The datasets provided by the Linked Data community currently form the world's largest, freely available, decentralised and interlinked knowledge bases. However, to be able to benefit from this knowledge in a specific use-case, one typically needs to understand the modelling of the knowledge and formulate appropriate SPARQL queries. In order to ease this process, we developed an evolutionary algorithm that learns such SPARQL queries (graph patterns) for pairwise relations between source and target entities. Given a training list of source-targetpairs, our algorithm learns a predictive model, which given a new source entity predicts target entities analogously to the training examples. In this demo paper we present a high level overview over our graph pattern learner and show its application to simulate human associations (e.g., “fish - water”). In the demo users can choose a semantic entity (e.g., dbr:Fish) as stimulus and let the learned model predict humanlike responses (e.g., dbr:Water).
In recent years, many large, machine accessible and interlinked RDF datasets
have emerged from the Semantic Web [
Despite the availability of all this knowledge, actually using it typically requires non-trivial up-front work: SPARQL queries need to be formulated to extract relevant knowledge for the given use-case.
Hence, we developed a graph pattern learning algorithm [
New Data ?source dbr:Fish Training Data d?bsro:Durocge d?btar:rCgaett dbr:Summer dbr:Winter .d.br:Paris .d.br:France ESnPdApRoQinLt
New Data ?source dbr:Fish
Trained Model Graph Patterns Fusion Model Graph Pattern
Learner Pattern Learner Fusion Training Trained Model Graph Patterns Fusion Model
Application Phase (Demo) in two main aspects: (i) it learns an ensemble model that can (and will) consist of many queries and (ii) it doesn’t try to learn queries that reproduce a given list of entities, but it learns queries that represent a relation R between entity source-target-pairs (s; t) 2 R. By learning queries for a relation between pairs of entities, the graph pattern learner can generate a predictive model, that given a new source entity s0 can predict targets ft0j(s0; t0) 2 Rg.
In this demo paper we show one such predictive trained model that has been generated by our graph pattern learner as detailed in the following sections. 2
Graph Pattern Learner System Overview
5 Datasets available at http://w3id.org/associations/#datasets.
6 For further details, see set-up instructions at: https://joernhees.de/blog/2015/
11/23/setting-up-a-linked-data-mirror
7 See [
(a) Stimulus auto-complete input-box (b) Fused Prediction Results (c) Filtered graph patterns highlighting those that generated the target dbr:
Fishing to an arbitrary source s0 (e.g. with a VALUES (?source) {(s0)} block). This results in one list of target candidates per pattern that need to be ranked to yield the set of predicted targets. The fusion training component uses several strategies to generate (late) fusion models for this purpose.
In order to fuse such resulting target-lists for a provided new source node, the graph pattern learner includes a fusion training component that generates late fusion machine learning models. The fusion models vary in complexity from basic to full-fledged machine learning models themselves. For example, we provide basic target-occurrence ranking over all queries (called “target occurrences”) potentially normalised so that each pattern only has a total vote of 1 (called “precisions”). As full machine learning models, we provide amongst others KNN, SVM, Logistic Regression and RankSVM models estimating relevance based on target candidate vectors wrt. the generating queries. In the demo the user can switch between these fusion models with a simple drop-down as shown in Section 3.
After training, the set of graph patterns and fusion model form a predictive trained model that is used in the application phase. Given a new source node (e.g. dbr:Fish) the trained model uses all learned graph patterns to issue SELECT ?target queries in which the ?source variable is bound to the new source node against the SPARQL endpoint, and fuses the individual target result lists into an overall ranked list of target predictions. 3
The main screen of our online demo8 starts with the stimulus auto-complete input-box on top (Figure 2a) asking the user to enter a stimulus. The auto8 https://w3id.org/associations/gp_learner/demo/predict.html complete is realised via the Wikipedia OpenSearch API9, allowing a fuzzy search for matching Wikipedia Articles, including the resolving of redirects.
After selecting one of the Wikipedia articles from the auto-complete box, the URI is transformed to the corresponding DBpedia resource and the prediction started. The fused prediction results are then presented in the “Fused Prediction” tab (Figure 2b), in which the user can provide feedback about the generated targets (logged and used for future improvements). The user can also click the explain button to gain insight on why a target was predicted. All graph patterns that played a role in predicting this target will be highlighted and expanded in the “Graph Patterns” tab (Figure 2c). 4
In this demo paper we presented a high level overview over the graph pattern learner and show its application to simulate human associations. The algorithm, used datasets and interactive visualisation of the results are available online.
https://w3id.org/associations/gp_learner/demo/predict.html. This work was supported by the University of Kaiserslautern CS PhD scholarship program and the BMBF project MOM (Grant 01IW15002). 9 https://www.mediawiki.org/wiki/API:Opensearch