Prompting Session-Based Context For Personalized Meal Recommendations: A Generative Model Perspective With Insights From Chain-of-Recommendations

In information-seeking perspectives, users’ intermediate interactions reflect rapidly changing intents, as seen in session-based recommender systems (SBRS). These systems adapt to intermediate recommendations but exceptionally result in context misalignments. This paper leverages the reasoning of generative models, with Chain-of-Thought (CoT), in Chain-of-Recommendations (CoR) Prompts with the goal to understand and synthesize sequential user interactions, contextual data, and cross-session dependencies to generate coherent and personalized meal recommendation sequences. Further, we evaluate the efficacy of generative models, Llama and Gemini, in contextual SBRS under the ’CoR’ framework. It focuses on integrating nutritional assessment, context awareness (e.g., weekly meal synthesis for users with specific diseases), and alternative food suggestions that meet mineral thresholds. Using metrics like RMSE, MAE, and MSE, the study assesses model performance across contexts, Prompts, session complexities, and chaining scenarios. Results demonstrate that Gemini outperforms Llama, achieving higher compliance in meal plan personalization.