Semantic Reasoning for Smog Disaster Analysis

Abstract

Smog disaster is a severe global problem. Although it has been investigated for decades in environmental sciences, the analysis of smog data recently becomes an open problem in elds like big data and arti cial intelligence. In this paper, we present our study of utilizing semantic reasoning techniques for accurate and explanatory smog disaster prediction. To this end, we enriched the smog data streams with background knowledge by ontology modeling, inferred underlying knowledge like semantic assertions and rules, built consistent prediction models by embedding the knowledge (i.e., assertions and rules) in machine learning algorithms, and nally provided explanations by rule-based reasoning.

Keywords

• Smog Disaster
• Semantic Reasoning
• Ontology
• OWL
• Rule

Introduction

Smog disaster is a kind of severe air pollution event that negatively in uences people's health and damages the environment[8]. In the past decades, it has attacked a large part of the population, especially in the fast developing economies like China and India[4]. To deal with smog disasters, they are widely studied in environment sciences with domain methods, e.g., chemical model and satellite remote sensing. In these studies, prediction of air pollutants is one of the most important problem because of its signi cance in real world applications.

With widely deployed physical sensors and big urban data, predictive analytics for smog disasters becomes an open reseach problem in the communities of data mining and machine learning[1, 2, 10, 11]. For example, the study U-Air[10] predicted the air pollution index for those urban areas where there are no air quality stations using correlation analysis, feature extraction and multi-view learning. Di erent from those in environment sciences, these studies model the prediction problem in perspectives of data science and arti cial intelligence. However, they apply the background knowledge with manual exploratory analysis and feature engineering, ignoring knowledge representation and reasoning. This disables their capability of automatically incorporating the underlying knowledge with the prediction model, which limits their generalization to other contexts. Meanwhile, the pure machine learning based prediction models usually lack of explanation to the results.

On the other hand, semantic reasoning has recently been applied for the predictive analytics of spatio-temporal data[3, 5{7]. They implemented some widely used prediction techniques e.g., association rule mining and auto-correlation analysis on the semantic enriched data, thus utilizing background knowledge and reasoning for semantic enhanced prediction. For example, in the study of semantic tra c analytics[5], the researchers ( 1 ) interpreted the tra c related time-series into ontology stream, ( 2 ) inferred assertions and axioms for each portion, also known as ontology stream snapshot, ( 3 ) calculated the auto-correlation across snapshots, ( 4 ) mined semantic rules from snapshots that are semantically similar to the testing snapshot, and nally ( 5 ) predictively inferred the tra c congestion status with explanations.

In this study, we aim at bridging the gap between semantic reasoning and machine learning in the context of predictive smog disaster analysis. To this end, we rst semantically enhanced the smog related time-series by modeling the domain knowledge with Web Ontology Language (OWL), and then inferred the underlying knowledge i.e., assertions and rules. We nally embedded these knowledge into basic machine learning algorithms to realize consistent sampling and automatic feature extraction. In brief, this study contributes to both application and methology: (i) it builds a more accurate and explanatory prediction model for smog disasters; (ii) it proposes a framework for incorporating semantic web techniques with machine learning algorithms, thus enhancing traditional prediction models with knowledge representation and reasoning. 2

Context

Multiple heterogenous time-series observed from both physical sensors and a Chinese microblogging website, also known as Sina Weibo are used for this study as shown in Table 1. All the records are tagged with geographical position i.e., latitude and longitude. The application aims at predicting a position's air pollution after a period of time e.g., 12 hours and 24 hours with all the current observations. We model the problem as a classi cation problem, where air pollution status is divided into 6 ranges (i.e., Good, Moderate, Unhealthy, Very Unhealthy, Hazardous and Emergent) according to a US standard based on AQI (Air Quality Index) metric and air pollution's health impact.

Datasets air quality meteorology weather forecast

Record # Stream # 78.53M 9 150.1M 11 101.2M 7

Coverage 945 stations in 190 cities

Datasets POI checkin tweet

Record # 23K 2M 23.36M

Coverage

Beijing &

Shanghai

knowledge with OWL2 ontology (i.e., TBox) using description logic fragment ALC and transform the time-series into streaming facts (i.e., ABox), ( 3 ) infer underlying assertions through entailments with streaming reasoner e.g., TrOWL streaming[9], ( 4 ) mine SWRL (Semantic Web Rule Language) rules across snapshots and calculate their con dence and support[5], ( 5 )( 6 )( 7 ) embed facts, assertions and rules as consistent vectors and feature vectors, with both of which ML models are trained by Stochastic Gradient Descent (SGD) algorithm, ( 8 ) apply the SWRL rules in reasoning with the matched rules being the explanations, and nally ( 9 ) ensemble the results.

( 2 )

Background   Modeling

Knowledge

Ontology   (OWL2  ALC) ( 2 )Transformation WSeobc Diaal t a (tPerxetp(r1mo)icneisnsg) HuTHmruamaffinac  On Cb MosneodrbivDialioDtyno I nInnd Idneedxxe  x ( 5 )Assertion Embedding ( 6 )Rule&Assertion Embedding PShDeynasstiacoarl    Trmaa(nt2si)foonr StrFeaacmtsing   ERnetaas(io3ln)mienngt AssFearc Dtso ns (4R)uAlsesoMciinaitnigon Consistent  Vectors   Feature  Vectors   Al(g7o)rSiGtDhm CMoLn Msisotednetl   R e s u l t s

Fac ts  

Asser  Dons

 

Rules  (SWRL)

( 8 )  Rule-­‐based  Reasoner   (Explanatory  Model)   ( 9 )  Results  

Ensemble Explanations

Consistent Vector. A consistent vector Vci = (Vci( 1 ); Vci( 2 ); :::; Vci(dc)) represents the true or false value of dc classi cation assertions in ith snapshot which is also known as a training example. The element Vci(k) is assigned to 1 if kth assertion is positive (e.g., GoodAir(a)) in that snapshot, and to 0 if kth assertion is negative (e.g., :Cloudy(m)). We transform the consistent vector into autocorrelation weight of the training example by counting the equal elements, and then incorporate the weight with the model using weighted SGD algorithm. The built consistent model has been proven to solve the concept shift problem in supervised learning and achieve higher accuracy than pure ML models. Feature Vector. A feature vector Vfi = (Vfi( 1 ); Vfi( 2 ); :::; Vfi(df )) represents the real value of df attribute facts or classi cation assertions in ith snapshot. An attribute fact (e.g., hasAQIV alue(a; 70)) produces a real value feature, while a classi cation assertion (e.g., Cloudy(m)) generates a discrete value feature. A technique called rule embedding is developed for automatic feature extraction. The rule with high con dence and support, also known as an prominent rule indicates strong predictive information and its pre xes are used to infer e ective features. For example, if the rule Emergent(at+12) Hazardous(at)^Cloudy(mt) is prominent, the attribute facts hasCloudV alue and hasAQIV alue are used for real value features, and a new concept combing Cloudy and Hazardous is constructed for a discrete value feature. 4 Acknowledgement This work is funded by NSFC 61473260, national key S&T Special projects 2015ZX03003012, and supported by the Fundamental Research Funds for the Central Universities.

References

1. Chen , J. , Chen , H. , Hu , D. , Pan , J.Z. , Zhou , Y. : Smog disaster forecasting using social web data and physical sensor data . In: Big Data (Big Data) , 2015 IEEE International Conference on . pp. 991 { 998 . IEEE ( 2015 )

2. Djuric , N. , Kansakar , L. , Vucetic , S. : Semi-supervised learning for integration of aerosol predictions from multiple satellite instruments . In: IJCAI 2013, Proceedings of the 23rd International Joint Conference on Arti cial Intelligence , Beijing, China, August 3-9 , 2013 ( 2013 )

3. Galarraga , L.A. , Te ioudi , C. , Hose , K. , Suchanek , F. : AMIE: Association rule mining under incomplete evidence in ontological knowledge bases . In: Proceedings of the 22nd international conference on World Wide Web . pp. 413 { 422 . International World Wide Web Conferences Steering Committee ( 2013 )

4. Konkel , L. : The view from Afar: Satellite-derived estimates of global PM2.5. Environmental health perspectives 123(2 ), A43 ( 2015 )

5. Lecue , F. , Pan , J.Z. : Predicting knowledge in an ontology stream . In: Proceedings of the Twenty-Third international joint conference on Arti cial Intelligence . pp. 2662 { 2669 . AAAI Press ( 2013 )

6. Lecue , F. , Pan , J.Z. : Consistent knowledge discovery from evolving ontologies . In: Twenty-Ninth AAAI Conference on Arti cial Intelligence . pp. 189 { 195 ( 2015 )

7. Lecue , F. , Tallevi-Diotallevi , S. , Hayes , J. , Tucker , R. , Bicer , V. , Sbodio , M. , Tommasi , P. : Smart tra c analytics in the semantic web with STAR-CITY: scenarios, system and lessons learned in Dublin City . Web Semantics: Science, Services and Agents on the World Wide Web 27 , 26 { 33 ( 2014 )

8. Pope III , C. A ., Dockery , D.W. : Health e ects of ne particulate air pollution: lines that connect . Journal of the air & waste management association 56(6) , 709 { 742 ( 2006 )

9. Ren , Y. , Pan , J.Z. : Optimising ontology stream reasoning with truth maintenance system . In: Proceedings of the 20th ACM international conference on Information and knowledge management . pp. 831 { 836 . ACM ( 2011 )

10. Zheng , Y. , Liu , F. , Hsieh , H.P. : U-Air: When urban air quality inference meets big data . In: Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . pp. 1436 { 1444 . ACM ( 2013 )

11. Zheng , Y. , Yi , X. , Li , M. , Li , R. , Shan , Z. , Chang , E. , Li , T. : Forecasting negrained air quality based on big data . In: Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . pp. 2267 { 2276 . ACM ( 2015 )