Linked data: P redicting missing properties Klemen Simonic, Jan Rupnik, Primoz Skraba {klemen.simonic, jan.rupnik,

Overview 1.Linked Data (Motivation for the work) 2.Problem Definition 3.Approaches 4.Results

An example

Linked Data -connect related data that was not previously linked -practice for exposing, sharing, and connecting pieces of data and information How: -URI (Uniform Resource Identifier) -RDF (Resource Description Framework) (description of how to model/present the data)

Linked Data, tiny example

ResourcePredicate / PropertyResource / Literal

Linked Data, one dataset -Nodes are resources -Edges are relations -Edge Labels are properties

Linked Data cloud diagram

DBpedia DBpedia extracted the information from the infoboxes from the Wikipedia websites Resource Properties Literal en.wikipedia.org/wiki/University_of_LjubljanaLocationhttp://en.wikipedia.org/wiki/Ljubljana en.wikipedia.org/wiki/University_of_LjubljanaEstablished“1919”

DBpedia DBraw contains all the properties from all the infoboxes within the English Wikipedia articles DBmapped the properties are unified (mapped onto a DBpedia ontology). Semantic of properties: PlaceOfBirth = BirthPlace The data is much cleaner and is better structured than the raw properties dataset.

Freebase An entity graph of people, places and things, built by people. -Colloborative knowledge base -Property schemas -Google Knowledge graph

Scale of Datasets #nodes#edges#objects#propertiesavgDeg DBmapped5M17M2M DBraw11M47M3M Freebase 141M607M 23M DBpedia 3.7 version (additional properties and resources may be added in the meanwhile) Largest and most structured dataset (Large number of edges and objects, and relatively small number of properties) Mesy and noisy dataset (Large number of different properties because they are not unified )

Missing properties Problem: What are the missing properties for Fiat? For a given resource, we want a rank of missing properties by likelihood.

Approach -Similar objects -Measure of similarity -Neighborhood -Ranking function

Approach Ranking = weighted average of the k nearest-neighbor objects’ property frequency vectors. General framework (Kernel smoother): We can replace d with normalized kernel function. (More math on this topic is in the paper.) The function g(o) depends on the choice of measure of closeness d(o,o i ).

Evaluation protocol The evaluation procedure: 1.For a given object, we delete one or more of its properties, denoting (o, {p 1, …, p k } ) 2.Run the recommendation algorithm for the object 3.Compute several evaluation metrics

Evaluation metrics -Inverse rank (IRank) = -Top 5 = -Top 10 =

Measure of Closeness -Local Measures: local graph properties -Baselines: -Random Objects -Objects with Common Properties -Property Co-occurrence -Global Measures: global graph properties -Exogenous Measures: external information (text)

Local Graph Measures We focus on a local description, based on the property distributions: -PropertyCount -DirPropertyCount -NeighbDirProperyCount

Random objects Choose uniformly at random some number of objects in the network

Objects with common properties Take the objects which share a minimum number of properties with the query object The number of shared properties is taken as the weight for the object

Property Co-occurence Approximate resource similarities through property co-occurrence patterns Only pairwise co-occurrences are considered for the purposes of scalability and feasibility of estimation

Our method Each object is described by DirPropertyCount vector The similarity is determined by the computing the dot product between DirPropertyCount vectors

Comparison

Other Measure of Closeness -Local Measures: local graph properties -Baselines: -Random Objects -Objects with Common Properties -Property Co-occurrence -Global Measures: global graph properties -Exogenous Measures: external (no graph) information

Global Graph Measures We use two global measures of closeness based on graph geodesics and graph diffusion: (We treat the graph as a simple undirected graph. We also remove all the literals and constants from the set of nodes to remove unintuitive paths.) -Shortest path length -The length of a shortest path between two objects -We calculate the distances corresponding to the k nearest objects -Exponential diffusion kernel -Based on computing the matrix exponential of the graph adjacency matrix A -Parameter α controls how local/global the similarities are -Takes into account both the total number of paths between nodes as well as their respective lengths -Robust measure

Exogenous Measures -Independent of the graph structure -Rely on additional external information about the objects -Helpful for nodes with little connections in the graph Textual information: -For some of the objects, we have extended abstracts describing the objects -TF-IDF weighting + cosine similarity

Results - IRank

Results - Top10

In vs. Out properties

Deleting several properties Method: DirPropertyCount vector Dataset: DBraw We remove a fixed fraction of in and out properties

Degradation – nodes / edges The negative effect of deleting a fraction of edges or nodes from the network

Degradation – properties The effect of deleting K most frequent properties from the network

Conclusion -Method for predicting missing properties -Use kernel smoother -Measure similarity in a number of different ways: -Local properties -Global graph structure -External data (text) -Extensive experimentation -Investigate more on combining measures -More details about the research is in the paper: -Linked data: Predicting missing properties [machine learning] -Predicting Instance Properties in Linked Data [semantics of data]

Take home message -Big redundancy / regularity in the data -Local measures perform well -Scale changes the structure -> we need different method

What’s Your Message? Questions ?