Sentimental Analysis on YouTube Video Platform

Meda Sai Harini Rekha

, L. Rajeswari

, V. Leena Parimala

B. Charitha

and A. Divya Hari Chandana

Department of CSE(AI), Ravindra College of Engineering for Women, Pasupula Village Nandikotkur Road 518002,

Andhra Pradesh, India

Department of CSE, Ravindra College of Engineering for Women, Pasupula Village Nandikotkur Road 518002, Andhra

Pradesh, India

Keywords: YouTube Analytics, Machine Learning, Audience Engagement, Video Performance, NLP, Content Trends,

Data‑Driven Framework.

Abstract: Due to its great success rate, YouTube is increasingly used as a platform for content creation and consumption.

Now, it has become more of a necessity to identify user behavior, content trends, and engagement patterns.

This study has a main feature which is a machine learning technique that uses YouTube data to provide deep

insights into video performance, audience preferences, and the growth strategies of channels. The machine

model is then used through which various features such as views, likes; comments, watch time, and video

metadata are extracted to predict the key metrics of the success of the video, the engagement rates, and the

audience retention. The system also makes use of some clustering techniques so as to define content categories

and user behavior patterns. On the other hand, regression and classification models rely on data with a track

record to assist them in predicting the success of a video. Similarly, the text analysis on comments completed

by users allows them to know their feelings, which in turn, helps them improve their content strategies. And

the deep learning methods of Natural Language Processing (NLP) are perfected for keyword optimization

based on the video descriptions and titles. To confirm the validity of our approach, we run some experiments

that actually show the models capability to be very useful for content creators, advertisers, and platform

admins. This take uses a data-driven framework that helps in the process of optimizing content creation, which

in turn aims to enhance viewer engagement and improve the total user experience on YouTube.

1 INTRODUCTION

YouTube has become one of the largest platforms for

video content creation and consumption, attracting

billions of users worldwide and thus enjoying big

growth. This content to YouTube's rapid expansion

would be the best avenue for business holders and

media managers seeking proper advertisement

structures. It is also the means of finding a solution to

a universal rating system for videos, make videos

easier to be found in the query or automatically

generate tags for videos.

This research article is dedicated to looking into

the application of artificial intelligence (AI)

techniques specifically aimed at YouTube data to

discover patterns and forecast statistics like the

number of times a video is watched, user retention,

and engagement. According to the results, we will

offer a way or model to assist content creators and

platform administrations in making data-driven

decisions. Our research also involves the application

of Natural Language Processing (NLP) to perfect the

content through a thorough study of the keyword

analysis done in the video descriptions and titles. The

video metadata includes details such as both location

and language chosen. Regarding user interactions, the

likes, comments, watch time, and sentiment feedback

provide information on the user's perception and

engagement. we can compare the content of this

description with that of another video under the same

category and the content of the target video and the

correlation of textual features of the two videos which

can help us make the right conclusions about the

similarities and differences between them.

An increased population has brought the youth

into a competitive environment. YouTube is a space

for e-learning presenting inexpensive and quality

course materials from educational institutions. Young

minds like the available free content on YouTube as

842

Rekha, M. S. H., Rajeswari, L., Parimala, V. L., Charitha, B. and Chandana, A. D. H.

Sentimental Analysis on YouTube Video Platform.

DOI: 10.5220/0013874400004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

842-849

ISBN: 978-989-758-777-1

it is to the contrary of the amount needed to join

coaching institutes. Some educational tutorial series

or marathons are accessible only on YouTube and

may be the case that one of the students prefers this

type of content to others. This is based on the prior

knowledge of the specific situation. The students who

have already watched the series or marathon and they

are beginners, intermediate and professional people

tell their thoughts on the quality and the usefulness of

the visual. The manner of comments posted by the

viewers, the number of likes, and the views of the

videos are all factors in the assessment. The project

can figure out the best videos of YouTube as there are

sentiments of comments, number of comments,

number of views, and number of likes to enable the

personalized sorting of YouTube according to its

ranking. We have received data on the videos and

extracted data like comments, number of comments,

likes and views with the help of the YouTube API.

Through its growing popularity YouTube has

become a crucial learning platform for e-learning

purposes that provide uncharged and superior

educational materials. The platform provides students

with free educational tutorials and courses which

serve as a suitable alternative to costly coaching

institutes. The exploration of educational media

content depends on the assessment of user reaction

alongside video performance analytics and comment-

based feedback evaluations. Several quantitative

measures along with sentiment tracking serve to

determine educationally successful YouTube videos.

2 RELATED WORKS

Many researches have been done which took the help

of YouTube and machine learning to spread out the

aspects of data analysis. One of these techniques is

sentiment analysis, which has been utilized largely to

comprehend the opinions of users about videos.

Between the lines, Wang et al. (2018) inserted a

sentiment analysis tool to YouTube comments to

grasp the information on the audience and the

reactions to the content. Like-wise, other researchers

have applied various clustering algorithms to transit

YouTube content, such as the tags, title, and the user

actions related to the video (e.g., likes, shares, and

comments) to a different cluster. Besides, machine

learning modeling from regression has been engaged

in the prediction of video popularity and engagement

by the historical performance metrics (Sharma &

Hoda, 2020). Additionally, the NLP methods were

exploited to scrutinize the titles and descriptions of

the videos for the keyword optimization purpose (Das

et al., 2021). Although a lot has been accomplished in

the area of YouTube data analysis, the present study

incorporates multiple machine learning strategies and

techniques, including clustering, regression,

classification, and sentiment analysis, to come up

with a holistic approach to content optimization and

user engagement analysis.

Data Collection and Sentiment Analysis: The

collection of data originates from YouTube through

its API. The data obtained includes video metadata

which comprises titles, descriptions, and tags along

with views, likes, comments, and watch time. We

cleaned and prepared the gathered data through

normalization techniques and handled all cases of

missing information. Sentiment analysis permits

discrimination of user comments into positive and

negative and neutral categories after writing the

comments through this approach. Subsequently we

utilize teaching tools that include classic BERT as

well as Support Vector Machines (SVM) or

Predictive Analytics.

Success predictions along with the organization of

content and clustering process guide our approach:

Our analytics include K-Means and DBSCAN

clustering method to connect similar published videos

through their metadata features as well as user

interaction data. Our system uses this method to

determine how parallel or functionally similar videos

perform. The team develops predictive models which

generate success calculations for videos. The team

builds two different types of predictive models where

the first one predicts watch time and views while the

second one uses user engagement to determine

successful videos.

Evaluation Keyword Optimization and Model

Evaluation Using Natural Language Processing

technology the system finds vital keywords that

appear in video title and description content to

optimize video accessibility. The Word2Vec

technology enables the system to create additional

keywords to boost search engine rankings. The

assessment model includes accuracy and precision

tests as well as metric evaluation of recall and F1-

Score scores.

3 METHODOLOGY

3.1 Overview

Modern machine learning techniques process

advanced sentiment analysis of YouTube information

to detect emotional responses from users together

with their active user metrics. The combination of

Sentimental Analysis on YouTube Video Platform

843

text-based comments with video frame visuals

enhances system sentiment recognition abilities to

better help YouTube content creators and

administrators create better content strategies.

3.2 Theoretical Structure

Figure 1: Sentiment Analysis Pipeline Using Machine

Learning.

3.3 Data Collection

As Figure 1 explains the first step, extracting all

mentioned data points constitutes the YouTube data

collection process. Sentiment analysis of user

emotions relies on the system to extract comments

made by users in their video uploads. The system uses

video metadata from combined metrics including

views and likes together with shares and watch time

data that applies to channels specifically. The

sentiment analysis benefits from video descriptions

and their titles as textual information. The video

frames function as specific visual material used for

performing multimodal sentiment analysis.

3.4 Preprocessing

Multiple critical operations must be performed during

the first step of data preprocessing. The text data

processing method begins with tokenizing comments

and descriptions and then eliminating stop words

before establishing text cleanliness. The extraction

methods periodically take video frames to acquire

meaningful visual elements that influence sentiment

analysis results.

3.4.1 Feature Extraction

Three sequential text extraction procedures transform

the features by using TF-IDF along with Word2Vec

analytical techniques for deep semantic processing.

Visual features in video frames are extracted using

pre- trained Convolutional Neural Networks (CNNs)

as the network-based tool for image features

extraction.

3.5 Sentiment Analysis Models

The system employs a combination of models for

improved sentiment classification:

Transformer-Based Models: BERT and T5 along

with GPT-3 undertake YouTube-specific data

training for analyzing sentiment expressions within

user comments and video descriptions. The detection

of complex sentences combined with both subtle

emotional expressions and verbalization that

approaches sarcastic tones makes these models

outstanding performers. Through its Attention

Mechanism function the system identifies crucial

document portions which enhance its detection

accuracy of sentiment.

3.5.1 Multimodal Sentiment Analysis

Through CLIP (Contrastive Language-Image

Pertaining) both text data and video data merge to

allow the connection of video frames with the

matching YouTube comments. This system delivers

enhanced sentiment comprehension through text and

image analysis since visual content influence how

people interpret sentiment in specific video materials.

3.5.2 Reinforcement Learning

User-created feedback activates reinforcement

learning optimization of analyses by employing

instantly activated reactions including comment likes

as well as dislikes and sharing functionalities and

direct messaging responses to achieve repeated use

accuracy enhancement. Through this feature the

model identifies how user needs change and what

fresh local behavior patterns develop.

3.6 Sentiment Classification

The text sentiment classification system employs an

LSTM network with attention mechanisms to extract

positive and negative along with neutral sentiments as

well as joy and anger and surprise emotional

outcomes. Sentiment classification with CLIP

enables the system to examine the visual cues in video

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

844

frames along with textual sentiment in comments for

an integrated content feedback analytical approach.

3.7 Output and Visualization

All video feedback is split into four sentimental

sections by the system which delivers unique

assessment of individual emotions through detailed

analytic reports. Users who utilize the tool receive

critical points of understanding regarding content

reactions through reports that merge crowd

engagement metrics with sentiment data extracted

from comments while displaying temporal sentiment

changes. Content creators and advertising staff along

with administrators access actionable strategies

because of sentiment analysis and viewer engagement

which help them improve their content direction and

develop better user experiences for elevated

engagement metrics.

3.8 Real-Time Processing

The system issues real-time analyses about YouTube

data directly to content creators through its immediate

operational mechanism. Both new material intake and

audience sentiment alteration and emerging trend

detection receive essential monitoring services

through this system.

3.9 Benefits of System

The proposed sentiment analysis system generates

important advantages that benefit both platform

management staff as well as content producers and

advertisers. The following details the main

advantages for each stakeholder segment:

3.9.1 For Content Creators

Content creators use audience sentiment analysis to

receive better content direction since they better

understand what establishes maximum viewer

engagement for creating future content accordingly.

Yahoo produces two advantages for content creators

who receive both helpful feedbacks alongside

audience motivation that eventually enhances their

future video production.

The system provides immediate sentiment

evaluation to content makers who get real-time

viewer feedback about their newly posted content.

The system allows fast upcoming content updates

combined with immediate responses to viewer

feedback through comments.

Content creators benefit from audience emotion

identification to build future content strategies

between joy and surprise experiences. Viewer

emotional understanding helps content creators to

establish better ways of connecting with their

audience. Content creators need to detect elements in

their videos that produce positive emotions for

building lasting viewer loyalty.

3.9.2 For Advertisers

The ability to conduct sentiment analysis helps

advertisers show targeted ads to viewers by assessing

positive reactions from emotions that result from

video content. Advertisers benefit from YouTube

sentiment analysis because they can evaluate their

brand perception from user comment sentiment

evaluations on their uploaded videos. Marketing

plans and approaches benefit from the obtained

information to modify themselves in real-time

operations. Advertisers employ sentiment trends in

conjunction with video emotional responses to

determine their precise ad placement opportunities for

finding the most receptive audience at their ideal

times.

3.9.3 For Platform Administrators

Administrators can discover concerning or

inappropriate content through Managed Content

Moderation by using patterns of negative sentiment

analysis and detection of harmful user contributions.

Such methods enable the development of positive

supportive conditions which spread throughout

YouTube.

The video sentiment monitoring function of

platform administrators detects mounting trends by

showing changes across content areas and indicates

which areas require promotional action or additional

monitoring initiatives. The platform operates with

authentic connectivity to altering user preferences

through this system. System administrators track user

emotional reactions to produce data that aids

YouTube UX strategy improvements. Platform

administrators should use system modifications along

with improved recommendations and content quality

evaluations for their platform improvements.

The suggestions function of YouTube depends on

sentiment analysis to refine how users find content. A

platform provides better video suggestions for users

by analyzing emotions since this approach helps

recommend content aligned with personal emotional

outcomes.

Sentimental Analysis on YouTube Video Platform

845

3.9.4 For Viewers

Through sentiment analysis optimization content

creators join platform administrators to develop better

curated content which fulfills personal and emotional

interests of viewers.

The recognition value of viewer feedback

(including comments and reactions) will enhance

through better understanding because content creators

and brands learn the underlying sentiments behind

reactions and comments.

Users who take content recommendations based

on sentiment find videos that match their emotional

tastes and preferences better thereby improving their

viewing session.

3.9.5 For Data Scientists and Researchers

Researchers can conduct superior analysis of user

activities when sentiment data tracks how users

interact with videos as well as their emotional

responses and behavioral tendencies. The exploration

of new research fields focuses on both studying user

activities on the internet and comprehending social

media operational functions.

YouTube and social media networks receive

enhanced sentiment-analysis capabilities through

joint implementations of transformer-based models

with multimodal analysis systems that researchers

develop to create effective solutions in this domain.

4 RESULTS AND EVALUATION

4.1 Model Evaluation

A collection of experiments has been performed to

confirm the effectiveness of the implemented

methodology. The regression and classification

models undergo K-fold cross- validation as part of

their evaluation to achieve general validity while

preventing model over fitting. The appropriate

metrics analyze the performance of the created

models. The metrics of Mean Squared Error (MSE)

together with R- squared assess the prediction

accuracy when measuring continuous metrics such as

views and watch time in the Regression model. The

success prediction models for videos are evaluated by

means of Accuracy and Precision, Recall, and F1-

Score metrics. The content clustering quality will be

assessed using the Silhouette score and Davies-

Bouldin index. The system's genuine effect on

YouTube channels gets measured through hands-on

tests executed on selected YouTube pages. The

usefulness of the system for video enhancement and

user engagement receives evaluation through surveys

conducted against content creators and advertisers.

Different experiments have been conducted to

verify the performance of the implemented method

for optimizing YouTube video success rates. All

regression models and classification models follow

K-fold cross-validation as their evaluation method.

General validity emerges from this testing procedure

because the models evaluate different data subsets

which protects against over fitting and establishes

reliable performance outcomes. The results of the

model performance evaluation through K-fold cross-

validation depend on different training and validation

sets to confirm the models do not yield biased

conclusions from a single training dataset. The

evaluation of regression models relies on prediction

of continuous performance metrics including views

together with watch time and multiple other

quantitative video metrics. The prediction accuracy

assessment includes an evaluation using Mean

Squared Error (MSE) along with R- squared (R²).

Using MSE allows researchers to calculate average

squared deviations between actual observations and

predicted data thus measuring model precision. R-

squared helps researchers determine the extent which

independent variables explain dependent variable

variability because it calculates the portion of

dependent variable variance which is predictable

from independent variables.

Classification models utilized for video success

prediction are evaluated through multiple

classification metrics which determine their

performance levels. The effectiveness of models can

be measured by Accuracy, Precision, Recall, and F1-

Score since each metric delivers a distinct aspect of

performance assessment. Accuracy measures correct

predictions as a whole while Precision identifies the

number of accurate positive results among all

potential positive results. Providing a balanced

assessment for unbalanced classes the F1-Score

calculates Precision and Recall to generate their

harmonic mean while Recall determines model

capability to find true positives among possible

positives. The evaluation of content clustering quality

depends on unsupervised learning metrics including

the Silhouette score and Davies-Bouldin index that

measure clustering quality in video segmentations for

similar audience attraction. The Silhouette score

assesses cluster separation quality by considering

how similar cluster objects are within a group and

how different cluster groups remain from one

another. Figure 3 talks about the key word

optimization.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

846

The Davies-Bouldin index evaluates cluster

quality by comparing intra-cluster distances to inter-

cluster distances and generates better clustering when

the score decreases shown in Figure 5. The YouTube

channel analysis moves beyond theoretical evaluation

because researchers test the system directly through

manual evaluations of chosen YouTube video pages.

Temperate assessments examine how the system

operates effectively in actual performance and

produces better channel engagement combined with

raised view counts and improved channel metrics.

Applying system-generated predictions to video

content allows for assessment of changes in audience

retention rates and channel growth performance and

click-through rate performance. The utility

assessment of enhancement capabilities and user

engagement for system users and content creators

involves survey analysis. The surveys evaluate the

perceived value of system predictions that help

content strategists improve their approach and boost

video visibility while engaging their audience better.

Survey responses provide qualitative findings about

system performance while measuring how well it

serves YouTube creators and advertisers in their

objectives so researchers can identify its practical

value through a complete evaluation shown in Figure

4. The combination of statistical methods and

machine learning model assessment together with

practical field tests creates a methodology that

provides scientists and practitioners with sound and

efficient results for both video publishers and

advertisers on YouTube. Figure 2 shows the top 10

trending videos in YouTube.

Figure 2: Top 10 Trending Youtube Videos.

Figure 3: Tells About the Key Words Optimization.

Figure 4: Engagement Metrics.

Figure 5: Clustering of the Youtube Video.

5 DISCUSSION

Research reveals that people are increasingly

watching online video material which causes them to

use conventional TV less frequently. The shift toward

digital media has different rates of adoption because

population segments from various economic

backgrounds together with age-based choices shape

how quickly the process unfolds.

The results show that online video quality does

not impact users' plan to drop traditional TV but

strong evidence exists that increased digital

engagement leads them to choose interactive media.

Standard viewers opt for entertainment content that

adjusts to their preferences and offers interactive

elements and matches their viewing needs according

to this research. Users who spend more time with

Sentimental Analysis on YouTube Video Platform

847

online content develop interests in video creation and

content co-operation together with interactive media

consumption.

The research demonstrates that conventional

television maintains some presence in Yemen despite

its inadequate adjustment to modern technological

developments. Televisual networks have encountered

additional problems because they have been unable to

create a presence on digital platforms. Traditional

broadcasters must integrate with new media trends

since their survival depends on it. Modern media

trends indicate that TV viewing will likely decrease

because people are attracted to customizable digital

media content.

Restrictive economic factors have substantial

impact on this transformation. The slower rate of

conventional television decline in Yemen results

from its limited disposable income and cultural

preference to watch traditional broadcasting through

older generations. The data indicates that although

people watch more online videos each year the

medium has not achieved dominance in the Yemeni

market.

Traditional TV networks need to transform their

content approach for youth viewers through

innovations which appear on their preferred

platforms. The popularity of flexible online video

platforms which offer wide content diversity

motivates conventional broadcasters to transform

their business models. Platform evolution through

technological change demonstrates that new

technologies do not replace existing ones but instead

create competition against their market standing.

Traditional TV in Yemen will survive only

through implementing digital strategies and

interactive features and successfully targeting

changing audience preferences. If traditional TV

systems neglect integration with digital strategies

their presence will eventually fade out while digital

media claim the entire market in the forthcoming

years.

6 CONCLUSIONS

The study develops a whole system based on machine

learning methods to explore YouTube information

that addresses sentiment analysis and content

optimization alongside engagement prediction.

Research outcomes demonstrate how machine

learning algorithms deliver important analytical

information which benefits creators together with

marketers and YouTube platform administrators.

Stakeholders who use these insights will be able to

develop better content tactics and create more

successful user relationships and enhance their video

performance metrics. The research restrictions today

create prospects for upcoming advancements in

YouTube analytics analysis and content improvement

methods.

The combination of machine learning analysis for

user data and emotional detection enables both

marketing staff and YouTube staff and video creators

to improve their productivity. Conclusions from the

system analysis boost strategies for content along

with marketing capabilities and audience connection

methods. The application of sentiment analysis

allows video personalization so the system performs

better and provides users with more customized

content selections. The implementation of deep

learning platforms evaluating various platforms

offers better recommendations to develop future

systems. Through the management system content

creators obtain complete control that lets them create

improved bonds with their audiences.

REFERENCES

Almeida, T. A., & Silva, A. (2013). Text Classification in

the Twitter Domain. Proceedings of the 2013

International Conference on Computer Science and

Software Engineering, 213–219.

Almeida, T. A., & Silva, A. (2013). Text Classification in

the Twitter Domain Proceedings of the 2013 Internatio

nal Conference on Computer Science and Software

Engineering, 213–219.

Boulanger, P., & Poncelet, P. (2016). Sentiment Analysis:

A Literature Survey. Journal of Computer Science &

Technology, 31(2), 413–426. Preoţiuc-Pietro, D., &

Cohn, T. (2016). Affective Text: The Role of Emotions

in Text Analysis. Proceedings of the 54th Annual

Meeting of the Association for Computational

Linguistics (ACL 2016), 1–10.

Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018).

BERT: Pre-training of Deep Bidirectional

Transformers for Language Understanding. arXiv

preprint arXiv: 1810.04805.

Hu, M., & Liu, B. (2004). Mining and Summarizing

Customer Reviews. Proceedings of the 10th ACM

SIGKDD International Conference on Knowledge

Discovery and Data Mining, 168–177.

Johnson, R. D., & Zhang, Y. (2015). Effective Use of

Word Order for Text Categorization with Convolution

al Neural Networks. Proceedings of the 24th Internatio

nal Conference on Machine Learning (ICML), 2459-

2467.

Joulin, A., Grave, E., Mikolov, T., & Bojanowski, P. (201

7). Bag of Tricks for Efficient Text Classification.arXi

v preprint arXiv: 1607.01759.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

848

Pang, B., & Lee, L. (2008). Opinion Mining and Sentimen

t Analysis. Foundations and Trends in Information

Retrieval, 2(1-2), 1–135.

Preoţiuc-Pietro, D., & Cohn, T. (2016). Affective Text: The

Role of Emotions in Text Analysis. Proceedings of the

54th Annual Meeting of the Association for

Computational Linguistics (ACL 2016), 1–10.

R. F. Alhujaili and W. M. Yafooz, "Sentiment analysis for

YouTube videos with user comments", 2021 Internatio

nal Conference on Artificial Intelligence and Smart

Systems (ICAIS).

S. Siersdorfer, S. Chelaru, W. Nejdl and J. San Pedro,

"How useful are your comments, analyzing and predic

ting YouTube comments and comment ratings", Proce

edings of the 19th international conference on World

wide web.

Vader Sentiment (2014). A Sentiment Analysis Tool for

Social Media. Proceedings of the 8th International

Conference on Weblogs and Social Media (ICWSM).

Vinyals, & Le, Q. V. (2015). A Neural Network for Machi

ne Translation, at Scale. Proceedings of the 33rd Inter

national Conference on Machine Learning (ICML)

Zhang, Y., & Wallace, B. C. (2015). A Sensitivity Analysi

s of (and Practitioners’ Guide to) Convolutional

Neural Networks for Sentence Classification. Proceedi

ngs of the 8th International Conference on Empirical

Methods in Natural Language Processing (EMNLP).

Zhao, T. (2020). Sentiment Analysis and Text Classificati

on: A Survey Machine Learning and Knowledge Extra

ction, 2(1), 1–35.

Sentimental Analysis on YouTube Video Platform

849