Graph Pattern Learner
(Work in progress...)
In this repository you find the code for a graph pattern learner. Given a list of source-target-pairs and a SPARQL endpoint, it will try to learn SPARQL patterns. Given a source, the learned patterns will try to lead you to the right target.
The algorithm was first developed on a list of human associations that had been mapped to DBpedia entities, as can be seen in data/gt_associations.csv:
As you can immediately see, associations don't only follow a single pattern. Our algorithm is designed to be able to deal with this. It will try to learn several patterns, which in combination model your input list of source-target-pairs. If your list of source-target-pairs is less complicated, the algorithm will happily terminate earlier.
You can find more information about the algorithm and learning patterns for human associations on https://w3id.org/associations . The page also includes publications, as well as the resulting patterns learned for human associations from a local DBpedia endpoint including wikilinks.
To run the graph pattern learner, we recommend:
- 8 cores (for parallel execution)
- more than 8 GB free RAM
- Linux 64 bit with Python 2.7
For now, the suggested installation method is via git clone (also allows easier contributions):
git clone https://github.com/RDFLib/graph-pattern-learner.git cd graph-pattern-learner
Afterwards, to setup the virtual environment and install all dependencies in it:
virtualenv venv && . venv/bin/activate && pip install --upgrade pip setuptools && pip install -r requirements.txt && deactivate
Running the learner
Before actually running the evolutionary algorithm, please consider that it will issue a lot of queries to the endpoint you're specifying. Please don't run this against public endpoints without asking the providers first. It is likely that you will disrupt their service or get blacklisted. I suggest running against an own local endpoint filled with the datasets you're interested in. If you really want to run this against public endpoints, at least don't run the multi-process version, but restrict yourself to one process.
Always feel free to reach out for help or feedback via the issue tracker or via associations at joernhees de. We might even run the learner for you ;)
To get a list of all available options run:
. venv/bin/activate && python run.py --help ; deactivate
Don't be scared by the length, most options use sane defaults, but it's nice to be able to change things once you become more familiar with your data and the learner.
The options you will definitely be interested are:
--associations_filename (defaults to ./data/gt_associations.csv) --sparql_endpoint (defaults to http://localhost:8890/sparql)
To run a full training cycle, you probably might want to execute this:
./run_create_bundle.sh --processes=8 --sparql_endpoint=... --visualise \ ./results/your_bundle_name \ --associations_filename=... # ... other-options ...
The algorithm will then by default randomly split your input list of
source-target-pairs into a training and a test set, train on the training set,
visualise the resulting learned patterns in
before evaluating predictions on first the training- and then the test-set.
To use a learned model for prediction, you can run:
. venv/bin/activate && \ PYTHONIOENCODING=utf-8 python \ -m scoop -n8 run.py --associations_filename=... --sparql_endpoint=... \ --RES_DIR=./results/your_bundle_name/results \ --predict=manual ; \ deactivate
- Jörn Hees
- Rouven Bauer (visualise code)