Automatic Annotation in Multirelational Information Networks

Many networks are completely encapsulated using a single node type and a single edge type. Often a more complicated model composed of multiple distinct node and edge types can be constructed to create a more informative network [2]. We call the former homogeneous networks and the latter heterogeneous. The ability to homogenize networks varies wildly, dependent on the network under analysis and the problem being solved. We present a class of networks multirelational information networks where the heterogeneous structure is necessary for performing node classi cation and detecting missing information in the network. A multirelational information network is a network G with nodes V that map to real world objects and concepts which we call entities, and with edges E which represent relations between these entities. We use nodes and edges interchangeably with entities and relations. For the multirelational networks we will consider, there exist a true vertex labeling function lv : V → 2\{}, where Σ is an alphabet of node types and an observed vertex labeling function l̂v : V → 2. Informally, entity types will exist in a hierarchy and we require that the labeling only map a vertex to a set of types T such that for any pair (ti,tj), i 6= j, ti is a descendant of tj in the hierarchy or vice versa. Similarly, there exists labeling functions l̂e and le for the edges, but we will not be exploring the existence of an edge hierarchy in this work. Our goal is to learn the true vertex labeling function from the observed multirelational information network and labels. Speci cally, given network G = (V,E), hierarchy T , and observed labeling functionl̂v, we will learn the true labeling function lv. Additionally, we predict missing data elds in nodes that have incomplete relations. Namely, the edgeset E that we observe is not the true and complete set of relations that exist in the world. For example, the Mustang entity and the Camaro entity have many features in common, but only the Camaro has the model years relation. We would like to automatically recommend relations that may exist for a node based on our estimation of the true set of relations E∗. More precisely, given G = (V,E), we would like to provide A : V− > E s.t. A takes a node v and outputs a subset of edges that it should be incident upon v. The heterogeneous structure of the network is vital to this task, as we are essentially determining the missing pieces of the heterogeneous structure. Being able to discover entity types and missing relations will provide a method for database maintainers to discover and correct missing information in their entities. We can directly present our algorithm to database providers to automatically discover incompleteness in their data. We will provide a speci c example of usefulness in 2.1.