I'm working on a machine learning problem involving inventory (i.e. physical retail stock), however through the cleaning (outlier removal) process some of the items (via their corresponding transactions) will be removed. Therefore, I thought of using KNN to group similar items into respective categories.

There are 1245 items

The info for each item is

  1. Average Weighted Price
  2. Total Quantity Sold
  3. Total Revenue Achieved
  4. Min Sold per Transaction
  5. Max Sold per Transaction
  6. Min Sell Price
  7. Max Sell Price
  8. Number of Unique Transactions

Am I right in thinking that KNN is a good option - and if so, how do I decide on the number of clusters?

  • $\begingroup$ Do you mean K-NN or K-means? The former one usually refers to the classifier which classifies a new instance based on the features that have close proximity. I guess you want is the second one to cluster your data into k clusters. $\endgroup$ – Grzegorz Aug 22 '20 at 13:43

So your question is on the effectiveness of KNN to categories items based on features you have listed above.

As you might already know, KNN is a unsupervised clustering algorithm which creates K clusters with a minimal intra-cluster variation. This is method can be particularly use for when you know what the number of groups K you need. Also, it is particularly handy if you do not have any labels for categories for all examples.

At the same time, this method isn’t deterministic, which means that groupings do vary after each execution.

From this information, you might get a better idea for yourself as to whether KNN would be useful for this task.

  • $\begingroup$ Thanks for the insight. I only have 'category' categorical labels for about 20% of the items and therefore need an unsupervised approach. Now thinking k means with elbow method. $\endgroup$ – tristar8 Jul 21 '20 at 7:41
  • $\begingroup$ That then sounds reasonable to use KNN. I would recommend using the labelled subset to then validate the clusters which result from clustering. $\endgroup$ – shepan6 Jul 23 '20 at 7:41

Training: You can use a distance metric to compute the distance between all observations along the dimensions of your observed variables (Avg. Weight. Price, Tot. Quant. Sold, etc.). For each observation or row or sample i, the point with the smallest distance from that observation is the nearest neighbor. The point with the second smallest distance is the 2nd nearest neighbor, and so on.

Prediction: You can find the nearest neighbors for new data by calculating their distances to each point in the training data as above. A predicted label is then assigned, usually by taking the most common label amongst the test data points' k nearest neighbors. Hence k-NN classification:

from sklearn.neighbors import KNeighborsClassifier

knn = KNeighborsClassifier(algorithm='auto', 
                           metric='minkowski', # pick a distance metric
                           n_neighbors=5, # take the majority label from the 5-nearest neighbors
                           p=2, # a hyperparameter required for 'minkowski' distance metric

knn.fit(train_data, train_labels) 

# Find the predicted class of the test data:

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.