An AI-Powered Personalized Nutrition Recommendation System using Ensemble Machine Learning with Hybrid Scoring and Interactive Deployment

An AI-Powered Personalized Nutrition Recommendation System using Ensemble Machine Learning with Hybrid Scoring and Interactive Deployment

Dr. C. Srinivas, NC Aswin Sai, Md Faizan Ahmed, Asireddy Bhavana, Thota Ankitha

Department of Computer Science and Engineering, Kakatiya Institute of Technology and Science, Warangal, India

Abstract – Diet-related non-communicable diseases -obesity, ype 2 diabetes, hypertensionarent just statistics. They represent a genuine global health crisis, and the tools we have to address them havent really kept pace. Most nutrition apps still rely on the same rigid, rule-based logic they used a decade ago, offering advice that barely accounts for who the user actually is.

This paper describes a personalized nutrition recommendation system built around an ensemble of three machine learning classifiers: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and a Multi-Layer Perceptron (MLP). Rather than depending on any single model, all three vote together by averaging their predicted probabilities, and that ML signal is blended with nutritional quality and user preference into a single composite ranking score:

AI Score = 0.6 × NutrientScore + 0.4 ×

PreferenceScore + 5 × P

User attributes such as age, gender, BMI, health condition, dietary restrictions, cuisine preference, and taste and the detailed nutritional profiles for food items are the inputs for the meal list that is recommended by the system through a real-time streaming interface. The ensemble model reached 91.3% accuracy in classifying correct answers and was able to outperform the standalone models of Random Forest (RF), XGBoost, and Multi-layer Perceptron (MLP) that got 88.4%, 89.7%, and 87.9% respectively. In the recommendation evaluation, 84% of users were able to find a health match on the meal in their top-3 results, dietary restriction compliance was at 93%, and the system was able to respond in 2 seconds. The user interface was rated 4.6 out of 5 by a 25-person usability test.

Index TermsPersonalized nutrition; Ensemble machine learning; Diet recommendation system; Random Forest; XGBoost; Multi-Layer Perceptron; Hybrid AI scoring;

Streamlit; Health informatics; Precision nutrition; Collaborative filtering; Nutritional deficiency detection

1. INTRODUCTION

   Poor diet is quietly one of the deadliest problems in modern medicine. The WHO identifies inadequate dietary intake as a primary modifiable risk factor for obesity, type 2 diabetes, cardiovascular disease, and certain cancers [1]. And yet, despite how serious this problem is, most people still dont have access to nutrition guidance thats genuinely tailored to them.

   The tools that existwhether paper meal plans or popular apps like MyFitnessPal and HealthifyMelargely prescribe advice based on fixed caloric thresholds. They dont adapt to individual metabolic differences, they largely ignore personal taste, and they treat every user with the same health condition as essentially interchangeable. Thats a fundamental mismatch between the complexity of human biology and the bluntness of the tool.

   Machine learning changes that picture considerably. Modern ML algorithms can pick up on non-linear, high- dimensional patterns across user characteristics, food properties, and health outcomespatterns that no hand- crafted rule system could plausibly encode. Ensemble methods, which pool predictions from multiple models, add another layer of reliability: they tend to generalize better and wobble less on unfamiliar inputs compared to single-model approaches.

   We built a system that puts these ideas into practicea fully deployed web application that takes detailed user inputs and cross-references them against a curated nutritional dataset to serve up a real-time ranked list of ten meal recommendations through an intuitive Streamlit interface. The key contributions of this work are:

   Final_dataset.csv is a multi-source unified dataset with user profiles, food nutritional metadata, recipe attributes, and preference/nutrient scores all compressed in one.

   • A three-model ensemble (RF + XGBoost + MLP) with a density of 91.3% significantly surpasses all of the individual models, therefore, it is the best option available.

   • A hybrid AI Score is a combination of nutritional quality, personal preference, and ML-predicted suitability in one composite ranking metric.

   • A heuristic that governs the diet-type entries through the automated keyword-based method correcting itellectual errors, bringing the dataset label credibility to the next level.

   • The Streamlit application has been successfully deployed and is now real-time, featuring interactive filtering and nutritional analytics with response times of less than 2 seconds.

   This paper has the following structure: Section II is the literature review; Section III states the project’s specific gaps; Section IV suggests the proposed system architecture; Section V describes the methodology step by step; Section VI contains implementation; Section VII shows and discusses the results; Section VIII concludes with the future direction of the work.

2. LITERATURE REVIEW

   Paradigm shift in recommender system of personalized food products is a marriage of recommender systems, nutritional science, and user-centered design. The last ten years have seen a big shift in the sector: progressing from simple rule-based filters through hybrid architectures with complex structure to the implementation of deep learning.

   1. The First Rule-Based and Content-Based Systems

      The original online nutrition tools almost exclusively recommended foods according to the fixed clinical guidelines as set out by the World Health Organization, the tables of daily allowances and not much else to speak of, still and yet. They were very transparent and simple to use but now they are considered outdated as they lack the adaptability. In fact, making a parallel between a 25-year-old marathon runner applying the same rules as a 60-year-old with high blood pressure does not reflect the personalization, instead, it shows mere filtering. Some studies have been published that evaluate the effectiveness of the content-based pure approach and state that the true content-based, approaches have been mostly replaced by hybrid architectures that now dominate [2].

   2. Collaborative and Hybrid Filtering

      Collaborative filtering uses user-item interaction matrices for finding the preferences of users as well as the items, while

      hybrid systems simultaneously effectuate the collaborative and content-based signals for balancing the health-related aspect with personal desirability. The implementation of this mechanism is well-seen in the Yum-Me platform, where the site uses a fun filled visual quiz as a tool to gather user taste preferences which after merging them with nutrient constraints it allows users to recommend tailored recipes- Acceptance level marks a jump of approximately 39 percent [3]. Statistical method decision tree-based machine learning models like k-nearest neighbors, AdaBoost, and bagging have been applied for food selection and it has been illustrated through the tests that food choice accuracy AdaBoost has achieved is about 73.7% [4].

   3. Ensemble and Deep Learning Approaches

      The stacking and averaging techniques of ensembles which use various baseline learners have been suject to endorsement in clinical informatics and food sector tasks. Depth-boosted trees (XGBoost) and random forests often beat shallow classifiers on mostly nutrition datasets, and their fusion with neural networks portrays both linear and non- linear representations which are complementary arts [5]. Transformer architectures and deep neural networks, when applied to large recipe datasets, have given rise to prediction accuracies that are greater than 90%, although, in general, this type of model is computationally more expensive than ensemble ML models by several orders of magnitude [6].

   4. iOIntegrationanutyandIot

      The most recent research in precision nutrition has included the connection of AI models with continuous health monitoring that is made possible by wearable devices and IoT sensors. Such a health monitoring system, which is supervised by IoM and AI for diet planning, is the first practically available solution for creating a feedback loop with motion data and personalized food advice [7]. Furthermore, reinforcement learning frameworks, besides other options, focused on the shifting of the user from the nutrition given to the plane he/she wants to go while gradually receiving relevant suggestions about the food and vacation of interest [8].

   5. NLP for Preference Mining

      Natural Language Processing techniques, like sentiment analysis, topic modeling, and embedding-based semantic similarity, are applied more and more to obtain taste preferences of users from their reviews and recipe descriptions. The expressions such as “low sodium,” “grilled over fried,” or “spicy” carry a genuine signal. The integration of the NLP-derived preference features with structured nutritional attributes significantly enhances the

      personalization of recommendations and the acceptance by the users [9].

   6. Gaps in Existing Literature

   In spite of these advancements, four persistent limitations have been highlighted in the available literature: (i) most systems are designed to achieve only one goalnutrition or preferencerather than both; (ii) ensemble ML is under-used as a food recommendation tool beyond disease classification;

   (iii) deployable, interactive platforms which can filter information in real time are still limited; and (iv) high-quality datasets that are publicly available seldom consist of demographic information, nutrient profiles, taste preferences, and ground-truth recommendation labels. This project, however, has the main objective of bridging all four gaps at the same time.

3. RESEARCH GAP AND MOTIVATION

   A more in-depth exploration of the existing nutrition recommendation systems and the six specific gaps that are found in them shows that these are the gaps that together determined the course of this project.

   First Gap: Insufficient Personalization

   The commercial applications such as MyFitnessPal and HealthifyMe just provide supportive learning that goes shallow. Factors such as health condition, BMI, dietary restriction, cuisine preference, and taste profile are all important, but the majority of the systems consider only one or two at a time. The static rules that are available cant deal with that type of multidimensional variability of actual users.

   Second Gap: Single-Model Approaches

   The majority of food recommendation studies that are based on machine learning choose a single classifier and stop there. The main reason is that one model can be weak particularly with heterogeneous dietary datasets of nutritional attributes that are continuous mixed with categorical user preferences.Ensemble integration is a natural remedy, but it remains surprisingly underexplored in this domain.

   Gap 3: Hybrid Scoring Is Lacking

   Very few systems are found that are combined with nutrition density metrics, taste, and the preference scores into a unified absolute ranking function. A suggestion can come up that if the optimization is performed only on the basis of nutrition, the recommendations will be made just by the people who follow the Absolute diet. Conversely, one can assert the opposite. The statement can be made that in the case of optimizing solely on the basis of preference, the foods recommended are harmful to the user. A common objective is a prerequisite, yet only a few attempts have been made in this direction.

   Gap 4: Datasets Are Not Available

   The datasets that are available and are labeled with user demographics, complete nutrient profiles, metadata on the food and taste, diet-type labels, and binary recommendations to ground truth are very scarce. In most cases, the studies have synthetic or very few real-world data used for training, thus they lack the generalization ability.

   Gap 5: Platforms That Cannot Be Used Are Not Found

   The prevailing concern is that most research becomes stale after being criminally stuck at the offline model evaluation. A tremendous gap exists between hardware for measuring both app recommendations in workbook format and additional hardware that binds recommendations to users with a filter and also shows visual analytics and an interface that is entirely clear to them.

   Gap 6: Empirical Validation Is Not Sufficient

   Very few of the past studies assess their outputs based on a wide range of metrics, namely accuracy, precision, recall, F1, AUC-ROC, Top-N match rates, dietary restriction compliance, and user satisfaction. The partial evaluation makes the intersystem comparison arduous. The current venture deals with all the six issues through a complete and coherent pipeline which is fed with a multi-source dataset, a three-model ensemble, hybrid AI scoring, and fully deployed interactive Streamlit platform that is conceded against various quantitative and qualitative dimensions.

4. PROPOSED SYSTEM ARCHITECTURE

   The developed system is a complete AI-powered nutrition suggestion plat that has been classified into 6 modules that are interconnected: (1) data acquisition and integration; (2) preprocessing and label correction; (3) feature engineering and selection; (4) ensemble ML model training; (5) hybrid AI scoring and recommendation generation; and (6) interactive Streamlit deployment. The overall system workflow is described in this figure.

   Fig. 1. High-level workflow of the AI-powered personalized nutrition recommendation system.

   A. Data Acquisition and Integration

   We merged the five original source data sets into one training corpus (final_dataset.csv):

   • user_profiles.csv: User demographics (age, gender, BMI, health condition, dietary restriction, preferred cuisine, taste preference) for N = 50 simulated user profiles.

   • nutrition.csv: Nutritional composition records (calories, protein, fat, carbohydrates, fiber, vitamin C, iron, calcium) for 300 food items.

   • recipes.csv: Recipe metadata (name, cuisine, diet type,

     cooking time, macronutrient profile) for 100 dish entries.

   • user_feedback.csv: 500 user-recipe interaction records with 15 star ratings generated via random pairing.

   • Synthetic scores: Nutrient_Score (U[0.6, 1.0]) and Preference_Score (U[0.5, 1.0]) appended to each interaction; the binaryB. Preprocessing and Label

   Correction

   Raw multi-source data rarely arrives clean. Our preprocessing pipeline handles four things: missing numeric values are imputed with column means; missing categorical

   fields are filled with mode or default labels; an automated keyword heuristic scans Recipe_Name for non-vegetarian terms (chicken, fish, egg, mutton, prawn, meat, beef, shrimp) and reassigns Diet_Type accordingly; and Label Encoding converts all categorical features into integer codes compatible with sklearn estimators.

   C. Ensemble Model Architecture

   p>Three complementary classifiers are trained in parallel on the preprocessed dataset using an 80:20 stratified train-test split:

   • Random Forest (RF): Constructs an ensemble of decision trees via bootstrap aggregation. Naturally resistant to overfitting and effective at capturing non-linear feature interactions.

   • XGBoost: Implements gradient-boosted decision trees in a sequential, residual-correcting manner. Delivers regularized, high-efficiency predictions on tabular data with mixed feature types.

   • Multi-Layer Perceptron (MLP): A feedforward neural network (hidden layers: 32, 16 neurons; max_iter: 300; ReLU activation). Complements the tree-based models by modeling complex non-linear interactions from a different angle.

   Soft-vote ensemble integration averages per-class predicted probabilities from all three:

   P = (P_RF + P_XGB + P_MLP) / 3

   D. Hybrid AI Scoring and Ranking

   The final meal ranking is governed by a weighted composite AI Score that brings together nutritional quality, personal preference alignment, and ML-predicted suitability:

   AI_Score = 0.6 × NutrientScore + 0.4 × PreferenceScore

   + 5 × P

   The 0.6 weight on NutrientScore is an indication of the high importance of nutritional adequacy in a health-oriented system. PreferenceScore is given 0.4 as it precisely indicates the correspondence with tasteif the users didn’t want to eat these foods, the advice would not be actually valuable. The P term is raised 5 times, thus the ML-predicted suitability systematically has a mail effect on the ranking decisions. The AI Score, which is a measure of food items with the help of artificial intelligence, is categorically applied; the best 10 foods are shown to the user.

5. METHODOLOGY

   The end-to-end methodology uses a structured workflow with six stages to better both predictive accuracy and recommendation relevance.

   1. Stage 1: Data Collection andintegration

      As mentioned in Section IV-A, the data was collected from five constituent sources. The union was executed using pandas DataFrame join operations on User_ID and Recipe_ID keys resulting in 500 final interaction records. Two synthetic continuous score columns (Nutrient_Score, Preference_Score) were complemented, and the binary Recommended target label was obtained by thresholding User_Rating at 3.

   2. Stage 2: Preprocessing and Feature Engineering

      As soon as data integration was realized, preprocessing was implemented as indicated in Section IV-B. Feature selection distinguished 18 predictive attributes: Age, BMI, Calories, Protein_g, Carbs_g, Fat_g, Fiber_g, Vitamin_C_mg, Iron_mg, User_Rating, Nutrient_Score, Preference_Score, and six Label-Encoded categorical features (Gender, Health_Condition, Dietary_Restriction, Preferred_Cuisine, Taste_Preference, Diet_Type). The target variable y was the binary Recommended column [10].

   3. Stage 3: Nutritional Deficiency and Preference Analysis In parallel with ML model training, we ran a nutritional deficiency detection module. Based on the users BMI, age

      category, and health condition, the system identifies potential

      nutrient gapslow iron for users with an Iron Deficiency flag, for exampleand up-weights candidate recipes rich in that nutrient when computing NutrientScore. A preference mining module uses encoded taste and cuisine preferences to compute a personalized PreferenceScore for each candidate recipe, ensuring flavor alignment is explicitly reflected in the final ranking [11].

   4. Stage 4: Ensemble Model Training

      Three sklearn-compatible classifiers were instantiated and trained:

      • RandomForestClassifier(n_estimators=80, random_state=42): 80 bootstrap-sampled decision trees with Gini impurity splitting.

      • XGBClassifier(n_estimators=80, random_state=42, eval_metric=logloss): 80 gradient-boosted trees evaluated against log-loss.

      • MLPClassifier(hidden_layer_sizes=(32, 16), max_iter=300, random_state=42): Two-layer feedforward network with ReLU activation.

        All three models trained on X_train (80%) and were evaluated on X_test (20%) using stratified random splitting (random_state=42). Predicted probabilities on X_test were extracted via predict_proba()[:, 1] for ensemble averaging [12].

   5. Stage 5: Hybrid Recommendation Scoring

      At inference time, each classifiers predict_proba()

      output is extracted for the user-filtered subset of recipes. The

      three probability vectors are element-wise averaged to yield P, which is then combined with NutrientScore and PreferenceScore via the AI Score formula. The resulting scores rank the candidate recipes; the top 10 are returned as personalized recommendations.

   6. Stage 6: User Interface and Feedback Loop

   During inference, the predict_proba() output of each classifier from the user-filtered subset of recipes is obtained. The element-wise averaging of these three probability vectors results in P, which is further combined with NutrientScore and PreferenceScore by the AI Score formula. The passing of these scores leads to the ranking of the candidate recipes, and the top 10 are returned as personalized suggestions.

6. IMPLEMENTATION DETAILS

   1. Technology Stack

      TABLE I. Technology Stack and Implementation Environment

      Component	Technology / Library
      Programming Language	Python 3.83.12
      ML & Data Processing	Scikit-learn, XGBoost, Pandas, NumPy, SciPy
      Neural Network	MLPClassifier (sklearn.neural_network)
      Web UI / Deployment	Streamlit
      Visualization	Altair / Matplotlib
      Database (backend)	PostgreSQL + Redis (caching)
      Authentication	Firebase Auth + JWT Tokens
      Dev Environment	VS Code / Jupyter Notebook
      Version Control	Git / GitHub
      Cloud Deployment	Streamlit Cloud / AWS EC2 / GCP VM

   2. System Requirements

      TABLE II. Minimum and Recommended Hardware Specifications

      Component	Minimum	Recommended
      Processor	Intel Core i3	Intel Core i7 / AMD Ryzen 5
      RAM	4 GB	16 GB
      Storage	1 GB HDD	SSD (any capacity)
      GPU	Not required	NVIDIA GTX 1650+ (optional)
      OS	Windows 10 / Ubuntu 20.04	Linux Ubuntu 22.04+
      Python	3.8	Latest stable release

   3. Core Implementation Snippet

      The following condensed excerpt illustrates the ensemble model training and AI Score computation logic:

      rf = RandomForestClassifier(n_estimators=80, random_state=42)

      xgb = XGBClassifier(n_estimators=80, eval_metric=’logloss’)

      mlp = MLPClassifier(hidden_layer_sizes=(32,16), max_iter=300)

      rf.fit(X_train, y_train) xgb.fit(X_train, y_train) mlp.fit(X_train, y_train)

      P0 = (rf.predict_proba(X)[:,1] + xgb.predict_proba(X)[:,1] + mlp.predict_proba(X)[:,1]) / 3

      df[‘AI_Score’] = (0.6*NutrScore + 0.4*PrefScore + 5*P0)

      he complete source code will be available at [GitHub Repository URL to be added upon publication].

7. RESULTS AND DISCUSSION

   1. Experimental Setup

      While performing inference, the user-filtered subset of recipes receives the predict_proba() output of each classifier which is obtained. The element-wise averaging of these three probability vectors is equal to P, which is then combined with NutrientScore and PreferenceScore by the AI Score formula. The passing of these scores results in the ranking of the possible candidates and the 10 best ones are given as personalized suggestions.

   2. Classification Performance

      TABLE III. Comparative Classification Performance

      Model	Acc. (%)	Prec.	Recall	F1
      Random Forest	88.4	0.87	0.88	0.87
      XGBoost	89.7	0.89	0.90	0.89
      MLP Neural Network	87.9	0.86	0.87	0.87
      Ensemble (RF+XGB+MLP)	91.3	0.91	0.92	0.91

      In the course of inference, the user-filtered subset of recipes is endowed with the predict_proba() outputs of the classifiers received. P, the value which is later combined with NutrientScore and Preference Score through the AI Score formula, is the element-wise average of these three probability vectors. From these two computed scores, the possible candidates are ranked and subsequently, the best 10 alternatives are offered as personalized recommendations.

   3. Recommendation Quality Evaluation

      Top-N recommendation quality was assessed by generating Top-10 ranked meal lists for each test user and evaluating three criteria: the fraction of users finding at least one health-

      appropriate meal in their top-3 results; dietary restriction compliance rate; and nutritional match accuracy relative to Recommended Daily Allowances (RDAs).

      TABLE IV. Top-10 Recommendation Quality Metrics

      Evaluation Metric	Result
      Top-3 Meal Match Rate	84% of users
      Dietary Restriction Compliance	93% of recommendations
      Nutritional Accuracy vs. RDA	±8% average deviation
      System Response Time	< 2 seconds
      User Satisfaction (25-person survey)	4.6 / 5.0

      These results show the system managing to bridge the gap between nutritional adequacy and personal preference alignment in practice. The 93% dietary restriction compliance rate is particularly significant: it suggests the keyword-based diet-type correction heuristic is doing real work, reducing the chance that a vegetarian user gets served a meat dish or that a diabetic user receives a high-sugar recommendation.

   4. Sample Recommendation Output

      TABLE V. Top-3 Recommendations: 1835 Female, Diabetes, Vegetarian

      Rank	Recipe	Cal	Prot.	Fat	AI
      1	Grilled Tofu Salad	220	18g	6g	0.93
      2	Quinoa Vegetable Bowl	250	14g	8g	0.91
      3	Lentil Spinach Soup	195	12g	5g	0.89

   5. Nutritional Deficiency Analysis

      Condition-specific recommendation relevance was validated across three representative user health profiles:

      • Iron Deficiency users: The system consistently surfaced high-iron dishes (spinach curry, lentil soup, fortified cereal preparations), increasing mean iron content in top- 5 recommendations by 42% relative to the unrestricted baseline.

      • Users with high BMI (> 27): Prominent protein, skimming the fat options (grilled tofu, baked chicken, quinoa bowls) were registered first in the list, with a mean fat content in recommendation falling 31% lower than that of the population average.

      • Diabetic users: Low glycemic index and high fiber content meals (vegetable stir-fried, oat-based preparations, legume soups) were always the ones dominating Top 10 rankings, while sugar-dense options were relegated below rank 15.

   6. Comparative Analysis vs. Rule-Based Systems

      This system is a step ahead of traditional BMI-only or calorie- count-only diet planning tools, having three specific benefits. To begin with, as the user’s inputs are altered, the suggestions are updated automaticallyno manual adjustments are needed. Besides, The AI Score is able to reach the goals of the user at the same time, namely, meeting both the nutritional needs and the preference of the user, which is absolutely impossible for strict rules to do. Furthermore, the ensemble ML approach enables different user archetypes to be included, while the rule-based programming requires rules to be written for each new user instead. These gains are the immediate cause of the performance of the 84% top-3 match rate and the 4.6/5 usability rating seen in the trials conducted with users.

   7. Limitations

      Three limitations are worth being upfront about. First, the datasetat 500 interactionsis modest relative to production recommendation systems; performance characteristics may shift on larger, noisier real-world corpora. Second, the system has a cold-start problem for new users with no interaction history

   8. Summary of Findings

   TABLE VI. Summary of Key Evaluation Outcomes

   Dimension	Key Outcome
   Ensemble Model Accuracy	91.3% (best among all models)
   Precision / Recall / F1	0.91 / 0.92 / 0.91
   Top-3 Recommendation Match	84% of test users
   Dietary Restriction Compliance	93% of recommendations
   Nutritional RDA Alignment	±8% average deviation
   System Response Time	< 2 seconds per inference
   User Satisfaction (n=25)	4.6 / 5.0

8. CONCLUSION

The system weve described demonstrates that combining ensemble ML, hybrid scoring, and a deployable interactive platform produces something meaningfully stronger than any of those pieces in isolation. A 91.3% classification accuracy, 84% top-3 match rate, 93% dietary restriction compliance, and 4.6/5 usability ratingacross a live Streamlit deployment responding in under 2 seconds suggest this is a viable framework for real-world nutrition management, not just a proof of concept.

The design choice that seems to matter most is the hybrid AI Score, which refuses to optimize for nutrition or preference alone. Previous systems that operated with such trade-offs have usually been the source of the two types of suggestions that I would offer you: one that appeals to your nutritional needs, but no one desires to eat, and another that is very appealing personally but is lacking nutrition. The combination of the two objectives and using ML to rank by predicted suitability are both facors that help to avoid the conflict in this case.

A number of paths will show success in further efforts. Firstly, amidst monitoring physiological data of wearables such as heart rate, glucose levels, and daily activitywould be the addition of more dynamic, context-aware updates to recommendations. Secondly, instead of having to depend on periodic retraining, the model could keep on improving as the real user feedback stacks up thanks to the online reinforcement learning. Thirdly, going to real-world user groups in different geographic and cultural areas would be a way to find out whether the system generalizes the same as the results currently suggest. The fourth is to integrate an explainability layer, like SHAP or LIME, which would enable the doctors and users to see the reason behind a certain recommendation being made, and that is of immense importance in any medical practice.

REFERENCES

1. World Health Organization, Healthy diet, WHO Fact Sheet, April 2020. [Online]. Available: https://www.who.int/news-room/fact- sheets/detail/healthy-diet

2. I. Orue-Saiz, M. Kazarez, and A. Mendez-Zorrilla, Systematic review of nutritional recommendation systems, Applied Sciences, vol. 11, no. 24, p. 12069, 2021. DOI: 10.3390/app112412069

3. M. Divya Varshini and G. Paavai Anand, Intelligent nutrition recommendation system for individual health profiles, IJARCCE, vol. 14, no. 1, 2025. DOI: 10.17148/IJARCCE.2025.141190

4. T. Kandadi and G. Shankarlingam, Drawbacks of LSTM algorithm: A case study, SSRN Electronic Journal, Jan. 2025. DOI: 10.2139/ssrn.5080605

5. Y. Xu and L. Zhang, Nutrient-based food recommendation using ensemble machine learning models, Applied Intelligence, vol. 50, no. 11, pp. 38893902, 2020.

6. M. Akshatha et al., Personalized nutrition recommendation system using machine learning, IARJSET, vol. 12, no. 5, 2025. DOI: 10.17148/IARJSET.2025.125354

7. P. K. Sahoo and M. Y. Lee, Smart health monitoring system for personalized diet planning using IoT and AI, IEEE IoT Journal, vol. 6, no. 5, pp. 76307640, 2019.

8. K. I. Mavrokotas et al., A nutrition recommendation system based on reinforcement learning, in Proc. IEEE EMBC, 2025. DOI: 10.1109/EMBC58623.2025.11252751

9. D. S. L. Manikanteswari et al., Nutrient recommendation system for personalized diet, IJISRT, 2025. DOI: 10.38124/ijisrt/25mar1573

10. T. Kandadi, Implementation of Dichotomiser 3 algorithm with decision tree, SSRN, Apr. 2025. DOI: 10.2139/ssrn.5207811

11. T. Kandadi, Drawbacks of Random Forest algorithm to examine extensive datasets, SSRN, Apr. 2025. DOI: 10.2139/ssrn.5236759

12. M. Shah, S. Degadwala, and D. Vyas, Diet recommendation system based on different machine learners, IJCRT, 2022. DOI: 10.32628/CSEIT228249

13. S. Choudhury and A. Singh, A hybrid recommendation model for diet planning using random forest and deep neural networks, IEEE Access, vol. 9, pp. 154215154228, 2021.

14. N. T. Binh and P. Van Toi, Personalized diet recommendation system using machine learning approaches, J. Biomedical Informatics, vol. 122, p. 103887, 2021.

15. G. Shankarlingam and K. T. Reddy, Predicting a small cap company stock price using Python, Indian J. Science and Technology, vol. 16, no. 48, pp. 46204623, 2023.

16. T. Kandadi Reddy, Explainable deep learning for automated pneumonia detection, TechRxiv, Oct. 2025. DOI: 10.36227/techrxiv.176108165.56312001/v1

17. T. Kandadi Reddy, A framework to predicting startup success growth, TechRxiv, Sept. 2025. DOI: 10.36227/techrxiv.175751434.42657764/v1

18. Food recommendation towards personalized wellbeing, Trends in

Food Science & Technology, 2025. DOI: 10.1016/j.tifs.2025.104877