Freshmen URD Track; Entrepreneurship and Innovation
In today's world, the vast variety of restaurants can overwhelm people, making it difficult and time-consuming to constantly seek out new and tasty dining options. This dilemma is further intensified for those with busy schedules, who may struggle to find suitable dining choices quickly, often leading to skipped meals, affecting their physical and mental health due to irregular eating habits.
Solution Approach
Our AIUAN: personalized restaurant recommendation project aims to simplify consumers’ daily task of finding the right restaurant. To achieve the concept of ‘A Friend who knows what you want to eat’, our approach focuses on three pivotal areas: more factors in recommendation, a simpler user interface, and an innovative chatbot feature. We have developed a prototype using Figma, based on comprehensive use case analysis, to ensure that our application is not only functional but also intuitive and user-friendly. The UI is designed to facilitate easy navigation and quick access to personalized restaurant recommendations.
Data Collection
After processing 15,561 restaurant listings from the BMA dataset, our data enrichment efforts successfully enhanced the feature sets of 10,568 restaurants. This represents a 67.91% success rate, covering a diverse range of eateries across 50 districts in Bangkok, Thailand. This significant accomplishment not only improves the comprehensiveness of our database but also ensures a robust representation of Bangkok's culinary landscape.Specifically for different data sources, Kaggle has 1,427 matches out of 15,561 restaurants, showing an inadequate 9.17% success rate. Google Place API achieved 11,559 matches out of 15,561 restaurants, representing a decent 74% hit rate.After model training is complete, we use a testing set (20% of the entire dataset) to evaluate the model and calculate F1 score from the result. In the final evaluation, shown in Figure 38, our NER model achieved an F1 score (SCORE) of 0.34, Precision (ENTS_P) of 54.55, and Recall (ENTS_R) of 25.00, indicating poor performance. This is largely due to the small dataset of only 100 labeled examples, which is insufficient for optimal training. Although the model successfully detects and categorizes some texts (figure 39), it lacks consistent accuracy and flexibility in others (figure 40).
Summary of Accomplishments
● Restaurant Data Collection: Successfully collected and enriched the data,with 67.91% success rate, for 10,568 out of 15,561 restaurants from the BMA dataset, covering a wide range of eateries in Bangkok.
● Recommender Model Prototyping: Developed a first version of prototype recommendation model using 20 manually verified restaurants. The model effectively cross-compared user and restaurant profiles, particularly focusing on food type preferences. The results showed high relevance and applicability of recommended restaurants to user preferences.
● Chatbot Prototyping: Prototyped a Named Entity Recognition (NER) model, albeit with a modest performance indicated by an accuracy score of 0.34. The model demonstrated potential in detecting and categorizing texts but also highlighted the need for a larger training dataset for improved accuracy and flexibility.
● UX/UI Design: Successfully designed high-fidelity user interface, as per the planned functional diagram. These include the landing page, authentication pages (Sign In & Sign Up), a questionnaire interface, home page with dark and light modes, a history page, detailed restaurant pages, a review and comment section, and a direction page with integrated mapping. Additionally, an interactive design prototype was developed.
Future Directions
In the next semester, we will first improve the capabilities of the recommender model. One approach is to utilize other techniques for recommendation. Specifically, we will develop the model with deep neural networks. Another approach would be to try a hybrid recommender model, by combining collaborative filtering. Most importantly, as more restaurant data is now available, we will use vastly more data in the recommender model.
