Research on Online Stream Media Marketing Strategy Based on Data
Mining
Yihan Jia
Beijing Institute of Fashion Technology, China
Keywords: Data Mining, Marketing Strategy, Machine Learning, Live Streamers.
Abstracts: With the rapid development of science and technology and economy in China, the online stream industry
relying on the Internet platforms is becoming more and more prosperous. As a new media industry, online
stream industry has characteristics such as low entry thresholds, image virtualization, function positioning,
and entertainment positioning. This paper uses two machine learning algorithms to analyze the potential
relationship between the characteristics of network anchors and the sales volume, and makes descriptive
statistical analysis and correlation test with the sales volume. A sales evaluation model is then built to adjust
the precision marketing strategy based on these characteristics. It is expected to provide a more effective way
to increase the income of enterprises and promote the economic development of online stream industry.
1 INTRODUCTION
"In 2021, the number of webcast users in China has
reached 635 million, and it continues to rise." The
statistics of China's Internet Network Information
Center are enough to prove that the online stream
industry is creating myths. The success of Li Ziqi, Li
Jiaqi, Feng Timo and others has made more and more
Chinese people imagine starting from scratch to
becoming rich. At present, live streamers have
become a "phenomenal" profession. Streamers in
different fields transmit and disseminate relevant
content information. It can be said that the
commercial value of the live streamers online stream
industry is becoming increasingly apparent, and it has
become an important link in the business chain of
enterprises to carry out brand building, consumption
insight and marketing promotion, and is a sharp tool
for contemporary corporate marketing.
This article will study the influence of the
behavioral characteristics of live streaming e-
commerce on its sales volume. On the other hand,
starting from sales and behavior data, combined with
the differentiated needs of groups, segmenting public
customer groups, and finally quantifying the
evaluation index system. The primary research topic
is how to measure and obtain behavioral
characteristics and sales data of live streaming e-
commerce
[1]
. The second thing to do is to extract and
analyze the behaviors of goods, activities and topics
released by live streamers, and conduct in-depth
research on the logic and laws behind the behaviors;
The third is based on the discovery of the second
point, a behavioral feature database is constructed,
and a machine learning model is constructed based on
the basic characteristics of Internet celebrities to
explore the relationship between each behavioral
characteristics of Internet celebrities and sales levels.
Finally, rationalization suggestions are put
forward, hoping to provide some theoretical
references for online celebrities' own operations,
platforms and merchants when choosing resources
for cooperation.
2 ELEMENTS AND
CHARACTERISTICS OF
ONLINE LIVE STREAMING
PLATFORMS
2.1 Analysis of the Characteristics of
Online Live Streaming Platforms
As a live streaming carrier, the live streaming
platforms is a basic and important part of the online
stream industry. The live streaming platforms uses
the media characteristics of digital media to combine
440
Jia, Y.
Research on Online Stream Media Marketing Strategy Based on Data Mining.
DOI: 10.5220/0011913500003613
In Proceedings of the 2nd International Conference on New Media Development and Modernized Education (NMDME 2022), pages 440-444
ISBN: 978-989-758-630-9
Copyright
c
 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

live streaming, small video, pictures, animation and
audio into a new form of streaming. In terms of
communication effect, it is more immersive and
interactive, and it is more likely to be loved by
everyone. It is worth mentioning that the
development and improvement of any emerging
things requires a long-term process, especially in
today's prevailing digital media era, online live
streaming as an emerging thing still has a lot of room
for development, and the country's regulatory
mechanism is relatively not mature and perfect.
2.2 Image Characteristics of Live
Streamers
Live streamers are the main force in the online stream
industry of live streamers from the perspective of
digital media, and the image construction of
streamers plays an important role in enhancing their
commercial value. The streaming content of
entertainment streamers is biased towards real-time
entertainment consumption, mainly through the
streamers's appearance, talent and fan interaction to
attract the attention of users, obtain rewards from
fans, and make profits by commercial activities and
product promotion after accumulating a certain fan
base. E-commerce streamers are similar to the Key
Opinion Leader of the segment, carry out product
promotion in the live streaming room, and
comprehensively show the advantages of the product
and promote sales through the streamers's personal
tasting, try-on, and trial (GUO 2020).
The powerful development of live streaming is
indispensable to the boost of brokerage companies.
MCN (MCN, Multi-Channel Network) is an
emerging business model and Internet format, which
is a new thing brought by online media live streaming
marketing and live streaming e-commerce, and is also
the inevitable result of Internet specialization and
segmentation.
3 RESEARCH ON THE
MARKETING STRATEGY OF
LIVE STREAMERS LIVE
MEDIA FROM THE
PERSPECTIVE OF DIGITAL
INTELLIGENCE
3.1 Data Acquisition and Feature
Extraction
Sources of Data: Considering the selection of market
segments and the definition of research problems, this
article will select the top 102 streamers who appeared
on the daily sales list of Douyin in September from
the third-party data detection platform as the analysis
objects, match their Douyin data, and analyze their
personal data information and behavioral data
information (JIN 2020).
Acquisition of Data: Combined with the list of
Internet celebrities that have been obtained, this
article collects and organizes the data through Python
scripts (Serdar 2017).
3.2 Sales Evaluation Model
3.2.1 Construction of Live Streamers
Features
Before constructing the model, the correlation
between each feature quantity and sales volume is
directly analyzed, and after the normality test of the
sales data, the sales data does not conform to the
normal distribution, so the Spearman correlation
coefficient is used in the correlation test of each
feature quantity (Xie 2019). (See Table 1)
Table 1 Correlation test between the characteristic quantity of goods and sales
Feature classification Feature volume Correlation coefficient
Characteristics of the behavior of
bringing goods
Number of videos -0.015
Video ratio -0.089
Number of video reviews -0.125
Number of video forwards -0.114
Number of video likes 0.078
Research on Online Stream Media Marketing Strategy Based on Data Mining
441

3.2.2 The K-NN Model Evaluates Sales
K-NN algorithm model construction
The K-NN model is mainly used in the classification
of feature space, so the K-NN model mainly has the
following three points: distance measurement, k-
value selection, and classification decision rules (Xi
2017).
Distance metric: Suppose the feature space X is
the n-dimensional real vector space Rn, 𝑥

,𝑥


∈ X,
𝐾


𝑥

,𝑋



=

𝛴


=𝑋


𝑙

−𝑥


𝑙






(1)
For the adjacent k training instance points
forming the set N, the misclassification rate is:
∑𝑥𝑖∈𝑁𝑘(𝑥) 𝐼(𝑦𝑖 ≠ 𝑐𝑗) = 1 - 𝑘1 ∑𝑥𝑖∈𝑁𝑘(𝑥) 𝐼(𝑦𝑖 = 𝑐𝑗)
(2)
In practical applications, if the value of k value
selected is small, it is equivalent to selecting the
training data in a smaller field for classification
prediction, and the approximate error of machine
learning will be small（Yu2004）; Therefore, only
data sets with inputs close to the actual amount of
data will have an effect on the predicted results of the
variables; If the value of k value is large, the learning
error will decrease, but the approximate error will
increase, so that the prediction effect will not play a
role in predicting the distant instance points, making
the prediction result wrong.
3.2.3 K-NN Model Evaluation Results
In table 2, the accuracy of the simple marketing
feature model and the personal attribute feature
model are 0.69 and 0.68, respectively, indicating that
relying solely on marketing characteristics and
personal characteristics of Internet celebrities cannot
comprehensively measure the ability of Internet
celebrities to bring goods (Sheng 2019).
Table 2 K-NN Evaluation indicator output results
Full-time full-
feature model
Personal attribute
characteristic
model
Daily behavior
characteristic
model
Full feature model
of marketing
month
Single
marketing
feature
model
The recall rate 0.44 0.54 0.41 0.39 0.54
Accurate rate 0.71 0.78 0.68 0.65 0.61
F1 0.41 0.36 0.49 0.41 0.39
AUC 0.39 0.41 0.55 0.42 0.50
The daily behavior characteristics model of
Internet celebrities, the accuracy rate is 0.75,
indicating that the daily sharing behavior and past
behavior characteristics of Internet celebrities can
also act on their sales at the same time, so e-
commerce Internet celebrities should not only pay
attention to their marketing behavior, but also pay
attention to the accumulation of their daily behavior
and the maintenance of fan relationships.
3.3 Logistic Regression Model to
Evaluate Sales
3.3.1 Logistic Regression Model
Construction
Sigmoid (𝑝) = 𝑙𝑜𝑔𝑒 (𝑝1-𝑝) (3)
Sigmoid(p)=m, there is:
𝑝=



(4)
The correlation test of the 55 eigens obtained in the
previous article found that the correlation coefficient
between some independent variables was greater than
0.8, and there was obvious collinearity, so the factor
analysis was used to reduce the dimensionality of the
data, the value of KMO was 0.714, according to the
KMO metric, the original variable was suitable for
factor analysis, and the probability of Bartlett's
sphericity test was 0.000, rejecting the null
hypothesis. Based on the total explanatory variance,
when the number of factors is 12, the cumulative
percentage of variance is 84.459%, so new variables
are generated to be placed into the logistic regression
model for sales level prediction.
After each feature is multiplied by a regression
coefficient, the result of multiplying the feature and
the regression coefficient is added into the Sigmoid
function to obtain the value between 0 and 1. If the
data result is greater than 0.5, it is divided into class
1; if the data result is less than 0.5, it is class 0. (See
table 3)
NMDME 2022 - The International Conference on New Media Development and Modernized Education
442

Table 3 KMO and Bartlett Test
KMO sampling
suitability
quantity
Bartlett's
sphericity test
The
approximate
chi-square
Degrees of
freedom
Significant
0.714
17653.148
578
0
3.3.2 Logistic Regression Model Evaluation
Through the introduction of the principle and
algorithm of logistic regression above, this section
still uses python's sklearn to put the feature quantity
after dimensionality reduction into the model to
estimate the sales level and evaluate the predicted
results. The evaluation indicators are as follows:
Fig 1 Logistic regression ROC curve
1. The accuracy rate, recall rate and F1_score of
dichotomy were 0.85, 0.64 and 0.68 respectively. The
accuracy rate of the model is slightly lower than that
of the full-feature model of the marketing month with
the highest accuracy in the KNN evaluation model. It
can be seen that the evaluation effect of KNN is better
than that of the logistic regression model when there
is a large number of features.
2. The ROC curve of the logistic regression model
is shown in Figure 1 below, and its AUC value is
0.85, slightly lower than the full feature of the
marketing month in the KNN model. Therefore,
under this classification, the KNN evaluation effect
of the data cited in this paper is better than that of the
logistic regression model.
Through the K-NN algorithm and logistic
regression to evaluate the sales of Internet celebrities,
and give the evaluation effect, collated to say that the
classification effect of logistic regression under this
classification is higher, the KNN model for each
group of comparative experiments finally reached the
same conclusion, that is, the marketing characteristics
of Internet celebrities and personal characteristics
information complement each other and act together
on the sales of Internet celebrities.
3.4 Problems in the Marketing Process
of Bringing Goods
Data falsification, loss of audience trust：Nowadays,
business transactions such as powder increase, brush
volume, interaction, and likes have become a gray
industry in the industry. And developed functions
such as commenting, liking, and interacting on the
corresponding software according to the rules of each
live streaming platforms, and it was even called
"shaping the perfect live streaming room data".
Brushing traffic and popularity are of great benefit to
streamers and platforms to some extent, so some
platforms will also take the initiative to carry out
capital operations to improve the traffic data in the
live streaming room, so a large amount of spam data
and fake streamers will appear frequently, and will
eventually lose the trust of the audience.
Fakes are frequent：Some streamers one-sidedly
pursue commercial interests in the live streaming
room, especially on some small-scale and poorly
regulated live streaming platforms, and do not fully
verify the quality and efficacy of the promoted goods,
resulting in the continuous emergence of news about
the sale of "three no products" and "fake goods" in
the live streaming room, and in the short term,
businesses and streamers have profited, but in the
long run, it has seriously affected the image of live
streamers in the minds of the audience, disintegrated
the trust of the audience, and adversely affected the
harmonious and orderly development of the Internet
business environment. This has led to damage to the
overall image of the online stream industry.
The number of fans, set the amount of praise,
praise powder ratio are ranked in the top ten
characteristics of the importance of live streamers,
the number of fans represents the personal influence
of live streamers, set the amount of praise, praise
powder ratio represents the quality of the work of live
streamers, and the activity of fans, therefore,
enterprise marketing should pay more attention to the
personal influence of live streamers when choosing
cooperative resources, Internet celebrities and MCN
organizations should also pay attention to the quality
of their works in the daily operation process, and
create more works with high originality and high
quality, so as to enhance the interaction with fans and
enhance the activity and stickiness of fans.
Research on Online Stream Media Marketing Strategy Based on Data Mining
443

4 CONCLUSIONS
This paper classifies the characteristics of Internet
celebrities' marketing behavior, personal attribute
information and other characteristics, and analyzes
the correlation between each characteristic quantity
and sales volume. It is found that the number of
livestream-related videos and the number of fans in
the marketing month have a high correlation with the
sales volume of e-commerce Internet celebrities, but
the correlation is still weak, which is not enough to
use a single variable to predict sales volume.
Based on various data reports and third-party
monitoring platforms, the list of e-commerce
influencers on Douyin platform that meet the research
conditions was obtained. Based on data crawler and
manual recording methods, the personal information
and behavioral data of these influencers on Douyin
platform were obtained as the main research objects.
According to the evaluation results of the
previous model, the marketing characteristics and
personal attributes of the month complement each
other and jointly affect the sales volume of live
streamers. Therefore, when choosing resource
cooperation, enterprises or businesses should not
only consider the current marketing methods of
Internet celebrities, but also refer to their historical
behavior data and personal influence. Internet
celebrities and MCN organizations should also pay
attention to the accumulation of daily behaviors and
the improvement of work quality in the daily
operation process, so as to increase the fans'
stickiness and promote the growth of their
commercial value.
It is necessary to strengthen the image
construction of the streamers themselves in order to
make great progress in live streaming. Only by
adhering to professional ethics, adhering to the
bottom line of morality, disseminating high-quality
content, establishing a positive image among
netizens, and guiding the public to establish healthy
consumption concepts and aesthetic appeals, can
streamers positively promote the green development
of the online stream industry, and can they also
maximize the commercial value of live streamers and
form a good Internet "business ecosystem".
REFERENCES
Guo Peilin, Lian Yueqi. Analysis of the impact of online
celebrity e-commerce on the purchase decision process
in the Internet environment--to Douyin platform as an
example [J]. Modern Marketing (Business Edition),
2020, 4:76-77
Jin Lin. Research on User Behavior Analysis Based on Data
Mining [J]. E-Commerce, 2020(04):41-42.
Sheng Ye. Analysis and prediction of business impact of
social behavior of live streamers [D]. Beijing: Beijing
Jiaotong University, 2019.
Serdar Oktay．An analytical study to identify and determine
the usage frequency of sales and marketing strategies
for 5 star hotels in the Antalya region [J].
ProcediaComputer Science, 2017, 120.
Xie Huimin. Research on Marketing Management of E-
commerce Enterprises Based on Data Mining [J].
Brand, 2019, 000(016):35-36.
Xi Y．Leung, Billy Bai, Mehmet Erdem．Hotel social media
marketing: a study on message strategy and its
effectiveness [J]. Journal of Hospitality and Tourism
Technology, 2017, 8(2)
YU Li, LIU Lu, LUO Zhanghua. Comparative analysis of
e-commerce recommendation strategy in China [J].
Systems Engineering-Theory & Practice, 2004(8):96-
101.)
NMDME 2022 - The International Conference on New Media Development and Modernized Education
444