This year I'm one of the PC Chairs for SIAM Data Mining (along with
Jieping Ye), and so I've been spending time in decidedly-not-sunny Vancouver. Being a PC Chair, or even being on a PC, is an exercise in constant
deja vu: I hear a talk and think "Where have I heard this before" before realizing that I've probably reviewed the paper, or looked at its reviews or meta-reviews.
Being the PC chair means though that I can float around the conference freely without feeling the pressure to attend talks, network or otherwise be social: I've earned my keep :). Following standard
conference networking maxims though, I made an effort to meet at least
one coffee shop that I've met before and introduce myself to at least
one new coffee shop, and
another.
But I am attending talks ! And I enjoyed listening to some work on
tensor spectral clustering by Benson, Gleich and Leskovec. It got me thinking about the larger issue of modeling higher-order interactions and what appear to be many different ways of modeling the problem.
The problem.
Imagine that you have a number of interacting entities. These could be points in a space, or vertices in a graph, or even dimensions of a point (aka variables). The easiest way to model a collection of such entities is to assume they're independent of each other. For example, I might draw i.i.d samples from a distribution. Or I might be looking at a collection of independent features describing an object, and so on.
Independence assumptions are powerful. They allow us to "factorize" structure into combinations of individual elements, which either leads to simple summations directly, or eventually after a log transformation of a product. This means we can deal with entities independently, and the "inference complexity blowup" is linear in the number of entities. A good example of this is the Naive Bayes approach to learning, where assuming all entities are independent leads to a likelihood cost function that's just a sum of terms, one for each entity.
I'm necessarily being rather loose with my statements here. Making them precise is possible in specific contexts, but it gets unwieldy.
But independence is too restrictive an assumption. It limits modeling power, and therefore will be unable to capture interactions that might make understanding structure a lot easier. For one thing, you'd never find correlations if you assumed that all features are independent.
Graphs.
The easiest form of interaction is a pairwise interaction. Modeling pairwise interactions gets us to a graph, and who doesn't love a graph ! More importantly for what follows,
a graph and its associated structures is a rich representation of a system of pairwise interactions
in that we have a colorful vocabulary and an arsenal of algorithms for talking about pairwise interactions and structures built on them.
Of course we've paid a price - in complexity. Instead of the linear cost incurred by independent entities, we now have quadratically many potential pairwise interactions to model. But (and here's the key), we can interpret a sparse graph as capturing weak interactions, and it's still a rich model to model different phenomena.
Higher-order interactions.
But what happens if we want to model interactions that aren't just pairwise ? What is the correct higher-order structure to model such interactions as effectively as graphs ? It turns out that there are many different ways to do this, and they all can be reduced to a sentence (
pace Saunders MacLane) of the form
A graph is just a very special kind of X
for different values of X.
1. The graphical model view.
A graph is just a special kind of clique intersection structure, if you only have 2-cliques.
One way to manage a collection of higher order interactions is to factorize them in a more general way. This is the basis for the clique-tree idea in graphical models, where the interaction structure is viewed as a set of complete joint interactions (aka cliques) all connected together in a tree structure for easy inference. Another name for this would be the 'bounded treewidth' model, but it misses the fact that we are allowing higher-order interactions, but in a controlled way.
The advantage of this representation is that in a true parametrized complexity way, it isolates where the true complexity is coming from (the size of each clique) from the overall complexity (the size of the graph).
A spectral perspective.
When graphs arise from natural sources of data (social networks and the like), we have to deal with noise and spurious signals. And the simple language of connectivity and paths is no longer robust enough. For example a graph might be connected, but only because there's one edge connecting two huge components. If this edge was spurious, we've just made a huge mistake on modeling this graph structure.
Spectral methods are currently our best way of dealing with noisy interactions. By focusing not on the topological structure of connectivity but on the amount of connectivity measured via cuts, spectral analysis of graphs has become perhaps the best way of finding structures in large graphs.
The spectral lens sees a graph through random walks on the edges. This is great for modeling a collection of pairwise interactions between entities, but not for modeling interactions among sets of entities. We have to be careful here. Spectral methods are actually quite good at finding community structure in graphs (i.e a partition into sets of vertices). What they can't do is find higher order partitionings in graphs (i.e sets of triangles or sets of special 4-vertex structures). And that's where the next three higher-order methods enter the picture
2. The algebraic topology view.
A graph is just the 1-skeleton of a simplicial complex.
If we're looking to model higher-order interactions, we need a language for describing a collection of well-defined higher order structures. That's what a simplicial complex is. I'll skip the formal definition, but the basic idea is that if you have a simplex (an interacting group of entities), then all subsets must be simplices as well. But you declare the simplices first, which means that the simplex $\{a,b,c\}$ is different from the three simplices $\{a,b\}, \{b, c\}, \{c, a\}$, even though the first must contain the second.
A simplicial complex is a topological object. It generalizes a graph because a graph is what you get if you limit yourself to simplices of size at most 2. Because it's a discrete topological object, you can now play with it using all the tools of topology, and in particular very powerful tools like homology and homotopy that reveal all kinds of hidden structure not accessible via a graph metaphor.
While simplicial complexes allow you to express higher order interactions (tunnels! holes!), they don't remove the problem of noise: one edge/simplex can still change the structure in nontrivial ways. There are two approaches that researchers have taken to this problem: one spectral, and one not.
The non-spectral approach is by far the most well-established one. It is based on the idea of persistence: a way to determine from the homology groups of the simplicial complex what structures are interesting and which ones are not. Persistence is the dominant weapon in the arsenal of topological data analysis, and I'll say no more about it here, so as to keep focus on spectral methods.
The spectral approach is less well developed, but is quite interesting. The idea is to generalize the notion of expansion from a graph to higher-order simplices, as well as generalizing the Laplacian operator to higher order simplices (or homology groups). Then a random walk on the simplicial complex can be linked to the Laplacian operator, and the eigenstructure of the operator can be linked to the existence (or nonexistence) of certain homology groups. Two places to start with this topic are one on
capturing generalizations of edge expansion, and another on
building random walks on simplicial complexes and connecting them to a combinatorial Laplacian.
3. The differential geometry view.
A graph is just a discrete approximation of (scalar functions on) a manifold.
The graph Laplacian is a discrete approximation of the Laplacian second order differential operator, and more generally the Laplace-Beltrami operator on a manifold. Indeed, one way to build intuition for what the graph Laplacian means is that it's capturing heat diffusion on an implicit manifold that the graph is merely approximating.
The Laplace-Beltrami operator is a "zeroth-order" operator in that it applies to the zero-dimensional entities on a manifold: namely, scalar fields over the points of the manifold. Suppose you were to build a vector field instead over the manifold, and wished to reason about it. Then the generalization of the L-B operator that you'd need is called the
Laplace-de Rham operator which formally acts like a Laplacian on the higher order differential forms defined over the manifold (formally, on sections of the tangent bundle). Writing down the L-R operator is a little tricky: it involves a combination of the exterior derivative and its dual (via the
Hodge * operator). But one useful observation is that the L-R operator on graphs amounts to a Laplacian on the set of
edges, rather than vertices.
This means that you can now treat edges as first-class objects for grouping, rather than vertices. And
this is useful for higher-order clustering. Whether this can be generalized even further remains to be seen.
4. The linear algebraic view.
A graph (adjacency matrix) is just a special case of a tensor structure on the entities.
This is perhaps the most well-known of the different higher-order approaches to modeling interactions, and is making the most waves right now. The idea is very simple. If we think of a graph in terms of its adjacency matrix, then each entry encodes a relation between
two basic entities. If we wished to encode relations between (say) three entities, then we need a "3D matrix", or more precisely, a
tensor.
Of course a tensor is more than just a way to assemble a collection of triples into a box, just like a matrix is much more than just a grid of numbers. The most basic question in tensor modeling is factorization: just like we can use the SVD to write down a matrix as a linear combination of outer products of basis vectors, can we write a tensor as 3-way wedge product of basic vectors ? If so, then we've been able to identify the key 3-way factors controlling an interaction.
But can we cluster using tensors ? Or more generally, is there a spectral theory of tensors, and maybe even the analog of Cheeger's inequality ? It turns out that it's possible to define eigenvalues and eigenvectors of tensors, and at least some of the standard spectral theory can be made to carry over - for more on this see
these slides by Lek-Heng Lim. But we're still looking for a theory can truly work on hypergraphs or tensors in general. In the meantime, tensor-based approaches to clustering boil down to factorization and PCA-like methods, of which there are many.
5. Coda
Are these approaches in any sense equivalent ? It's hard to say in general, although for a particular problem there might be connections. Certainly, the de Rham cohomology yields a nice filtration over homology groups that could be wrestled into a persistence framework (research question !!!) thus allowing us to extract robust higher-order structures from both geometric and topological considerations. The tensor approach is closely linked to probabilistic models, not surprisingly. But whether the spectral perspective (and all its attendant intuition and value) will extend nicely across these different frameworks remains to be seen.
