Time Series Clustering in R

Introduction

• Purpose of the tutorial: To demonstrate a quick and straightforward implementation of time series clustering using the widyr package in R

• What is time series clustering?: Grouping time series data into clusters where data points in the same cluster group are more similar to each other than to those in other clusters. For example, if we have monthly sales data, time series clustering can help identify stores with similar sales patterns over time.

Load library

library(tidyverse)
library(widyr)

Import data

• About the data:

  • Fake dataset that can be downloaded from my GitHub.

  • Contains 832 rows & 3 columns

    • Columns:

      • year (<date>): Date information for each observation.

      • storecode (<chr>): Unique identifier for each store.

      • sales (<dbl>): Sales figures for each store.

• Importing data: Using read_csv()

store_list <- read_csv("https://raw.githubusercontent.com/zahiernasrudin/datasets/main/sample_store.csv")

• Glimpse of the dataset:

Clustering with widyr

Using widely_kmeans for time series clustering:

# Perform k-means clustering using widely_kmeans
cluster_group <-  store_list %>%
  widely_kmeans(item = storecode, 
                feature = year, 
                value = sales,
                k = 3)

# Join the clustering results back to the original data
store_list_with_cluster <- left_join(store_list, cluster_group)

• Define item:

  • Description: Item to cluster. In the context of our dataset, this would be the storecode

• Define feature:

  • Description: Feature column (dimension in clustering). In our case, the feature is the time component, which is represented by year column

• Define value:

  • Description: Value column. In our dataset, this would be the sales

• Define k:

  • Description: Number of clusters. This should be chosen based on the specific requirements of your analysis or determined using evaluation metrics. For the sake of simplicity in this tutorial, we will use 3 clusters.

• Joining Results: The clustering results are joined back to the original dataset.

Evaluating Clustering Results

• We can visualize the clustering results using ggplot2.

library(ggthemes)

store_list_with_cluster |> 
  ggplot(aes(x = year, y = sales, group = storecode, colour = cluster)) +
  geom_line(show.legend = F) +
  scale_y_continuous(labels = scales::comma) +
  facet_wrap(vars(cluster)) +
  scale_color_solarized()

• There you have it, a simple way to implement time series clustering using the widyr package in R. Of course, there is much more you can explore and refine in your clustering analysis. For comprehensive documentation and further exploration of the widyr package, visit the widyr page itself: widyr Documentation.