Hello to everyone. I am trying to implement KNN analysis to fix minPts in the DBSCAN clustering algorithm. My dataset is composed only of 4 variables and 935 observations. I have found that if k = 5 (no. of variables + 1) I get as output of DBASCAN 2 clusters: one of 911 observations and one of 8 observations. If I use a larger k, according to many papers as sqrt(no. of observations), I get 909 observation in only one cluster and the other are classified as noise points.
Both could be possible results, but their meaning is fundementally different. How can I get rid of this arbitrary choise of minPts hence k?
Thanks!
the boundary becomes smoother with increasing value of K. The training error rate and the validation error rate are two parameters we need to access different K-value.
To get the optimal value of K, you can segregate the training and validation from the initial dataset. Now plot the validation error curve to get the optimal value of K. This value of K should be used for all predictions.
The optimal K value usually found is the square root of N, where N is the total number of samples. Use an error plot or accuracy plot to find the most favorable K value. KNN performs well with multi-label classes, but you must be aware of the outliers.
https://www.analyticsvidhya.com/blog/2018/03/introduction-k-neighbours-algorithm-clustering/
Dear Vincenzo Guida,
There is heugh literature on this topic, see e.g., https://www.analyticsvidhya.com/blog/2018/03/introduction-k-neighbours-algorithm-clustering/ and https://realpython.com/knn-python/ for basic concepts.
Good Luck
It may vary for your dataset and problem domain. You can use a grid search like technique to find the optimal values of k.
Hi, Vincenzo Guida
There are various methods to choose the best k in KNN. I am listing a few below:
- Badges
- Science topic