Enhancing Post-Harvest Quality Prediction Models: A Synergistic Approach Integrating Temporal Boosting for Improved Performance with New Season's Datasets
Enhancing Post-Harvest Quality Prediction Models: A Synergistic Approach Integrating Temporal Boosting for Improved Performance with New Season's Datasets
Tamar Holder, Tel-Aviv University Advisors: Dr. Noam Koenigstein (TAU), Dr. Yael Salzer (Volcani Institute)
Abstract:
Approximately $162 billion worth of food goes uneaten yearly. Almost 30% of food loss takes place during post-harvest, processing, packaging, and distribution stages. This significant loss highlights the urgent need for innovative solutions within the industry. The current industry practice employs a "First In, First Out" (FIFO) logistic management approach, which may not be optimal for perishable goods. Recognizing the limitations of FIFO, transitioning to a "First Expired, First Out" (FEFO) strategy to minimize food loss has previously been proposed. To address this, previous efforts utilized recent years data sets to construct traditional machine learning models, with the aim of predicting the shelf life of produce. However, while the traditional models performed well when trained on randomly split data, gathered over the duration of one season, they did not perform well when making blind predictions for new unseen seasons. Our research embraces a more realistic scenario in which the test set comprises the most recent chronological data points available. This is vital for addressing real-world challenges in predicting future outcomes based on historical data. The present research strives to address the time-related constraint; Leveraging the power of boosting, where models sequentially learn and correct errors, Temporal Boosting introduces a temporal dimension. Each model aims to predict and refine the chronological errors of the previous model, leading to a progressively more accurate prediction over time. These methods were applied to large-scale datasets acquired over two seasons of high-throughput phenotyping analysis of the effects of pre-harvest and post-harvest features on the shelf-life quality of 'Orri' Mandarin. The research compared two baseline configurations of Linear Regression models to a Linear Regression with an implementation of the Temporal Boosting configuration. Results suggest that applying Temporal Boosting improved new unseen season prediction. Focusing on enhancing the models' capacity to forecast future data points, we aim to contribute to developing more potent and applicable solutions for mitigating food loss in the industry.
Bio:
Last year student in the “Direct Route for IE MSc” program, finished my BSc in Industrial Engineering last year, while simultaneously starting my MSc in Industrial Engineering, focused on Data Science. Before starting the program, I worked in data-oriented companies; Riskified and Intel’s AI Solutions Group – both helped me make the decision of going after a career in Data Science and pursuing my MSc degree in order to do so. During my thesis, I had the opportunity to work in collaboration with Volcani Institute, thus incorporating Machine Learning in the field of Agricultural Engineering and I was even given the opportunity to present my work in AgEng-2024 conference.
Contact:
• E-Mail: tamarholder31@gmail.com
• Linkedin: https://www.linkedin.com/in/tamar-holder/
