven't been seen during training and have had several advancements like StyleGAN, StyleGAN2, and StyleGAN2-adaptive discriminator augmentation (ADA). Contrastive language-image pre-training (CLIP), by OpenAI, is a visual linguistic model that has been trained to associate texts with images. Recently, new CLIP variants were developed, such as metadata-curated language-image pre-training (MetaCLIP), released by Facebook and trained on a larger dataset, and Multilinigual-CLIP, which adapts CLIP to multiple languages. We compare CLIP and its variants in text-to-face synthesis with a custom StyleGAN2-ADA model and a pre-trained StyleGAN2 model. Our training-free algorithm starts with an initial image latent code that is iteratively manipulated to match a given text description. It achieves this by minimizing the distance between the text and image embedding in the multi-modal embedding space of the CLIP models. An examination of CLIP and its variants showed that MetaCLIP outperformed its competitors in LPIPS similarity and closeness of the synthesized image to the actual prompt. CLIP produced the most realistic images with the best FID score and multilingual-CLIP presented a choice of input text language and generated decent images
Application of self-organizing map for modeling the Aquilaria malaccensis oil using chemical compound
Get PDFAgarwood oil, known as ‘black gold’ or the ‘wood of God,’ is a globally prized essential oil derived naturally from the Aquilaria tree. Despite its significance, the current non-standardized grading system varies worldwide, relying on subjective assessments. This paper addresses the need for a consistent classification model by presenting an overview of Aquilaria malaccensis oil quality using the self-organizing map (SOM) algorithm. Derived from the Thymelaeaceae family, Aquilaria malaccensis is a primary source of agarwood trees in the Malay Archipelago. Agarwood oil extraction involves traditional methods like solvent extraction and hydro-distillation, yielding a complex mixture of chromone derivatives, oxygenated sesquiterpenes, and sesquiterpene hydrocarbons. This study categorizes agarwood oil into high and low grades based on chemical compounds, utilizing the SOM algorithm with inputs of three specific compounds: β-agarofuran, α-agarofuran, and 10-epi-φ-eudesmol. Findings demonstrate the efficacy of SOM-based quality grading in distinguishing agarwood oil grades, offering a significant contribution to the field. The non-standardized grading system's inefficiency and subjectivity underscore the necessity for a standardized model, making this research crucial for the agarwood industry's advancement
Non-small cell lung cancer active compounds discovery holding on protein expression using machine learning models
Get PDFComputational methods have transformed the field of drug discovery, which significantly helped in the development of new treatments. Nowadays, researchers are exploring a wide ranger of opportunities to identify new compounds using machine learning. We conducted a comparative study between multiple models capable of predicting compounds to target non-small cell lung cancer, we focused on integrating protein expressions to identify potential compounds that exhibit a high efficacy in targeting lung cancer cells. A dataset was constructed based on the trials available in the ChEMBL database. Then, molecular descriptors were calculated to extract structure-activity relationships from the selected compounds and feed into several machine learning models to learn from. We compared the performance of various algorithms. The multilayer perceptron model exhibited the highest F1 score, achieving an outstanding value of 0,861. Moreover, we present a list of 10 drugs predicted as active in lung cancer, all of which are supported by relevant scientific evidence in the medical literature. Our study showcases the potential of combining protein expression analysis and machine learning techniques to identify novel drugs. Our analytical approach contributes to the drug discovery pipeline, and opens new opportunities to explore and identify new targeted therapies
Designing an intelligent system for vibration diagnosis of centrifugal water-cooling pumps using Bayesian networks
Get PDFImplementing monitoring methods is a viable method to reduce substantial damage to cooling water centrifugal pumps. Engaging in manual vibration analysis requires considerable time and a requisite level of competence. Small datasets pose challenges when applying classification systems that utilize linear classification models and deep learning. Given these issues, our proposal entails developing a system capable of autonomously, precisely, and accurately diagnosing vibrations using a limited dataset. The system is anticipated to possess the capability to detect multiple categories of mechanical defects, such as static imbalance, dynamic imbalance, misalignment, cavitation, looseness, and bearing corrosion. The Bayesian network (BN) structure was constructed using the MATLAB software. The input data parameters comprise vibration signals measured in the frequency domain and values representing phase differences. The constructed intelligent system was subsequently assessed using a dataset including 120 samples. The smart system can rapidly anticipate and precisely identify every form of harm with exceptional accuracy and sensitivity, relying on test outcomes. The test data analysis reveals that the intelligent system attained an average accuracy of 94.74%, precision of 95.32%, sensitivity (recall) of 93.67%, and F-score of 94.36%.
Investigation on low-performance tuned-regressor of inhibitory concentration targeting the SARS-CoV-2 polyprotein 1ab
Get PDFHyperparameter tuning is a key optimization strategy in machine learning (ML), often used with GridSearchCV to find optimal hyperparameter combinations. This study aimed to predict the half-maximal inhibitory concentration (IC50) of small molecules targeting the SARS-CoV-2 replicase polyprotein 1ab (pp1ab) by optimizing three ML algorithms: histogram gradient boosting regressor (HGBR), light gradient boosting regressor (LGBR), and random forest regressor (RFR). Bioactivity data, including duplicates, were processed using three approaches: untreated, aggregation of quantitative bioactivity, and duplicate removal. Molecular features were encoded using twelve types of molecular fingerprints. To optimize the models, hyperparameter tuning with GridSearchCV was applied across a broad parameter space. The results showed that the performance of the models was inconsistent, despite comprehensive hyperparameter tuning. Further analysis showed that the distribution of Murcko fragments was uneven between the training and testing datasets. Key fragments were underrepresented in the testing phase, leading to a mismatch in model predictions. The study demonstrates that hyperparameter tuning alone may not be sufficient to achieve high predictive performance when the distribution of molecular fragments is unbalanced between training and testing datasets. Ensuring fragment diversity across datasets is crucial for improving model reliability in drug discovery applications
Domain-specific knowledge and context in large language models: challenges, concerns, and solutions
Get PDFLarge language models (LLMs) are ubiquitous today with major usage in the fields of industry, research, and academia. LLMs involve unsupervised learning with large natural language data, obtained mostly from the internet. There are several challenges that arise because of these data sources. One such challenge is with respect to domain-specific knowledge and context. This paper deals with the major challenges faced by LLMs due to data sources, such as, lack of domain expertise, understanding specialized terminology, contextual understanding, data bias, and the limitations of transfer learning. This paper also discusses some solutions for the mitigation of these challenges such as pre-training LLMs on domain-specific corpora, expert annotations, improving transformer models with enhanced attention mechanisms, memory-augmented models, context-aware loss functions, balanced datasets, and the use of knowledge distillation techniques
A survey of missing data imputation techniques: statistical methods, machine learning models, and GAN-based approaches
Get PDFEfficiently addressing missing data is critical in data analysis across diverse domains. This study evaluates traditional statistical, machine learning, and generative adversarial network (GAN)-based imputation methods, emphasizing their strengths, limitations, and applicability to different data types and missing data mechanisms (missing completely at random (MCAR), missing at random (MAR), missing not at random (MNAR)). GAN-based models, including generative adversarial imputation network (GAIN), view imputation generative adversarial network (VIGAN), and SolarGAN, are highlighted for their adaptability and effectiveness in handling complex datasets, such as images and time series. Despite challenges like computational demands, GANs outperform conventional methods in capturing non-linear dependencies. Future work includes optimizing GAN architectures for broader data types and exploring hybrid models to enhance imputation accuracy and scalability in real-world applications
Pre-trained convolutional neural network-based algorithms: application for recognizing the age category
Get PDFCybercrime is a major issue in the current digital era, with one of its branches-cyber pornography-notably affecting Indonesia. Various efforts have been made to suppress or prevent this problem. One alternative solution involves using technological advances to recognize age ranges based on facial recognition. This age range recognition can be implemented to prevent users from accessing content that is not appropriate for their age. An optimal age-range recognition system is essential for this purpose. However, limited research has focused on this domain. Therefore, our research aimed to develop the best possible system. The proposed method applies a trained convolutional neural network (CNN) as a feature extractor to the artificial neural network (ANN) and k-nearest neighbor (K-NN) methods for age recognition based on facial images. By incorporating computational learning techniques, the system's performance is significantly enhanced, leveraging advanced algorithms to improve accuracy. The test results show that the performance of the pre-trained CNN-based ANN model is superior. This is indicated by the model's accuracy and F1-score, which were 11% and 0.11 higher, than the pre-trained CNN-based K-NN model. The error rate of the pre-trained CNN-based ANN model was also reduced by 0.11