Using Dynamic Linear Model to Impute Missing Values in a PM 2.5 Time Series

Описание: Missing values are really common in data science. However, learning a model is really difficult for time series with missing data. In this video, we are using DLM, which builds a model from single components of a time series, to analyze PM 2.5 series. DLM can be used for learning time series with missing data. Furthermore, DLM would also generate one-step ahead prediction. That is what we are going to use for imputation.

Note: in 18:54 , I said "V and T". What I really meant was "V and W" 🙏 The subtitle is already correct.

Link to the source code: https://github.com/stephanielees/time...

00:00 Introduction
01:14 What is PM 2.5?
03:00 Get the data
07:06 Data preprocessing
12:01 Time plot
15:00 Brief introduction of DLM
17:47 Trend component of DLM (local level mean model)
21:28 Fourier form representation of seasonality
35:36 Regression DLM
40:45 Adding up all model components
42:38 Filtering
44:24 Trying out some alternatives
46:50 Plotting the imputed and raw data together

A little background story:
I got the idea of making this video after completing a course called AI and the Public Health from DeepLearning.AI. The use case in that course is also about PM 2.5 in Bogotá, so I thought of a different approach to it.

#timeseries #pm25 #seasonal #trend #regression #bayesian #rprogramming