TravIS - Travel Intelligent System

Motivation

How do we decide where to spend our holidays? Sometimes we talk to our friends, who just returned from somewhere, happy and sun-tanned. Otherwise we go to these funny travel-agency shops with seashells on display, talk to the friendly agents and corteously buy a package where the agency has their best commision. The most rebellious of us do web search in an attempt to discover new fascinating destinations, so we are fighting with cookies, cliché descriptions written by graphomaniac bloggers, and garish advertisement pop-ups...

Vacations provide us with some of the best memories in our lifetime, so why should an organizing of it be so complicated and stressful? Holiday planning, like an appetizer, has to bring a big smile to your face and make you daydreaming about all the fun things you are going to do and experience.

The Concept

Through a user-friendly web interface, users can input keywords that are associated with their dream vacation. Based on these keywords, the system will identify and display the most suitable country for the user. In the next step, the system will search for the optimal region within the selected country that matches the user's preferences, and provide suggestions for hotels, apartments, or other accommodations. The system also takes into account travel planning, including transportation options. Additionally, the user will be presented with recommendations for possible leisure activities and attractions to enhance their stay.

Who we are and why we can make it

Alina brings more than 5 years of experience in hospitality industry.

Dominik brings more than 7 years of expertise in applied programming.

Together: 1.5 years of successful cooperation in student projects and labs in Esslingen University, Computer Engineering department.

Problem in the initial attempt

In our project, we faced the challenge of associating user-provided keywords with country- coherence specific tags in our database. In the first attempt, we used a similarity matrix to determine the best-matching country. However, we encountered a problem when dealing with highly exotic queries that contained words not present in our database, resulting in unsatisfactory results.

Proposed solution

To address this issue, we have thought a new approach. Our plan is to introduce a "Large Language Models" (LLM) layer between the user input and the matrix. This layer aims to identify similarities between the input and the existing tags in our database and find corresponding synonyms. To mitigate potential misinterpretations by the LLM, we will use different models such as Falcon-7B, GPT-4, and code-davinci-002. These models will cross-validate and confirm each other's outputs, to ensure that the generated results make sense. This approach enables us to better prevent misinterpretations and incorrect mappings.

The utilization of Large Language Models provides us with increased flexibility in query processing, particularly for "exotic queries" that include words not present in our database with 100% exact matches. By introducing a "synonym recognition layer, " we can retain valuable information that would otherwise be lost. This allows for a more precise mapping of user input to country-specific tags in our database.

Conclusion

Overall, we are confident that the failure of our initial attempt has revealed some weaknesses in the original plan, allow us to significantly improve the system in subsequent iterations and quickly move towards a robust, functional version.

The project itself has been enjoyable, involving planning, research, and implementation. It has provided us with insights into working with datasets and modern technologies. It has also shown the importance of learning from failures and utilizing them as stepping stones for progress.

We would like to express our heartfelt thanks to our instructors, Jan Seyler and Dionysios Satikidis, who have supported us throughout the semester and provided valuable information for both this project and our future.

We also extend our gratitude to the jury, who have provided us with good and constructive feedback.

In summary, this project has been a valuable learning experience, emphasizing the significance of adapting and improving our approaches.