Some machine learning/artificial intelligence/natural language processing algorithms implemented in PHP. Note that in almost all cases, PHP as it stands today is the wrong tool for most machine learning jobs. This library provides a pedagogical introduction to these tools more than it is a recommendation that it is used for day-to-day development.
Copyright (C) 2011-2015 Giuseppe Burtini joe@truephp.com and contributors as appropriate.
In general, you'll want to grab just the "required" features from this repository for your project -- a lot of the individual methods are standalone (or only dependent on the accessory directory). Browse the lib/ directory and decide which techniques you are interested in.
- DBScan (dbscan.php) - Density Based Clustering [1][2] - a clustering/unsupervised classification algorithm based on the idea of "density reachability." This algorithm is a win over the others because one does not need to specify the number of clusters a priori. The parameters are $e(psilon), the size of a neighborhood to visit (a noise threshold) and $minimumPoints, the minimum number of points to form a cluster.
- K Means (kmeans.php) - the standard clustering algorithm which breaks data in to k "most different" groups. The technique is simply to reposition the "centroid" to the average of all points until it doesn't move any longer.
- K Nearest Neighbors (knn.php) - similar to K Means, except "flipped on its head" - a clustering algorithm which builds the best clusters that are of size k (rather than building k clusters). (Wikipedia)
- Markov Chain (markovchain.php) - a n-order Markov Chain implementation - takes in a list of values to train and computes probabilities simply from observations.
- Anomaly Detection (anomaly_detection.php) - assume a normal distribution, train data (n-dimensional) and then test to see if a given record is an "outlier" (less likely than a given percent, given the distribution). Assumes semi-stationarity (training can happen online with testing if you wish).
- Naive Bayes (naivebayes.php)
- Regression (regression.php) including optimization implementations for gradient descent ("take a step in the right direction"), stochastic gradient descent, normal equations and a logistic regression implementation.
- Simulated Annealing (sann.php) - an implementation of simulated annealing, a probabilistic metaheuristic for finding global optima with no assumptions about smoothness or size of the search space. Written by Graeme Douglas (2014).
- Mann-Whitney (mann_whitney.php) - provides the test of Pettitt (1979) based on the Mann and Whitney "U" test (or rank-sum test, Wikipedia). The function provided accepts a list and a threshold (critical point) and returns a list of locations where changepoints were detected according to the test.
- Page-Hinkley (page_hinkley.php) - provides the test used in Mouss et al. (2004) and Hartland et al. (2007) among others based on Page's (1953) idea of a cumulative sum (Wikipedia). The test is parameterized with $alpha, the minimum amplitude of a change and $lambda, a parameter proportional to the false positive rate. An extension to Page-Hinkley as provided in Ikonomovska (2012) is also provided which automatically calibrates $alpha to be the standard deviation of the data.
- Epsilon Greedy Bandit (EpsilonBandit.php) - explores (purely randomly) e fraction of the time, exploits (maximizes estimated reward) the rest.
- UCB1 Bandit (UCB1Bandit.php) - implements the UCB1 (upper confidence bound) algorithm, as described by Auer et al. (2002). Provides an average and padding function implementation that can easily be extended to other UCB variants.
- [1] Domenica Arlia, Massimo Coppola. "Experiments in Parallel Clustering with DBSCAN". Euro-Par 2001: Parallel Processing: 7th International Euro-Par Conference Manchester, UK August 28–31, 2001, Proceedings. Springer Berlin.
- [2] Hans-Peter Kriegel, Peer Kröger, Jörg Sander, Arthur Zimek (2011). "Density-based Clustering". WIREs Data Mining and Knowledge Discovery 1 (3): 231–240. doi:10.1002/widm.30.
- [3] Auer, Peter, Nicolo Cesa-Bianchi, and Paul Fischer. "Finite-time analysis of the multiarmed bandit problem." Machine learning 47.2-3 (2002): 235-256.
- Update to PHP 5.x, change to use namespaces instead of messy function names.
- Lots of missing documentation - most public facing methods are currently undocumented; bad.
- Build TF-IDF class / simple vector search space class.
- Build MonteCarlo class (with callbacks)
- OOize all appropriate algorithms (use "train" and "test" when possible).
- Complete tests for algorithms that do not have them.
- Consider adding a pathfinding/graph search algorithm set.
- Ensemble and boosting learning methods like random forests / CART / BART
- Neural networks and HMMs.
- NLP work, specifically a class for using WordNet and the Stanford Core NLP library; eventually, NLP work should probably be forked as its own project.
For effective use, a lot of this library will have to be customized. This is a largely academic project. In many cases I've opted to write clearer code in favor of faster code, and in other cases, I've excluded useful features for "real world" applications (like training, saving the trained data to a file, and then running the actual "estimates" after in a separate location).
If you would like to deploy this in a real world application, I would be happy to discuss work on any machine learning problems you do have: contact me at joe@truephp.com.
Most of the code in this library is designed such that the rest of the library can just be thrown away if you would like to use it. In parametric/ and unsupervised/ each "type" of learning is implemented in a file of its own (though, regression stuff gets a whole directory!) so as to be useful without loading the rest of the library.
There are many things that can be improved and there are many known properties (and optimization techniques) that can be used to improve the performance of these algorithms that have not been implemented here. This is very much a "first run" at implementing a lot of these algorithms in PHP and should be looked at as a possible starting point for learning algorithms in PHP, not necessarily a deployable library.
In many cases, the right answer will be to implement the learning algorithms in a faster language and use PHP only to evaluate their probabilities / compute results from the existing estimates.
If you wish to cite this work your work, you can cite it as
Burtini, G. (2011). Learning Library for PHP. Canada. <_http://github.com/gburtini/Learning-Library-for-PHP_>
If you wish to have your work using the Learning Library for PHP in a machine learning, artificial intelligence, bandits or other context, please contact me with details and I will add it here. Actually, even if you don't wish to have your work listed here, I'd love to hear about how you're using the library or what you have learned.
- Buchmann et al. Personal Information Dashboard: Putting the Individual Back in Control. Knowledge and Privacy Analytics Engine. Digital Enlightenment Yearbook (2013) p. 139-164, Iso Press, September 2013. ISBN 978-1-61499-294-3 (print) | 978-1-61499-295-0 (online).
- Various private projects.
This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this program; if not, write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
If you need to work with the Learning Library in an environment that is not conducive to the GPL, please contact me at joe@truephp.com and we can discuss alternative licensing terms.