Using Neural Networks to aggregate Linked Data rules
Download as PPTX, PDF1 like1,191 views
The document discusses the application of artificial neural networks (ANNs) to aggregate linked data rules, focusing on improving the explanations of clustered researchers based on co-authorship. It details the methodology used for training and testing an ANN model on various datasets, achieving performance enhancements in predicting rule combinations. The results indicate that this ANN-based approach is more efficient compared to traditional methods in terms of speed, accuracy, and scalability.
Using Neural Networks to aggregate Linked Data rules
- 1. Using Neural Networks to aggregate Linked Data rules Ilaria Tiddi, Mathieu d’Aquin, Enrico Motta
- 2. Knowledge Discovery patterns • Useful, explicit information about some collected data. raw data • Problems clean data Quantity - too many Quality - not interesting Knowledge PATTERNS Prepro cessing Mining Interpreting
- 3. Dedalo: interpreting patterns with Linked Data • Given the patterns of a clustering process • Find explanations for one cluster with information from Linked Data KMi researchers grouped together according to their co-authorship
- 4. Dedalo: interpreting patterns with Linked Data • Linked Data can help the non experts in understanding the pattern “Researchers working on projects led by someone interested in Semantic Web”
- 5. Dedalo: the bottleneck • Explanations are atomic. “People working with Enrico Motta” F1=66% “People working interested in Semantic Web” F1=66% “People working interested in Ontologies” F1=66% • We want to aggregate them to improve the explanation of the cluster “People working with Enrico Motta OR interested in Semantic Web” F1=93% “People interested in Semantic Web AND Ontologies” F1=86%
- 6. Rule aggregation state of the art • Pre-production of patterns • Machine Learning techniques • Artificial Neural Networks (ANN) for features reduction • Patterns post-processing • Interestingness measures (IR, Statistics…) • Semantic knowledge (ontologies, taxonomies) Proposition: use ANNs for post-processing patterns (i.e. for aggregating Linked Data rules)
- 7. Using ANNs to combine rules • We want to know if two rules r1 and r2 are worth combining • There must be a relationship between features of two rules that can help in deciding their combination • We know their F-score, hence Precision and Recall …(r1) …(r2) …(r1) UNION(r1,r2) INTERSECTION(r1,r2)
- 8. Model training and testing • Model configuration • Model : Feedforward multilayer perceptron • Neurons structure: 9 – 12 – 2 • Inputs : P(r1), P(r2), R(r1), R(r2), F1(r1), F1(r2), and their absolute differences • Training and test set • 30,000 automatically labeled combinations (unions and intersections) for training • Boolean label if the F1 of the combination has increased • 30,000 combinations for testing • Error rate: 0.24 (MSE rate)
- 9. Predicting combinations • A process to combine rules • a large set of ranked rules • the nnet learned model • a prediction indicator p(r1,r2)= nnet(r1,r2)*max(f(r1),f(r2)) • Start from the top(H) rule • predict p(top(H),ri) for each rule in H • combine rules if p(top(H),ri) above a given threshold • add the new rule
- 10. Experiments - datasets • KMiA – authors clustered according to the papers written together • KMiP – papers clustered according to the abstracts words • Huds – students clustered according to the books they borrowed Data Rules RAM Time (sec) Initial top(H) Ending top(H) KMiA1 369 4G 60’’ 71.1% 86.3% KMiA2 511 4G 60’’ 60.6% 63.9% KMiP1 747 4G 75’’ 54.9% 63.9% KMiP2 1746 4G 160’’ 30.6% 84.1% Hud1 11,937 10G 2,500’’ 20.2% 66.9% Hud2 11,151 10G 3,000’’ 13.3% 67.3%
- 11. Experiments - strategies Compare the NNET process with other strategies Random baseline combine a random rule with the top(H) AllComb baseline combine everything in H with everything in H Top100 naïve combine the first 100 rules in H only First naïve always combine the top(H) with H Delta combine all rules above a threshold NNET combine any pair predicted NNET50 combine if prediction is higher than 50% of the highest score at the current iteration
- 12. Experiments - results 0.7 0.65 0.6 0.55 0.5 Random AllComb Top100 First Delta NNET NNET50 0.2 Huds1 example. 0.45 0.4 0.35 0.3 0.25 0 500 1000 1500 2000 2500
- 13. Experiments - results 1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0 10 20 30 40 50 60 0.88 0.86 0.84 0.82 0.8 0.78 0.76 0.74 0.72 0.7 0 10 20 30 40 50 60 KMiP2 example. Huds2 example. 0 10 20 30 40 50 60 70 KMiA2 example. 0.63 0.61 0.59 0.57 0.55 KMiP1 example. 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 500 1000 1500 2000 2500 3000 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0 30 60 90 120 150 KMiA1 example.
- 14. Experiments - performance • Comparing performances Speed Accuracy Scalability Random ++ - ++ AllComb + ++ -- Top100 + -- + First + -- -- Delta + -- -- NNET / NNET50 ++ ++ ++
- 15. Conclusions and future work • An approach to predict rule combination based on Artificial Neural Networks • Trained model on the information (Precision, Recall, F-score) about the rules • Save time and computational costs (vs. other) • Evaluating Dedalo on Google Trends: why is a trend popular according to Linked Data? http://linkedu.eu/dedalo/eval/
- 16. THANKS FOR YOUR ATTENTION ilaria.tiddi@open.ac.uk @IlaTiddi Questions?