Sample Data and Replication Code for: Mega or Micro? Influencer Selection Using Follower Elasticity

In the sample data folder, we provide a small sample of hashtags we collected from TikTok Discover page and some videos under them. In the code folder, we show how we 1) Extracted multi-modal video features from the original videos and save them into a local database (under database/generate) from which we generated the training and test data for the SVAE model (under database/output) 2) Train the SVAE model to get a 256-D latent vector representation for each video based on the learned feature weights (under SVAE) 3) Combine the content representation in the above step with other video covariates (under video_info) as the input for our causal inference (under DeepIV) 4) Estimate the DeepIV model to obtain the average and heterogeneous treatment effects (under DeepIV/treatment_effects) Finally, supplementary plots and tests are provided under DeepIV/distribution_plots and mis.

在示例数据文件夹中，我们提供了从TikTok发现页采集的话题标签（hashtag）及其关联视频的小型样本集。在代码文件夹中，我们展示了完整的实施流程：1）从原始视频中提取多模态视频特征，并将其存储至本地数据库（路径为database/generate），随后基于该数据库生成SVAE模型所需的训练与测试数据集，数据集存储于database/output路径下；2）训练SVAE模型，基于学习得到的特征权重为每一条视频生成256维的隐向量（latent vector）表征，相关文件存储于SVAE目录下；3）将上述步骤得到的内容表征与其他视频协变量（covariates）相结合，作为因果推断（causal inference）模型的输入数据，输入数据存储于video_info目录下；4）拟合DeepIV模型，以获取平均处理效应与异质性处理效应，相关结果存储于DeepIV/treatment_effects目录下。最后，补充绘图文件与测试脚本存储于DeepIV/distribution_plots与mis目录中。