This is somewhere in my code library, but here it is again, as I might use it for the Massive Version of Black Tree, which should be ready early to mid-June. All it does is test accuracy as a function delta, rather than run the original supervision step. The idea here is, you produce a single value that generates the highest accuracy, or the desired number of rejections.
Massive Supervised Delta
This algorithm is a more efficient analog of the algorithm that calculates delta for my Supervised Delta Algorithm. More to come.
Updated Sort-Based Classification
This is exactly the same technique I posted the other day, that uses sorting to achieve an approximation of the nearest neighbor method. Then, clusters are generated by beginning at a given index in the sorted list, and including rows to the left and right of the index, iterating through increasing boundary indexes, simulating expanding the radius of a sphere (i.e., as you include more rows to the left and right of a given index, you are effectively including all rows within a sphere of increasing radius from the origin / index). The only substantive change in this case, is the use of volume in calculating confidence (the other changes were corrections). Specifically, uncertainty increases as a function of volume, for the simple reason that, if you have a cluster contained within a large volume, and few rows in the cluster, then you have more uncertainty (c.f., a large number of rows contained in a small volume). In my original confidence-based classification algorithm, the spheres are all fixed sizes, so you could argue it didn’t matter in that case (especially since it works). I came up with yet another method, that makes use of a combinatorial measure of uncertainty that I mentioned in Footnote 14 of Information, Knowledge, and Uncertainty, which also should work for variable sized clusters, which I should be done with soon.
This particular algorithm is incredibly fast, classifying 4,500 testing rows, over 25,500 training rows, in about two seconds, with accuracy that often peaks at close to perfect (i.e., accuracy increases as a function of confidence). The chart below shows accuracies and runtimes for this method, on average over 100 iterations. All of the testing percentages are fixed at 15% (i.e., 30,000 rows implies 4,500 Testing Rows and 25,500 Training Rows). The Raw Accuracy is the accuracy of the method on its own, with no confidence filtering, which is almost always terrible. The Max Accuracies are given for both the information-based confidence metric, which diminishes as a function of volume, and the size-based confidence metric, which does not, and simply looks to the number of elements in each cluster (i.e., the more cluster elements, the higher the confidence). Confidence filtering is achieved by simply ignoring all predictions that carry a confidence of less than 
|
Dataset |
Raw Acc. |
Max Acc. (Inf.) |
Max Acc. (Size) |
Runtime (Seconds) |
No. Rows |
|
UCI Credit |
38.07% |
78.30% |
100.0% |
3.290 |
30,000 |
|
UCI Abalone |
17.37% |
83.50% |
71.30% |
0.322 |
4,177 |
|
UCI Spam |
40.83% |
98.61% |
100.0% |
0.469 |
4,601 |
|
UCI Ion |
53.00% |
94.30% |
100.0% |
0.057 |
351 |
|
UCI Iris |
32.50% |
100.0% |
86.20% |
0.024 |
150 |
|
UCI Parkinsons |
47.20% |
95.60% |
100.0% |
0.035 |
195 |
|
UCI Wine |
37.50% |
94.20% |
100.0% |
0.029 |
178 |
|
UCI Sonar |
20.16% |
58.11% |
34.80% |
0.048 |
208 |
There’s more to come, with simple variations on this overall method, that just reapplies my core work using a sorted list. Once that’s complete, I’ll write something more formal about these methods.
Here’s the code, you can find the normalization algorithm a few posts below. The Command Line is set up as a function, for convenience (MASSTempFunction), so just change the dataset path in that file, and you’re good to go.
Chaos and Information
Chaos is a word that gets thrown around a lot, and I know nothing about Chaos Theory, which is at this point probably a serious branch of mathematics, but again, I know nothing about it. However, it dawned on me, that you can think about chaos in terms of information, using Cellular Automata as the intuition. Specifically, consider the initial conditions of a binary automaton, which together form a single vector of 
Said in words, 
Updated Sort-Based Classification
This is basically the same as the original algorithm on the topic, though I fixed a few bugs, and provided a command line that makes it easy to generate a confidence / accuracy distribution. The runtime is astonishing, about 2 seconds per 30,000 rows. More to come soon.
Supervised Clustering Algorithm
Here’s another version of my supervised clustering algorithm that saves a few steps using sorting, but it turns out it doesn’t matter much (it seems to be about twice as fast).
Massive Clustering / Prediction Algorithm (Second Version)
Here’s another version of the same algorithm from the previous post, there are some slight changes, and this one seems to have higher accuracy in general. Again, I’ll post something explaining how they work sometime soon.
Massive Clustering / Prediction Algorithm
This is an extremely fast algorithm similar to the one I posted the other day, I’ll write something more fulsome on it shortly.
VeGa – Updated Draft
I won’t publish the complete draft, because it is borderline revolutionary, but this is charming, and harmless, so you should read it:
Using Sorting in Clustering
I use sorting all the time in my algorithms to simulate running nearest neighbor (see Section 2.4 of Vectorized Deep Learning), and what just dawned on me, is that I actually proved formally, that a list is sorted if and only if its adjacent entries have the minimum possible distance (see Theorem 2.1 of Sorting, Information, and Recursion). This implies, the resultant sorted list, provides you with the nearest neighbors of each element in the list. This in turn allows for a trivial adaptation of my core algorithms, where rather than take the norm of the distance between a given vector and all others, you simply take the norm of the difference between a given vector and the vectors in the order in which they’re sorted in the list. The advantage in that case, is that if you’re not running the algorithms truly in parallel (which is the case on consumer devices when you have too many rows), then you’re only performing one operation per comparison. Attached is an example using my supervised clustering algorithm, which increases the radius of a sphere, until it hits its first error, which in this case means simply increasing the index of a sorted list, until you encounter a classifier that is unequal to the classifier in question (i.e., the origin of the sphere). This produces really fast runtimes, running in about 10 seconds given 100,000 rows with 15 columns – This is pretty serious stuff, and will be included in the Massive Version of Black Tree AutoML, for just $999. A mutually exclusive version (i.e., non-intersecting clusters) would typically produce even faster runtimes, since the size of the effective dataset can reduce each iteration.
For a testing dataset, you could simply combine the training and testing datasets, store the entries of the testing rows, and then go out some radius from each testing row by checking the classifiers of the rows to the left and right of each testing row. Applying the attached approach (i.e., first error), you would proceed until you encountered more than one class. You could instead proceed by no more than some fixed distance, or some fixed number of entries. You could report the modal class, or simply report the entire cluster of classes as a prediction. This will be extremely fast, since you’re operating only on the testing rows and the adjacent training rows, rather than the entire training dataset (save for the sorting step). I’ve attached code that implements this method, which seems to work really well, though more testing is required. I’ve included a basic confidence metric that also seems to work, in that accuracy increases as a function of confidence. This code is applied to the MNIST Fashion Dataset, which makes use of image preprocessing algorithms you can find in my A.I. Library on ResearchGate, but you can also simply ignore the preprocessing, as everything past the heading, “Runs Prediction”, is generalized and requires only a dataset.
Here is a plot of accuracy as a function of confidence over the MNIST Fashion Dataset:
Information Overload 