Investigation on the Self-Improving Algorithm of TikTok Based on
Extensive User Interactions
Xiaoxing Chen
a
Department of Computer and Science Engineering, The Ohio State University, Ohio, U.S.A.
Keywords: Tiktok, Social Media, Self-Growing, Tag Analyze, Time-Dependent Content Recommendation.
Abstract: The ubiquity of short video apps in contemporary society is epitomized by the widespread adoption of TikTok
on mobile devices. The platform's escalating user rates and engagement duration are indicative of its growing
influence. This paper investigates the TikTok algorithm's ability to process extensive data sets to curate and
recommend user-preferred content. Conducted through a series of surveys and analytical studies across
various age demographics within university populations, this research emphasizes the pivotal role of metadata
tags and the platform's autonomous enhancement algorithms. By harnessing advanced machine learning and
artificial intelligence technologies—such as Graph Neural Networks (GNN), Reinforcement Learning (RL),
Temporal Convolutional Networks (TCN), Natural Language Processing (NLP), Generative Adversarial
Networks (GANs), and Attention Mechanisms—TikTok effectively tailors its algorithmic learning to user
interactions. This strategic integration allows for the progressive refinement of user recommendations,
fostering personalized content delivery while ensuring privacy, and enhancing the overall quality of content
and user engagement. The study's findings reveal that these technological integrations enable TikTok to more
accurately discern user preferences, thus facilitating the delivery of more engaging and relevant content.
Ultimately, these improvements have substantial implications for the enrichment of user experience on the
platform.
1 INTRODUCTION
TikTok was started in September 2016 and expanded
to the United States in 2017 before spreading to 150
countries. TikTok has experienced a meteoric rise in
its user base. By February 2024 TikTok had amassed
4.7 billion downloads. The platform's short video
format resonates with users by capturing moments.
Hashtags play a role in driving trends and content
recommendations on TikTok, which evolve over
time. Continuous research and analysis are crucial for
understanding these dynamics. Through an
examination of label usage trends during periods I
have personally observed shifts in content
suggestions, on TikTok. Reflecting on my experience
of encountering food related videos on the app
prompted me to delve into TikTok intricate data
algorithms. In today's world it's common to see
people of all ages. Whether young people, to adults,
or even children. Spending an amount of time on
social media platforms like video apps. According to
a
https://orcid.org/0009-0000-2464-8117
data from Datareportal it is projected that by 2023
individuals will allocate about 35.8% of their time to
media activities (DataReportal, 2024). Surprisingly
this surpasses the time spent on sleeping or working
for some individuals. Handling data requires robust
algorithms for support. An algorithm can be
described as a procedure embedded within a
computer system that aids in problem solving and
calculations (Amjady, 2001). The algorithms utilized
in media platforms are advanced and efficient(Ma,
2014). Platforms like TikTok utilize user data such as
information, engagement levels, location details,
viewing habits and other relevant data to tailor
content for users with precision(Ma, 2014).
2 DTAT SURVEY
This paper finds that more than 110 college students
who have been using social media for at least 5 years,
random school professors and ordinary passers-by to
Chen, X.
Investigation on the Self-Improving Algorithm of TikTok Based on Extensive User Interactions.
DOI: 10.5220/0012925100004508
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 227-233
ISBN: 978-989-758-713-9
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
227

conduct a questionnaire survey to get their opinions
on what TikTok’s big data analysis is based on. In the
questionnaire, the time spent on social media is
included, do you authorize the location of the social
media you use, do you think the videos TikTok sends
you are your favorite and so on. For anyone who has
been on social media for less than a year (and very
few), this paper chose to ignore their responses. They
also ignored extreme responses, such as 99 years of
social media use. The questions included in the
questionnaire include age range, gender, daily use
time, main use period, main use purpose, algorithm
accuracy, algorithm basis, understanding of
algorithm working principle, data and privacy
processing, and an open question about how Tik Tok
will improve its algorithm or user experience in the
future.
This paper analyzes the data gathered from a
survey questionnaire as shown in Fig.1. Focusing on
the sections previously introduced in Data collection
let's delve into each of them individually. Firstly，
regarding age range is the first part of the
questionnaire. The majority of participants (70%) fall
within the 18-24 age bracket aligning with college
student demographics. Secondly in terms of gender
distribution is next. It's fairly balanced with 45%
male, 45% female, 10% genders and those preferring
not to disclose. Thirdly, let’s can look at usage time.
The highest percentage falls within the range of 30
minutes to 2 hours indicating usage patterns among
college students. Fourthly, it is the social media usage
time. Most TikTok activity occurs during the night
(30%) followed by afternoon and morning slots.
Fifthly, let’s analyze the primary usage purposes
which after collect. Entertainment (40%) and
educational information seeking (30%) are
highlighted as reasons for using TikTok—a platform
that serves both informative and entertaining roles
effectively. Sixth, algorithm accuracy assessment is
the part people care so much. A large portion finds
the algorithm accurate" (40%) or "average" (30%)
suggesting high user satisfaction, with the
recommended content. Seventhly, the algorithms
foundation lies in user engagement (35%). Time
spent viewing content (30%) which are considered
the influences on algorithmic suggestions.
Additionally, when it comes to comprehending how
algorithms operate the majority of individuals possess
a " understanding" (35%) or a "moderate
understanding" (25%) suggesting a restricted
awareness of algorithm functioning. Lastly, data and
privacy handling is the thing that have to analyze.
TikTok’s handling of user data and privacy is mostly
"average" (40%), indicating that there is room for
improvement.
Figure 1: Data Chart of questionnaire survey.
This paper also collects and analyzes a question at the
end of the questionnaire. This paper divided them into
10 broad categories and analyzed them briefly. Let us
expand it one by one. The first thing that need to know
is More personalized content recommendations.
Students may want algorithms to better understand
their interests and needs and provide more relevant
video content. The second one is Increased privacy
protection which is the thing that people care so
much. With the increasing emphasis on personal
privacy, many students may expect platforms to
provide stronger privacy Settings and transparency.
The third one is Reduce advertising and promotional
content. Students may want to encounter fewer ads
while scrolling through videos to provide a smoother
user experience. The fourth one is Optimize the
interface and interaction design which is intuitive
display in front of the user. Some users may mention
EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence
228

that they want a more intuitive and user-friendly
interface, including improved search and video
classification. The fifth one is Provide more
education and knowledge content. As students, they
may want to see more education related content, such
as professional knowledge, study skills, etc. The sixth
one is Add content creation tools and features which
can attract more target users. Expect more innovative
video editing tools
and creation features to encourage
creative expression. The seventh one is Strengthen
community and interactive features. Students may
want better comment and interactive features to
promote community building. The eighth one is to
Improve fairness and diversity of algorithms.
Algorithms are required to fairly promote creators
from different backgrounds and cultures. The ninth
one is to Improve content moderation systems. Avoid
blocking or restricting harmless content by mistake
and remove harmful content quickly. The last one is
to support multilingual and cross-cultural
communication because the world today is very
diverse. Multilingual users may request improved
translation and subtitling features to better enjoy
internationa1l content.
3 TIK TOK ALGORITHMS
3.1 The Role of Tag in Massive
Algorithms
Based on the above data analysis, this paper found
that people's opinions about TikTok’s algorithm are
mainly based on the definition of labels. In fact, tags
must occupy the core foundation of TikTok’s
algorithm(Klug, 2021). Users generally believe that
cumulative tags, such as hashtags, are one of the
factors that influence algorithm recommendations.
The study empirically confirmed that video
interaction, release time, and the use of tags are the
key factors that improve the probability of video
recommendation. TikTok assigns descriptive tags to
new videos based on computer vision analysis, tag
mentions, post descriptions, sound effects, and
embedded text. These tags are then used to align
videos with groups of users whose interests match,
optimizing the recommendation process(Klug, 2021).
According to Fig.2, the data suggests that the best
times to post on TikTok are Tuesday, Wednesday and
Thursday from 2pm to 5pm (Sprout Social, 2024). In
this way, uploaded videos have a better chance to be
seen by more people, because it happens that more
than 50 percent of people in the questionnaire survey
use short video apps such as TikTok in the afternoon
and evening.
Use 15-point type for the title, aligned to the center,
line space exactly at 17-point with a bold font style
and initial letters are capitalized. No formulas or
special characters of any form or language are
allowed in the title.
Figure 2: TikTok Global Engagement chart from
SproutSocial.
3.2 Time-Dependent Content
Recommendation
Understanding the Complexities of TikTok's Time-
Dependent Content Recommendation Strategy. In
today's digital age, social media platforms have
become an integral part of people’s lives, with
TikTok leading the way in short-form video content.
Behind the scenes, TikTok's algorithm works
tirelessly to provide users with personalized and
engaging content. However, the intricacies of its
time-dependent content recommendation strategy go
beyond what meets the eye. To effectively curate
content based on users' preferences, TikTok's
algorithm incorporates various disciplines such as
data science, sociology, marketing, and computer
science. In a recent study, seven key aspects were
identified that significantly impact the platform's
time-dependent content recommendations.
User Behavior and Interaction on TikTok involve
a complex algorithm which takes into account users'
past activities and viewing habits. By analyzing your
history on the platform, including the types of content
you engage with and when, the algorithm adjusts the
push content accordingly. This explains why you
might find yourself scrolling through mouth watering
food videos during dinner time.
Time-Sensitive Content Strategy is content
creators on TikTok understand the significance of
Investigation on the Self-Improving Algorithm of TikTok Based on Extensive User Interactions
229

timing. They recognize the topics that are likely to
resonate with users during specific time periods. For
instance, pushing food-related content during
mealtimes takes advantage of users' daily activity
patterns and their interest in culinary delights.
Cultural Habits and Algorithm Design is the thing
that must pay attention on. Food cultures vary across
regions, and TikTok acknowledges this diversity.
Algorithms are designed to reflect cultural habits,
recommending specific types of content at specific
times. This ensures that users are more likely to
explore and engage with content that aligns with their
cultural preferences.
Personalized Recommendations and Group
Behavior on TikTok's algorithm not only considers
individual user behavior but also analyzes large-scale
data to identify broader patterns of user behavior. By
understanding that most people crave food content in
the evening, the algorithm pushes relevant videos
during this time, catering to the collective interests of
the user community.
Real-Time Trends and Content
Recommendations can be expanded in the algorithm
captures real-time trends and adjusts the
recommended content accordingly. If a particular
food topic suddenly gains popularity in the evening,
the algorithm quickly increases the feed of videos
related to that topic, ensuring users stay up to date
with the latest trends.
Social Media Marketing Strategies is merchants
and brands strategically deliver food-related
marketing content during peak meal times. TikTok's
algorithm prioritizes this content to attract potential
customers, benefiting both businesses and users
seeking relevant recommendations.
Multimodal Data Analysis on TikTok's
algorithms go beyond text analysis. They incorporate
image recognition, analyzing food images, time
stamps, and geolocation data to provide more precise
and personalized content recommendations. This
level of analysis ensures that users receive content
tailored to their individual tastes and preferences.
3.3 The Self-Growing of Tik Tok
The self-growing nature of TikTok's algorithm is
another fascinating aspect. It constantly optimizes
itself through user interactions, utilizing the
principles of big data and machine learning. While the
specific algorithms used by TikTok remain a closely
guarded secret, it is likely that advanced methods
from the fields of artificial intelligence, machine
learning, data mining, and user behavior analysis
contribute to the platform's content optimization and
user engagement strategies.
Although the exact details of TikTok's algorithm
may remain undisclosed, the platform undoubtedly
relies on cutting-edge technology to provide a
seamless and captivating user experience. As users
continue to engage with the platform, TikTok's
algorithm evolves and adapts, ensuring that content
recommendations remain fresh, relevant, and highly
addictive. The Evolution of TikTok Algorithms:
Unleashing the Power of AI.
TikTok, the popular social media platform, has
captivated millions of users worldwide with its short-
form videos and endless entertainment. Behind the
scenes, the TikTok algorithm works tirelessly to
curate personalized content streams for users, keeping
them engaged and coming back for more. The
secret to TikTok's success lies in its cutting-edge use
of artificial intelligence (AI) and machine learning
techniques. In this article, author will explore some of
the key AI techniques that power the self-growing
TikTok algorithm.
Graph Neural Networks (GNNs) are
revolutionizing the way social networks analyze user
interactions and content. With their ability to simulate
complex relationships and interactions, GNNs enable
TikTok to understand user preferences by analyzing
the global structure of user interactions. This allows
the algorithm to predict which content is most likely
to be relevant and engaging, enhancing the user
experience(Wu et al, 2021).
Reinforcement Learning (RL) on TikTok
leverages reinforcement learning algorithms to
dynamically adjust content recommendations based
on user feedback in real-time. By analyzing metrics
such as likes, shares, and viewing time, the algorithm
learns to optimize content suggestions and maximize
user engagement. Through trial and error, the RL
algorithm continuously fine-tunes its
recommendations, ensuring a personalized and
addictive user experience(Sutton and Barto, 2018).
Federated Learning is the part that TikTok use.
Privacy concerns have become a major issue in
today's digital landscape. TikTok addresses this by
employing federated learning, a decentralized
approach that enables machine learning models to be
trained on user devices without compromising
privacy. With federated learning, TikTok can
personalize content recommendations without relying
on centralized user data, preserving privacy while
delivering tailored recommendations(Konečný et al,
2016).
Temporal Convolutional Networks (TCNs) can
Predicting user behavior over time is crucial for
EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence
230

TikTok's algorithm. TCNs excel at sequence
prediction problems, making them ideal for analyzing
past user activity and predicting future interactions.
By utilizing TCNs, TikTok improves the timing of
content recommendations, ensuring that users are
presented with the right content at the right
moment(Bai et al, 2018).
Zero-Sample Learning is a way to enhance
content discovery and diversify recommendations,
TikTok employs zero-sample learning. This
technique allows the algorithm to recognize objects
or patterns that it has not encountered during training.
By recommending new or niche content without
relying solely on historical interaction data, TikTok
expands the horizons of its content discovery,
surprising and delighting users with fresh and
exciting recommendations(Xian et al, 2017).
Natural Language Processing (NLP) and
Transformer Models are powerful tools. Beyond
video analysis, TikTok harnesses advanced NLP
techniques to understand reviews, descriptions, and
labels associated with content. By utilizing
transformer models, renowned for their effectiveness
in processing linguistic data, TikTok gains a deeper
understanding of the text environment surrounding
videos. This enhances the recommendation engine
and ensures that content suggestions align with user
preferences(Vaswani et al, 2017).
Generative Adversarial Networks (GANs) is
typically associated with content creation, can also
play a role in TikTok's content planning process. By
generating synthetic data, GANs can train more
robust recommendation models while enhancing
user-generated content in innovative ways. This
opens up new avenues for creativity and expands
TikTok's content ecosystem(Goodfellow et al, 2014).
Attention Mechanisms is integrating attention
mechanisms into deep learning models empowers
TikTok's algorithm to prioritize crucial parts of the
data. Whether it's a particular frame in a video or a
paragraph in a user's interaction history, attention
mechanisms help determine what is most important
for making accurate predictions. By focusing on the
most relevant user interactions, TikTok refines the
personalization of content streams, ensuring an
engaging and tailored user experience(Bahdanau et
al, 2014).
Through the application of these AI techniques,
TikTok's algorithms achieve self-growth and
optimization. The continuous evolution of the
algorithm, driven by cutting-edge technologies,
enables TikTok to deliver personalized, addictive,
and captivating content to its users. As TikTok
continues to innovate and explore new frontiers in AI,
users can expect even more exciting experiences and
discoveries on this ever-evolving platform. A
Comprehensive Look at TikTok's Recommendation
Algorithm. TikTok, the popular social media
platform, has gained immense popularity due to its
highly accurate and engaging content
recommendations. Behind the scenes, numerous
technologies work together seamlessly to create a
personalized experience for each user. In this article,
author will explore how these technologies
collaborate to solve problems and enhance user
satisfaction.
The first step in TikTok's recommendation
algorithm is to collect and understand data. Graph
Neural Networks (GNN) play a crucial role in this
process. By simulating complex relationships
between users and content, GNN helps to understand
the interaction structure within social networks(Wu et
al, 2021). This understanding allows the algorithm to
predict with greater accuracy which content will be
most relevant to a particular user's preferences.
Natural Language Processing (NLP) and
transformation models are also vital components of
TikTok's algorithm. NLP is used to parse and
understand user-generated text such as comments and
tags, which provide context for the relevance of
conten(Vaswani et al, 2017). Transformation models
like BERT or GPT dive deeper into these texts,
providing richer data input to the recommendation
system.
Once the data is collected and understood, the
algorithm focuses on real-time optimization and
adaptation. Reinforcement Learning (RL) plays a
significant role in this stage(Sutton and Barto, 2018).
By adjusting recommendations in real-time based on
user feedback, such as viewing time, likes, and
shares, the algorithm increases user engagement. It
learns which content types are most appealing to
users and adjusts its recommendation strategy
accordingly. Temporal Convolutional Networks
(TCN) are also instrumental in this process(Bai et al,
2018). By analyzing user behavior over time, TCN
predicts future interactions and optimizes the timing
of content recommendations.
Personalization and privacy protection are crucial
aspects of TikTok's algorithm. Federated Learning
allows the algorithm to make personalized
recommendations without direct access to user data.
By training models locally on user devices, TikTok
enhances privacy protections while still providing
personalized content recommendations. Diversity of
content and user engagement are essential
considerations for TikTok's recommendation
algorithm. Zero-shot learning allows the algorithm to
Investigation on the Self-Improving Algorithm of TikTok Based on Extensive User Interactions
231

recommend new or niche content that does not appear
in the training data, thereby increasing the diversity
of content discovery. Additionally, Generative
Adversarial Networks (GANs) can generate synthetic
data to train recommendation systems or enhance
user-generated content in novel ways, improving the
diversity and appeal of content. To refine
recommendations further, attention mechanisms
come into play. These mechanisms determine the
most important parts of the data and enable models to
predict users’ interests more accurately. This, in turn,
enables more personalized content recommendations.
By combining all these technologies, TikTok's
algorithms continually grow and optimize themselves
in different dimensions. The algorithm learns from
user behavior, feedback, and interaction patterns,
constantly adapting and improving its
recommendation mechanism with new data inputs.
This dynamic process aims to enhance user
satisfaction and engagement.
TikTok's recommendation algorithm utilizes a
wide range of technologies to provide users with a
highly personalized and engaging experience. From
data collection and understanding to real-time
optimization, personalization, and diversity of
content, each component plays a crucial role in
ensuring that users are presented with content that
aligns with their preferences and interests. As TikTok
continues to evolve, its algorithm will undoubtedly
become even more sophisticated, enhancing user
satisfaction and cementing its position as a leading
social media platform.
4 CONCLUSIONS
In the realm of personalized content recommendation
systems, TikTok has established itself as a true leader.
Leveraging cutting-edge machine learning and
artificial intelligence technologies, TikTok has
revolutionized the way users discover and engage
with content. Through extensive analysis and the
implementation of various advanced techniques, such
as graph neural networks, reinforcement learning,
federated learning, temporal convolutional networks,
zero-sample learning, natural language processing,
generative adversarial networks, and attention
mechanisms, TikTok has elevated the user experience
to new heights. Not only does this optimize user
satisfaction, but it also increases overall engagement
with the app. The combination of these innovative
technologies allows TikTok to gain a deep
understanding of user preferences, ensuring that the
content served is not only relevant but also highly
engaging. This commitment to personalization has
led to a significant diversification and interactivity of
content, all while prioritizing user privacy. As time
progresses and technology advances, TikTok's
algorithmic self-growth mechanism ensures that it
remains at the forefront of the short video platform
competition. By providing users with a personalized
and highly participatory platform, TikTok has
solidified its position as a pioneer in the industry. This
achievement not only showcases TikTok's
exceptional ability to handle vast amounts of data and
optimize recommendation algorithms, but it also
emphasizes its pivotal role in driving innovation and
progress in the digital media landscape. Finally,
TikTok's personalized content recommendation
system has truly revolutionized the way users interact
with short videos. By utilizing state-of-the-art
technologies, TikTok continues to push boundaries
and provide users with a platform that is both tailored
to their preferences and highly engaging. Its
commitment to privacy, innovation, and progress
cements its position as a leader in the ever-evolving
digital media landscape.
REFERENCES
Fangfang., 2011, Research on power load forecasting based
on Improved BP neural network. Harbin Institute of
Technology.
DataReportal., 2024, The time we spend on social media.
Retrieved from https://datareportal.com/reports/digital-
2024-deep-dive-the-time-we-spend-on-social-media
Amjady, N., 2001, Short-term hourly load forecasting using
time series modeling with peak load estimation
capability. IEEE Transactions on Power Systems,
16(4), 798-805.
Ma, K., 2014, Short term distributed load forecasting
method based on big data. Changsha: Hunan
University.
Klug, D. et al., 2021, "An Empirical Investigation of
Personalization Factors on TikTok." arXiv preprint
arXiv:2201.12271.
Sprout Social., 2024, TikTok algorithm. Retrieved [March
29, 2024], from https://sproutsocial.com/insights
/tiktok-algorithm/
Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S.,
2021, "A Comprehensive Survey on Graph Neural
Networks." IEEE Transactions on Neural Networks and
Learning Systems, 32(1), 4-24.
Sutton, R. S., & Barto, A. G., 2018, "Reinforcement
Learning: An Introduction." Second Edition, The MIT
Press.
Konečný, J., McMahan, H. B., Yu, F., Richtárik, P., Suresh,
A. T., & Bacon, D., 2016, "Federated Learning:
Strategies for Improving Communication Efficiency."
arXiv preprint arXiv:1610.05492.
EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence
232

Bai, S., Kolter, J. Z., & Koltun, V., 2018, "An Empirical
Evaluation of Generic Convolutional and Recurrent
Networks for Sequence Modeling." arXiv preprint
arXiv:1803.01271.
Xian, Y., Lampert, C. H., Schiele, B., & Akata, Z., 2017,
"Zero-Shot Learning - A Comprehensive Evaluation of
the Good, the Bad and the Ugly." arXiv preprint
arXiv:1707.00600.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,
L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I., 2017,
"Attention Is All You Need." In Advances in Neural
Information Processing Systems (pp. 5998-6008).
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., & Bengio,
Y., 2014, "Generative Adversarial Nets." In Advances
in Neural Information Processing Systems (pp. 2672-
2680).
Bahdanau, D., Cho, K., & Bengio, Y., 2014, "Neural
Machine Translation by Jointly Learning to Align and
Translate." arXiv preprint arXiv:1409.0473.
Investigation on the Self-Improving Algorithm of TikTok Based on Extensive User Interactions
233