SRLBoost: ⚡ Fast implementations of boosted relational dependency networks and Markov logic networks.

Getting Started

SRLBoost can be built as a maven package. For example, on Windows:

git clone
cd .\SRLBoost\
mvn package

Then learning should feel familiar if you’re familiar with other distributions like BoostSRL. After switching out X.Y.Z with the latest version:

java -jar .\target\srlboost-X.Y.Z-jar-with-dependencies.jar -l -train .\data\Toy-Cancer\train\ -target cancer

Full notes are available with the repository:


  • I was one of the main people behind releasing “BoostSRL,” but wanted to go in a different direction with the software.
  • At one point there was discussion around developing a “BoostSRL-Lite” implementation. But this didn’t really go anywhere (and as you’ll see in the benchmark, it wasn’t especially lite).

SRLBoost aims to be a small and fast core—not to implement every possible feature.


Size Comparison

“BoostSRL-Lite” cut around 6,000 lines of Java out of “BoostSRL.”

“SRLBoost” cut close to 50,000 lines of code.

This graph was made at SRLBoost commit cb952a4, BoostSRL-Lite commit e198b76, and BoostSRLv1.1.1 at commit 4f0ad2b. Lines of code were measured with cloc-1.84, and Java files in each source directory were counted.

Time Comparison

  • “BoostSRL” and “BoostSRL-Lite” are nearly indistinguishable in terms of runtime
  • “SRLBoost” is at least twice as fast

The following diagram compares the learning time (in seconds) for the three implementations on three benchmark datasets. On larger datasets like imdb, SRLBoost took an average of 5 seconds while the other two implementations took close to 20 seconds:

Each of these datasets has 4-5 cross validation folds. This measured the amount of time that it took to learn a boosted relational dependency network on each fold. This procedure repeated 10 times, and the 40-50 runs for each method were averaged to estimate average learning time. The "box-and-whisker" plots show median runtime, interquartile range for how much the total learning time varied, and uses diamonds to show outliers. SRLBoost (blue) is always the fastest.

On large datasets with lots of relations (like cora), this difference is even more pronounced. SRLBoost is so much faster that it’s difficult to visualize the difference on a linear scale:

The tiny bar on the left shows that the average SRLBoost time for cora is around 17 seconds, compared to 270 seconds for BoostSRL and BoostSRL-Lite.

Are there any downsides?

Metrics are indistinguishable on the first three datasets. But on the cora benchmark, being 15x faster also led to differences in some key metrics. Specifically, AUC-ROC decreased by 0.04 and AUC-PR decreased by 0.01.

BoostSRL-v1.1.1 appeared to have significantly worse F1 compared to the other two implementations, but it’s unclear why.1

Implementation cora mean AUC ROC cora mean AUC PR cora mean CLL cora mean F1
SRLBoost 0.61 0.93 -0.27 0.96
BoostSRL-Lite 0.65 0.94 -0.29 0.96
BoostSRLv1.1.1 0.65 0.94 -0.29 0.78


I’m implementing this as the core for srlearn, so most of the user interfaces for using SRLBoost are documented there.

  1. My best guess is that this is a bug introduced when thresholding changed between v1.0 and v1.1 (See commit 5a91ba0). If this is this case, there’s might exist a threshold setting that makes these two the same.