Abstract
Probabilistic Inductive Logic Programming (PILP) is gaining attention for its capability of modeling complex domains containing uncertain relationships among entities. Among PILP systems, cplint provides inference and learning algorithms competitive with the state of the art. Besides parameter learning, cplint provides one of the few structure learning algorithms for PLP, SLIPCOVER. Moreover, an online version was recently developed, cplint on SWISH, that allows users to experiment with the system using just a web browser. In this demo we illustrate cplint on SWISH concentrating on structure learning with SLIPCOVER. cplint on SWISH also includes many examples and a step-by-step tutorial.
Access provided by CONRICYT-eBooks. Download conference paper PDF
Similar content being viewed by others
Keywords
- Probabilistic Inductive Logic Programming (PILP)
- Slipcover
- Complex Domain Models
- Uncertain Relationship
- Structure Learning Algorithm
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.
1 Introduction
Probabilistic Inductive Logic Programming (PILP) [3] uses Probabilistic Logic Programming (PLP) for modeling in domain characterized by uncertain relationships among entities.
One of most successful approaches to PLP is based on the distribution semantics [8] where a probabilistic program defines a probability distribution over non probabilistic programs, called worlds. The probability of a query is simply the sum of the probability of the worlds where the query is true. Various languages follow the distribution semantics such as Probabilistic Logic Programs, Logic Programs with Annotated Disjunctions (LPADs), CP-logic and ProbLog.
Many systems for performing inference and learning with these languages have been proposed in the past 20 years. Among them, cplint provides an interesting mix of algorithms, including structure learning algorithms.
cplint on SWISH [6] is a web application for running cplint with just a web browser: the algorithms run on a server and the user can post queries and see the results in his browser. cplint on SWISH is available at http://cplint.lamping.unife.it and Fig. 1 shows its interface.
cplint on SWISH interface.
cplint on SWISH is based on SWISH, a web framework for Logic Programming using features and packages of SWI-Prolog and its Pengines library. SWISH allows the user to write a Logic Program in a browser window and ask a query over it. The query and the program are sent to a server using JavaScript. The server then builds a Pengine (Prolog Engine) that evaluates the query and returns answers for it to the user. Both the web server and the inference back-end are run entirely within SWI-Prolog.
cplint on SWISH uses the language of LPADs and includes two inference algorithms: PITA, that uses knowledge compilation, and MCINTYRE, that uses Monte Carlo sampling. For parameter learning EMBLEM [1] is available while SLIPCOVER [2] can be used for structure learning.
cplint on SWISH is similar to ProbLog2 [4] that has also an online versionFootnote 1. Problog2 offer inference and learning algorithms for ProbLog. In the online version, users are allowed to write programs and run algorithms on a server with a browser. ProbLog2 is written in Python, runs in an Python HTTP server and exploits the ACE editorFootnote 2 which is written in JavaScript.
The main difference between cplint on SWISH and ProbLog2 is that the first offers also structure learning. Moreover, cplint on SWISH uses a Prolog-only software stack. ProbLog2, instead, relies on several different technologies, including JavaScript, Python 3 and the DSHARP compiler. In particular, it writes intermediate files to disk in order to call external programs such as DSHARP while we work in main memory only.
With both cplint on SWISH and ProbLog2 users who want to experiment with PILP can do it without the need to install a system, a procedure which is often complex, error prone and limited mainly to the Linux platform. However, since it is impossible to predict the load of the server, the system is more targeted at development, while for production it is recommended to use the standalone version of cplint.
One of the main objectives of cplint on SWISH is to reach out to a wider audience and popularize PILP, similarly to what is done for the functional probabilistic language Church [5], which is equipped with the webchurch system for compiling Church programs into JavaScript. To try to achieve this goal, cplint on SWISH includes various learning examplesFootnote 3: Machines (shown below), Registration, Bongard, Shop, Hidden Markov Model, Mutagenesis, University. Moreover, a complete online tutorial is available [7] at http://ds.ing.unife.it/~gcota/plptutorial/.
2 EMBLEM and SLIPCOVER
EMBLEM and SLIPCOVER in cplint on SWISH take as input a program divided in five parts: (1) a preamble where all the parameters (such as the maximum number of iterations and the verbosity level) are set, (2) a background knowledge that contains information valid for all interpretations, (3) an initial LPAD if there is one, (4) a language bias for guiding the learning phase. In particular, for EMBLEM it contains the declarations of the input and output predicates. Basically, input predicates are those for which we do not want to learn parameters, while we want to learn parameters for output predicates. SLIPCOVER uses also bias declaration in the style of Progol and Aleph: atoms for modeh/2 specify the literals that can appear in the head of clauses while atoms for modeb/2 specify the atoms that can appear in the body of clauses. Moreover, SLIPCOVER requires the use of the determination/2 predicate, as in Aleph, for specifying which predicates can appear in the body of clauses.
Using the machine dataset of the ACE data mining systemFootnote 4 as a running exampleFootnote 5 we have:
Finally, (5) the last part contains example interpretations. Here, we can use two different representations shown below, models or keys, as in ACE. The first specifies an example interpretation as a list of Prolog facts surrounded by begin(model(<name>)) and end(model(<name>)). In the latter the facts can be directly listed using the first argument as the example name.
We can also define folds and which examples are included so that learning can be performed by just asking the query induce(<folds>,P). The learned program will be returned in variable P.
cplint on SWISH has also facilities for testing the learned models: predicate test/7 takes as input the learned program and the testing folds and returns the log-likelihood, the areas under the precision-recall and receiver operating characteristics curves together with the set of points forming the curves themselves. These set of points can also be drawn on the screen as the respective curves, as shown in Fig. 2. cplint on SWISH offers also a separate AUC calculatorFootnote 6 that takes as input the list of examples together with their sign and probability.
ROC curve for the mutagenesis dataset.
References
Bellodi, E., Riguzzi, F.: Expectation maximization over binary decision diagrams for probabilistic logic programs. Intell. Data Anal. 17(2), 343–363 (2013)
Bellodi, E., Riguzzi, F.: Structure learning of probabilistic logic programs by searching the clause space. Theor. Pract. Log. Program. 15(2), 169–212 (2015)
Raedt, L., Kersting, K.: Probabilistic inductive logic programming. In: Raedt, L., Frasconi, P., Kersting, K., Muggleton, S. (eds.) Probabilistic Inductive Logic Programming. LNCS, vol. 4911, pp. 1–27. Springer, Heidelberg (2008). doi:10.1007/978-3-540-78652-8_1
Fierens, D., den Broeck, G.V., Renkens, J., Shterionov, D.S., Gutmann, B., Thon, I., Janssens, G., De Raedt, L.: Inference and learning in probabilistic logic programs using weighted boolean formulas. Theor. Pract. Log. Program. 15(3), 358–401 (2015)
Goodman, N.D., Tenenbaum, J.B.: Probabilistic Models of Cognition. http://probmods.org
Riguzzi, F., Bellodi, E., Lamma, E., Zese, R., Cota, G.: Probabilistic logic programming on the web. Softw. Pract. Exp. 46(10), 1381–1396 (2016)
Riguzzi, F., Cota, G.: Probabilistic logic programming tutorial. Assoc. Log. Program. Newsl. 29(1) (2016). http://www.cs.nmsu.edu/ALP/2016/03/probabilistic-logic-programming-tutorial/
Sato, T.: A statistical learning method for logic programs with distribution semantics. In: Sterling, L. (ed.) ICLP-95, pp. 715–729. MIT Press, Cambridge (1995)
Acknowledgement
This work was supported by the “GNCS-INdAM”.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2017 Springer International Publishing AG
About this paper
Cite this paper
Riguzzi, F., Zese, R., Cota, G. (2017). Probabilistic Inductive Logic Programming on the Web. In: Ciancarini, P., et al. Knowledge Engineering and Knowledge Management. EKAW 2016. Lecture Notes in Computer Science(), vol 10180. Springer, Cham. https://doi.org/10.1007/978-3-319-58694-6_25
Download citation
DOI: https://doi.org/10.1007/978-3-319-58694-6_25
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-58693-9
Online ISBN: 978-3-319-58694-6
eBook Packages: Computer ScienceComputer Science (R0)