2024 IEEE 2nd International Conference on Control, Electronics and Computer Technology (ICCECT)
Research on development and application of AI
agent for travel recommendation driven by large
language model
Meirong Lin
2024 IEEE 2nd International Conference on Control, Electronics and Computer Technology (ICCECT) | 979-8-3503-8095-8/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICCECT60629.2024.10545805
Information Engineering College
Hainan Vocational University of Science and Technology
Haikou, China
[email protected] Abstract—Based on the large language model and ZeLinAI
platform, the construction and application of travel II. PROJECT ANALYSIS
recommendation AI agents are carried out. And the deep The goal of the travel recommendation AI agent is to
learning and sentiment analysis methods of the large model are provide users with personalized travel plans and trip
integrated to further explore the needs of users, provide users recommendation services. The agent can intelligently
with more accurate and personalized travel recommendation
recommend suitable travel routes according to user needs
services, and continuously optimize travel recommendation AI
agents by optimizing prompt and introducing custom training
and scenarios. According to the folk customs of the tourist
sets. destination. It provides users with reasonable diet,
accommodation arrangements and other precautions. Based
Keywords—large language model, Artificial Intelligence on large language model technologies such as machine
agent, travel recommendation, prompt learning and deep learning, intelligent optimization of user
origin, destination, time and budget is carried out to improve
user experience and satisfaction. In order to carry out more
I. INTRODUCTION
in-depth research and analysis of AI agents, the author built
Travel recommendation is an interdisciplinary research and applied a travel recommendation AI agent based on the
field, which combines the theories and technologies of large language model and ZeLinAI platform. The agent can
computer science, artificial intelligence, behavioral obtain customized travel recommendations according to user
psychology, tourism science and other fields. In recent years, preferences, interest points, budget and other conditions.
it has received extensive attention. Travel recommendation ZeLinAI platform is a platform for building enterprise AI
aims to provide personalized travel product and service applications based on zero code, which helps users access
recommendations by analyzing users' personal preferences, large models quickly, safely and efficiently [7]. The platform
historical behavioral data and real-time contextual uses different agents to train customized data through
information. Current research on travel recommendation debugging Prompt and individual agent parameters to
focuses on improving the accuracy and transparency of achieve stable output effects, helping to open up the last
recommendation algorithms [1][2][3], enhancing user kilometer of the landing of the large model to achieve cost
interaction experience [4], ensuring data security and privacy reduction and efficiency.
protection [5][6], and so on. This paper proposes a travel
recommendation AI agent driven by a large language model. III. PROJECT DESIGN
Through the construction and application of the travel
recommendation AI agent, the deep learning and sentiment On the basis of detailed demand analysis, the main
analysis methods of a large model are integrated to further functions of this project to be completed are clarified to
explore the needs of users and provide users with more ensure that appropriate travel recommendation services are
accurate and personalized travel recommendation services. provided to users. It mainly relies on the ZeLinAI platform
AI Agent (Artificial Intelligence Agent) is an entity or for development, which can train the self-built data set
software system that can autonomously sense, make according to requirements, associate the trained data set with
decisions, and take actions in a particular environment to the created agent, and improve the accuracy of the agent data.
achieve a specific goal. AI agents simulate human intelligent Deploy the travel recommendation AI agent to the H5
behaviors such as learning, reasoning, and planning through application and share the agent. Users can use the AI agent
algorithms and models, which can be applied in a variety of through the mobile and computer terminals, and give
scenarios, such as smart homes, autonomous vehicles, and feedback on the problems and suggestions in the use process.
virtual assistants. It is one of the mainstream forms of large The project team will continue to optimize the travel
language models landing in business scenarios. The large recommendation AI agent according to the feedback. The
language model is the core brain of the AI agent, which can entire project flow chart is shown in Figure 1.
decompose complex tasks into achievable sub-tasks to form
an agent with independent thinking, decision-making and
execution of tasks.
979-8-3503-8095-8/24/$31.00 ©2024 IEEE 92 April 26-28, 2024 Jilin, China
Authorized licensed use limited to: Sai Vidya Institute of Technology. Downloaded on April 10,2025 at 10:58:43 UTC from IEEE Xplore. Restrictions apply.
recommendation AI agent. Enter the tourist destination and
other content in the input bar on the lower right side of the
form, and the agent can plan the tourist route and
recommend the diet and accommodation according to your
input content, as shown in Figure 3.
Fig. 3. Agent testing.
Fig. 1. The flow chart of the project.
B. Optimizing Travel Recommendation AI Agents
1) Optimize Prompt: Prompt is the key to the
IV. PROJECT IMPLEMENTATION development of our agent. It needs to be defined accurately
and confirmed well in order to have an output that meets our
A. Building a travel recommendation AI agent needs. Usually this step requires repeated iteration and
First sign up and log in to the ZeLinAI platform. After optimization. Below, the project team conducted structured
entering ZeLinAI platform, select "Agent" in the list on the optimization of Prompt of the travel recommendation AI
left side of the platform, and click "Create Agent" button to agent [8]. The structured Prompt has a variety of templates.
customize and create the agent after entering the agent The project team chose the well-known CRISPE framework
interface. There are many parameters on the Create Agent [9], and the optimized content is shown in Figure 4.
page. You need to set Agent Name, Model, and Role. Set Structured Prompt can be directed to awake the depth ability
other parameters by default. Subsequent can continue to of large models and improve semantic cognition [10].
optimize the other parameter which is set according to the
demand.
1) Configure the agent name: Users can customize the
agent's name, such as "Travel Recommendation assistant".
2) Model type configuration: The model number is
selected from the drop-down menu. At present, the ZeLinAI
platform supports models including Microsoft-3.5,
Microsoft-4, ERNIE Bot, SparkDesk, MJ drawing, SD
drawing, DALLE3 and other large models. Different large
language models have different effects on the results
returned by agents due to different algorithms. The user can
debug the selection of different large language models
during testing. If the results returned by the agent are not
ideal, the model can be changed to optimize.
3) Role location configuration: Role positioning defines
the agent's role (Prompt), which we will write as prompt.
The Prompt we defined in the travel recommendation AI
agent is shown in Figure 2.
Fig. 4. Optimized structured Prompt.
2) Import training set: In order to make the agent better
match our project team's own business, we can also
introduce a custom training set to associate the agent with
the training set. The current platform training set supports
Fig. 2. Prompt defined by the travel recommendation assistant. text, pictures, videos, multimedia libraries and other forms.
4) Agent testing: After completing the configuration of a) Create training set: Select "Training" from the list
"Agent Name", "Model" and "Role positioning", click on the left side of the platform to enter the training window.
"Confirm" to complete the creation of the travel Two kinds of training sets can be created: intelligent
93
Authorized licensed use limited to: Sai Vidya Institute of Technology. Downloaded on April 10,2025 at 10:58:43 UTC from IEEE Xplore. Restrictions apply.
training sets and ordinary training sets. The intelligent the project team hopes to share it to more users, and it is
training set supports two training methods: convergence and necessary to package the built agent as an AI application.
divergence of input content, while the ordinary training set The platform supports the creation of three types of
is the Q&A form. Create intelligent training set, select applications: H5 applications, API applications, and BOT
divergence mode, can be associative learning of input applications. In this project, we adopted the method of H5
content. Text training and document training can be used in application, which can be opened and used directly on
intelligent training set. Text training directly inputs the computers and mobile phones.
content of the training. Document training uses drag and 1) Create an application: Select "Application" from the
drop to add the documents that need to be trained to the list on the left of the platform. In the application window,
training set. Currently, the platform supports documents in select "Create H5 Application". In this window, enter the
three formats, such as.txt, .docx and.PDF. application name, application icon, application Description,
b) Associative training set: Click the Add training set and select Agent to complete the application creation. In the
content setting bar, select the agent that needs to be shared --
button in the Create Agent window, check the name of the
the travel recommendation assistant, and after realizing the
new training set in the pop-up window, click the training set
association between the application and the agent, click
after completion, and automatically add the training set in "Complete and share" to enter the window of the sharing
the "role positioning" box, completing the association application.
between the self-built training set and the agent. After the
creation of the agent that introduces the self-built training 2) Sharing application: The sharing application window
set, enter the content related to the tourist destination that has three sharing methods: link sharing, two-dimensional
you need to know in the input field on the lower right side code sharing and poster sharing. You can log in and use AI
of the form, the agent can answer the question according to by scanning a poster or QR code on your phone. As shown
your input content, and mark the paragraph below the in Figure 5, it is the operation effect of the poster travel
answer that references the training set. recommendation AI agent scanned by the user through the
mobile phone.
C. Application of travel recommendation AI agents
After the creation of the travel recommendation agent,
Fig. 5. The operation effect of AI agent on mobile phone.
built by the project team, and rich travel recommendation
V. CONCLUSION cases can be obtained from the data set to meet the
This project has three advantages: First, it is efficient and personalized needs of users at different levels. Third, it is
intelligent. Using the current advanced artificial intelligence comprehensive and easy to use. With a method powerful and
technology and large language model, it can quickly and easy to use, users can complete travel search at any time
accurately provide users with travel recommendation support through the mobile terminal, improve efficiency and user
and services, whether it is itinerary recommendation or route satisfaction.
optimization, showing the characteristics of efficiency and Through the development and application of travel
intelligence. Second, there are plenty of cases. Training can recommendation AI agents, the project team deeply felt the
be conducted based on the travel recommendation data set next improvement direction of the agent project: first,
94
Authorized licensed use limited to: Sai Vidya Institute of Technology. Downloaded on April 10,2025 at 10:58:43 UTC from IEEE Xplore. Restrictions apply.
continuous optimization of Prompt words. Good prompt [2] Adomavicius, G., & Tuzhilin, A, “Towards the next generation of
words can make the model correctly understand the user recommender systems: A survey of the state-of-the-art and possible
extensions,” User Modeling and User-Adapted Interaction, 2005,
needs and give the expected results. The quality of prompt 15(1-2), pp. 33-70.
words determines the output quality of the agent. In the next [3] Burke, R, “Hybrid recommender systems: Survey and experiments,”
step, the project team will write prompt words with clear User Modeling and User-Adapted Interaction, 2002, 12(4), pp. 331-
semantics, clear logic and reasonable results as much as 370.
possible through more learning and practice, so as to guide [4] Lops, P., de Gemmis, M., & Semeraro, G, “Privacy-awareness in
high-quality responses of large models. The second is to recommender systems: State-of-the-art and challenges,” In
continuously enrich the training set, ensure the diversity of Proceedings of the 5th ACM conference on Recommender systems,
2011, pp. 29-36.
the training set, and enrich the exploration space of users as
[5] McSherry, F, “ Measuring information disclosure in recommender
much as possible. systems,” In Proceedings of the 1st international workshop on
Privacy enhancing technologies (PET 2005), 2005, pp. 8-14.
ACKNOWLEDGMENT [6] Smith, M. D., “Two decades of privacy enhancing technologies,”
Computer, 2011, 44(1), pp. 13-14.
This work was supported by Hainan Provincial Natural
[7] ZeLinAi, https://www.zelinai.com/, 2023.
Science Foundation of China(621RC1084) .
[8] lijigang, How to write Prompt: Structure,
https://www.lijigang.com/posts/chatgpt-prompt-structure/, 2023.
REFERENCES [9] CRISPE framework, https://github.com/mattnigh/ChatGPT3-Free-
[1] Bobadilla, J., Ortega, F., Hernando, A., & Gutiérrez, A, “Knowledge- Prompt-List.
based recommender systems,” Recommender systems survey, [10] The river tree in the cloud, https://www.zhihu.com/people/zphyix,
Springer, 2013. 2022.
95
Authorized licensed use limited to: Sai Vidya Institute of Technology. Downloaded on April 10,2025 at 10:58:43 UTC from IEEE Xplore. Restrictions apply.
