Fans_Kalman

📁 Dependencies

☁️ Introduction

A Kalman filter is utilized to predict the fans of any vtuber. The Kalman impl is inherited and modified from here:

We expect that there is an update of the follower number every 5 minutes. If there isn't, then the Kalman filter will only perform the time-update, while the measurement-update will not be performed.

⚡️ Quickstart

Run:

python3 fans_kalman.py

to see the prediction result.

Run:

python3 grid_search.py

to find the two optimized parameters:

1. sigma_w: The variance of the process noise
2. sigma_v: The variance of the measurement noise

Nevertheless, the current performance is ok even if these two parameters are randomly chosen.

🎉 Results

The model is 'trained' on the data from Kizuna ai, and it is tested on the data from mea. Here are the results:

Mean prediction error on the training set (kizuna.csv):

Mean prediction error on the test set (mea.csv):

🚧 Usage example

First, create a kalman filter with your favorite parameters. Note that input_data[0] is the initial value of the follower number:

kalman_filter = fans_kalman(input_data[0], sigma_w = sigma_w, sigma_v = sigma_v)    

Then, every 5 minutes, if the new data is avialable, run:

new_prediction = kalman_filter.predict_interface(new_obtained_value)

to get the prediction of the next time step (next 5 minutes). Note that new_prediction is a number in this case.

Otherwise, run:

new_prediction = kalman_filter.predict_interface(None)

The model will not be updated by any measurement, but it will still predict the follower number of the next time step (next 5 minutes).