International Journal of Engineering and Management Research

1. Introduction

Food is an essential component of human survival, acting as a source of energy and sustenance as well as being vital to people's general health and well-being. It gives the body the vital nutrients that are needed for development, repair, and optimal operation. Food has cultural, social, and emotional value in addition to its biological relevance, affecting daily life and customs throughout. Every nation has their own food culture. Additionally, Sri Lanka has a rich culinary tradition. Food is a crucial factor in the context of health. The hectic life cycle of human beings and negligence of rich nutrient foods are the main causes for non-communicable diseases. According to the World Health Organization (WHO), a healthy diet can aid in the prevention of diabetes, heart disease, stroke, and cancer like Non-Communicable Diseases (NCDs) [1].

It is a challenge to figure out healthy food menus/dishes based on our preference, cultural background and disease information. However, with the busy life we do not have time to get instructions from nutritionists about healthy diets. This is the place where food recommendation systems come into play. The food recommendation system can overcome all these challenges by helping us find the right foods to eat. They provide a workable answer for people who struggle to strike a balance between a hectic schedule and the need for a balanced diet by bridging the gap between nutritional science and daily meal planning.

When it comes to Sri Lankan context, there are few food recommendation systems are available. However, they do not align with the user’s specific needs and preferences. The main goal of this study is to develop a solution that can recommend Sri Lankan diet based on user preference, health condition and cultural background. In order to recommend diets in the context of Sri Lanka, we particularly developed an ontology-based machine-learning model. Here we developed ontology, based on food and nutrient information. Furthermore, user-specific data is included into the ontology to improve the relevance and accuracy of dietary suggestions catered to user preference, health condition and cultural background. We have developed a Recurrent Neural Network (RNN) model that calculates target nutrient values by considering factors such as Body Mass Index (BMI), Basal Metabolic Rate (BMR), age, and specific health conditions.

When individual predictions are combined into a single prediction, content information is added to the collaborative model, the weighted average of content and collaborative recommendations is calculated, or final recommendations are derived from the combined rankings, hybrid recommendation combines the best features of both approaches [6].

2.2 Ontology Development

Ontologies and the semantic web are two exciting new technologies that can be used to represent knowledge and draw conclusions for the creation of intelligent systems. Similarity concepts serve as the foundation for the semantic web. This idea of similarity is applied in several study areas. A set of ideas or classes that represent the objects or entities found in the domain make up an ontology most frequently. A number of interconnects these ideas different relationships that show how they relate to one another [7].

According to Ying Ding (2002), the topic of ontology is one that is quickly expanding and holds enormous potential for improving information management, organization, and comprehension. It is necessary for the Semantic Web, the next phase of the Web's evolution, because it enables content-based access, communications, interoperability, and the provision of significantly higher service levels. Ontology engineers are still required to build the knowledge base and ontology for a specific job or area, as well as to manage and update the ontology to keep it current and relevant. Ontologies that are manually created take a lot of time, labor, and mistakes. There are different ontology generation techniques. They can be summed up as follows: middle-out, which starts with the most crucial concepts and moves toward generalization and specialization (Eg: Enterprise and Methodology ontologies), bottom-up: from specification to generalization, top-down, which starts with generalization and ends with specification (Eg: KACTUS ontology) and bottom-up: from specification to generalization. An ontology can be created from scratch, utilizing just global or local ontologies that already exist, a corpus of data sources, or a combination of the latter two techniques. Three are three ways as fully manual, semi-automated and fully automated [8].

2.3 Major Developments in Recommendation System with Machine Learning Techniques

The evolution of recommendation systems powered

They created this recommendation system for diabetic patients by substituting foods according to nutrition and food parameters. Eighteen (18) nutrient values are considered initially. Then based on nutritionist marks and rank, eight (8) nutrients were selected as most important nutrients for diabetics. The SOM is built and trained first, and then it is clustered using the K-mean method. The food clusters that show which foods in the same category contain roughly the same quantity of each of the eight nutrients are the end result. Nutritionists have assessed the suggested approach, and they find it to be very effective and beneficial for the field of nutrition [11].

Future developments in ontology-based recommendation systems using machine learning are expected to include enhanced data integration, hybrid recommendation techniques, contextual and explainable recommendations, and cross-domain personalization. Advanced ML models such as deep neural networks, CNNs, RNNs, and reinforcement learning can improve the interpretation of complex user behavior and multimedia data. Techniques like sentiment analysis and NLP may be used to refine user profiles by extracting dietary preferences from reviews or social content. Cross-domain recommendation (E.g. integrating fitness and dietary data) can enable comprehensive lifestyle guidance, supported by multi-task learning and complex ontologies. Context-aware systems that consider factors like location, weather, or health status can further improve recommendation relevance. Matrix Factorization, particularly Explainable Matrix Factorization (EMF), is also gaining attention for generating top-n explainable recommendations when limited side data is available, using metrics like Mean Explainability Precision and Recall [12].

3. Methodology

This study proposes an ontology-based personalized diet recommendation system tailored to Sri Lankan individuals, considering cultural food habits, medical conditions, and nutritional needs. The methodology encompasses user data collection, ontology development, and implementation of a machine learning engine for personalized diet recommendations. And the overall architecture of the ontology-based personalized diet recommendation system is illustrated in Figure 1.

The second dataset comprises information about various foods along with their nutritional values and dietary guidelines tailored for individuals with specific diseases information such as high blood pressure, cholesterol issues, obesity, and diabetes. This data set is collected from nutritionists through interviews and gathered some data from public database and Sri Lanka food websites.

3.2 Data Preprocessing

Data preprocessing is a key step which cleans and transforms the input data to ensure its suitability for further analysis. We used one hot encoding technique (e.g. gender, food preference) and multi hot encoding (e.g. user health condition) to convert categorical variables as numerical values. Additionally, we performed missing value analysis to increase the accuracy of the dataset. Missing values were handled by replacing them with appropriate statistical measures such as the mean for numerical attributes

3.3 Model Selection and Development

An assessment of machine learning techniques for personalized diet recommendation required the selection of robust regression model capable of predicting individual nutrient requirements based on user health profiles. Here we have chosen Recurrent Neural Network (RNN) enhanced with LSTM. As a first step by using age, sex, height and weight values in dataset, we have calculated the BMI and BMR values. Then by getting the help of nutritionists, target nutrient values calculated for each user by considering their BMI and health status. After that, we performed feature extraction. Several important health indicators are included in our dataset throughout the feature extraction stage. We used programs to extract features from our data, including age, blood pressure, diabetes status, cholesterol levels, body mass index, and basal metabolic rate. These characteristics were combined into a multidimensional array called X, where each row and one of the previously described health indicators represent a user by each column. We simultaneously created the goal variables, which represent the estimated nutritional demands and contain the protein, fat, and carbohydrate targets. To make sure that each nutritional objective corresponds with the associated user data in X, these targets were arranged into another array, Y. In addition to guaranteeing

Figure 2: Classes hierarchy of the ontology

Object properties in an ontology define relationships between two instances of classes. Data properties assign data values to individual instances. There are few data properties and object properties available in our ontology. Domain and the range are the two main concepts associated with object properties and data properties. The following table represents the object properties available in the ontology with their domain and range.

Table 1: Object properties in the ontology

Moreover, the ontology includes a variety of other data properties that are crucial for constructing comprehensive user profiles. These properties encompass age, Basal Metabolic Rate (BMR) values, BMI class, cultural background, gender, and specific disease-related information. Each of these data properties plays a vital role in the recommendation system by ensuring that all user-specific factors are considered when generating diet plans. Figure 3 shows the data properties available in the ontology.

The system considers predicted protein, fat, and carbohydrate targets from the RNN model, along with user preferences, disease history, and cultural background. Diets are recommended for breakfast, lunch, and dinner. The process begins with user profile generation, which is then integrated into the ontology. Nutritional and food data are retrieved from the ontology using SPARQL queries, after loading it into the Jupiter environment. To generate recommendations, cosine similarity is computed between the predicted nutrient targets and the nutrient content of foods in the ontology. These similarity scores are then refined using user preferences, cultural context, and disease information to provide a personalized list of suitable foods.

4. Result and Evaluation

4.1 Results of Recurrent Neural Network

Several evaluation metrics are used to evaluate the performance of the RNN LSTM model. Raksha Pawar and his research team (2001) suggested the K-Nearest neighbors (KNN) algorithm to find the necessary nutrients based on user needs [13]. To evaluate the performance of the RNN-LSTM algorithm suggested by this study, we developed a KNN model using the same dataset for comparison. By implementing both models on identical data, we aim to rigorously compare their predictive accuracies and overall performance. The following table presents a comprehensive comparison between the performance metrics of the RNN-LSTM and KNN models.

Table 2: RNN- LSTM and KNN model Comparison

The RNN-LSTM model has a significantly lower MSE compared to the KNN model, indicating that the RNN-LSTM model's predictions are closer to the actual values on average. In addition, the lower RMSE of the RNN-LSTM model further confirms its higher prediction accuracy, as RMSE penalizes larger errors more significantly. The RNN-LSTM model explains 97.04% ofthe variance in the target variable, compared to 80.67% explained by the KNN model, indicating a much better fit to the data.

4.2 Results of Recommendation System

Recommendation is the most valuable part of this study. Content based filtering and cosine similarity functions are used for recommendation. The implemented recommendation system was tested by using different test scenarios simulating different user profiles. Below are two representative test cases demonstrating the system’s ability to generate personalized dietary recommendations based on user preferences and health conditions.

Test Scenario 1:

Case Description:
  User preference: Non-Vegetarian
  Cultural Background: Sinhala
  Dietary Restriction: No
  Pressure: False
  Diabetes: False
  Cholesterol: False
  Parent Pressure condition: False
  Parent Diabetic condition: False
  Parent Cholesterol condition: False

Recommendation Output:

Figure 7: Recommendation interface for Test Scenario 1

Test Scenario 1:

Case Description:
  User preference: Vegetarian
  Cultural Background: Sinhala
  Dietary Restriction: No
  Pressure: False
  Diabetes: False
  Cholesterol: False
  Parent Pressure condition: False
  Parent Diabetic condition: False
  Parent Cholesterol condition: False

[3]  Y. Ding, & S. Foo. (2022). Ontology research and development. Part I - A review of ontology generation. Journal of Information Science, 28, 123–136. DOI: 10.1177/016555150202800204.

[4]  A. El-Rady. (2020). An ontological model to predict dropout students using machine learning techniques, pp. 1–5. DOI: 10.1109/ICCAIS48893.2020.9096743.

[5]  T. N. Trang Tran, M. Atas, A. Felfernig, & M. Stettinger. (2018). An overview of recommender systems in the healthy food domain. J Intell Inf Syst, 50(3), 501–526. DOI: 10.1007/s10844-017-0469-0.

[6]  C. Trattner, & D. Elsweiler. (2017). Food recommender systems: Important contributions, challenges and future research directions.

[7]  H. Ko, S. Lee, Y. Park, & A. Choi. (2022). A survey of recommendation systems: recommendation models, techniques, and application fields. Electronics, 11(1), Art. No. 1. DOI: 10.3390/electronics11010141.

[8]  “[PDF] Using Content-Based Filtering for Recommendation | Semantic Scholar.” Accessed: Jul. 29, 2025. Available: https://www.semanticscholar.org/paper/Using-Content-Based-Filtering-for-Recommendation-Meteren/4a57e0f0641b7a70fece89c14fbf5030869ededb

[9]  C. Obeid, I. Lahoud, H. El Khoury, & P.-A. Champin. (2018). Ontology-based Recommender System in Higher Education. In: Companion of the the Web Conference 2018 on the Web Conference, Lyon, France: ACM Press, pp. 1031–1034. DOI: 10.1145/3184558.3191533.

[10]  A. Ruiz-Iniesta, & O. Corcho. A review of ontologies for describing scholarly and scientific documents.

[11]  Ontology-based machine learning to predict diabetes patients. In: ResearchGate. DOI: 10.1007/978-3-030-91738-8_40.

[12]  Y. Zhang, & X. Chen. (2020). Explainable recommendation: A survey and new perspectives. INR, 14(1), 1–101. DOI: 10.1561/1500000066.

[13]  R. Pawar, S. Lardkhan, S. Jani, & K. Lakhi. (2021). NutriCure: A disease-based food recommender system, 6(6).