Comparative Analysis of AI Models in Managing Household Food Waste: OpenAI GPT-4, Google Gemini, Mistral, and Anthropic Claude

Comparative Analysis of AI Models in Managing Household Food Waste: OpenAI GPT-4, Google Gemini, Mistral, and Anthropic Claude

7ª Conferência Internacional de Sustentabilidade e Inovação

Organization:
Associação Portuguesa para o Empreendedorismo (EMPREEND) e a Universidade Estadual Paulista – UNESP, Campus Jaboticabal, São Paulo – Brasil

Transcript

1. 7ª Observatório 2º IIC Comparative Analysis of AI Models in Managing Household Food Waste: OpenAI GPT-4, Google Gemini, Mistral, and Anthropic Claude Ezequiel França dos Santos¹, Joaquim José de Almeida Soares Gonçalves² ¹ IADE, European University of Lisbon, Lisbon, Portugal. ² Polytechnic Institute of Cávado and Ave (IPCA), Barcelos, Portugal. 2024 Conferência Internacional de Sustentabilidade e Inovação
2. 7ª Observatório 2º IIC Introdução Objetivos Questão problema Revisão Teórica Metodologia Conclusão Análise e Discussão de Resultados Estrutura:
3. 7ª Observatório 2º IIC Food waste is a significant global challenge, with approximately one-third of food produced for human consumption (1.3 billion tons) wasted annually (Vilariño, Franco, & Quarrington, 2017; Nordin et al., 2020; Ishangulyyev, Kim, & Lee, 2019). Recent research highlights the potential of artificial intelligence (AI) and machine learning (ML) in addressing food waste and enhancing sustainability in the food industry. AI technologies can optimize food production, supply chains, and household food management (Onyeaka et al., 2023; Kler et al., 2022). Intelligent applications, such as smart packaging and appliances, can improve food storage and reduce waste (Liegeard & Manning, 2020). Introdução
4. 7ª Observatório 2º IIC This paper compares the performance of four leading AI models, OpenAI’s GPT-4, Google’s Gemini, Mistral, and Anthropic Claude, in managing household food waste. Specifically, we evaluate their ability to classify food items into appropriate food loss groups, provide accurate compositions of these items, and suggest relevant recipes. The aim is to understand each model’s strengths and limitations and identify areas for improvement in AI-assisted food waste management. Objetivos
5. 7ª Observatório 2º IIC • Which AI model is most effective in managing household food waste? Questão problema
6. 7ª Observatório 2º IIC The problem of food waste has been extensively studied in the literature. According to recent reports, a significant portion of global food is wasted at various stages, including production, distribution, and consumption (Food and Agriculture Organization [FAO], 2011; Gustavsson et al., 2011). Household food waste, in particular, contributes substantially to the overall food wastage, necessitating effective interventions at the consumer level. Various strategies have been proposed to address food waste, including improving food storage practices, increasing consumer awareness, and enhancing supply chain efficiencies. Recently, AI technologies have emerged as powerful tools for tackling food waste. AI can assist in predicting food spoilage, optimizing inventory management, and providing creative solutions for using leftover ingredients (Parfitt et al., 2010; Papargyropoulou et al., 2014). Revisão Teórica
7. 7ª Observatório 2º IIC Data was collected using a dataset of household ingredients categorized by avoidable and unavoidable waste (Secmeler & Ekinci, 2021). The tools used included Python and APIs for the GPT-4, Gemini, Mistral, and Claude models. Each model was queried to classify ingredients, generate compositions, and suggest recipes. Evaluation metrics consisted of ROUGE and BLEU scores for quality assessment, accuracy for classification, and statistical tests (ANOVA, t-tests) to compare results. Metodologia
8. 7ª Observatório 2º IIC • Classification Accuracy: Claude had the highest accuracy (75%), followed by Mistral (62.5%), GPT-4 (56.25%), and Gemini (31.25%) (Table 2). • Composition Quality: GPT-4 achieved the best ROUGE and BLEU scores, showing high-quality content generation (Table 3). • Recipe Suggestions: GPT-4 aligned closest with reference recipes, demonstrating strong text generation capabilities. • Statistical Analysis: Significant performance differences among models (p < 0.05) were found, supporting these rankings. Análise e Discussão de Resultados
9. 7ª Observatório 2º IIC Análise e Discussão de Resultados
10. 7ª Observatório 2º IIC The compositions provided by the models were evaluated against the reference compositions. The ROUGE and BLEU scores were computed to measure the similarity and fluency of the generated compositions. As shown in the Table 3, GPT-4 outperformed the other models in both ROUGE and BLEU scores, indicating that GPT-4 generated compositions that were more similar to the reference compositions. Análise e Discussão de Resultados
11. 7ª Observatório 2º IIC The quality of the recipe suggestions was also evaluated using ROUGE and BLEU scores. These metrics provide insights into how closely the model-generated recipes matched the reference recipes. The Table 4 shows that GPT-4 achieved higher ROUGE and BLEU scores than the other models, indicating that GPT-4 generated suggestions more aligned with the reference recipes. Análise e Discussão de Resultados
12. 7ª Observatório 2º IIC A paired sample t-test was conducted to compare the classification accuracy between GPT-4, Gemini, Mistral, and Claude to understand the performance differences further. The results indicated a statistically significant difference p < 0.05. Classification accuracy for all models is depicted in the figure below, with Claude demonstrating higher accuracy in classifying food loss groups. Análise e Discussão de Resultados
13. 7ª Observatório 2º IIC This study evaluated the performance of OpenAI's GPT-4, Google's Gemini, Mistral, and Anthropic Claude in managing household food waste by classifying ingredients, providing their compositions, and suggesting recipes. The results indicated that Claude outperformed the other models in classification accuracy, followed by Mistral, GPT-4, and Gemini. GPT-4 achieved the highest ROUGE and BLEU scores for composition and recipe suggestions, indicating its superior text generation capabilities. The statistical analysis confirmed significant differences in performance among the models. These findings highlight the strengths of each model and suggest areas for improvement in AI- assisted food waste management, contributing to sustainable food practices and reducing environmental impacts. Future work will explore enhancing the models' accuracy and robustness, integrating additional datasets, and developing more comprehensive AI-based solutions for household food waste management. Conclusão