Agenda

London | 16-17 October, 2019

The core Bayesian idea, when learning from data, is to inject information — however slight — from outside the data.  In real-world applications, meta-information is clearly needed — such as domain knowledge about the problem being addressed, what to optimize, what variables mean, their valid ranges, etc.  But even when estimating basic features (such as rates of rare events), even vague prior information can be very valuable. This key idea has been re-discovered in many fields, from the James-Stein estimator in mathematics and Ridge or Lasso Regression in machine learning, to Shrinkage in bio-statistics and “Optimal Brain Surgery” in neural networks.  It’s so effective — as I’ll illustrate for a simple technique useful for wide data, such as in text mining — that the Bayesian tribe has grown from being the oppressed minority to where we just may all be Bayesians now.

Predictive analytics has usually been the preserve of a small number of highly trained and qualified individuals. How do we harvest that knowledge and capability and package it in a form that all data workers can consume and deploy for their own needs and requirements? This session examines the issues involved and suggest one possible approach using Pyramid.

Everyone will tell you, Artificial Intelligence and Machine Learning are the next Big Thing. Data scientists job satisfaction and salaries are both at an all time high, and companies routinely quote data science as their biggest strategic priority and area of investment. So, where is my infallible and intelligent personal assistant yet? Successful applications that make our everyday life better don’t seem to quite match this hype and effort, in both abundance and quality.

Certainly, AI is hard: more advanced algorithms, better modelling of the data, better engineering approaches take time, effort, knowledge. But the reality of industrial AI is that success mostly relies on functions that have nothing to do with sophisticated algorithms, and all to do with product operationalisation. Operationalisation ranges from defining what problem to solve (market understanding, customer workflows, product requirements) to verifying that what was built actually solves it (quality assurance, product analytics, customer success) to ensuring that the result is a viable product (business intelligence, marketing and operations).

The requirements of success are classic but, in the case of AI, more complex to implement and requiring a higher level of data literacy and operational maturity from the implementing organizations. Let’s explore together some of the specifics of data operationalization, through the lens of how various organizations have driven the development of their Machine Learning platforms, and attempt to extract some general principles (lean, shift left, culture shift) and practices (devops and dataops, measure everything) on how to successfully operationalize data & ML.

Companies are spending more and more on AI. Is it seeing the ROI we expect? In this talk Michelle will discuss some of the major blockers in productionizing advanced techniques in AI and data science. Using case studies from both a legacy company and a startup, together you will explore how to overcome some of the technical and cultural challenges of implementing scalable AI. 

Today there are many generic as well as specialist tools available to aid in predictive analytic tasks.  But customers want to leverage their own business knowledge in predictive analytics and create their own unique IP.

We support a broad spectrum of analytics but have always focused on empowering the business to adopt a more self-sufficient model that does not rely heavily on specialists or IT.  This lets businesses quickly address these particular challenges and gives users the freedom  to adapt and incorporate relevant data at any given time. We will discuss how customers are doing this to their advantage.

Group Nine reaches nearly 45M Americans each day, totalling nearly 3 billion minutes of video engagement per month. Based on experience across multiple categories, Ashish and Juan Pablo will present a framework for identifying appropriate opportunities to deploy AI/ML/Data Science for video media analysis and discuss some of the technical choices they made (e.g. build vs. buy) in their decision making process. This is applicable to other industries. They will discuss the decisions they’ve made on what to automate and where to respect human creativity and processes, means of analysis to reach those decisions, as well as ways to promote implementation in organizations where human tasks are perceived as under threat. Finally, they will explain why they made the decision to stop short of full content automation.

Despite the massive quantities of behavioural data that E-commerce sites collect each day, the question persist: who are these online customers? Grasping the intentions and interests of online customer is one of the next steps for many e-commerce sites on their journey towards a data-driven culture. A website is a complex entity to be tamed, so how can we find a scalable approach to understand connections between pages, and how can we apply such insights for enhanced smart targeting strategies? In this talk, Benoit and Hassan  offer you a glimpse into some approaches they have taken to solve this challenge, along with corresponding insights and challenges.

Joost and Egge will jointly share experiences working with Data Science in Healthcare, sharing best practices from a large academic hospital (Erasmus MC) and an impactful Dutch Data Science scale-up in Healthcare (Pacmed), they will explain why unequal care is imminent and in fact desirable. The talk will cover lessons learned both in technical areas as well as process, focussing on balance between analytical rigour and explainability and leading to trusted results and enlarged adoption of innovative approaches. Joost and Egge will also share ideas on reshaping Healthcare (education) to invest it with the agility needed to face future disruptions.

In this session Yuri Katz will introduce a new method, based on the topological data analysis (TDA), to financial time series and detect early warning signals of approaching financial crashes. Analyzes of the time-series of daily log-returns of major US stock market indices and cryptocurrencies shows that in the vicinity of financial meltdowns, the Lp-norms of persistence topological landscapes exhibit strong growth. Remarkably, the average spectral density at low frequencies of the derived Lp-norms demonstrates a strong rising trend at least 100 trading days prior to either dotcom crash on 03/10/2000, or to the Lehman bankruptcy on 09/15/2008. The study suggests that TDA provides a new type of predictive analytics, which complements the standard statistical measures and ML-algorithms. The approach is very general and can be used beyond the analysis of financial time series.

Understanding text that appears on images in social media platforms is important not just for improving experiences such as the incorporation of text into screen readers for the visually impaired, but they also help keep the community safe by proactively identify inappropriate or harmful content in a way that pure object detection or NLP systems alone cannot. This talk describes the challenges behind building an industry-scale Optical Character Recognition (OCR) system at Facebook that processes over a billion images each day. Viswanath will cover the Deep Learning methods behind building models that perform text detection in arbitrary orientations with high-accuracy, and how simple convolutional models work extremely well for recognizing text in over 50 languages. A critical aspect of the work is scaling up these models for efficient server-side inference. He’ll dive into quantization methods to run neural networks with 8-bit integer weights and activations instead of 32-bit floating points, and the challenges involved in doing so.

CarLens was the most challenging project Krzysztof has ever worked on as a Product Owner. At the beginning, having almost no experience in Machine Learning, he had no clue how difficult it was going to be. He struggled for 7 months to train the model, create beautiful designs, and combine everything into a working app. He learned how important a quality data set is and how important is also the way you manage it. He built an internal data annotation tool, scraped half of the internet and, in the end, he have an app that almost works. But it’s not the end of this surprising story! If you would like to know how it ended, learn some tips for working with the uncertainty of ML projects, and how one simple decision can influence the future of your project – this is a talk for you. 

meet your fellow attendees.
See the registration desk for more information
Location: St Barts Brewery, 66 West Smithfield, London EC1A 9DY

A lot of companies are struggling with managing data science projects and get bogged into several unexpected difficulties. Newcomers on the data science field also face significant problems with hiring the specialists and organizing product teams efficiently. So, though, the data science is a well developed field, managing of data science projects is still quite challenging. 

Nikita is Chief Data Officer in Russian Airline S7 with over 100 planes in fleet. Three years ago the airline company launched the first data science projects with only two data science project managers involved. Now they have a team of around 40 specialists in the fields of product management, data engineering, data science, software development and business analysis. Over these three years they have successfully implemented a few state-of-the-art machine learning products in some core areas of their business. 

Nikita believes that their experience and ideas can help you and your organizations to launch and operate data science teams more effectively. In his talk he is going to cover the following areas of a data science management: roles and competences, organizational structure, product management frameworks, use of the Data Lake, Data Catalogue and Data-as-a-Service, discovery of new business cases and many other.

If your stakeholders have tasked you with moving beyond science and experimentation into creating enterprise data products or analytical decision support tools that will show some business or user/customer value, then whether you’re a designer or not, there are UX research techniques you can apply today to increase your success. While machine learning, AI, and advanced analytics remain in the “hype” cycle, data teams still struggle to design engaging, valuable decision support tools that customers love. Why? Solutions are too often data-first and human-second, leaving customers scratching their head. In this interactive session, Brian will share some UX & design strategies that can help leaders create more engaging, useful, and usable analytics software that is rooted in what users and stakeholders actually need. He will end the ssession with some research interview practice (in pairs) so that participants can hopefully avoid creating one of the “80% of analytics insights that won’t deliver outcomes through 2022” (according to Gartner!)

The first generation of Automated Machine Learning tools (from DataRobot, H20, Auger.AI and others) enables data scientists and business analysts to train with thousands of algorithm and hyperparameter combinations to generate the best possible predictive models. After uploading the data as a spreadsheet and waiting for training, the user selects the best model from the leaderboard and they are ready to do predictions.

Recently, several new AutoML products (from Google Cloud AutoML Tables, Microsoft Azure AutoML and Auger.AI’s open source A2ML API) have introduced the ability to automate the full AutoML process. Their APIs each support several phases in a pipeline: Importing Data, Training Against Algorithms and Hyperparameters, Evaluating Models, Deploying Models, Predicting Against New Data, and finally Reviewing the Performance of Models. These products all emphasize automated use by developers, not analysts uploading spreadsheets and viewing leaderboards.

Now that the full A2ML process can be automated new frontiers in exploiting AutoML’s capabilities are opened. Business logic in applications can be replaced by predictive models automatically generated from any data the developer has access to. Painstakingly coded sorting of results and lists of objects (accounts to manage, contacts to call, devices to maintain, trucks to route) can be ordered by a predictive model ranking. Complex cascades of if-then-else and switch statements (also known as business rules) derived from some “subject matter expert” or business person’s judgment can be replaced by the insights of a predictive model.

This use of AutoML has a far wider audience than just data scientists. Enterprise application developers can be far more productive and the amount of hard coded business logic in applications will steadily be reduced by use of predictive models. Software Has Eaten the World. With second generation AutoML, AI will now eat software.

This talk will describe in more detail just what second generation “Automated AutoML” entails. And describe several use cases where we have put this into effect for various applications and business problems.

With large data collected from assets, machine (supervised) learning is increasingly being used to develop predictive analytics solution for maintenance and failure prediction. Examples include failure prediction of main bearings in locomotive engines, troubleshooting of faults in healthcare assets like CT/MR scanners. For developing any supervised learning models, labeling of training data is imperative. In fact, error in labeling is the single biggest source of “bad” machine learning models. Typically, experts are consulted to label “interesting” events viz. failure of an asset, equipment downtime, part replacement etc. History of this “interesting” events resides in engineer/technician notes. The experts manually mine this notes and label the input data. In this presentation, Tapan will explain an unsupervised learning method which can mine through the technical notes and create a corpus of “interesting” events and label each data point to one of these events. It also allows for structured expert feedback to edit the labeling, if required.

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. There will be two discussion rounds of 20 minutes each. So choose your two most burning topics and discuss with your colleagues. 

Topics:

– Finding, selecting and retaining data science talent with Richard Downes
– How to  acquire new skills and develop professionally with Hassan ElHabbak
– Leveraging Auto-ML and other available packages and tools with Adam Blum
– Deploying analytic solutions to production environments with Yann Landrin-Schweitzer
– Tapping into third party data sources with Phil Winters
– Internet-based applications (e-commerce, media, etc) with Duncan Manhattan

Johannes will share insights into the recent advancements setting up operational analytics products at adidas, along with lessons learned along the way.

This talk will introduce pipe, an in-house Machine Learning Pipeline that has the capability to power thousands of ML models built on top of user data at Automattic, the company behind WordPress.com. A generalised machine learning pipeline, pipe serves the entire company and helps Automatticians seamlessly build and deploy machine learning models to predict the likelihood that a given event may occur, e.g., installing a plugin, purchasing a plan, or churning. The case studies will demonstrate how predictive models are used in marketing campaigns, discuss experiment setups for online evaluation, touch on uplift modeling, and most importantly, plot a journey to building models that are useful as opposed to just accurate.

For companies that want to learn and iterate quickly there is no substitution for scalable experimentation practices — and as machine learning techniques have become increasingly democratized the burden of rigorous work is and should increasingly shift to how effectively models are tested and validated. In the pure online publisher space this problem is relatively easy, and A/B testing practices are common enough that much of this problem has largely been “solved” — but what about when A/B testing doesn’t work? Sometimes real world release mechanics or marketplace effects make traditional online experimentation practices impractical or, worse, can lead teams to conclude the wrong results. 

This session will cover the importance of good experimentation practices, and some of the tools and techniques that Uber uses to run experimentation at scale.

AutoML is becoming a pervasive tool for data scientists and machine learning practitioners to quickly build accurate machine learning models. Recent products from Google, Microsoft, Auger.AI and others emphasize a programmatic API approach (versus a visual leaderboard) to applying AutoML. All of these products have a similar processing pipeline to achieve a deployed prediction capability: data importing, configuring training, executing training, evaluating winning models, deploying a model for predictions, and reviewing on-going accuracy. With “AutoAutoM”L, ML practitioners can automatically retrain those models based on changing business conditions and discovery of new algorithms. But they are often practically locked into a single AutoML product due to the work necessary to program that particular AutoML product’s API. This talk describes an open source multivendor project called A2ML that allows developers to embed AutoML into their applications with any machine learning vendor’s product.

There is a lot of visioning going on around AI/Augmented/Automated machine learning as well as some first good specialist topic examples. But what are the techniques being used and can they be applied to day to day tasks? This presentation will bring together the latest thinking on practical automated machine learning and apply it to a common task: ensuring relevant graphics are created. This is of course useful for non-data science professionals but has a particular value for data scientists who are taking an initial look at new very big very wide data to determine an approach. A summary of the status quo will be given along with practical open source examples to show the techniques and concepts in use.

Exceptional digital experiences are vital for success in the competitive online travel landscape, with highly engaged customers being 4 times more likely to refer brands to their friends, family and connections. With over 80% of holiday bookings happening online, improving digital customer experience is a vital objective for this industry. This session details the journey taken by TUI, which resulted in a model that enables TUI’s digital team to improve customer experience and predict revenue growth.

This is where the rubber meets the road and we find out if we came through on our promise. What was your most important learning? Did you make new valuable contacts? What will you tell your colleagues, your friends, your family when you get home?