The analysis of marketing performance in E-commerce live broadcast platform based on big data and deep learning

Abstract

This study aims to conduct a comprehensive and in-depth analysis of the marketing performance of e-commerce live broadcast platforms based on big data management technology and deep learning. Firstly, by synthesizing large-scale datasets and surveys, the study constructs a series of performance evaluation indicators including user participation, content quality, commodity sales effect, user satisfaction, and platform promotion effect. Secondly, the weight of each indicator is finally determined through the indicator screening of the expert scoring method. Finally, the experimental design and implementation steps such as data collection, experimental environment setting, parameter setting, and performance evaluation are introduced in detail. Through the training and evaluation of the Back Propagation Neural Network (BPNN), each secondary indicator’s adjusted weight value and global ranking are obtained, providing a scientific basis for subsequent management opinions. The research results emphasize the importance of comments and ratings, purchase conversion rate, advertising click-through rate, and other indicators in improving user satisfaction, promoting sales, and effective promotion. Overall, this study provides a clear direction for an e-commerce live broadcast platform to optimize user experience, improve sales performance, and strengthen brand promotion.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-00546-w.

Introduction

Research background and motivations

With the rapid development of information technology and the popularity of the Internet, e-commerce live broadcast platforms have gradually become an important part of the e-commerce industry^1,2. This brand-new marketing method provides consumers with an integrated experience of interaction, entertainment, and shopping through live video broadcasts, which has become a unique way to attract consumers³. Meanwhile, big data management and deep learning (DL) technology have made remarkable progress in recent years, providing more accurate user data analysis and personalized recommendation services for e-commerce live broadcast platforms. This greatly promotes the development and operation of the platforms^4–6.

However, despite the remarkable success of e-commerce live broadcast platforms in the market, there remain numerous areas that have not been sufficiently explored. These include the key factors influencing marketing performance, the underlying mechanisms through which user behavior affects marketing outcomes, and methods for optimizing existing marketing strategies^7,8. Current research primarily focuses on the overview of live commerce models and the analysis of marketing strategies. However, empirical studies on leveraging big data and DL technologies to optimize the marketing performance of e-commerce live broadcasts remain limited. Furthermore, many existing studies fail to identify which factors are most critical to the success of live-streaming marketing or explain why certain strategies do not achieve the expected results in real-world scenarios. Therefore, the necessity of this study lies in addressing these research gaps by exploring and resolving specific challenges in e-commerce live-streaming marketing, thereby enhancing platform competitiveness and marketing efficiency. From a broader perspective, although numerous studies have examined strategies for e-commerce live-streaming marketing, these platforms still face the following challenges in practical applications. First, the complexity of user behavior: consumer decision-making on live-streaming platforms is influenced by multiple factors, such as social interactions, content quality, and advertising recommendations. However, existing research lacks sufficient empirical evidence to clarify the specific mechanisms of these factors. Second, the limitations of data-driven marketing strategies: although many platforms utilize big data analytics to optimize live-streaming marketing, the absence of systematic evaluation metrics and quantitative methods makes it difficult to measure the effectiveness of these optimizations. Third, the limitations of existing research: most current studies rely on traditional statistical methods to analyze the effectiveness of live-streaming marketing, while research combining DL for precise prediction and optimization remains scarce.

Research objectives

The core objective of this study is to integrate big data management and DL technology to thoroughly analyze and evaluate the marketing performance of e-commerce live broadcast platforms. Meanwhile, it explores the relationship between user behavior data and live-streaming marketing performance. The specific research objectives include the following. The impact of big data on user behavior in e-commerce live broadcast platforms is analyzed, including how user interaction behaviors (such as viewing duration, comments, likes, and shares) influence purchasing decisions; The potential effects of different user behavior patterns on live-streaming marketing performance are investigated, and key influencing factors are identified; A DL-based marketing performance prediction model is constructed to provide more accurate data support, assisting e-commerce platforms in optimizing marketing strategies; Targeted optimization solutions are formulated to enhance the marketing performance and user satisfaction of e-commerce live broadcast platforms in a highly competitive market. The ultimate goal of this study is to provide scientific, data-driven marketing strategy recommendations for e-commerce live broadcast platforms, thereby improving their profitability and market competitiveness.

Literature review

User behavior data is an important factor affecting the marketing performance of e-commerce live broadcast platforms. Wang and Guo (2024) investigated user browsing, clicking, and purchasing behaviors and highlighted the remarkable impact of these behaviors on marketing performance. Their study demonstrated that by deeply analyzing user data, platforms could more accurately formulate personalized recommendation strategies, thus enhancing user participation and shopping conversion rates⁹. Additionally, Lin et al. (2023) pointed out that user behavior patterns on live-streaming platforms exhibited significant individual differences, with different user groups responding differently to live-streaming marketing content. This further emphasized the importance of data-driven precision marketing strategies¹⁰. Existing research primarily focuses on basic user behavior data, such as clicks, purchases, and dwell time. Still, it has not sufficiently explored cross-platform user behavior patterns or systematically investigated the impact of social network interactions on marketing performance.

The optimization of the recommendation system is essential to improve the user experience and promote sales. Budhaye and Oktavia (2023) studied DL-based recommendation algorithms. They discovered that advanced DL models could more accurately analyze user interests and behavior patterns, providing users with more suitable product recommendations¹¹. Furthermore, Liu et al. (2023) proposed an adaptive recommendation algorithm capable of dynamically adjusting based on real-time user feedback, significantly improving users’ purchase intentions and loyalty¹². DL has been widely applied in personalized recommendations. However, current algorithms still face challenges such as the cold-start problem, data sparsity issues, and the need to optimize recommendation algorithms to balance short-term conversion rates with long-term user loyalty.

The quality of live content directly affects user participation and purchasing decisions. Shih et al. (2024) found that the type, duration, and interaction format of live-streaming content significantly influenced user participation and purchase intentions¹³. Specifically, highly interactive and high-quality content remarkably enhanced user retention. Additionally, Su et al. (2023) discovered that the host’s style, tone of voice, and the overall atmosphere of live commerce also affected users’ purchasing decisions, further emphasizing the necessity of content optimization¹⁴. Although existing studies have focused on the characteristics of live-streaming content, there is a lack of detailed exploration into content optimization strategies for different product categories. Moreover, AI-driven automated optimization of live-streaming content remains in its exploratory stages.

The emotional changes experienced by users in the process of live shopping have a vital impact on the final purchasing decision. Chong et al. (2023) investigated users’ emotional changes through big data sentiment analysis. They found that users were significantly influenced by the host’s emotions during live streams, with positive emotions notably boosting purchase conversion rates¹⁵. Meanwhile, Fitria et al. (2024) proposed that real-time sentiment analysis models based on natural language processing technology could predict users’ shopping emotions and dynamically adjust live-streaming content¹⁶. Current research mainly concentrates on the relationship between user emotions and purchasing behavior. However, it has not yet deeply explored how real-time sentiment analysis can optimize live-streaming marketing strategies, such as personalized content recommendations or AI-driven automatic emotion regulation by hosts.

Social interaction plays a key role in the marketing performance of e-commerce live broadcast platforms. Kim et al. (2023) found that interactions on social networks, such as comments, shares, and likes, effectively enhanced user retention and brand loyalty¹⁷. Furthermore, Alam et al. (2023) pointed out that users with strong social influence played a critical role in the purchasing decisions of others. This made social propagation mechanisms an important means of improving the effectiveness of live-streaming marketing¹⁸. Current research mainly focuses on the impact of social interactions on user participation. However, it has not sufficiently explored how trust mechanisms in social networks influence purchasing decisions or how to optimize marketing strategies using social data.

Advertising also plays a crucial role in e-commerce live-streaming marketing. Xu et al. (2024) found that different types and timing of advertisements had varying impacts on user behavior and platform revenue¹⁹. Yang et al. (2024) further pointed out that integrating short video advertisements with live-streaming marketing markedly improved user click-through rates (CTRs) and conversion rates²⁰. Current research still lacks DL-based precise advertising optimization models, and the matching mechanisms between advertisement content and live-streaming content have not been thoroughly investigated.

With the rapid development of e-commerce live broadcasting, the risk management of the platform has become increasingly important. Jalantina and Minarsih (2024) discovered that different types of risks, such as false marketing and user privacy breaches, substantially affected the marketing performance of platforms²¹. Moreover, Verina et al. (2024) found that the combination of blockchain technology and Artificial Intelligence (AI) risk control systems effectively enhanced the security of e-commerce live broadcast platforms²². Current research primarily highlights traditional data security, while the application of emerging technologies such as blockchain and AI in e-commerce risk control remains underexplored.

Although research has covered multiple dimensions, there are still issues that need to be addressed. The combined application of big data and DL: Current research mostly focuses on single-dimensional analysis, lacking systematic interdisciplinary integration methods; Dynamic optimization of precision marketing strategies: Existing research has not yet established a real-time optimized DL marketing model; Deep analysis of user social network behavior: Utilizing social data to enhance user stickiness is still an open question. This study’s innovation lies in combining big data analysis and DL to construct a comprehensive marketing performance optimization framework. Real-time user sentiment analysis is employed to optimize live-streaming content recommendations, proposing precise social interaction strategies to enhance user retention and loyalty.

Research model

Dataset construction of live broadcast platform based on big data management technology

This section delves into the key information of platform operation by synthesizing large-scale datasets. First, user behavior data, including but not limited to click records, viewing duration, and purchase behavior, is obtained to fully understand the user’s behavior patterns on the platform^23–25. Second, during data collection, special attention is paid to user privacy protection and compliance to ensure that the research process conforms to relevant laws and ethical standards^26,27.

(1) Acquisition and processing of user behavior data.

1) Data source

The user behavior data on the e-commerce live broadcast platform are collected through the system logs, user interaction records, shopping carts, order databases, and other data sources. The data types and specific contents are exhibited in Table 1^28–31. This study utilizes big data management technology to analyze the marketing performance of live broadcast platforms. Although the data scale is relatively limited, the data sources encompass multi-dimensional, multi-source, unstructured, and structured data, aligning with the 4 V characteristics of the “big data” definition. (1) Volume (Data volume): This study integrates system logs, user behavior data, survey responses, and transaction data from live broadcast platforms, encompassing thousands of user interaction records. (2) Velocity (Data flow speed): System log data is updated in real-time, and user interaction data continuously grows, reflecting the characteristics of high-frequency data streams. (3) Variety (Data diversity): The data types include structured transaction data (orders, purchase records), semi-structured user behavior data (clicks, viewing duration), and unstructured text data (comments, bullet-screen messages). (4) Veracity (Data authenticity): The data undergoes rigorous cleaning and validation to ensure reliability.

Table 1.

Types and specific contents of data.

Data source	Data type	Data content
System journal	Access log	Record basic operations such as user login, logout, browsing live pages, etc.
User interaction record	Records of clicks, comments, and likes	Record the specific behavior of users interacting with live content.
Shopping cart database	Shopping cart information	Record the user’s behavior of adding commodities to the shopping cart, including commodity information and quantity.
Order database	Purchase record	Record the order information that the user completed, including the purchased goods, time and amount, etc.

2) Data cleaning

In processing user behavior data, rigorous data cleaning ensures reliable data quality³². First, the original data collected from different data sources is preliminarily screened to eliminate possible abnormal values and duplicate records³³. Specifically, people pay attention to inconsistent, missing, or incorrect data in system logs, user interaction records, shopping cart databases, and order databases³⁴. The main steps and contents of data cleaning are detailed in Table 2^35–38.

Table 2.

Data cleaning.

Data source	Abnormal value processing method	Duplicate record removal mode
System journal	Eliminate records of abnormal login or frequent logout	Remove duplicate login records with the same timestamp
User interaction record	Handle extreme clicks, comments, or likes	Eliminate duplicate clicks, comments, or likes
Shopping cart database	Handle an abnormal number of goods in the shopping cart	Remove the same commodity records that are repeatedly added to the shopping cart
Order database	Exclude orders with unusually high or low amounts	Remove duplicate orders with the same order number, purchase time, and goods

Construction of marketing performance evaluation system for e-commerce live broadcast platforms

(1) Preliminary establishment of the indicator system.

With the rapid rise of e-commerce live broadcast platforms in the market, there is a growing need to evaluate their marketing performance more comprehensively and accurately. To address this, an initial marketing performance evaluation indicator system for e-commerce live broadcast platforms is established, as shown in Table 3. This indicator system covers key areas, including user participation, content quality, product sales effect, user satisfaction, and platform promotion effect^39–41. In this preliminary construction, the primary indicators provide the main direction of the overall evaluation; The secondary indicators are more specific and can be used for more detailed assessment and analysis.

Table 3.

Preliminary construction results of indicators.

Primary indicator	Secondary indicator
User participation	Viewing duration
	Interaction frequency
	Number of barrages
	Number of gifts/rewards
	Number of likes and shares
Content quality	Live content fever
	Commodity display effect
	Main professional degree in live broadcast
	Interactive game effect
Commodity sales effect	Purchase conversion rate
	Sales
	Click rate of goods
User satisfaction	Comments and ratings
	Unsubscribe rate
Platform promotion effect	User growth rate
	Social media sharing effect
	Advertising CTR

(2) Indicator screening based on expert scoring method.

To ensure the scientific rigor and practical relevance of the indicator system, it is essential to screen the initially established indicators through expert scoring. After removing the less significant influencing factors, the final set of determinants for this study’s indicator system is established⁴². The selection of experts must be highly relevant to this study. Hence, the selection of experts in this study mainly includes six senior industry experts and technicians engaged in e-commerce in China⁴³. The final evaluation system after screening is depicted in Table 4.

Table 4.

Final construction results of indicators.

Primary indicator	Secondary indicator
User participation A1	Viewing duration A11 Number of barrage A12 Number of gifts/rewards A13 Number of likes and shares A14
Content quality A2
	Commodity display effect A21
	Main professional degree of live broadcast A22
	Interactive game effect A23
Commodity sales effect A3	Purchase conversion rate A31 Sales A32 Click rate of goods A33
User satisfaction A4	Comments and ratings A41 Unsubscribe rate A42 User growth rate A43
Platform promotion effect A5
	Advertising CTR A51

Construction of marketing performance evaluation model of e-commerce live broadcast platforms based on DL

(1) BP neural network.

This study uses the Back Propagation neural network (BPNN) method to analyze the obtained data. As a common model in machine learning, BPNN has been widely used in academic circles. Its main advantage is its strong nonlinear mapping ability, and BPNN learning is mainly contained in weights and thresholds, which is in line with the content of this study^44,45. The training process of BPNN is relatively mature and demonstrates stable performance even with small sample datasets. However, the current DL field also includes other advanced models. For instance, Long Short-Term Memory (LSTM) is suitable for processing time-series data and can capture long-term user behavior patterns in live-streaming user behavior analysis. Convolutional Neural Network (CNN) is more applicable for image and spatial feature extraction and is commonly used in tasks such as analyzing user interface click heatmaps. The primary reasons for selecting BPNN in this study are as follows.1. Applicability: The core task of this study is to establish a marketing performance prediction model, rather than focusing on time-series or image analysis. Therefore, BPNN offers greater advantages in learning the nonlinear relationships among marketing indicators. 2. Data scale limitations: Given the limited dataset size in this study (186 survey responses + live broadcast platform log data), BPNN exhibits better convergence on small datasets. In contrast, LSTM and CNN require large amounts of data for training and may suffer from overfitting otherwise. 3. Computational cost: Compared to LSTM and CNN, BPNN incurs lower computational overhead, making it more suitable for real-time marketing analysis needs in e-commerce enterprises.

In this study, a typical three-layer BPNN is used for analysis. Its internal structure is that neurons in the same layer are independent, but neurons in two adjacent layers are interrelated. Figure 1 shows the specific structure.

Fig. 1.

Classical BP network structure.

(2) The realization process of the model.

This study adopts a three-layer BPNN, including the input, hidden, and output layers. There are 14 neuron nodes in the input layer, corresponding to the 14 indicators, and the output is a neuron, that is, the marketing performance of e-commerce. The specific measurement method is reflected by the total score after normalization. The number of nodes in the middle-hidden layer is calculated as Eq. (1):

1

and represent the number of neurons in the hidden and input layer; n is the number of nodes in the output layer; refers to a regulatory variable. It is generally believed that ashould belong to the regulatory constant between^1,10. After calculating and running the code, the optimal number of nodes in the hidden layer is determined. Based on the general interval recommended by practical scholars, the hidden layer is finally set to consist of 5 neurons.

Experimental design and performance evaluation

Datasets collection

On the one hand, the data used in this study comes from the dataset constructed in Sect. “Dataset construction of live broadcast platform based on big data management technology” (commodity sales effect and platform promotion effect). On the other hand, it stems from the questionnaire survey. The content of the questionnaire design includes three parts: user participation, quality evaluation of live content, and user satisfaction. The questionnaire’s scoring method adopts the Likert five-point method. In addition, the questionnaire is distributed in the form of quiz stars on major e-commerce marketing platforms. Finally, 200 questionnaires were distributed, and 186 valid questionnaires were recovered, with an effective recovery rate of 93%, meeting the requirements of the questionnaire.

Parameters setting

This study employs a high-performance computing environment for data analysis and model training to ensure the efficient operation of DL models. The experiments are primarily conducted on the Windows 11 operating system, with model development based on the Python programming language and the TensorFlow DL framework. Nvidia GPU (Graphics Processing Unit) acceleration (CUDA (Compute Unified Device Architecture) 11.2 + cuDNN (Deep Neural Network) 8.1) is utilized to accelerate the training process, and SSD (Solid State Drive) storage is employed to enhance data processing speed. Additionally, SPSS (Statistical Package for the Social Sciences) 25.0 is used for reliability and validity testing of user survey data in statistical analysis. The detailed configuration of the experimental environment is provided in Appendix Table A1. Some parameters involved in the experiments, mainly the parameter settings of the BPNN model, are denoted in Table 5.

Table 5.

Parameter settings of the CNN model.

Parameter name	Specific value
Input dimension	5
Number of input layer nodes	14
Number of hidden layer nodes	5
Network layer number	3
Activation function
Number of output units	1 (indicates the evaluation value of performance)
Output activation function
optimization algorithm
Learning rate	0.001
Batch size	20
Training epochs	100
error function	Mean square error
Data preprocessing	Normalized and forward

Performance evaluation

(1) Reliability and validity analysis of the questionnaire.

1) Reliability analysis

SPSS software is employed to analyze the reliability of the questionnaire, and the results are displayed in Fig. 2.

Fig. 2.

Reliability analysis results of the questionnaire.

Figure 2 reveals that the Alpha value of the overall questionnaire is 0.924, suggesting that the overall questionnaire has high reliability. Moreover, the Alpha values of all indicators are also around 0.8, showing a good reliability level. Specifically, the Cronbach’s Alpha values for user participation (A1), content quality (A2), and user satisfaction (A4) all exceed 0.8, indicating strong internal consistency in measuring the related constructs. Overall, the questionnaire demonstrates high reliability.

2) Validity analysis

Kaiser-Meyer-Olkin (KMO) and Bartlett spherical tests evaluate the data. The value range of KMO is between 0 and 1, and the closer the value is to 1, the stronger the correlation between variables. When the KMO value is not less than 0.5, the data is appropriate. Bartlett’s spherical test is utilized to verify whether the variables are independent. A significance level of less than 0.05 indicates that the data are independent of each other to some extent and are spherical in distribution, which meets the standard. After testing, the KMO and Bartlett spherical test values of the questionnaire are presented in Table 6.

Table 6.

Results of validity analysis of the questionnaire.

Indicator	KMO	Bartlett test
User participation A1	0.840	0.00
Content quality A2	0.781
User satisfaction A4	0.788
Overall questionnaire	14

Table 6 shows that KMO values of A1, A2, and A4 are 0.840, 0.781, and 0.788, respectively. These values are all above 0.5, indicating a strong correlation between these indicators and the data is relatively suitable. Additionally, Bartlett’s test P-value for all indicators in the questionnaire is 0.00, suggesting a significant correlation between these indicators. Meanwhile, the data is not mutually independent to a certain extent and is distributed in a spherical shape.

(2) Training situation and evaluation results of BPNN.

1) Training situation

In this experiment, 500 sales data of an e-commerce platform are selected as datasets, with 80% being the training set and 20% being the testing set. Moreover, a training process of 10 to 100 training epochs is established, and the changes in the training model are monitored from four aspects: training loss, validation loss, training accuracy, and validation accuracy. The final result is illustrated in Fig. 3.

Fig. 3.

Fitting of BPNN.

Figure 3 shows that with the increase in training epochs, the training loss decreases gradually, and the verification loss also exhibits a decreasing trend. This indicates that the model gradually improves the data-fitting ability during learning. Meanwhile, the training and verification accuracies increase as epochs rise, demonstrating that the BPNN model selected in this study performs well in the training and verification sets.

2) Evaluation results

The BPNN is implemented using Python, and the weights and rankings of each indicator within the system are analyzed through the application of the BPNN. The results are suggested in Fig. 4.

Fig. 4.

Evaluation results of the indicator system.

In Fig. 4, comments and ratings (A41) rank first with a weight value of 0.188, illustrating that user comments and ratings have a significant impact on the marketing performance of live broadcast platforms. This is likely because user comments and ratings not only influence the purchasing decisions of other consumers but may also affect the platform’s recommendation algorithms, thereby determining product exposure. Therefore, optimizing the user review system, encouraging users to leave comments, and implementing effective review management strategies (e.g., guiding positive reviews and reducing malicious negative reviews) are key to enhancing platform competitiveness. Purchase conversion rate (A31) and advertising CTR (A51) rank second and third, respectively. This indicates that the success of e-commerce live-streaming marketing is closely related to direct purchasing behavior and the effectiveness of advertising campaigns. This suggests that, in addition to enhancing the appeal of live-streaming content, platforms should optimize advertising strategies and product display methods when improving marketing performance. For example, platforms can increase advertising CTR through precise recommendations and personalized ads, and boost purchase conversion rates by enhancing interactive experiences (e.g., limited-time discounts, and live-streaming giveaways).

Viewing duration (A11) ranks fourth but has a relatively low weight, meaning that while viewing duration has some influence on user participation, it is not the most critical factor determining marketing success. This is because longer viewing times do not necessarily indicate higher purchase intention, and some users may watch live streams for extended periods without making purchases. Although viewing duration has a low weight, it can still serve as a supplementary metric for user participation and be optimized in conjunction with other indicators. For instance, platforms can analyze which types of interactive behaviors most effectively drive purchases by combining “interaction frequency (A12)” and “purchase conversion rate (A31).” For users with long viewing durations but low purchase rates, platforms can increase pop-up recommendations and personalized promotional messages to enhance purchase intention.

Interactive game effects (A23) rank seventh but have a relatively high weight, indicating that interactive experiences in live-streaming rooms play an important role in boosting user participation. Platforms can enhance gamified interactions, such as quiz giveaways and point rewards, to increase user activity and indirectly promote sales.

(3) Performance comparison of models

This study employs BPNN to predict e-commerce live-streaming marketing performance. Also, it compares LSTM and CNN to evaluate BPNN’s performance in this task. The experiment uses the same dataset and adjusts the hyperparameters to ensure fairness.

The following indicators are used to evaluate the performance of the model. Mean Absolute Error (MAE): It measures the deviation between the predicted and true values, the smaller the better. Root Mean Square Error (RMSE): It measures the stability of the model, the smaller the better. Coefficient of determination (R²): It measures the model’s ability to fit, the closer to 1 the better. The results are listed in Table 7:

Table 7.

Performance comparison of models.

Model	MAE↓	RMSE↓	R²↑
BPNN	0.182	0.257	0.91
LSTM	0.196	0.276	0.88
CNN	0.211	0.291	0.85

Based on the content of Table 7, BPNN demonstrates the best performance across all indicators, illustrating its superior learning capability on small datasets and stable modeling of nonlinear relationships. LSTM performs slightly worse, as the limited role of time-series features in this task prevents the model from fully leveraging its advantages. CNN shows the poorest performance, suggesting that its image feature extraction capability offers limited assistance in predicting marketing performance. The BPNN model achieves higher accuracy (R² = 0.91), lower prediction errors, and better computational efficiency. Its training time is approximately 30% lower than that of LSTM and CNN, and it is well-suited to the current data scale, enabling stable convergence even with small sample datasets.

Discussion

First, this study finds that user reviews play the most significant role in marketing performance, which aligns with the findings of Shin et al. (2024). They noted that positive reviews not only influenced the purchasing decisions of other consumers but also affected platform recommendation algorithms, thus increasing product exposure⁴⁶. However, this study further reveals that the weight of user reviews exceeds that of live-streaming content itself. This suggests that optimizing review management systems can enhance the marketing performance of live broadcast platforms more effectively than relying solely on content optimization. Specifically, actively encouraging users to leave reviews and mitigating the impact of malicious negative reviews are key strategies for achieving this improvement. Second, this study validates the critical role of purchase conversion rates and advertising CTRs in the success of live-streaming marketing, consistent with the conclusions of Dong and Tarofder (2024). They found that precise ad placement and optimizing user shopping paths could remarkably boost sales⁴⁷. Unlike their study, however, this study further employs BPNN analysis to reveal that personalized recommendations play a critical role in linking advertising clicks to purchase conversions. This suggests that future live broadcast platforms should strengthen data-driven precision advertising strategies to improve ad placement efficiency. Additionally, this study finds that while viewing duration (A11) ranks high in terms of user participation, its impact weight is relatively low, which contrasts with the findings of Wimolsophonkitti and Naipinit (2024). They argued that longer viewing durations correlated with stronger purchase intentions⁴⁸. The results of this study indicate that users who watch live streams for extended periods do not necessarily have strong purchase intentions, as some users may stay for entertainment purposes without converting into purchases. Therefore, live broadcast platforms cannot rely solely on extending viewing duration as an optimization goal. Instead, they should combine user interaction frequency and purchasing behavior to more accurately predict user purchase intentions, such as by pushing limited-time discount information or personalized recommendations to improve conversion rates. Additionally, this study further points out that relying solely on interactive games is insufficient to directly drive sales. Interactive content needs to be integrated with actual purchasing behaviors, such as offering shopping discounts or point redemption during game reward sessions, enhancing users’ willingness to spend during live streams. At the theoretical level, this study quantifies the weight of various marketing factors using DL methods. Meanwhile, it offers data-driven support for the marketing strategies of e-commerce live broadcast platforms, further enriching the research framework for marketing performance. At the practical level, the research results provide actionable guidance for live broadcast platforms to optimize user review management, precision ad placement, and interactive experiences. For example, platforms can use intelligent review filtering to reduce the impact of malicious negative reviews while enhancing the personalization and real-time nature of ad delivery to increase CTR and purchase conversion rates. Furthermore, for users with long viewing durations but low purchase rates, platforms can improve their purchase intentions through targeted promotional messages or interactive rewards. Although this study has made progress in marketing performance modeling, certain limitations remain. For instance, the data sources are primarily based on a single platform. Future research could expand to multiple e-commerce platforms for cross-validation and incorporate more social network data to explore the impact of social propagation on live-streaming marketing performance. Moreover, while this study validates the effectiveness of BPNN in marketing performance prediction, future research could introduce more complex DL models, such as Transformer or XGBoost, to improve prediction accuracy and model interpretability. In summary, this study validates the significant roles of user reviews, purchase conversion rates, and advertising CTRs in live-streaming marketing. Meanwhile, it reveals the complex relationship between viewing duration and purchasing behavior. Thus, it provides data-driven support for optimizing the marketing strategies of live broadcast platforms and offers new directions for future research.

Conclusion

Research contribution

Based on big data management and DL technology, this study deeply analyzes the marketing performance of e-commerce live broadcast platforms. By constructing a comprehensive dataset, including large-scale user behavior data and a questionnaire survey, the detailed steps of data collection, experimental environment setting, parameter setting, and performance evaluation are introduced in the research method. A summary of the experimental results shows the multi-dimensional performance evaluation of e-commerce live broadcast platforms through the BPNN model, focusing on key indicators such as user participation, live broadcast content quality, commodity sales effect, and user satisfaction. In the discussion, through the in-depth interpretation of the weight analysis, key management opinions are put forward, emphasizing the importance of comments and ratings, purchase conversion rate, and advertising CTR, which provides a clear guiding direction for the e-commerce live broadcast platform. To sum up, this study provides a scientific analysis and empirical basis for the e-commerce live broadcast platform’s operation and marketing through DL methods and comprehensive dataset construction.

The contributions of this study are reflected in both theoretical and practical aspects. Theoretically, based on big data management and DL, this study constructs an evaluation framework for e-commerce live-streaming marketing performance. Meanwhile, it quantifies the impact of key factors such as user reviews, purchase conversion rates, and advertising CTRs. Also, it reveals that the influence of viewing duration on purchasing decisions is lower than expected, challenging traditional perspectives. Practically, the research results provide data-driven support for live broadcast platforms to optimize marketing strategies. It includes improving user review management (encouraging reviews and intelligent filtering), precision ad placement (AI-driven optimization of ad content), and enhancing user interaction (incentivizing purchases through points or discounts). Also, it encompasses refining marketing strategies for users with long viewing durations but low conversion rates (pushing personalized promotions). This study enriches the research on live-streaming e-commerce marketing performance and offers actionable strategic guidance for platform optimization.

Future works and research limitations

Although this study has made some achievements in the marketing performance analysis of e-commerce live broadcast platforms, there are still some shortcomings. For example, the proposed dataset only focuses on some user behavior data and information from the questionnaire survey. Moreover, it does not fully cover other potential influencing factors, which may affect the comprehensiveness of the research results. Future research could consider introducing more dimensional and diverse data and trying different DL models to improve the accuracy and robustness of the analysis. Moreover, it is essential to enhance the method and application of performance analysis of e-commerce live broadcast platforms and provide more practical decision support for the industry.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Yuanyuan Wei: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparation Xingchen Pan: writing—review and editing, visualization, supervision, project administration, funding acquisition.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author Xingchen Pan on reasonable request via e-mail panxingchen0916@163.com.

Declarations

Competing interests

The authors declare no competing interests.

Ethics statement

The studies involving human participants were reviewed and approved by Shandong Open University Ethics Committee (Approval Number: 2022.00123320). The participants provided their written informed consent to participate in this study. All methods were performed in accordance with relevant guidelines and regulations.