Schedules

The last few years have witnessed Artificial Intelligence (AI) making quantum leaps into several domains such as healthcare, customer analytics, education, manufacturing, and so forth. With the changes in business strategy, personalized product recommendation has become an integral part of the customer experience journey. There has been extensive research leveraging collaborative filtering, market basket analysis and various other machine- learning / deep-learning based solutions for developing recommendation engines. The features such as customer behavior, historically purchased items, characteristics of similar users, etc. are used as inputs to the machine learning pipelines. In this paper, we present a novel approach using Graph algorithms for building a product recommendation solution for a publishing company. The publishing company has footprints with its customized education content, software and services at many world-renowned institutes and universities across the world. However, the company is experiencing a significant churn rate for their online platform for learning which is subscribed by various institutes. The developed solution is an ensemble of insights generated from several graph algorithms for centrality (personalized PageRank, betweenness & degree algorithm) and community detection (Louvain, strongly connected components, etc.) to control churn with product recommendation being the key step. The approach for recommendation generation for courses and books is based on the overall influential power of various other institutes and associated instructors in the network. Simultaneously, the approach focuses on the popular books and courses inside a local community identified by the graph algorithms to generate recommendations.

Computational Linguistics or the study of Natural Language Processing is the modern approach for a machine to understand and interpret language including its grammar, semantics, phonetics etc. by using large datasets and computational tools and techniques. Back in the 1990s, statistical machine learning methods began to replace the classical top-down rule-based approach to interpret languages, primarily due to accurate results, speed of processing and robustness of the algorithms. The advent of faster processors in the 2010s exponentially improved the performance of NLP algorithms. The need for clean and labeled training data also grew along with the emergence of faster processors and robust algorithms to improve the accuracy of the models. Statistical approaches have turned another corner and is now strongly focused on the usage of deep neural networks to both perform inferences on specific linguistic tasks and for developing robust algorithms. This paper focuses on a machine learning approach to perform word segmentation and Hierarchy detection on medical documents (in form of any editable digital documents like PDFs). When text is extracted from digital pdf using python libraries such as PDFMiner, the output is in the form of Scriptio continua - a style of writing without spaces between the words or sentences. One of the ways to use this unseparated raw text in a meaningful way is known as word segmentation, which is a process to determine the word boundaries in a sentence. In this paper, we present a machine learning approach to build a word segmentation algorithm and also find the hierarchical structure in the text (for example Header1 or Header2 etc.). We leverage the English Wikipedia dataset to build and train this advanced sequence to sequence model. The technique used in this paper has been successfully tested on medical domain data. We have also observed significant improvement in model accuracy by further training the algorithm with domain specific documents (in form of PDFs)

Analyttica is a niche data science and advanced business analytics company focused on providing incremental business impact for its clients by developing custom innovative solutions for them in the predictive and prescriptive analytics space. With the advent of newer technology with greater computing power, big data and contemporary media channels, organizations have realized the importance of marketing mix modelling (MMM) and identified the benefits it entails that gives rise to cost saving opportunities and drives profitability. Marketers are under increasing pressure to move away from intuition based budgeting decisions to factual budgeting decisions, substantiated through quantitative evidence. In an attempt to understand how their marketing activities connect with real movements in sales and market share, the client, an Australian subsidiary of one of the world’s largest automotive manufacturer, wanted to conduct a MMM exercise and arrive at a clearer understanding of their Return on Marketing Investment. The client’s marketing strategy was designed around the themes of Media, Messaging and Brand Equity. Various levers were identified under each of these themes which are essentially the components of the marketing mix, to help assess the relative impact of marketing on enquiry and sales of the client’s product. Using historical marketing and sales data under each of the exhaustive components identified, Analyttica helped estimate the relative influence of the various components of the marketing mix, while controlling for other sales drivers such as seasonality. Proper capturing of the event and longevity of effect of the media channels on the outcome (in this case unit sales), is paramount for the success of any MMM application, which is heavily dependent on the suitable selection of the transformations such as decay (simple, logarithmic, exponential etc.), logistic, AdStock, and gamma. The appropriate selection of these transformation functions is highly contextual and driven by the skills, experience, knowledge and judgment of the modeler. As a surrogate to the knowledge, skill and experience of the modeler, we developed a heuristic optimization methodology for selection of the transformation functions for the media channels. The methodology considered six transformation functions competing against each other to arrive at the most appropriate one, that reduced the error of the model. This method is automated and has been developed into a prototype that can be applied in similar situations towards the process of measuring “Above the Line” (ATL) marketing effectiveness and optimization. The above exercise highlighted the potential costs and benefits of all the components, which was then weighed against one another in order to build a media mix solution, to arrive at the effectiveness of each of the ATL channels, to enable the decisions around objective allocation of the marketing budget across channels.

In world of Artificial Intelligence, Machine learning, etc., it is becoming simple for Business User to get required Insights, information or take action on complex Enterprise environment. But it still needs some effort from users to achieve those along with maintaining required data protection of underlying system and get information on finger tips. This is due to multiple complexities of Enterprise landscapes which needs some manual effort to achieve required results. Even for testing of such cloud solution for Enterprise, we have to consider multiple combination of complexity and deployments which could be possible using different Authentication services (ADFS or SAML SSO, etc.), AppServer, or any other 3rd party software or any On- premise solution for leveraging Hybrid environment benefits. In other case, where one cloud application has to connect with solutions on different Cloud platforms, regions, or applications to form complex deployment for which effort for manual setup for validation after every release is significant. It can be achieved by automating these custom manual actions using RPA (Robotic Process Automation) framework using Machine Learning enabled Natural Language Processing system for providing inputs to system using voice, text, etc.

The agenda of the talk has been broken down into two parts: 1. Query Understanding [30mins]: [Sonu Sharma] • NLP based Deep Learning Models for finding the intent of a Query in a particular taxonomy/categories: Description and Jupyter-notebook demonstration [20mins]: o Multi-Label/Multi-Class Classification Model from scratch in Keras o Feature Engineering in Spark Scala and pandas o Keras Functional APIs details in TF 2.0 o ImageNet moment of NLP - Latest invention in Word embeddings – ELMO and BeRT o Understanding Deep Neural Networks like Bi-directional Long Short-term Memory (BiLSTM) and character embeddings for language modeling • NLP based Deep Learning Models for Query Tagging with entities like Brand, Color, Nutrition, product quantity, etc. using Named Entity Recognition: [10mins]: o Building custom model in Tensorflow Estimator API o Traditional Word Embeddings like Glove, Fasttext etc. o Query (Text) Preprocessing o Sequence modeling using Convolutional Random Fields (CRF) o Saving and Restore heavy model in TF using SavedModel concept 2. Related Searches [20mins]: [Atul Agarwal] • NLP based Deep Learning Model for predicting Next Search Keyword – Model Description and Jupyter Notebook demonstration[20mins]: o Building Sequence to Sequence (Seq2Seq) model using Long Short Term Memory (LSTM) concept of Deep Neural Network in Keras o comparing different word Embeddings e.g. word2vec, fasttext, glove, etc. in popular AI framework such as gensim. o Keras Sequential APIs details in Keras o Similarity Search based on Facebook AI Research aka FAISS.

NLP is evolving like a ripple in the ocean of Machine Learning. Today most of the companies invest a lot in creating NLP models for their needs. It is customized for their needs. So Scaling the NLP model to other teams is always a challenge. Each time a team has to analyses the data and write NLP algorithm for the data. Solution: To address this problem, I am proposing a solution that creates Standard NLP framework that just needs the data of the customer. Customers do not need to invest more in creating NLP models and algorithm. Instead, they can just submit their data to NLP framework that automatically learns the intent And develop an algorithm for the particular data.

One of the major challenges in Educational NLP is to assess student responses to questions. Automatic short answer grading (ASAG) is the task of assessing short natural language responses questions using computational methods. Short Answers tests the ability of the student’s recall using natural language, unlike multiple-choice questions that evaluate only recognition. However, there is no easy way to evaluate short answers, resulting in manual evaluation and feedback putting an enormous burden on teachers in resource constraint countries like India. In this paper, we are presenting our research into this aspect by combining conventional and advanced Natural Language Processing (NLP) methods which include various embeddings that range from word level, sentence level and contextual. We are also presenting a comprehensive evaluation of various embedding techniques (word2vec, FastText, ELMo, Skip-Thoughts, Quick-Thoughts, FLAIR embeddings, InferSent, Google’s Universal Sentence Encoder and BERT) with respect to short text similarity. Our method helps in evaluating natural language-based short answers resulting in instant feedback for students. We also combined Explainable AI (XAI) insights on the logic behind scoring an answer thus providing feedback and improving our models.

To date, natural language processing has been extensively deployed in the domain of online media – Twitter, Facebook, News, and other widely available text resources such as IMDb movie reviews, Reuter’s data, etc. Different problems such as topic modelling, entity recognition, sentiment analysis have been tested and benchmarked on such standard “generic” datasets. However, Specialized domains, such as biomedical text, have their own complexity. Biomedical text typically comprises of two genres - scientific journal articles such as PubMed, and clinical documents. These data sources have a number of characteristics that make it difficult for using NLP – such as presence of parenthesized text, lack of tagged standard data for aforementioned NLP problems, and therapeutic area variance. Additionally, crucial information is present in tables and figures, which are in general difficult for NLP applications to handle. At ZS, the Advanced Data Science Team is at the forefront of solving these challenges and building new applications to drive efficiency in trial operations. In this workshop, we will walk participants through how AI (and NLP in particular) is transforming the pharma R&D landscape. We will then deep-dive into a specific challenge - biomedical entity recognition. Biomedical NER involves identifying biomedical entities such as diseases, drug and chemical compounds, p-values for statistical tests etc. Participants will be acquainted with multiple state of the art methods such as Conditional Random Fields, Bi-LSTM and Tree LSTM co-training that are currently being used to do Biomedical NER.

1. Introduction to Neural Networks and CNN 2. Hands on CNN using TensorFlow and Keras on Google Colab Platform. 3. Various CNN architecture and Transfer Learning Techniques for pretrained CNN model. 4. Hands on Transfer Learning Techniques. 5. Convolutional Neural Networks for Object Detection and Segmentation.

The globally accepted solution, that solves climate change problem, is increased adoption of renewable energy sources for generation of electricity. This is leading to numerous wind power plants being set up all over the world. These power plants rely on the available wind energy for power generation. To harness all the possible energy from the wind, the rotor needs to be perpendicularly aligned with respect to the direction in which the wind is blowing. This is done through a control system called Yaw mechanism. The wind turbine is said to have a yaw error, if the rotor is not perpendicular to the wind. The share of energy that can be harnessed from the wind, drops at the rate of cosine of the yaw error. Thus, the correction of yaw errors is necessary to ensure lowered losses and optimized generation. The usual detection techniques include intuitive guesses made by expert site engineers working at the wind farm and through observation of generation patterns. These techniques are at their best, approximations of the actual scenario and only detectable when there are really large yaw errors. The delay in detection in these manual methods amounts to approximately 12% of loss in energy generation. This paper explains how machine learning is being used to detect yaw error right after its onset, identify the root cause and quantify the consequent energy lost. The detection of yaw misalignment is done by analyzing the wind direction, nacelle position and yaw angle data captured by the Supervisory Control and Data Acquisition (SCADA) system installed at the wind power plant. This analysis is then augmented with statistical tests to rule out errors in measurement and ascertain that the data is indeed indicating yaw misalignment. When the results prove to be statistically significant, the next step is to identify the root cause for the misalignment, so that an appropriate corrective action can be recommended to the site engineers. The probable root causes are presence of a faulty sensor, wrongly calibrated sensor or low yaw speed. These root causes are identified through pattern recognition algorithms, that look for different signatures, that each one leaves in the data. A non-linear regression model is then trained to learn the pattern of energy generation of a turbine in healthy condition. This model is then used to predict the energy generation of the turbine for which yaw misalignment has been observed. The difference in the actual and predicted values is the estimate of loss in power generation due to yaw misalignment. The adoption of this algorithm has helped in early identification of yaw errors and take corrective actions in near real time, thus preventing the loss in energy generation of up to 7-9% annually.

The onset of cancerous cells caused by mutants in the foetal stages is an evolving mechanism that can be traced to pathological as well as clinical evidence that would require interpretations on a Machine Learning (ML) matrix. Random psychosomatic states in the antenatal stages have profound influences on the embryonic cellular development. These changes have the imprint on the pathological parameters of the patient and the progressive evolution of changes triggered by the mental states of relative well-being. Hormonal configurations are the essential markers in the psychosomatic reference frames that are fed into the ML model. The mathematical models explore the parametric influences of the pathological changes, the hormonal changes, the relative coordinates of the fetus, the relative growth ratio of the brain and the body and finally the functional MRI derivatives for cell morphology and cytoplasmic changes that signal potential mutants in the embryonic development evolution. The algorithm development factors in the measurable variables in the development cycle of the fetus and crystallize the probabilistic computations of predicting the impact on the cytoplasmic growth trajectory with the potential of conversion into mutants. The elements of fMRI are valuable parametric measures that help crystallize the predictive models albeit these are on theoretical models yet requiring the investigative rigors of applied clinical research for cancer and a host of psychosomatic diseases.