The fruit industry in Chile has achieved global recognition for its productivity and leadership in fruit exportation, being the main exporter in the Southern Hemisphere, especially of cherries, grapes, and blueberries. Agricultural automation is a growing trend aimed at reducing laborious work and the consumption of time and personnel. Advances in artificial intelligence are enabling the automation of various processes, such as fruit categorization, though there are still gaps in the precision of classifying fruits in good and bad condition, particularly when considering specialized multimodal models. This work addresses this gap by combining convolutional neural network models and the multimodal CLIP technique, evaluating the effectiveness of convolutional architectures such as ResNet50, Xception, and MobileNet. The experiments show interesting results among different architectures, with ViT-B/16 model standing out for its higher precision in this task.

This work investigates the causal correlation between the resolution of a programming guide and academic performance in an introductory programming course at Andrés Bello University in Chile. Specifically, we explore whether completing a guide comprising fifty-two exercises can predict first-year students' performance on the initial test of the course. In particular, we propose utilizing causal modelling framework to analyze and comprehend the impact of programming guides on student performance. The research encompasses a review of pertinent literature, a descriptive examination of collected data, and a discussion on both practical and theoretical implications. The findings aim to enhance strategies for student support and inform decision-making regarding the educational utility of guides

Higher education is a critical driver of social mobility and economic development globally. Access to universities and the factors influencing academic performance are essential for shaping the future workforce and ensuring equitable opportunities. In Chile, the higher education system plays a pivotal role in providing pathways for students from diverse socioeconomic backgrounds. This study focuses on the discovery of causal relationships influencing student admission to higher education in Chile, using data from the Higher Education Access Test of 2021. The research aims to identify the factors affecting academic performance and access to university by applying causal inference algorithms, specifically PC, GES, and LINGAM. These algorithms help uncover directed acyclic graphs (DAGs) from observational data, revealing the underlying causal structure among variables like socioeconomic background and academic scores. The results highlight the potential causal relationships between these factors, providing critical insights for educational policy-making. Key findings demonstrate the value of such causal models in understanding the dynamics that affect educational outcomes. Future work should explore the application of additional sensitivity analyses and broader datasets to further validate and refine these causal models.

Chile, Japan, and Mexico are among the world’s most seismically active regions, each representing a distinct subduction setting. This work extends prior studies on earthquake classification and evaluates the effectiveness of sequence-based methods across these different subduction zones. We frame the task as classifying events into foreshock, mainshock, and aftershock classes. Our pipeline comprises three stages: first, spatio-temporal clustering to group related seismic events; second, labeling each event to identify the mainshock and assign foreshock/aftershock roles; and finally, sequence classification using deep learning models, including Long Short-Term Memory (LSTM), Transformer, and Spatio-Temporal Transformer (STTN) architectures. Our results across the three regions show that the approach is effective: mainshock detection is consistently strong, aftershock classification achieves intermediate performance, and foreshocks remain the most challenging class. Among the tested architectures, the Transformer model exhibits the most consistent and competitive performance across regions. These findings underscore the effectiveness of the approach across distinct subduction zones while highlighting persistent challenges in foreshock recognition and opportunities for further improvement.

In today’s digital economy, where personalization has become a cornerstone of effective marketing strategies, companies face the dual challenge of increasing advertising impact while safeguarding sensitive customer information. Despite the rapid progress of large language models (LLMs), existing commercial solutions often neglect the integration of synthetic data to reduce privacy risks and enhance adaptability, leaving organizations dependent on external providers. To address this gap, our work fine-tunes open-source LLMs (LlaMa2, Mistral, and Zephyr) with synthetic datasets generated via GPT, aiming to produce customized marketing emails tailored to demographic and behavioral features. This thesis demonstrates not only the feasibility but also the competitiveness of such models by evaluating outputs with standard metrics (BLEU, ROUGE) and human-like scoring through GPT-4, showing that open-source models can approximate the performance of proprietary alternatives at significantly lower cost. The results confirm that fine-tuned LLMs with synthetic data represent a viable solution for enterprises seeking efficiency, personalization, and internal control of data.

This paper focuses on short-term traffic speed projections, with the goal of estimating vehicle speeds within specific time intervals using advanced machine learning models. The research proposes the use of graph-based neural models for the prediction of short-term speeds of public buses on a segment of Avenida Alameda Libertador Bernardo O'Higgins in Santiago, Chile. This paper describes the methodology employed, including data collection, preprocessing, model training, and performance evaluation. The findings and recommendations derived from the models have the potential to offer valuable information to improve traffic efficiency and urban mobility planning. The study highlights the importance of leveraging technological advances and local data to develop more accurate and efficient traffic forecasting models, ultimately benefiting public and private entities involved in transportation management.

Chile is among the world’s most seismically active countries, with an annual average of over 1,000 seismic events exceeding moment magnitude () 4.0. In the past 20 years, the country has experienced two major events surpassing 8.0. While deep neural network models have been widely employed to detect patterns in seismic data, the classification of seismic events into foreshocks, mainshocks, and aftershocks remains a challenging task. This study proposes a hybrid approach for the classification of earthquakes in Chile. The methodology comprises three main steps: first, a spatio-temporal density-based clustering algorithm is applied to group seismic events based on their spatial and temporal similarities; second, the seismic events within each cluster are labeled as foreshocks, mainshocks, or aftershocks; and finally, deep neural networks, including Long Short-Term Memory (LSTM) and Transformer models, are employed to classify earthquakes. Features such as longitude, latitude, magnitude, depth, and distances between events are used as inputs. For aftershock classification, the LSTM model achieves the highest accuracy at 0.8. Meanwhile, for precursor event classification, the Transformer network outperforms the LSTM, achieving a recall of 0.6. Future work will focus on a more detailed exploration of the precursor class and the incorporation of additional seismic data from other countries to enhance the model’s generalization.

Sales prediction is crucial for business intelligence, aiding in workforce management or resource allocation. Accurate sales forecasting is vital for financial planning and predicting both short-term and long-term company performance. In this work, we propose the use of adaptive ensembles of classification models to accommodate different trends within the data, unlike typically used machine learning models. Our approach is based on a Mixture of Experts (MoE) model using LSTM networks, with block cross-validation. We compare our proposal to various standard models in prediction tasks. Experiments show that our model achieves greater generalization on unseen stores compared to other models. As future work, we plan to extend this model to Transformer models.

The rapid growth of e-commerce and the increasing need for logistical optimization in highly congested urban environments require advanced models for vehicle speed prediction. Traditional models often overlook the influence of the geographic environment and rely solely on historical speed data, limiting their accuracy in dynamic scenarios. In addition, most approaches use square grid structures, which introduce spatial distortions and fail to capture the connectivity of road networks effectively. In this work, we propose a multimodal model that integrates spatio-temporal information from GPS sensors with satellite imagery, leveraging HexConvLSTM and MLP neural networks to enhance predictive robustness. Unlike conventional methods, our approach utilizes a hexagonal grid representation, which provides a more uniform spatial structure and improved neighborhood representation that aligns better with road topology than conventional square grids for modeling multidirectional traffic dynamics. This paper presents the implementation and evaluation of the model, highlighting its effectiveness in improving the accuracy of route planning for freight transportation in Santiago Centro. The results show that the multimodal approach significantly reduces the mean absolute error (MAE) to 2.296 in test dataset, outperforming a baseline model based solely on spatiotemporal data by 8.3%. This research validates the benefits of incorporating visual data and hexagonal grid-based spatial modeling into traffic prediction and suggests exploring its applicability in other urban settings.

Recognizing variable stars is a task of interest in the astronomy community. Currently, this task has taken advantage of deep learning algorithms. However, these algorithms require a large amount of data to achieve high levels of precision. In this work, self-supervised learning is proposed to improve the classification of variable stars considering a reduced amount of data using recurrent networks. The experiments in Gaia dataset show that the proposed approach allows to improve performance, when compared with traditional initialization schemes, up to 7% and 13% in real databases in semi-supervised learning scenarios. In future work, we propose considering experiments with other variable star databases.