k-nearest neighbors algorithm

It has everything.

It's simple, it's fast, it's easy to implement, and it's one of my favorite machine learning algorithms, K-Nearest Neighbors.

I sometimes do contracting and consulting work for clients, and the number of times I've had to tell somebody that their problem really doesn't require deep learning or neural networks is substantial.

So when I see something super simple like this, doing text classification for sentiment analysis with K nearest neighbors for the classification and gzip for the compression distances, which is kind of being used as your feature vectors, I really just have to try that for myself.

But this is due to, I think, my simplistic methods.

I think I'm calculating NCD different and all that.

And the fact that with my methods were at 70% accuracy with, again, only 500 samples up to this point, Even that alone, I really just want to stress how absolutely staggering that is to me, that that is even possible, that there is something there for K-nearest neighbors to nibble on and make some progress.

I still can't believe it, and yes, I sat on this because I still just, I feel like there's got to be a problem, and I didn't really want to release a video about on this because something has to be wrong here um uh and i still suspect there is and maybe someone can can point it out and get us back to what i believe should be the worthy accuracy for this model which is likely 50 percent now on the flip side what actually is going on here right i can logically think through why normalized compression distances might work for comparing strings in some ways.

There's been over two million downloads since preview launched at Google Cloud Next '26.

The database combines SQL, graph, key value, full text search, and vector search in a single multi-model engine with native support for KNN and ANN vector search, and Spanner graph for AI agentic application workloads.

The general availability adds enterprise features needed for production, including TLS encryption, audit logging, backup restores, dedicated worker nodes for offloading background operations, and direct Google Cloud customer care support.

Licensing is split into two tiers, a free developer edition with nine-day license or perpetual for single server deployments, uh, for the vCPUs or less, excluding backup and worker nodes, and a paid commercial edition using a vCPU based annual subscription model with discounted proof of concept pricing available to you.

There were six.

One of them was LLM and the other five were Bayesian, KNN, multiple regression, penalized regression, whatever.

But then you know, and then you know the whole cycle of the accounting.

Since you are filling up the information for the closing, Until you are preparing your financial projections.

And you have to identify where you can play or you can use different modes.

But you know where you can imagine things and use LLMs at the end in projections because they are projections.

But you have to use KNN for plot detection and making sure the information is correct or just detect for whatever agent what may be wrong.

Then when you understand where the different models, stochastic models cannot be applied at the different financial, then you are better off.