Popularity, face and voice: Predicting and interpreting livestreamers' retail performance using machine learning techniques

Xiong Xiong, Fan Yang, Li SuarXiv 2023 · DOI · arXiv

Abstract

Livestreaming commerce, a hybrid of e-commerce and self-media, has expanded the broad spectrum of traditional sales performance determinants. To investigate the factors that contribute to the success of livestreaming commerce, we construct a longitudinal firm-level database with 19,175 observations, covering an entire livestreaming subsector. By comparing the forecasting accuracy of eight machine learning models, we identify a random forest model that provides the best prediction of gross merchandise volume (GMV). Furthermore, we utilize explainable artificial intelligence to open the black-box of machine learning model, discovering four new facts: 1) variables representing the popularity of livestreaming events are crucial features in predicting GMV. And voice attributes are more important than appearance; 2) popularity is a major determinant of sales for female hosts, while vocal aesthetics is more decisive for their male counterparts; 3) merits and drawbacks of the voice are not equally valued in the livestreaming market; 4) based on changes of comments, page views and likes, sales growth can be divided into three stages. Finally, we innovatively propose a 3D-SHAP diagram that demonstrates the relationship between predicting feature importance, target variable, and its predictors. This diagram identifies bottlenecks for both beginner and top livestreamers, providing insights into ways to optimize their sales performance.

Introduction

Livestreaming commerce is a multi-billion dollar hybrid of e-commerce and social media, yet the "secret sauce" for success remains largely anecdotal. Most previous research has relied on small-scale surveys or subjective theories to explain why some hosts move millions in merchandise while others struggle. The authors address the lack of rigorous, large-scale empirical testing in this field, aiming to move beyond "gut feelings" by using high-dimensional data to answer two main questions: Can we accurately predict a livestreamer's revenue (GMV) before the cameras even roll? And, more importantly, what specific features—from facial aesthetics to vocal pitch—actually drive those numbers?

Methods

The study compares eight different machine learning algorithms to find the most accurate predictor of GMV: Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Extra Trees (ET), Logistic Regression (LR), K-Nearest Neighbor (KNN), AdaBoost, and Gradient Boosting Regression Tree (GBRT). To quantify "soft" traits, the authors used specialized tools: the So-Young facial scoring system for physical attractiveness and the Praat software for acoustic analysis (measuring pitch, intensity, and bandwidth). To "open the black box" of the winning model, they employed SHAP (Shapley Additive Explanations) to rank feature importance and ALE (Accumulated Local Effects) to see how specific variables influence predictions. They also introduced a novel 3D-SHAP diagram to visualize complex, non-linear relationships.

Datasets

The authors constructed a massive longitudinal firm-level database sourced from Taobao’s "Shopping Globally" sector. It contains 19,175 observations covering 75 streamers over 579 days (April 2020 to November 2021). The dataset includes 30 features across four main categories: popularity metrics (likes, views, fan growth), physical traits (facial "golden triangle," youthfulness), vocal attributes (intensity, bandwidth), and service quality.

Correlation matrix showing the relationships between GMV and the 30 features analyzed in the study.

Key Insights and Results

The Random Forest (RF) model was the clear winner, achieving a Mean Absolute Percentage Error (MAPE) of 7.53%, suggesting a prediction accuracy of 92.5%. The top four predictors of GMV are all popularity-related: Comments, Page Views, Likes, and Fan Growth.

A surprising finding is that voice trumps face. In a highly visual medium, vocal attributes (like intensity and bandwidth) are more decisive for sales than physical appearance. For instance, "minimum intensity" of the voice ranked 6th in importance, significantly higher than "Physical Appearance" (10th).

Chord diagrams illustrating the relationships between GMV and four feature categories: (a) popularity, (b) physical traits, (c) vocal traits, and (d) service/other factors.

The study also revealed a gender divide: for female hosts, popularity and social interaction are the primary determinants, while for male hosts, "vocal aesthetics" are more influential. Furthermore, sales growth was found to be non-linear, following three stages: a "trough" for beginners, a "fast-growth" middle stage, and a "bottleneck" for top-tier streamers where increasing popularity yields diminishing returns.

Limitations and Future Work

The study is primarily limited by its focus on a single subsector on Taobao. Consumer behavior might differ in other categories or on platforms like TikTok/Douyin. Future work should explore deep learning architectures to further refine predictions and investigate the "relationship theory" between hosts and consumers more deeply.

Summary updated

Abstract

Livestreaming commerce, a hybrid of e-commerce and self-media, has expanded the broad spectrum of traditional sales performance determinants. To investigate the factors that contribute to the success of livestreaming commerce, we construct a longitudinal firm-level database with 19,175 observations, covering an entire livestreaming subsector. By comparing the forecasting accuracy of eight machine learning models, we identify a random forest model that provides the best prediction of gross merchandise volume (GMV). Furthermore, we utilize explainable artificial intelligence to open the black-box of machine learning model, discovering four new facts: 1) variables representing the popularity of livestreaming events are crucial features in predicting GMV. And voice attributes are more important than appearance; 2) popularity is a major determinant of sales for female hosts, while vocal aesthetics is more decisive for their male counterparts; 3) merits and drawbacks of the voice are not equally valued in the livestreaming market; 4) based on changes of comments, page views and likes, sales growth can be divided into three stages. Finally, we innovatively propose a 3D-SHAP diagram that demonstrates the relationship between predicting feature importance, target variable, and its predictors. This diagram identifies bottlenecks for both beginner and top livestreamers, providing insights into ways to optimize their sales performance.
BibTeX
@misc{xiong2023popularity,
  author = {Xiong Xiong and Fan Yang and Li Su},
  title = {{Popularity, face and voice: Predicting and interpreting livestreamers' retail performance using machine learning techniques}},
  year = {2023},
  eprint = {2310.19200},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.19200},
}
Figure