Schedule 2019

Most organizations understand the predictive power and the potential gains from AIML, but AI and ML are still now a black box technology for them. While deep learning and neural networks can provide excellent inputs to businesses, leaders are challenged to use them because of the complete blind faith required to ‘trust’ AI. In this talk we will use the latest technological developments from researchers, the US defense department, and the industry to unbox the black box and provide businesses a clear understanding of the policy levers that they can pull, why, and by how much, to make effective decisions?

There are a lot of megatrends that have surfaced in the last couple of years like Digital Transformation, Big Data, AI, Blockchain, Customer Experience, Personalization, Industrial IoT, or Digital Factory, and all of that. These mega trends have led our clients to invest a huge amount of money in different initiatives, so they had created different centers of excellences, invest in a lot in talent, they are hiring data scientists, data engineers, technology folks and many more. They are investing a lot in creating the infrastructure to support these use cases or initiatives, but these clients have not seen a lot of success. They are sharing with us a lot of frustrations in the sense that they are not seeing a lot of tangible RoI, they are not seeing a lot of adoption from the end business users, and even hiring and engaging the talent that they are recruiting right now is a big pain point. Knowledge management is a big pain point. Governance is a big pain point. So, how do we solve for all of that? Having heard all of these frustrations from our client organizations, we would be sharing some frameworks and case studies which had helped our client network to overcome these challenges.

This session will provide a glimpse of Aditya Birla Group DNA’s efforts to deploy AI/ML solutions in production. It will showcase industrial applications for computer vision (perceiving the seen) and forecasting techniques (foreseeing the unseen). The technical session will do a deep dive into the technical challenges encountered and how they were overcome.

The talk is followed by keynote speech of Kaushik Dey, where Data Scientists will guide the audiences to implement a Federated Learning Algorithm on anonymized real data and compare the performances of Centralized vs Federated models. The session will cover the following aspects of the topic: o Introduction to de-centralized model – federated learning o Technical architecture of federated learning o Introduction to PyTorch o Illustration of data generated from 4G eNodeB o Business importance of predicting network KPIs o Developing centralized models of throughput degradation based on 4G network performance KPIs o Develop FL models using deep neural nets o Performance comparison of centralized and de-centralized models o Challenges and open areas of Federated learning

Suppose that a customer who has given a high rating about a mobile phone writes the following review about the product: The front camera of the phone is excellent! Truly speaking, this is the best front camera I have experienced so far. From this review, we can understand two things. First, the customer holds a positive opinion about the phone. Secondly, the front camera of the phone is the targeted feature on which the opinions have been expressed in the review. In this workshop, we will be particularly interested in discovering patterns as indicated in the second case. We will discuss a framework that enables us to first discover the targets on which the opinions have been expressed in a review and then determine the polarity of the opinions. This kind of detailed analysis helps us to discover the components or features of the products which the customers have liked or disliked and thus help us to better summarize the information.

The most common question marketers ask now is - “Did my ad campaign cause the user to convert and generate more revenue for my brand or would that have happened anyway?” The Complexity of the ad-tech ecosystem is constantly growing with brands running marketing activities across multiple channels, new targeting capabilities, and formats. Due to this, traditional digital marketing measurement metrics like cost per click, return on investment, cost per conversion, etc just scratch the surface while measuring the impact of marketing strategies. Lack of a mathematical approach to differentiate between correlation and causation leads us to look at the incremental lift as a metric to measure the impact of marketing strategy. In this session, we will introduce incremental lift measurement, talk about common challenges faced in calculating incremental lift and different approaches to tackling them in the digital marketing ecosystem. Our methodologies cover concepts of various A/B testing environment setups, randomization, bias handling, delivering outputs at scale and how we use these outputs for marketing strategy planning and optimizations helping us achieve higher campaign efficiency.

An insurance company wants to identify photo-journalists at risk, an oilfield services company wants to identify personnel at risk at their work locations, a chemical company wants to improve plant growth - and all of them are adopting innovative techniques to solve these problems, using the emergence of AI & Deep Learning. It has now become possible for organizations to exploit rich unstructured data to gain insights. These include video stream data from CCTV and drone footage, audio data from customer calls, and text from social media interactions. This talk will describe in technical detail how image classification, object detection, and segmentation techniques are being used to solve such a diverse range of problems. It will also expand on the common roadblocks encountered such as lack of sufficient training data, incompleteness of architectures, which often lead to poor accuracy.

Managing customer churn is vital to business performance. As is researched and proven, a 5% improvement in customer retention rates increases profitability by 25 to close to 95%, as it is many times more expensive to acquire new customers than it is to retain an existing one. Understanding controllable and uncontrollable drivers of customer churn is essential in order to increase customer engagement and designing short- and long-term retention strategies. Below is a “Customer Retention Framework” we developed for one of our F1000 clients in the Logistics domain: Define Churn: As a first step, we defined churn for our client, keeping their business in context and the category of customers they wish to retain. Identify Customer Segments: We used RFM-LP (Recency, Frequency, Monetary Value, Longevity and Profitability) as the key metric to segment their customers into Bronze, Silver and Gold. This metric was used as a proxy to “Loyalty”. Develop a Hypothesis Map: There are multiple reasons, controllable and uncontrollable, why customers switch. We leveraged our experience working in the industry, secondary research and client (business and field) interviews to generate a list of potential churn triggers. We call this collection of hypotheses a hypothesis map. This map should be MECE (mutually exclusive and collectively exhaustive) and is the basis for predictive model(s) to be developed in the next phase of the project lifecycle. Build a Predictive Model: Leveraging the hypothesis map, Evalueserve developed a data-mart which had all the relevant data harmonized and in the required format. A predictive model was then developed to calculate a risk score for all the customers. The model does not only help the client identify customers at-risk of attrition but also understand the triggers behind the same. Create an Action Item Matrix: Actionable insight is key to a successful analytics project. For this purpose, we developed an action item matrix (a framework that brings together the results of the model developed and specific KPIs such as probability of churn and customer lifetime value) to help our client prioritize retention efforts and determine ROI for that effort. The matrix further helped them in identifying up-selling and cross-selling opportunities. Operationalization Strategy: Evalueserve helped the client in designing a retention framework and then in designing a pilot to test out the same. Once the pilot was deemed successful, the framework was rolled out enterprise wide. In my presentation, I will focus on (i) Problem Formulation (ii) Solution Design (iii) Insights to Action Framework and (iv) Operationalization Strategy.

Businesses are collecting more and more data every day. Vast amounts of third-party data is also publicly available. It is almost impossible for businesses to analyze all the data available at their disposal. Enterprises need a better system to generate insights across various aspects of the business, easily analyze different scenarios and get prescriptive recommendations. ‘Autonomous analytics’ much like ‘Autonomous cars’ helps in identifying anomalies as they occur, uncover opportunities as they unfold, find root cause and suggest necessary actions. These analytics are adaptive, instantly evaluate outcomes and help re-align or auto-correct to achieve desired outcomes based on the business situation.

What happens when a race team has all the insights to make fast, accurate decisions, working as a fully connected unit that adapts to complex, shifting conditions? Podium-winning performances and an instinctive racing team.

Manu will talk about his experiences with using some of the recent advances in text analytics/ NLP/ NMT including BERT, GPT2 and XLNet

While failures of software engineering projects had been understood and contained, data science is exposing project failures that are causing some to pause. Some say 90%+ of Data Science projects never see the light of the day. In this techno-strategic talk, we look at on-the-ground reasons why and provide a blueprint to avoid your project becoming one of that statistic. Build Different. Be Different.

Ronald Fisher considered as father of Statistics discovered Discriminant Analysis(DA) in the year 1936 as a scoring model to achieve the largest possible separation between classes. The model involved a linear discriminant equation of input variables called the Z score to classify the individual records. Support Vector Machines(SVM) which was originally invented in 1963 by Vladimir N. Vapnik and Alexey Ya. Chervonenkis has the same goal of achieving the largest separation between classes specifically binary, but can be extended to multiple classes. SVM separates the two classes by drawing the best hyperplane because many hyperplanes can produce the solution. Both DA and SVM use hyperplane as the criterion to separate the classes. While DA finds the best hyperplane that maximizes the distance between the group or class centroids subject to the variance of the hyperplane set equal to one, SVM finds the best hyperplane that maximizes the distance margin between the closest data points to the hyperplane on either side of it. The talk will focus on comparing and contrasting DA and SVM as classifiers from the perspective of Analytics. The similarities, differences, explanatory power, and accuracy will be delineated threadbare for both DA and SVM so that analytics professionals can leverage on their strengths for better insights in the context of classification. The entire philosophical convergence between SVM and DA will be demonstrated with excellent and relevant data sets.