A Study on Prediction of New Followers of Social Media Bloggers Based on Principal Component Clustering and Temporal Difference Regression

Zhiqing Guo, Junpeng Yuan, Hang Zeng, Youwen Qiu, Lionel Z. Wang, Jiaxin Huang, Chih-Yi Wen, Xiaolu Zou, Jie Lan, Tingyan Wang, Manni Li, Zhe LiJournal of Computer Science and Artificial Intelligence 2025 · DOI

Abstract: In recent years, social media platforms have profoundly affected people's social interactions and information acquisition. Precisely analyzing users' needs to achieve efficient matching of content supply promotes positive ecological cycles and enhances platform competitiveness. This paper focuses on predicting the number of new followers for individual bloggers. We integrate dynamic Bayesian networks and collaborative filtering concepts to construct a "blogger-user interaction matrix," extracting 9-dimensional sliding window differential features (viewing, liking, commenting) from the preceding three days. High-dimensional user features were compressed using principal component analysis (PCA). Subsequently, K-means clustering divided 42 bloggers into three groups, and time-difference OLS regression models were built for each. The adjusted $R^2$ for all three groups exceeded 0.78, verifying the hypothesis that "similar bloggers have similar attention growth patterns." After robustness and sensitivity testing, the model predicted blogger B21 would gain the most new followers on July 21 (554 people). These results provide a quantitative basis for platform push algorithms and blogger operation strategies.

Introduction

In the hyper-competitive world of social media platforms like Xiaohongshu or TikTok, platforms struggle to efficiently match content with users. A key metric for success is "new follower acquisition." However, predicting this is difficult because user behavior is non-linear and varies significantly across different blogger "niches."

Existing recommendation algorithms often struggle with the "cold start" problem for smaller bloggers or fail to capture the immediate temporal dynamics of viral content. This research aims to bridge this gap by creating a model that analyzes the pathway of user interactions—how viewing, liking, and commenting today translates into a "follow" tomorrow.

Methodology

The study proposes a three-stage "fusion framework":

  1. Dimensionality Reduction (PCA): Starting with a high-dimensional blogger-user interaction matrix, PCA was used to compress features into a 29-dimensional set (26 principal components + 3 temporal features) to eliminate noise and reduce computational complexity.
  2. Blogger Clustering (K-means): Recognizing that different niches (e.g., food vs. tech) have distinct "fan signatures," K-means clustering divided 42 bloggers into three distinct groups to capture niche-specific growth patterns.
  3. Temporal Difference OLS Regression: For each cluster, an Ordinary Least Squares (OLS) regression model was built. Independent variables included "sliding window" features—specifically, the counts of views, likes, and comments from the previous three days—to capture profile momentum.

Datasets

The study utilized a real-world dataset from a social platform containing user behavior records from July 11 to July 20, 2024.

  • Data Points: User ID, Blogger ID, Timestamp, and Behavior Type (watch, like, comment, follow).
  • Scope: The model was trained on the first nine days of data to predict the "new followers" count for July 21.

Key Results and Insights

The core of the paper is the verification that bloggers in the same cluster exhibit similar attention growth patterns. By training separate models for each cluster rather than one global model, the researchers significantly improved accuracy.

Regression coefficients and model performance metrics for the three blogger categories.
  • Predictive Power: The adjusted $R^2$ values for the three blogger clusters were 0.8052, 0.7802, and 0.9017, indicating that the model explains a vast majority of the variance in follower growth.
  • Ranking Accuracy: The Spearman correlation coefficient between the predicted and actual rankings reached 0.91, outperforming traditional collaborative filtering baselines.
  • Specific Prediction: The model successfully identified blogger B21 as the top gainer for July 21, predicting 554 new followers based on high interaction volume in preceding days.
Predicted top 5 bloggers by new follower count for July 21.
  • Interaction Pathways: The model quantifies the "funnel" from casual viewing to following. Viewing is a "foundational motivator," while comments represent "in-depth interaction" that strongly reinforces the connection. Sensitivity analysis revealed that a 1% increase in likes over a three-day window leads to a roughly 0.24% net increase in followers the following day.

Conclusion

This framework can help platforms reduce "invalid exposure" by an estimated 30% by tilting resources toward bloggers with high growth momentum before peak traffic. While robust, future work should integrate external factors (cross-platform trends) and Natural Language Processing (NLP) to distinguish sentiment in comments, which likely impacts follow rates differently.

Summary updated

Abstract

In recent years, social media platforms have profoundly affected people's social interactions and information acquisition, precisely analysing users' needs to achieve efficient matching of content supply, promoting positive ecological cycles, and ultimately enhancing the competitiveness of platforms and commercial value. This paper focuses on the prediction task of "the number of new followers of each blogger", integrates dynamic Bayesian network and collaborative filtering concepts, constructs the "blogger-user interaction matrix", extracts 9-dimensional sliding window differential features of viewing, liking, commenting, etc. in the past three days, and The high-dimensional user features were compressed to 42 dimensions using principal component analysis. Subsequently, K-means was used to divide the 42 bloggers into three groups, and the time-difference OLS regression models were built separately, and the adjusted R² of the three groups was higher than 0.78, which verified the hypothesis that "similar bloggers have similar attention growth patterns". After the model was tested for robustness and sensitivity, the blogger with the most new followers on 21 July was predicted to be B21 (554 people). The research results can provide a quantitative basis for the platform's accurate push and blogger's operation strategy. This paper is better in data dynamic modelling, reviewing a large amount of literature to establish the best model, with more thorough consideration in the problem. The model passes the robustness test and sensitivity test, and has reference value in platform push.
BibTeX
@article{guo2025study,
  author = {Zhiqing Guo and Junpeng Yuan and Hang Zeng and Youwen Qiu and Lionel Z. Wang and Jiaxin Huang and Chih-Yi Wen and Xiaolu Zou and Jie Lan and Tingyan Wang and Manni Li and Zhe Li},
  title = {{A Study on Prediction of New Followers of Social Media Bloggers Based on Principal Component Clustering and Temporal Difference Regression}},
  journal = {Journal of Computer Science and Artificial Intelligence},
  year = {2025},
  doi = {10.54097/628c0y48},
}
Figure