Muhammad Kashif Hanif

Accurate crop yield prediction is important to minimize uncertainty for informed decision- making and resource allocation. A variety of machine learning models are used in yield prediction; however, the available benchmarking literature offers limited insight to achieve a balance between predictive accuracy and model interpretability of different models. Therefore, this study was conducted to evaluates popular machine learning models for U.S. soybean yield prediction using a multi-source spatiotemporal dataset comprising weather, soil, and management features. The model performance was evaluated using root mean squared error (RMSE) metric, and feature impact was explained using Shapley Additive Explanations (SHAP) for interpretability. The findings indicate that Random Forest is the best model that achieved least RMSE of 5.07 and highest correlation coefficient of 90.36% on test set. SHAP results revealed that precipitation and solar radiation are leading yield determinants, while soil properties, such as soil pH and bulk density, exerted moderate effects. The contribution of this work is fourfold: (i) a rigorous benchmarking of ML models using accuracy metrics for yield prediction, (ii) evidence based affirming the model superiority for complex agronomic dataset, (iii) systematic assessment of global feature importance connecting yield affecting climatic and edaphic factors, and (iv) application of SHAP as a means for interpretation and explainability. The results bring together predictive performance and explanation, providing insights into the advancement of smart agriculture through informed decision-making for irrigation planning, efficient input application, and climate-resilient strategy formulation. Key words: Crop Yield Prediction, Informed decision-making, Machine Learning, Smart Agriculture, SHAP Interpretability, Explainable AI

Enzymes play a critical role in catalyzing biochemical reactions, and accurately predicting their substrate specificity is essential for applications in drug discovery, metabolic engineering, and biotechnology. However, traditional experimental methods are time-consuming and resource-intensive, while existing computational approaches often rely on predicting Enzyme Commission (EC) numbers as coarse proxies for substrate specificity, failing to capture the many-to-many relationships between enzymes and their substrates. In this study, we present a hybrid deep learning framework for prediction of enzyme-substrate interactions. The proposed model integrates ProtT5 protein language model embeddings with extended connectivity fingerprints (ECFP4) and 3D molecular descriptors through a dedicated geometric attention mechanism that dynamically weights features based on interaction patterns. The key findings demonstrate that proposed approach achieves state-of-the-art performance, with an overall accuracy of 88.2% and F1-score of 0.823 on the BRENDA benchmark dataset. The model shows particular strength in predicting interactions for promiscuous enzymes (18.7% of the dataset), improving recall by 22% compared to EC-based prediction methods. The hybrid attention mechanism provides a 9.2% AUC improvement over static feature concatenation. The proposed methodology can be extended to other multi-label classification problems in biological systems, paving the way for future advancements in computational biology.

Time series analysis is a critical task across various scientific and industrial domains, enabling the extraction of valuable insights from temporal data. High dimensionality of time series data can lead to computational inefficiencies and increased complexity. Dimensionality reduction techniques play a crucial role in handling the high-dimensional nature of time series data essential information while reducing computational complexity. This study explores the impact of various dimensionality reduction techniques and machine learning models on enhancing the accuracy and efficiency of time series datasets. The effectiveness of different dimensionality reduction methods is evaluated based on their ability to simplify data while preserving essential features. Subsequently, several machine learning models—such as Autoregressive Integrated Moving Average, k-Nearest Neighbors, Random Forest, Least Absolute Shrinkage and Selection Operator, and Extreme Gradient Boosting are applied to the transformed data. The findings reveal that the choice of dimensionality reduction technique significantly influences the performance of these models. Certain methods excel at uncovering underlying patterns and improving predictive accuracy, while others offer computational advantages. These results highlight the importance of selecting an appropriate combination of dimensionality reduction and machine learning techniques based on the specific characteristics of the time series data. This can contributes to a deeper understanding of how these methods can be effectively integrated, thereby enhancing decision-making in areas such as finance, meteorology, and operational planning.

The diverse types of fake text generation practices by spammer make spam detection challenging. Existing works use manually designed discrete textual or behavior features, which cannot capture complex global semantics of text and reviews. Some studies use limited features while neglecting other significant features. However, in case of a large number of features set, the selection of all features leads to overfitting the model and expensive computation. The problem statement of this research paper revolves around addressing challenges concerning feature selection and evolving spammer behavior and linguistic features, with the goal of devising an efficient model for spam detection. The primary objective of this endeavor was to identify the most efficacious subset of features and patterns for the task of spam detection. Spammer behavior features and linguistic features often exhibit complex relationships that influence the nature of spam reviews. The unified representation of features is another challenging task in spam detection. Various deep learning approaches have been proposed for spam detection and classification but these methods are specialized in extracting the features but lack to capture feature dependencies effectively with other features but there is a lack of comprehensive models that integrate linguistic and behavioral features to improve the accuracy of spam detection. The proposed spam detection framework SD-FSL-CLSTM used the fusion of spammer behavior features and linguistic features which automatically detect and classify the spam reviews. Fusion enables the proposed model to automatically learn the interactions between the features during the training process, allowing it to capture complex relationships and make predictions based on both types of features. SD-FSL-CLSTM framework apparently shows the promising result by obtaining a minimum accuracy 97%.

In today’s digital world, social media platforms generate a plethora of unstructured information. However, for low-resource languages like Urdu, there is a scarcity of well-structured data for specific tasks such as event classification. Urdu, a language prominent in South Asia, has boasted a complex morphological structure with unique features but has lacked standard linguistic resources like datasets. Long-text classification has demanded more effort than short-text classification due to its expansive vocabulary, information redundancy, and noise. Text processing has been the latest trend in research, with many machine learning and deep learning techniques widely used for it. Multiclass classification has been utilized to classify different languages for various purposes. In this research, a multiclass classification for the Urdu language was performed using a text dataset taken from five different social media platforms including Geo News, Samaa News, Dawn News, Express News, and Urdu Blogs totaling 103,771 sentences. We used sentence-level classification to categorize sentences including terrorist attacks, national news, sports, entertainment, politics, safety, earthquakes, fraud and corruption, sexual assault, weather, accidents, forces, inflation, murder and death, education, and international news. Deep learning, transformer-based and machine learning classifiers are used for event classification. The SMFCNN classifier achieved the greatest accuracy of 88.29%. We incorporated transformer-based models, with the proposed XLM-R+ model demonstrating superior performance with an accuracy of 89.8%. Our results were compared to previously reported techniques that used traditional models, highlighting the significant improvements offered by our approaches. The novelty of this research lies in the inclusion of 16 event categories to broaden coverage and the implementation of the SMFCNN and transformer-based algorithms. This study highlights the potential of deep learning and transformer-based models in enhancing the accuracy and generalizability of multiclass classification in low-resource languages Urdu.

One of the biggest threats to international trade, journalism, and democracy is fake news on social media, which also causes substantial collateral damage. The manipulation of digital news to spread misinformation has become a common practice for personal benefits or relief. Therefore, developing an automated system that can detect fake news before publication is crucial. This study proposes a three-level methodology with a new model called Multi-Kernel Optimized Convolutional Neural Network (MOCNN) to investigate its effectiveness for fake news detection. The parameters of the proposed model have been optimized using the grid search technique. We evaluated ten different deep learning models on two benchmark datasets of fake news articles and compared their performance with the proposed model. Finding the best model with good accuracy performance is the primary goal of this paper. F-measure and accuracy are used to evaluate and compare the classification performance of these deep learning models. Our proposed model achieves 85.8% and 68.2% accuracy and 85.8% and 67.7% F1-measure on UFN and BET, respectively. Experimental results confirm that the proposed model performs better than other models on UFN and BET datasets.

Aspect-context sentiment classification aims to classify the sentiments about an aspect that corresponds to its context. Typically, machine learning models considers the aspect and context separately. They do not execute the aspect and context in parallel. To model the contexts and aspects separately, most of the methods with attention mechanisms typically employ the Long Short Term Memory network approach. Attention mechanisms, on the other hand, take this into account and compute the parallel sequencing of the aspects-context. The interactive attention mechanism extracts features of a specific aspect regarding its context in the sequence, which means aspects are considered when generating context sequence representations. However, when determining the relationship between words in a sentence, the interactive attention mechanism does not consider semantic dependency information. Moreover, the attention mechanisms did not capture the polysemous words. Normally conventional embedding models, such as GloVe word vectors, have been used . In this study, transformers are embedded into the attention mechanism approaches to overcome the semantic relationship problem. For this reason, the BERT pre-train language model is used to capture the relationship among the words in a sentence. The interactive attention mechanism is then applied to the model’s distribution of that word. The final sequence-to-sequence representation in terms of context and aspect is used into general machine learning classifiers for aspect-level sentiment classification. The proposed model was evaluated on the two datasets, i.e., Restaurant and Laptop review. The proposed approach has state-of-the-art results with all attention mechanisms and attained significantly better performance than the existing ones.