I am running k-means clustering on ~1 million items (each represented as a ~100-feature vector). I have run the clustering for various k, and now want to evaluate the different results with the silhouette score implemented in sklearn. Attempting to run it with no sampling seems unfeasible and takes a prohibitively long time, so I assume I need to use sampling, i.e.:
metrics.silhouette_score(feature_matrix, cluster_labels, metric='euclidean',sample_size=???)
I don't have a good sense of what an appropriate sampling approach is, however. Is there a rule of thumb for what size sample to use given the size of my matrix? Is it better to take the largest sample my analysis machine can handle, or to take the average of more smaller samples?
I ask in large part because my preliminary test (with sample_size=10000) has produced some really really unintuitive results.
I'm also open to alternative, more scalable evaluation metrics.
Editing to visualize the issue: The plot shows, for varying sample sizes, the silhouette score as a function of the number of clusters
What's not weird is that increasing sample size seems to reduce noise. What is weird, given that I have 1 million, very heterogenous vectors, that 2 or 3 is the "best" number of clusters. In other words, what's unintuitive is that I would find a more-or-less monotonic decreases in silhouette score as I increase the number of clusters.
Other metrics
Elbow method: Compute the % variance explained for each K, and choose the K where the plot starts to level off. (a good description is here https://en.wikipedia.org/wiki/Determining_the_number_of_clusters_in_a_data_set). Obviously if you have k == number of data points, you can explain 100% of the variance. The question is where do the improvements in variance explained start to level off.
Information theory: If you can calculate a likelihood for a given K, then you can use the AIC, AICc, or BIC (or any other information-theoretic approach). E.g. for the AICc, it just balances the increase in likelihood as you increase K with the increase in the number of parameters you need. In practice all you do is choose the K that minimises the AICc.
You may be able to get a feel for a roughly appropriate K by running alternative methods that give you back an estimate of the number of clusters, like DBSCAN. Though I haven't seen this approach used to estimate K, and it is likely inadvisable to rely on it like this. However, if DBSCAN also gave you a small number of clusters here, then there's likely something about your data that you might not be appreciating (i.e. not as many clusters are you're expecting).
How much to sample
It looks like you've answered this from your plot: no matter what your sampling you get the same pattern in silhouette score. So that patterns seems very robust to sampling assumptions.
kmeans converge to local minima. Starting positions plays a crucial role in optimal number of clusters. It would be a good idea often to reduce the noise and dimensions using PCA or any other dimension reduction techniques to proceed with kmeans.
Just to add for the sake of completeness. It might be a good idea to get optimal number of clusters by "partition around medoids". It is equivalent to using silhouette method.
Reason for the weird observations could be different starting points for different sized samples.
Having said all the above, it is important to evaluate clusterability of the dataset in hand. Tractable means is by Worst Pair ratio as discussed here Clusterability.
Since there is no widely-accepted best approach to determine the optimal number of clusters, all evaluation techniques, including Silhouette Score, Gap Statistic, etc. fundamentally rely on some form of heuristic/trial&error argument. So to me, the best approach is to try out multiple techniques and to NOT develop over-confidence in any.
In your case, the ideal and most accurate score should be calculated on the entire data set. However, if you need to use partial samples to speed up the computation, you should use largest possible sample size your machine can handle. The rationale is the same as getting as many data points as possible out of the population of interest.
One more thig is that the sklearn implementation of Silhouette Score uses random (non-stratified) sampling. You can repeat the calculation multiple time using the same sample size (say sample_size=50000) to get a sensing on whether the sample size is large enough to produce consistent results.
Related
This Question is more Theoretical, and not specifically trying to problem-solve.
I recently was introduced to the K-Means Clustering algorithm, and unsupervised machine learning algorithm, and I was intrigued by the though that one some sets of data, even if completely random, the average centroids drawn could keep changing through each iteration.
Example:
What I am trying to show here, is, imagine if the program flipped between iteration 6, to iteration 9, and kept doing this forever.
I have had my code randomly hang before using K-Means, so I don't believe this is impossible, but please let me know if this is a known occurrence, or if it is impossible due to the nature of the algorithm.
If you need more information just ask me in a comment. Using Python 3.7
tl;dr No, a K-means algorithm always has an end point if the algorithm is coded correctly.
Explanation:
The ideal way to think about this is not in the sense of what datapoints would cause issues, but rather about how kmeans is working in the broader sense of things. The k-means algorithm is always working in a finite space. For N data points, there are only N ^ k distinct arrangements for the data points. (This number can be pretty large, but is still finite)
Secondly, a k-means algorithm is always optimizing a loss function, based on the sum of squared distances between each data point and it's assigned cluster center. This means two very important things: Each of the N ^ k distinct arrangements can be arranged in an ascending/descending order of minimum loss to maximum loss. Also, the K-means algorithm will never go from a state of lower net loss to a higher net loss.
These two conditions guarantee that the algorithm will always tend towards the minimum loss arrangement in a finite space, thus ensuring that it has an end.
The last edge case: What if more than one minimum state has equal loss? This is a highly unlikely scenario, but can cause issues if and only if the algorithm is coded poorly for tie breakers. Essentially, the only way this can cause a cycle is if a data point has equal distance for two clusters, and is allowed to change clusters away from it's current cluster even on equal distance. Suffice to say, the algorithms are generally coded so that the data points never swap on a tie, or in some other deterministic manner, thus avoiding this scenario entirely.
As seen above,how to build clusters that are approximately balanced in size in sklearn?I have a question，clustering is done according to certain rules,Why can we specify the number in cluster?Anyway, I want to know how to achieve this step.
I have another idea about it.Calculate the number of each label, then calculate the variance,and get the one with the smallest variance
Some methods (for example, non-sklearn's HDBSCAN: https://hdbscan.readthedocs.io/en/latest/parameter_selection.html) have parameters like minimal_cluster_size. Probably, sklearn's DBSCAN's min_samples will work the similar way. It will not give you exact 'balanced' clusters but may help.
But in my opinion, sometimes it is more reasonable to run clusterization algorithms with different parameters and select 'more balanced' output by your hands. In this case you can see what points are not separable and probably add more data (calculate additional distance matrix, for example) or change target metric.
Why can we specify the number in cluster?
Because the tasks 'find clusters' and 'balance them' are a bit opposite in their meaning in the most cases. I'm not even speaking about algorithms when you need to specify the number of clusters.
I am a PhD student who is trying to use the NEAT algorithm as a controller for a robot and I am having some accuracy issues with it. I am working with Python 2.7 and for it and am using two NEAT python implementations:
The NEAT which is in this GitHub repository: https://github.com/CodeReclaimers/neat-python
Searching in Google, it looks like it has been used in some projects with succed.
The multiNEAT library developed by Peter Chervenski and Shane Ryan: http://www.multineat.com/index.html.
Which appears in the "official" software web page of NEAT software catalog.
While testing the first one, I've found that my program converges quickly to a solution, but this solution is not precise enough. As lack of precision I want to say a deviation of a minimum of 3-5% in the median and average related to the "perfect" solution at the end of the evolution (Depending on the complexity of the problem, an error around 10% is normal for my solutions. Furthermore, I could said that I've "never" seen an error value under the 1% between the solution given by the NEAT and the solution that it is the correct one). I must said that I've tried a lot of different parameter combinations and configurations (this is an old problem for me).
Due to that, I tested the second library. The MultiNEAT library converges quickly and easier that the previous one. (I assume that is due to the C++ implementation instead the pure Python) I get similar results, but I still have the same problem; lack of accuracy. This second library has different configuration parameters too, and I haven't found a proper combination of them to improve the performance of the problem.
My question is:
Is it normal to have this lack of accuracy in the NEAT results? It achieves good solutions, but not good enough for controlling a robot arm, which is what I want to use it for.
I'll write what I am doing in case someone sees some conceptual or technical mistake in the way I set out my problem:
To simplify the problem, I'll show a very simple example: I have a very simple problem to solve, I want a NN that may calculate the following function: y = x^2 (similar results are found with y=x^3 or y = x^2 + x^3 or similar functions)
The steps that I follow to develop the program are:
"Y" are the inputs to the network and "X" the outputs. The
activation functions of the neural net are sigmoid functions.
I create a data set of "n" samples given values to "X" between the
xmin = 0.0 and the xmax = 10.0
As I am using sigmoid functions, I make a normalization of the "Y"
and "X" values:
"Y" is normalized linearly between (Ymin, Ymax) and (-2.0, 2.0) (input range of sigmoid).
"X" is normalized linearly between (Xmin, Xmax) and (0.0, 1.0) (the output range of sigmoid).
After creating the data set, I subdivide in in a train sample (70%
percent of the total amount), a validation sample and a test sample
(15% each one).
At this point, I create a population of individuals for doing
evolution. Each individual of the population is evaluated in all the
train samples. Each position is evaluated as:
eval_pos = xmax - abs(xtarget - xobtained)
And the fitness of the individual is the average value of all the train positions (I've selected the minimum too but it gives me worse performance).
After the whole evaluation, I test the best obtained individual
against the test sample. And here is where I obtained those
"un-precise values". Moreover, during the evaluation process, the
maximum value where "abs(xtarget - xobtained) = 0" is never
obtained.
Furthermore, I assume that how I manipulate the data is right because, I use the same data set for training a neural network in Keras and I get much better results than with NEAT (an error less than a 1% is achievable after 1000 epochs in a layer with 5 neurons).
At this point, I would like to know if what is happened is normal because I shouldn't use a data set of data for developing the controller, it must be learned "online" and NEAT looks like a suitable solution for my problem.
Thanks in advance.
EDITED POST:
Firstly, Thanks for comment nick.
I'll answer your questions below::
I am using the NEAT algorithm.
Yes, I've carried out experiments increasing the number of individuals in the population and the generations number. A typical graph that I get is like this:
Although the population size in this example is not such big, I've obtained similar results in experiments incrementing the number of individuals or the number of generations. Populations of 500 in individuals and 500 generations, for example. In this experiments, the he algorithm converge fast to a solution, but once there, the best solution is stucked and it does not improve any more.
As I mentioned in my previous post, I've tried several experiments with many different parameters configurations... and the graphics are more or less similar to the previous showed.
Furthermore, other two experiments that I've tried were: once the evolution reach the point where the maximum value and the median converge, I generate other population based on that genome with new configuration parameters where:
The mutation parameters change with a high probability of mutation (weight and neuron probability) in order to find new solutions with the aim to "jumping" from the current genome to other better.
The neuron mutation is reduced to 0, while the weight "mutation probability" increase for "mutate weight" in a lower range in order to get slightly modifications with the aim to get a better adjustment of the weights. (trying to get a "similar" functionality as backprop. making slighty changes in the weights)
This two experiments didn't work as I expected and the best genome of the population was also the same of the previous population.
I am sorry, but I do not understand very well what do you want to say with "applying your own weighted penalties and rewards in your fitness function". What do you mean with including weight penalities in the fitness function?
Regards!
Disclaimer: I have contributed to these libraries.
Have you tried increasing the population size to speed up the search and increasing the number of generations? I use it for a trading task, and by increasing the population size my champions were found much sooner.
Another thing to think about is applying your own weighted penalties and rewards in your fitness function, so that anything that doesn't get very close right away is "killed off" sooner and the correct genome is found faster. It should be noted that neat uses a fitness function to learn as a opposed to gradient descent so it wont converge in the same way and its possible you may have to train a bit longer.
Last question, are you using the neat or hyperneat algo from multineat?
Are there any types of clustering algorithms that focus on forming specific sized clusters? This can be thought of us as a grouping algorithm more than a clustering algorithm.
Basically, given n data points, and fixed groups of a certain size k, find the optimal distribution of points to sets based upon certain classifiers, that will hopefully minimize the distance of classifiers for each point in a given group.
This problem seems to be pretty similar to a clustering problem, but the main difference is that we are concerned with a specific cluster size, but not concerned about the number of clusters.
There is a tutorial on how to implement such an algorithm in ELKI:
http://elki.dbs.ifi.lmu.de/wiki/Tutorial/SameSizeKMeans
Also have a look at constraint clustering algorithms; although usually these algorithms only support "Must link" and "cannot link" constraints, not size constraints.
You should be able to do a similar modification where you first specify the group sizes, then assign points randomly, and swap cluster members as long as your objective function improves; similar to k-means / k-medoids. As you may get stuck in local minima, restart a number of times and only keep the best.
See also earlier questions, e.g.
K-means algorithm variation with equal cluster size
and
Group n points in k clusters of equal size
The problem that you are posing is a combinatorial optimization problem. It is very important to know if you need an exact solution, or that can you settle for an approximate one?
If you need exact solutions, there is a body of work that focuses on clustering with different types of constraints. The constraint that you mentioned can be encoded in this framework. However, you should now that this approach scales up to a datasets with a certain size.
I have written code in python to implement DBSCAN clustering algorithm.
My dataset consists of 14k users with each user represented by 10 features.
I am unable to decide what exactly to keep as the value of Min_samples and epsilon as input
How should I decide that?
Similarity measure is euclidean distance.(Hence it becomes even more tough to decide.) Any pointers?
DBSCAN is pretty often hard to estimate its parameters.
Did you think about the OPTICS algorithm? You only need in this case Min_samples which would correspond to the minimal cluster size.
Otherwise for DBSCAN I've done it in the past by trial and error : try some values and see what happens. A general rule to follow is that if your dataset is noisy, you should have a larger value, and it is also correlated with the number of dimensions (10 in this case).