Deep Learning is nowadays the standard tool for classification tasks and is used not only for differentiating cats and dogs but also for industry and private life applications (e.g. insurance document input management, autonomous driving, molecule folding, …)
For tasks beyond classification, more layers of information are required: Named Entities.
Named entities on insurance documents are usually IBANs, addresses, customer numbers, specific dates and amounts etc… Successful extraction of named entities enables more precise classification tasks and automatized document processing for instance. For instance distinguishing between company address, customers address and an address of the local company outlet.
In this presentation, we discuss some promising approaches we developed for and inside ERGO to extract those named entities.
Further, we elaborate on the occurring challenges (not only inside primary insurers) of generating a labeled data set, train scalable models and the corresponding model performance.
TALKS
3 Optionen:
Prio 1: Failures aus Mobilfunk-Daten – Fail Drops | Case mit einer extrem hohen Anzahl von Daten | Analyse von komprimierter Pipeline, Diversifikation, Clusterung & Bilddaten | Zielfrage: Wo gibt es Probleme im Netzwerk?
Sehr spannend in Bezug auf Data Science Methoden und Architektur
Prio 2: Predictive Maintenance im Festnetz der DT (Glasfaser oder Kupfer-Kabel, Unterschiedliche Schadens- und Problemfälle)
Prio 3: Einsatz von Knime (Data Citizens, Upscaling, Wo wäre ein Vorteil, bei Einsatz von Knime (Gerne auch BarCamp Session)
Vertrags- oder Bedingungswerke im Versicherungskontext sind teils mehr als 50 Seiten lang und enthalten hunderte komplexe Klauseln. Ich werde einige aktuelle Projekte rund um Knowledge Discovery in solchen Dokumenten beschreiben – von der Digitalisierung über Metadaten-Anreicherung bis zum KI-unterstützten semantischen Vergleich. Dabei ist nicht nur die Auswahl der richtigen Technologien wichtig, sondern auch der regelmäßige Austausch mit den Fachabteilungen unabdingbar.
We at REWE digital have successfully managed the transition from a monolithic legacy system to a distributed, cloud-native software architecture. However, this transformation has brought new challenges such as distributed data sources which hampered the development speed of our machine learning teams due to impeded availability of required data sources. To cope, we have begun to adopt the philosophy of the Data Mesh, whereby teams consider their data as a product and publish them in analytical databases (e.g., Google BigQuery) for consumption by other teams. Eventually, moving towards a Data Mesh architecture enabled us to design, explore, and develop machine learning systems in a faster way. Yet, the development of such machine learning systems introduces other inherent engineering challenges, particularly due to uncertainties concerning the data (e.g., changing data) and concerning the model (e.g., behavior of the model is not deterministic and may change over time). Thus, we have also adopted MLOps practices to deliver robust and reliable machine learning systems. In this presentation, we will explore the adoption of the Data Mesh using Google DataFlow and BigQuery, implementation of MLOPs practices with Google Vertex AI, and demonstrate a pertaining case study from our last-mile delivery domain.
Planning and organisation of last-mile delivery offers a lot of challenges. Traditional planning based purely on mathematical optimization often lacks important real-life aspects and thus does not satisfy relevant operational requirements. Experienced couriers have tacit knowledge about the delivery area and its customers, enabling them to choose more efficient routes than the originally planned ones. This in turn renders predictions of arrival times very imprecise because those predictions can only be based on planned routes. This courier’s tacit knowledge is almost impossible to collect and maintain, let alone to incorporate in optimization and prediction algorithms. Thus, we at Deutsche Post DHL developed a novel, more holistic approach. We implicitly learn this tacit knowledge from historical tours and combine this with optimization algorithms to plan routes that an experienced courier would choose. Based on these routes, the delivery time predictions are done using machine learning trained on a large amount of past delivery events. In this talk we will present details of our algorithm, which incorporates machine learning, statistics, and optimization in a novel way. Furthermore, we show how it impacted last-mile delivery planning at Deutsche Post after its rollout across Germany.
Vectorization algorithms like TF-IDF enable powerful similarity comparisons, searches, and recommendations on texts. However, often relations between documents are latent factors that cannot be considered in this approach. Using a practical example I want to showcase the use of Siamese Graph Convolutional Networks and invite you to take a peek at what it is that makes such networks worth implementing.
Interactive notebooks like Jupyter have become more and more popular in the recent past and build the core of many data scientist’s workplace. Being accessed via web browser they allow scientists to easily structure their work by combining code and documentation.
However, notebooks often lead to isolated, hardly reproducible and disposable analysis artefacts…
In this talk we will walk through the OSEMN ( “awesome”) data science process and see, how data scientist’s (sometimes hidden) needs can be easily addressed by building your own OSEMN data science platform based on state of the art containerized tools.
Otto.de is running its own system for product recommendations. We use Thompson Sampling as a means of reinforcement learning in order to quickly explore the relative performances of different algorithms and their variations.
I will walk you through our implementation of real-time reinforcement learning and share with you some of the insights we gathered as well as some of the pitfalls and stumbling blocks we tripped over.
As always nothing is as easy as it looks. In our experience reinforcement learning did not work out of the box, but had to be adjusted to the statistical and technical specifics of the feature at hand. In the end we achieved an excellent tool that enables us to test and learn in faster iterations than with just classical A/B-Tests.
For global companies like Continental, Advanced Analytics and AI promise numerous applications and significant opportunities.
But how can these be seized quickly by global organizations?
Continental’s Tire division achieved this by working towards several targets in parallel:
– setting up a use case portfolio
– building agile analytics teams and establishing agile development processes
– aquiring know how for data scientists and the broader organisation
– building the right infrastructure that uses containerization and automated machine learning;
Our talk will present the Continental approach and key findings with examples along their way to implement a Data Science organization.