Agenda 2020

November 16-17, 2020 - Virtual

When RTL startet to broadcast german content in 1984 it was received initially by just a handful of roughly 200.000 households. In the past years Mediengruppe RTL became the biggest broadcaster in Germany. However, with the advent of large scale VOD streaming services viewing behaviour started to change. In 2018 we launched our own VOD streaming service TVNOW. To support the growth of TVNOW we are leveraging state of the art statistical analyses, AI and machine learning techniques. We are developing models to gain insights about the effectiveness of our communication channels and formats in generating inflow. We are building in house recommender systems to increase the click through rate of our videos – just to name a few of our big data projects. In total we do see the future and these developments will contribute significantly to our transformation in becoming a major future streaming broadcaster and as such repeat the success of our predominantly linear past.

Speed Networking at a virtual conference? Yes, it is possible AND fun. All attendees will get instruction with their login details.

The handling of data represents a new challenge for many companies – this applies not only from a technical but also from a legal point of view. The use of data can generate added value and exploit potential. The question therefore arises as to whether and how this added value should be reflected in the balance sheet. The presentation provides an answer to this question and introduces the accounting principles and prerequisites for capitalizing data. Where the tax balance sheet is based on the commercial balance sheet, this question also directly affects the tax treatment of data. In this context, transfer pricing aspects must be taken into account, especially since the internal exchange and use of data is particularly relevant for internationally active group companies. Accordingly, the question is currently also being investigated by the OECD, which wants to have group-internal transactions remunerated at third price.

Building an AI solution, if the data is unlabeled and the labeling of the full data set is too expensive, is a more then complex task. In order to overcome this challenge, GfK uses a two-layer approach similar to active learning. In the first step we build a model to propose a relatively small subset of the data that should be annotated by the market experts that will work with the solution. Then, to further reduce the needed involvement we build a second model on the annotations to minimize their involvement for the future. The presentation will showcase how two-layer approach helped GfK to increase data quality while minimizing the needed human labelling effort. Furthermore, we will discuss the challenges and benefits of this approach. Finally, there will be a deep dive into the code, the architecture and continuous evolution pipeline for the model.

Financial time series forecasting is a challenging problem. Deep learning approaches, such as recurrent neural networks (RNNs), have proven powerful in modelling the volatility of financial stocks and other assets, as they are able to capture non-linearities in sequential data. Recent studies have shown that RNNs have surpassed well-known autoregressive forecasting models (Siami-Namini, 2018). Besides forecasting the next value of a stock, creating an optimal portfolio is equally important. Deep portfolio theory (Heaton et.al.,2018) uses autoencoders to model the non-linearity of the time series to accurately predict returns. Within this session, Julian will extend this approach by first performing a 10-day ahead forecast and then train an autoencoder model to construct an ideal portfolio that incorporates previous and future stock market information. You will be provided with a theoretical understanding of how RNNs and autoencoders work and how to apply them on multivariate timeseries forecasting and portfolio optimization problems. The new model  is then applied on financial stock data and compared to traditional portfolio optimization methods. Finally, the session will end with presenting an overview of key challenges and current research topics within that field.

The acceptance of machine learning algorithms in predictive analytics heavily depends on their interpretability. With the growing complexity of deep learning and machine learning models, the need for explanability techniques is continually increasing. Model-agnostic explanability techniques are a topic of on-going research in the field of explainable AI, since they allow to compare different modeling approaches within one concept. This deep dive presents SHAP, a model-agnostic game theoretic approach. SHAP unifies existing concepts like LIME and Shapley regression. State-of-the-art software implementation, highly-tuned algorithms for tree-based and deep learning models, as well as unified explanations of both global and local feature importance made SHAP the new gold standard in the field of XAI. The session will demonstrate the advantages and elegance of SHAP in the field of predictive analytics, and show why it belongs in every data scientist’s toolbox.

There have been multiple instances when a machine learning model was found to discriminate against a particular section of society, be it rejecting female candidates during hiring, systemically disapproving loans to working women, or having a high rejection rate for darker color candidates. Recently, it was found that facial recognition algorithms that are available as open-source have lower accuracy on female faces with darker skin color than vice versa. In another instance, research by CMU showed how a Google ad showed an ad for high-income jobs to men more often than women. Using credit risk data where we wanted to predict the probability of someone defaulting on a loan, we were able to shortlist features that were discriminatory in nature.

Obtaining cookie profiles and valid information about daily user behaviour on the internet is equally important for websites and advertisers. In Germany, this is the responsibility of agof (Arbeitsgemeinschaft Online Forschung) with Kantar Analytics Deutschland being a key service provider. The process links cookie-based technical measurements with survey and projection data from traditional market research. This deep dive will show how such a multi-method model works in principle, how REST-APIs, SQL databases and neural networks come into play and how the models are regularly updated. The presentation will additionally describe how innovations and improvements can be implemented even during ongoing operations.

It’s hard to find a field where machine learning didn’t make its impact already. But despite all the buzzwords and marketing, machine learning in finance is still an open case. The main reason is that machine learning practitioners treat financial data as it is regular routinely practice in computer vision or NLP, not taking into account stochastic underlying nature of the data and related decisions. This deep dive presentation shows on real examples the top-mistakes that practitioners do with financial time series, how these mistakes can be fixed and how drastically change the related results.

There are few frameworks on effective AI Project Management. Industry-standard frameworks for data analysis projects, like CRISP-DM, exist but none are effective for managing the development of AI products from deployment to production. The result is that many data science teams are focused on outputting one-off analytical projects, rather than building long-term, maintainable products that directly drive business processes and goals. In this talk we propose an effective project management framework for building production data science systems based on our experience constructing such systems at large tech companies, late-stage start-ups, and early-stage ventures. Specifically, we propose that a blend of CRISP-DM and the Extreme Programming agile development methodology can serve as a starting point for building organization-specific data science product development processes.

For most data science applications, the algorithms do not matter nearly as much as data integration and data preparation. But sometimes, we really need to bring out the big guns: the most sophisticated and complex algorithms to solve difficult problems. This talk will describe situations where complex algorithms matter and when one might want to choose Random Forests, XGBoost or Deep Learning networks to solve those difficult problems.

In a fast-growing company like HRS, knowing how to use Data Analytics, Data Science and Machine-Learning is not only about hiring the right data people and work on the right problem but also how to setup the function and spread best knowledge and practices related to data and associated technology across the organization. In this talk I will use concrete examples of problems faced in setting up a large data function and how we are solving it at HRS, in particular in times like during the COVID-19 crisis. The audience will be given the following key take aways: How to set up your data function for it to scale as fast as your company grows? How can you support your team leads and teams with data in this critical period of growth? What are best current practices?

Have you ever received a letter from a debt collection agency? If you have not, then you are among a lucky few people. Failing to pay a bill happens even to those who are financially well-off. For instance, a letter with an open bill can easily slip one’s attention when one is travelling frequently. This is only one reason among many why people don’t pay. Ignoring these can seriously jeopardizes the good relationship with valued customers or even lead to substantial problems. Creditreform has teamed up with Simply Rational, a spin-off from the Max Planck Institute, to design and implement a novel system for debt collection combing Machine Learning and Behavioral Science. The approach has proven highly effective in large-scale field-testing and promises to revolutionize debt collection by taking the individual needs of debtors seriously.

When every app or website has a search box, it is crucial to level up your search engine to rise past the competition and increase your bottom line. Introducing machine learning to your search engine is complicated but worth it – so let’s break it down. This talk details the transition from a simple weight-based search engine to a machine learning powered, data driven setup; and how to achieve the “Always be testing” state with rapid iteration cycles. It covers an end-to-end search engine architecture, from data logging from Microservices, processing with Apache Spark, training with LambdaMART, and deploying your models with ONNX on top of ElasticSearch or Solr.

From the start Predictive Analytics World has been the place to discuss and share our common problems. These are your people – they understand your situation. Often rated the best session of all, sharing your problems with like-minded professionals is your path to answers and a stronger professional network.

• Pros and Cons of Data Warehousing for Data Science with Christian Schäfer
• MLOps in Real Life with Timo Böhm
• Will Kill Regulations AI for FinTech & InsurTech? with Tom Alby
• Demand Forecasting in Times of Corona: Challenges & Solutions with Sven Crone
• Data Science: Easy to Start, Hard to Scale with Norbert Wirth

Artificial Intelligence and Machine Learning (AI/ML) reign high on the Gartner hype cycle, promising new business models, and capturing the imagination of executives well outside the digital industries of Facebook, Google and Uber. However, despite dozens of POC studies, in demand planning AI/ML is struggling to prove significant increases in accuracy. A limiting factor is that the underlying data is fundamentally different. After a brief introduction into AI/ML in forecasting we will explore the differences in data between image recognition and demand planning, where the dataset size, structure and labels are often sparse. Companies are not drowning in data, but rather sitting in a puddle! As this limits AI/ML algorithms, we present a case study at one of the worlds leading pharmaceutical manufacturers Janssen (a Johnson & Johnson company) where enhanced datasets with increased sample frequency and automatic feature generation promise break-through increases in accuracy by using AI/ML.

Fraud impacts banks that provide online payments service and challenges audit companies and financial players. Driven by growing B2C online trade, an increasing number of credit card transaction evaluations and credit decisions have to be made in real time. Manual processes won’t scale and are error prone. The challenge is to implement real-time transaction assessment and improve the accuracy of fraud detection. The overall objective is to reduce decision cycles, improve customer experience, automate the transaction and lending process, to recognize fraudulent transactions and transactions involving a credit line that won’t be covered later on. This case study will provide an example application of behavioral fraud analytics to credit card transactions data provided by YapiKredi. Based on roughly 30 million transactions this case study will present a technical deep dive comparing behavioral approaches in combination with machine learning against purely statistical/machine learning approaches.

Text summarization is a highly useful tool for extracting key information from text, which helps businesses speed up processes dramatically. With the use of bidirectional encoders, such as BERT, RoBERTa or BART, automatic production of human-like summaries has become easier to achieve than ever. For some of the inputs, however, the quadratic time complexity of bidirectional encoders results in a specific token limit, acting as a constraint for processing large text sequences. In her talk, Nataliia will show how to produce high quality summaries for large text inputs, guiding you through the data preparation process and suggesting some heuristics to deal with the token limit. She will compare the performance of different state-of-the-art bidirectional encoders on large text sequences on the example of consumer complaint data and show which architectures ensure producing the best summaries.

The time when it was okay for Data Science to just be cool is over. Nowadays, the expectation is to integrate models into production seamlessly. At Helvetia, the third biggest insurance company in Switzerland, data scientists can prototype on their personal sandbox and deploy cutting-edge data products effortlessly. They created a cloud-based working environment together that takes out the burden of deployment, allowing the data scientists to focus on what they can do best. This talk will explain how the Data Science Workbench was designed and discuss the key ideas and components: AWS AMIs to pre-configure EC2 instances equipped with key data science tools. Users manage their instances comfortably via Slack. Julia and Christian will also discuss our experiences and other insights gained during development.

Classic TV commercials are a building block for creating brand awareness. But some companies also run TV commercials to directly attract users to their website and close a deal. In order to estimate the impact of TV commercials on a website, the expected number of web sessions are compared to the number of web sessions right after the commercial was broadcast. But it’s not as simple as it sounds. TV commercials might be broadcast in parallel on multiple small TV stations, so signals are feeble and overlapping. In this talk, Christian shows how to model the problem as a non-parametric estimation of intensity measures of multiple Poisson point processes where only the sum of the processes is observed. You will see how to calibrate the estimation procedure in the absence of ground truth and run simulation studies to ensure that the tuned method provides unbiased results. The presented methodology has supported marketing decision makers to adjust media spend on a variety of large direct TV campaigns.

Governments worldwide are rushing to come up with ideas to regulate so-called Artificial Intelligence, automated and “autonomous” systems, and predictive analytics applications. From the EU’s AI White Paper to the newly founded Partnership on Artificial Intelligence (GPAI), comprising 15 states (including the EU and the US), the question is how to deal with the powerful promises and high risks of optimisation technologies and probabilistic decision-support systems. Companies react by hurriedly drafting “AI ethics guidelines”, hoping this will save them from regulatory oversight. In my keynote, I will delineate the state of play of both governments’ and businesses’ activities and sketch out ideas of how to tackle the regulatory and ethical challenges ahead.