Quantifying Challenges in the Application of Graph Representation Learning

Abstract

Graph Representation Learning (GRL) has experienced significant progress as a means to extract structural information in a meaningful way for subsequent learning tasks. Current approaches including shallow embeddings and Graph Neural Networks have mostly been tested with node classification and link prediction tasks. In this work, we provide an application oriented perspective to a set of popular embedding approaches and evaluate their representational power with respect to realworld graph properties. We implement an extensive empirical data-driven framework to challenge existing norms regarding the expressive power of embedding approaches in graphs with varying patterns along with a theoretical analysis of the limitations we discovered in this process. Our results suggest that "one-to-fit-all" GRL approaches are hard to define in real-world scenarios and as new methods are being introduced they should be explicit about their ability to capture graph properties and their applicability in datasets with non-trivial structural differences.

Mapping urban socioeconomic inequalities in developing countries through Facebook advertising data

“…communities). Other methods focus on measuring interpretability of node embeddings with respect to node labels [4] and node centralities [25]. In [2] global interpretations are given as a hierarchy of graph partitions, but they do not focus on interpreting single dimensions.…”

“…• GEMSEC [29], a variation of DEEPWALK which jointly learns node embeddings and node clusters. We train the model 4 with the same configuration as DEEPWALK, plus the number of clusters fixed equal to the number of dimensions.…”

“…However, one of their shortcomings is the innate lack of interpretability of the latent vector spaces they exist in. Specifically, each of the learned latent dimensions does not have a corresponding realizable interpretation in the input graphs [2]- [4]. This paper aims to fill this critical gap by proposing a method to retrofit "already learned" non-interpretable embeddings into a new and interpretable vector space without compromising the task performance.…”

“…The meaning of individual latent embedding dimensions is hard to define and determine [2]- [4]. We operate on a general, yet powerful notion of grounding the interpretation of each dimension to understandable sub-structures of the input graphs, e.g., communities, subgraphs, etc.…”

“…Specifically, the meaning of a latent dimension can still not be explained using these local methods. Secondly, in the presence of ground-truth node labels [4], [11] measure the interpretability of embedding dimensions in terms of their association strength with these labels without providing an explicit explanation like our approach. These approaches pre-suppose a certain interpretable mapping and are not flexible.…”