Abstract: With the rapid development of mobile Internet, more and more attention has been paid to the new form of online social media represented by WeChat official accounts. However, these operators do not have clear ideas for enhancing WeChat heat, and clear and effective guidance on how to positively influence users' information behavior. Therefore, this paper combines the psychology neutral Stimulus-Response theory with the user information behavior research to construct a second-order user information behavior model. Then, through the official accounts data provided by the Sootoo Network, we collected a total of more than 10,000 push data of nine accounts from January to December, 2017. Based on the empirical research, we find the main influencing factors of user information behavior are identified, and construct a push hotspot prediction model. Also, this paper provides account operator with practical guidance significance in enhancing user information behavior.
Introduction
The WeChat Official Account (WOA) ecosystem has evolved into a primary "we-media" platform, boasting over 10 million accounts. Despite this scale, many operators lack data-driven insights into enhancing engagement or "heat." This research, presented at the 2018 International Conference on Electronic Business (ICEB), utilizes psychological Stimulus-Response (S-R) theory to analyze user behavior and predict "hotspots"—posts that achieve top-tier engagement.
Theoretical Framework: The S-R Model
The study constructs a second-order user information behavior model based on S-R theory. This framework posits that user engagement occurs in two distinct stages:
- First-Order Response: Account attributes and title characteristics (stimuli) trigger the initial "reading" action.
- Second-Order Response: The article's content and visual elements (stimuli) trigger deeper engagement, such as sharing, liking, or following.
Data and Methodology
The researchers analyzed data from the Sootoo Network, encompassing nine representative accounts across three categories: "Emotional Psychology," "Video Entertainment," and "Science & Technology Finance." The dataset included 10,507 push messages from 2017, representing 250 million reads and 4.9 million shares.
Statistical tools like one-way ANOVA were used to identify significant variables, while machine learning models—specifically Backpropagation (BP) Neural Networks and Random Forests—were trained to predict hotspots (defined as the top 10% of engagement).
Key Insights: Beyond Clickbait
The empirical analysis revealed several critical nuances in user behavior:
- The "Title Party" Paradox: While interrogative or exclamatory titles (clickbait) boost reading rates for leisure accounts (psychology/entertainment), they negatively impact professional accounts (finance/tech). Professional audiences value objectivity and rational descriptions.
- The Push Index: The position of an article in a multi-post message is the strongest predictor of initial reads. The "top push" receives the most attention, though this advantage does not translate directly to sharing rates.
- Content vs. Titles: Titles drive the first click, but content quality and the number of images are the primary drivers for sharing.
- Temporal Effects: Weekend posts benefit leisure accounts but see decreased engagement for professional content, reflecting users' desire to disconnect from work-related topics during time off.
Predictive Modeling Results
The study compared BP Neural Networks (BPNN) and Random Forests (RF) across two feature sets: E1 (all variables) and E2 (statistically significant variables only).
- Performance: BPNN achieved the highest accuracy (0.873) and precision (0.768) using the E2 feature set.
- Recommendation: While Random Forests identified more total hotspots (higher recall), BPNN's higher precision makes it more practical for operators who need to identify high-potential content with minimal false positives.
Conclusion
This research provides a quantitative foundation for WeChat operations. By understanding the divergent needs of different audience types and utilizing predictive modeling, operators can move beyond intuition to strategically influence user information behavior. Future work will look to incorporate Natural Language Processing (NLP) to analyze semantic content and explore more granular timing data.
Summary updated