This challenge is about predicting if cryptocurrencies are affected by 24-hour or 7-day price changes by using unsupervised learning (PCA and KMeans) in Python.
Following tasks were completed:
- Loaded the crypto_market_data.csv into a dataframe, observed the summary statistics and plotted the data to see what the data looks like.
-
Prepared the data by normalizing it using StandardScaler() module from scikit-learn. Created a DataFrame with the scaled data and set the "coin_id" index from the original DataFrame as the index for the new DataFrame.
-
Used the elbow method to find the best value for k which came out to be 4.
- Clustered the cryptocurrencies for the best value for k on the original scaled data and created a scatter plot using hvPlot.
-
Using the original scaled DataFrame, performed a PCA and reduced the features to three principal components.
-
Retrieved the explained variance to determine how much information can be attributed to each principal component. What is the total explained variance of the three principal components?
-
Used the elbow method to find the Best Value for k Using the PCA Data.
Question: What is the best value for k when using the PCA data?
Answer: 4 is the best value for k when using the PCA data.
Question: Does it differ from the best k value found using the original data?
Answer: No, both methods suggeted the same value k=4. -
Clustered Cryptocurrencies with K-means Using the PCA Data and created a scatter plot using hvPlot What is the impact of using fewer features to cluster the data using K-Means?
The elbow curve 2 for the PCA data shows lower inertia value compared to the elbow curve for the original data elbow curve 1. This indicates that using less features reduces the sum of squared distances within the clusters. A lower inertia value suggests better clustering and tighter grouping of data points within each cluster.
Cluster plot 1 using the original data: Here the clusters (0, 1, 2) are overlapping, and there is no clear separation. This suggests that using the original data with more features may not effectively capture the underlying patterns of the data for clustering process.
Cluster plot 2 using PCA data: Here the clusters are not overlapping, and there are well-defined clusters as there is a clear separation. We can observe 4 distinct clusters. This indicates that using fewer features through PCA has potentially reduced noise and helped reveal more distinct patterns in the data, leading to more meaningful and distinct clusters.