When a data set is analyzed by a clustering algorithm in ELKI 0.5, the program produces a number of statistics: the Jaccard index, F1-Measures, etc. In order to calculate these statistics, there have to be 2 clusterings to compare. What is the clustering created by the algorithm compared to?
The automatic evaluation (note that you can configure the evaluation manually!) is based on labels in your data set. At least in the current version (why are you using 0.5 and not 0.6.0?) it should only automatically evaluate if it finds labels in the data set.
We currently have not published internal measures. There are some implementations, such as evaluation/clustering/internal/EvaluateSilhouette.java, some of which will be in the next release.
In my experiments, internal evaluation measures were badly misleading. For example on the Silhouette coefficient, the labeled "solution" would often even score a negative silhouette coefficient (i.e. worse than not clustering at all).
Also, these measures are not scalable. The silhouette coefficient is in O(n^2) to compute; which usually makes this evaluation more expensive than the actual clustering!
We do appreciate contributions!
You are more than welcome to contribute your favorite evaluation measure to ELKI, to share with others.
Related
I have a question regarding cross validation in Linear regression model.
From my understanding, in cross validation, we split the data into (say) 10 folds and train the data from 9 folds and the remaining folds we use for testing. We repeat this process until we test all of the folds, so that every folds are tested exactly once.
When we are training the model from 9 folds, should we not get a different model (may be slightly different from the model that we have created when using the whole dataset)? I know that we take an average of all the "n" performances.
But, what about the model? Shouldn't the resulting model also be taken as the average of all the "n" models? I see that the resulting model is same as the model which we created using whole of the dataset before cross-validation. If we are considering the overall model even after cross-validation (and not taking avg of all the models), then what's the point of calculating average performance from n different models (because they are trained from different folds of data and are supposed to be different, right?)
I apologize if my question is not clear or too funny.
Thanks for reading, though!
I think that there is some confusion in some of the answers proposed because of the use of the word "model" in the question asked. If I am guessing correctly, you are referring to the fact that in K-fold cross-validation we learn K-different predictors (or decision functions), which you call "model" (this is a bad idea because in machine learning we also do model selection which is choosing between families of predictors and this is something which can be done using cross-validation). Cross-validation is typically used for hyperparameter selection or to choose between different algorithms or different families of predictors. Once these chosen, the most common approach is to relearn a predictor with the selected hyperparameter and algorithm from all the data.
However, if the loss function which is optimized is convex with respect to the predictor, than it is possible to simply average the different predictors obtained from each fold.
This is because for a convex risk, the risk of the average of the predictor is always smaller than the average of the individual risks.
The PROs and CONs of averaging (vs retraining) are as follows
PROs: (1) In each fold, the evaluation that you made on the held out set gives you an unbiased estimate of the risk for those very predictors that you have obtained, and for these estimates the only source of uncertainty is due to the estimate of the empirical risk (the average of the loss function) on the held out data.
This should be contrasted with the logic which is used when you are retraining and which is that the cross-validation risk is an estimate of the "expected value of the risk of a given learning algorithm" (and not of a given predictor) so that if you relearn from data from the same distribution, you should have in average the same level of performance. But note that this is in average and when retraining from the whole data this could go up or down. In other words, there is an additional source of uncertainty due to the fact that you will retrain.
(2) The hyperparameters have been selected exactly for the number of datapoints that you used in each fold to learn. If you relearn from the whole dataset, the optimal value of the hyperparameter is in theory and in practice not the same anymore, and so in the idea of retraining, you really cross your fingers and hope that the hyperparameters that you have chosen are still fine for your larger dataset.
If you used leave-one-out, there is obviously no concern there, and if the number of data point is large with 10 fold-CV you should be fine. But if you are learning from 25 data points with 5 fold CV, the hyperparameters for 20 points are not really the same as for 25 points...
CONs: Well, intuitively you don't benefit from training with all the data at once
There are unfortunately very little thorough theory on this but the following two papers especially the second paper consider precisely the averaging or aggregation of the predictors from K-fold CV.
Jung, Y. (2016). Efficient Tuning Parameter Selection by Cross-Validated Score in High Dimensional Models. International Journal of Mathematical and Computational Sciences, 10(1), 19-25.
Maillard, G., Arlot, S., & Lerasle, M. (2019). Aggregated Hold-Out. arXiv preprint arXiv:1909.04890.
The answer is simple: you use the process of (repeated) cross validation (CV) to obtain a relatively stable performance estimate for a model instead of improving it.
Think of trying out different model types and parametrizations which are differently well suited for your problem. Using CV you obtain many different estimates on how each model type and parametrization would perform on unseen data. From those results you usually choose one well suited model type + parametrization which you will use, then train it again on all (training) data. The reason for doing this many times (different partitions with repeats, each using different partition splits) is to get a stable estimation of the performance - which will enable you to e.g. look at the mean/median performance and its spread (would give you information about how well the model usually performs and how likely it is to be lucky/unlucky and get better/worse results instead).
Two more things:
Usually, using CV will improve your results in the end - simply because you take a model that is better suited for the job.
You mentioned taking the "average" model. This actually exists as "model averaging", where you average the results of multiple, possibly differently trained models to obtain a single result. Its one way to use an ensemble of models instead of a single one. But also for those you want to use CV in the end for choosing reasonable model.
I like your thinking. I think you have just accidentally discovered Random Forest:
https://en.wikipedia.org/wiki/Random_forest
Without repeated cv your seemingly best model is likely to be only a mediocre model when you score it on new data...
I have used the ELKI implementation of DBSCAN to identify fire hot spot clusters from a fire data set and the results look quite good. The data set is spatial and the clusters are based on latitude, longitude. Basically, the DBSCAN parameters identify hot spot regions where there is a high concentration of fire points (defined by density). These are the fire hot spot regions.
My question is, after experimenting with several different parameters and finding a pair that gives a reasonable clustering result, how does one validate the clusters?
Is there a suitable formal validation method for my use case? Or is this subjective depending on the application domain?
ELKI contains a number of evaluation functions for clusterings.
Use the -evaluator parameter to enable them, from the evaluation.clustering.internal package.
Some of them will not automatically run because they have quadratic runtime cost - probably more than your clustering algorithm.
I do not trust these measures. They are designed for particular clustering algorithms; and are mostly useful for deciding the k parameter of k-means; not much more than that. If you blindly go by these measures, you end up with useless results most of the time. Also, these measures do not work with noise, with either of the strategies we tried.
The cheapest are the label-based evaluators. These will automatically run, but apparently your data does not have labels (or they are numeric, in which case you need to set the -parser.labelindex parameter accordingly). Personally, I prefer the Adjusted Rand Index to compare the similarity of two clusterings. All of these indexes are sensitive to noise so they don't work too well with DBSCAN, unless your reference has the same concept of noise as DBSCAN.
If you can afford it, a "subjective" evaluation is always best.
You want to solve a problem, not a number. That is the whole point of "data science", being problem oriented and solving the problem, not obsessed with minimizing some random quality number. If the results don't work in reality, you failed.
There are different methods to validate a DBSCAN clustering output. Generally we can distinguish between internal and external indices, depending if you have labeled data available or not. For DBSCAN there is a great internal validation indice called DBCV.
External Indices:
If you have some labeled data, external indices are great and can demonstrate how well the cluster did vs. the labeled data. One example indice is the RAND indice.https://en.wikipedia.org/wiki/Rand_index
Internal Indices:
If you don't have labeled data, then internal indices can be used to give the clustering result a score. In general the indices calculate the distance of points within the cluster and to other clusters and try to give you a score based on the compactness (how close are the points to each other in a cluster?) and
separability (how much distance is between the clusters?).
For DBSCAN, there is one great internal validation indice called DBCV by Moulavi et al. Paper is available here: https://epubs.siam.org/doi/pdf/10.1137/1.9781611973440.96
Python package: https://github.com/christopherjenness/DBCV
Can we use Hierarchical agglomerative clustering for clustering data in this format ?
"beirut,proff,email1"
"beirut,proff,email2"
"swiss,aproff,email1"
"france,instrc,email2"
"swiss,instrc,email2"
"beirut,proff,email1"
"swiss,instrc,email2"
"france,aproff,email2"
If not, what is the compatible clustering algorithm to cluster data with string values ?
Thank you for your help!
Any type of clustering requires a distance metric. If all you're willing to do with your strings is treat them as equal to each other or not equal to each other, the best you can really do is the field-wise Hamming distance... that is, the distance between "abc,def,ghi" and "uvw,xyz,ghi" is 2, and the distance between "abw,dez,ghi" is also 2. If you want to cluster similar strings within a particular field -- say clustering "Slovakia" and "Slovenia" because of the name similarity, or "Poland" and "Ukraine" because they border each other, you'll use more complex metrics. Given a distance metric, hierarchical agglomerative clustering should work fine.
All this assumes, however, that clustering is what you actually want to do. Your dataset seems like sort of an odd use-case for clustering.
Hierarchical clustering is a rather flexible clustering algorithm. Except for some linkages (Ward?) it does not have any requirement on the "distance" - it could be a similarity as well, usually negative values will work just as well, you don't need triangle inequality etc.
Other algorithms - such as k-means - are much more limited. K-means minimizes variance; so it can only handle (squared) Euclidean distance; and it needs to be able to compute means, thus the data needs to be in a continuous, fixed dimensionality vector space; and sparsity may be an issue.
One algorithm that probably is even more flexible is Generalized DBSCAN. Essentially, it needs a binary decision "x is a neighbor of y" (e.g. distance less than epsilon), and a predicate to measure "core point" (e.g. density). You can come up with arbitary complex such predicates, that may no longer be a single "distance" anymore.
Either way: If you can measure similarity of these records, hiearchical clustering should work. The question is, if you can get enough similarity out of that data, and not just 3 bit: "has the same email", "has the same name", "has the same location" -- 3 bit will not provide a very interesting hierarchy.
I have a set of weighted features for machine learning. I'd like to reduce the feature set and just use those with a very large or very small weight.
So given below image of sorted weights, I'd only like to use the features that have weights above the higher or below the lower yellow line.
What I'm looking for is some kind of slope change detection so I can discard all the features until the first/last slope coefficient increase/decrease.
While I (think I) know how to code this myself (with first and second numerical derivatives), I'm interested in any established methods. Perhaps there's some statistic or index that computes something like that, or anything I can use from SciPy?
Edit:
At the moment, I'm using 1.8*positive.std() as positive and 1.8*negative.std() as negative threshold (fast and simple), but I'm not mathematician enough to determine how robust this is. I don't think it is, though. ⍨
If the data are (approximately) Gaussian distributed, then just using a multiple
of the standard deviation is sensible.
If you are worried about heavier tails, then you may want to base your analysis on order
statistics.
Since you've plotted it, I'll assume you're willing to sort all of the
data.
Let N be the number of data points in your sample.
Let x[i] be the i'th value in the sorted list of values.
Then 0.5( x[int( 0.8413*N)]-x[int(0.1587*N)]) is an estimate of the standard deviation
which is more robust against outliers. This estimate of the std can be used as you
indicated above. (The magic numbers above are the fraction of data that are
less than [mean+1sigma] and [mean-1sigma] respectively).
There are also conditions where just keeping the highest 10% and lowest 10% would be
sensible as well; and these cutoffs are easily computed if you have the sorted data
on hand.
These are somewhat ad hoc approaches based on the content of your question.
The general sense of what you're trying to do is (a form of) anomaly detection,
and you can probably do a better job of it if you're careful in defining/estimating
what the shape of the distribution is near the middle, so that you can tell when
the features are getting anomalous.
What is the most popular text clustering algorithm which deals with large dimensions and huge dataset and is fast?
I am getting confused after reading so many papers and so many approaches..now just want to know which one is used most, to have a good starting point for writing a clustering application for documents.
To deal with the curse of dimensionality you can try to determine the blind sources (ie topics) that generated your dataset. You could use Principal Component Analysis or Factor Analysis to reduce the dimensionality of your feature set and to compute useful indexes.
PCA is what is used in Latent Semantic Indexing, since SVD can be demonstrated to be PCA : )
Remember that you can lose interpretation when you obtain the principal components of your dataset or its factors, so you maybe wanna go the Non-Negative Matrix Factorization route. (And here is the punch! K-Means is a particular NNMF!) In NNMF the dataset can be explained just by its additive, non-negative components.
There is no one size fits all approach. Hierarchical clustering is an option always. If you want to have distinct groups formed out of the data, you can go with K-means clustering (it is also supposedly computationally less intensive).
The two most popular document clustering approaches, are hierarchical clustering and k-means. k-means is faster as it is linear in the number of documents, as opposed to hierarchical, which is quadratic, but is generally believed to give better results. Each document in the dataset is usually represented as an n-dimensional vector (n is the number of words), with the magnitude of the dimension corresponding to each word equal to its term frequency-inverse document frequency score. The tf-idf score reduces the importance of high-frequency words in similarity calculation. The cosine similarity is often used as a similarity measure.
A paper comparing experimental results between hierarchical and bisecting k-means, a cousin algorithm to k-means, can be found here.
The simplest approaches to dimensionality reduction in document clustering are: a) throw out all rare and highly frequent words (say occuring in less than 1% and more than 60% of documents: this is somewhat arbitrary, you need to try different ranges for each dataset to see impact on results), b) stopping: throw out all words in a stop list of common english words: lists can be found online, and c) stemming, or removing suffixes to leave only word roots. The most common stemmer is a stemmer designed by Martin Porter. Implementations in many languages can be found here. Usually, this will reduce the number of unique words in a dataset to a few hundred or low thousands, and further dimensionality reduction may not be required. Otherwise, techniques like PCA could be used.
I will stick with kmedoids, since you can compute the distance from any point to anypoint at the beggining of the algorithm, You only need to do this one time, and it saves you time, specially if there are many dimensions. This algorithm works by choosing as a center of a cluster the point that is nearer to it, not a centroid calculated in base of the averages of the points belonging to that cluster. Therefore you have all possible distance calculations already done for you in this algorithm.
In the case where you aren't looking for semantic text clustering (I can't tell if this is a requirement or not from your original question), try using Levenshtein distance and building a similarity matrix with it. From this, you can use k-medoids to cluster and subsequently validate your clustering through use of silhouette coefficients. Unfortunately, Levensthein can be quite slow, but there are ways to speed it up through uses of thresholds and other methods.
Another way to deal with the curse of dimensionality would be to find 'contrasting sets,', conjunctions of attribute-value pairs that are more prominent in one group than in the rest. You can then use those contrasting sets as dimensions either in lieu of the original attributes or with a restricted number of attributes.