140 likes | 200 Views
Expert system. Data Mining for TRACES. Expert system - Plan. Purpose of the project Certificates and controls Models Effectiveness TRACES integration and planning. Expert system - Purpose.
E N D
Expert system Data Mining for TRACES
Expert system - Plan • Purpose of the project • Certificates and controls • Models • Effectiveness • TRACES integration and planning
Expert system - Purpose • Trade, regulated by a strict legislation defined by EC (animal welfare, transport safety, consumers’ health, disease control…). • TRACES expert system • Detect or prevent the consignments that do not follow the EC rules. • Be able to increase the control in a more risky period or decrease the control in a less risky period. • Regulate the workload balancing (number of controls per day) of BIP. • 3 kinds of controls • EC legislation • Personal suspicion of the BIP authority • Random! The aim of the project is to improve the random control. => Use of Data Mining.
Expert system – Data Mining • Data Mining, definition: • process of analyzing data and identifying existing patterns in information. • Diverse techniques exist: • Statistics • Database • pattern recognition • artificial intelligence etc. • For TRACES, we use a technology which • allows for fast build of predictive models, and • Provides for the integration with TRACES
Expert system – Certificates and Controls Certificates 3 kinds of certificates: • CVEDP (Common Veterinary Entry document for Animal Products), animals products export from a third country to the EC and EE area. • CVEDA (Common Veterinary Entry document for Animal), animals export from a third country to the EC and EE area. • INTRA (Community Trade of Animals and certain Animal Products), trade among Europe. All CVEDA are controlled! Predictive models => CVEDP and INTRA Controls 4 kinds of controls: • Documentary check, it is done systematically. • Identity check, it is done systematically. • Physical check, it is not systematic and depends on Commission Decision 94/360/EC, by random or suspicion. The predictive model will replace the random control. • Laboratory check, less frequent Use of predictive models. Expert system => Physical check and laboratory check
Expert system – Predictive models • The starting point is a single model for the European countries • 4 European models: cvedp physical, cvedp laboratory, intra physical, intra laboratory. • Consignments in Portugal and in Lithuania may not share the same patterns => Need a model per country. • 132 country models: 4 models for each of the 33 european countries. • In large countries, heterogeneous distribution of patterns, consignments in Marseille may not share the same patterns with those in Dunkerque or Paris. => Need a model per BIP. • 1428 BIP models: 4 models for each of the 357 BIP.
Expert system – Predictive models Which model to use? • 3 predictive models are called • European model, all TRACES certificates of the same type (Cvedp or Intra). • Country model, all TRACES certificates of the same type (Cvedp or Intra) and the same BIP country. • BIP model, all TRACES certificates of the same type (Cvedp or Intra) and the same BIP. • Predicted value • The predicted value ‘non satisfactory’ is sent as a response if there is at least one predicted value ‘non satisfactory’. • Otherwise, the value ‘satisfactory’ is sent.
Expert system – Effectiveness Predictive model on all EU CVEDP consignments,30% consignments checked => 85% of problematic consignments detected! Analysis of EU CVEDP consignments for 2007, 2008 and 2009
Expert system – Effectiveness Predictive model on France CVEDP consignments,30% consignments checked => 87% of problematic consignments detected! Analysis of France CVEDP consignments for 2007, 2008 and 2009
Expert system – Effectiveness Predictive model on Austria consignments,30% consignments checked=> 99% of problematic consignments detected! Analysis of Austria consignments for 2007, 2008 and 2009
Expert system – Effectiveness Analysis of TRACES consignments, 2007, 2008 and 2009 Effectiveness of the predictive models in case of 5% of problematic consignements. • A random control will have to check all the consignments to get the problematic 5%. • A perfect model will detect all the problematic consignments, so will check only 5% of consignments. For countries with very different heterogenous consignments as France or United Kingdom we will also need an analysis at the BIP level.
Expert system – TRACES Integration TRACES web UI mock-up Banner shown if above threshold Data mining result is always shown
Expert system – Next steps • Deployment:Q3 – Q4 2010 • Evaluation: • Collect and evaluate results: • Was the advise followed? • What was the result? • Adapt model as needed (requires critical mass of data) • Possible future adaptations: • Additional certificate models • Possibility to define specific model (cncode, species-class, species-family…)