Exploiting symmetries for scaling loopy belief propagation and relational training

Babak Ahmadi, Kristian Kersting, Martin Mladenov, Sriraam NatarajanMachine Learning 2013 · OpenReview

Abstract

Judging by the increasing impact of machine learning on large-scale data analysis in the last decade, one can anticipate a substantial growth in diversity of the machine learning applications for “big data” over the next decade. This exciting new opportunity, however, also raises many challenges. One of them is scaling inference within and training of graphical models. Typical ways to address this scaling issue are inference by approximate message passing, stochastic gradients, and MapReduce, among others. Often, we encounter inference and training problems with symmetries and redundancies in the graph structure. A prominent example are relational models that capture complexity. Exploiting these symmetries, however, has not been considered for scaling yet. In this paper, we show that inference and training can indeed benefit from exploiting symmetries. Specifically, we show that (loopy) belief propagation (BP) can be lifted. That is, a model is compressed by grouping nodes together that send and receive identical messages so that a modified BP running on the lifted graph yields the same marginals as BP on the original one, but often in a fraction of time. By establishing a link between lifting and radix sort, we show that lifting is MapReduce-able. Still, in many if not most situations training relational models will not benefit from this (scalable) lifting: symmetries within models easily break since variables become correlated by virtue of depending asymmetrically on evidence. An appealing idea for such situations is to train and recombine local models. This breaks long-range dependencies and allows to exploit lifting within and across the local training tasks. Moreover, it naturally paves the way for the first scalable lifted training approaches based on stochastic gradients, both in an online and a MapReduced fashion. On several datasets, the online training, for instance, converges to the same quality solution over an order of magnitude faster, simply because it starts optimizing long before having seen the entire mega-example even once.

Introduction

Graphical models, particularly relational ones like Markov Logic Networks (MLNs), are powerful because they can represent complex dependencies. However, they suffer from a "scaling wall" when applied to large-scale data. Standard inference techniques like Loopy Belief Propagation (BP) treat every variable as unique, even when many variables are structurally identical. Furthermore, when training on a single "mega-example," the introduction of evidence usually breaks these symmetries, causing traditional "lifted" (symmetry-aware) methods to fail. This paper bridges the gap between Statistical Learning, which scales via parallelization and stochastic gradients, and Statistical Relational Learning, which scales via symmetry.

Table 1: Categorization of existing work and the paper's contributions across different scaling paradigms.

Methods

The authors introduce three major technical contributions to exploit symmetries for scaling:

  1. Lifted Loopy BP (LBP): Utilizing a "color-passing" algorithm, nodes are "painted" based on initial evidence and iteratively exchange colors with neighbors. Nodes with identical color signatures are collapsed into "clusternodes," significantly reducing the graph size.
  2. MapReduced Lifting: The color-passing process is shown to be equivalent to a sorting problem. By linking lifting to Radix Sort, the authors demonstrate that graph compression can be distributed across a cluster using the MapReduce framework.
  3. Relational Tree Shattering: To address symmetry-breaking during training, the giant graph is "shattered" into smaller, local tree-structured pieces. This local view restores symmetries that were lost globally, allowing for Lifted Online Training via Stochastic Meta-Descent (SMD).

Experimental Results

The approach was evaluated on synthetic benchmarks and real-world relational datasets like CORA (entity resolution) and Wumpus World (imitation learning).

  • Inference Scaling: On synthetic grids, the MapReduce version (MR-CP) scaled linearly, outperforming single-core lifting by orders of magnitude as the number of variables reached 250,000.
  • Training Efficiency: In the Wumpus domain, the proposed Lifted Online Training converged in seconds, while standard Scaled Conjugate Gradient (SCG) failed to converge within 72 hours.
  • Compression: On the CORA dataset, piecewise lifting achieved significant network compression even when standard lifting failed to find any global symmetries.

Conclusion

This work demonstrates that lifting is not just a single-core optimization but is inherently parallelizable and compatible with modern distributed frameworks like MapReduce. By "shattering" graphs to recover local symmetries, the authors enable efficient online training for massive relational models. While the benefits depend on the presence of structural symmetries, the results show that exploiting these redundancies is a key path forward for scaling complex graphical models to "big data" applications.

Summary updated

Abstract

Judging by the increasing impact of machine learning on large-scale data analysis in the last decade, one can anticipate a substantial growth in diversity of the machine learning applications for “big data” over the next decade. This exciting new opportunity, however, also raises many challenges. One of them is scaling inference within and training of graphical models. Typical ways to address this scaling issue are inference by approximate message passing, stochastic gradients, and MapReduce, among others. Often, we encounter inference and training problems with symmetries and redundancies in the graph structure. A prominent example are relational models that capture complexity. Exploiting these symmetries, however, has not been considered for scaling yet. In this paper, we show that inference and training can indeed benefit from exploiting symmetries. Specifically, we show that (loopy) belief propagation (BP) can be lifted. That is, a model is compressed by grouping nodes together that send and receive identical messages so that a modified BP running on the lifted graph yields the same marginals as BP on the original one, but often in a fraction of time. By establishing a link between lifting and radix sort, we show that lifting is MapReduce-able. Still, in many if not most situations training relational models will not benefit from this (scalable) lifting: symmetries within models easily break since variables become correlated by virtue of depending asymmetrically on evidence. An appealing idea for such situations is to train and recombine local models. This breaks long-range dependencies and allows to exploit lifting within and across the local training tasks. Moreover, it naturally paves the way for the first scalable lifted training approaches based on stochastic gradients, both in an online and a MapReduced fashion. On several datasets, the online training, for instance, converges to the same quality solution over an order of magnitude faster, simply because it starts optimizing long before having seen the entire mega-example even once.
BibTeX
@article{ahmadi2013exploiting,
  author = {Babak Ahmadi and Kristian Kersting and Martin Mladenov and Sriraam Natarajan},
  title = {{Exploiting symmetries for scaling loopy belief propagation and relational training}},
  journal = {Machine Learning},
  year = {2013},
  url = {https://openreview.net/forum?id=_uHV7-IruA},
}
Figure