We present TEEN, an industry grade solution to the problem of time expression extraction and normalization (Timex). Extraction and normalization of temporal units is a challenging problem due to several factors, e.g., (i) same time units may be expressed in different ways, (ii) inherent ambiguity in natural languages leading to multiple interpretations, and (iii) context sensitive nature of natural languages. While various academic and industrial approaches have presented solutions towards Timex, building an industry strength solution involves additional challenges in the form of user expectations, need for delivering high precision, and lack of training corpora. We elaborate how TEEN carefully mitigates these challenges. We demonstrate how the proposed approach compares with various state-of-the-art baselines on textual data from finance industry. We further categorize inadequacies of these baselines in an industrial setting. Finally, we provide insights gathered through the observations we made and the lessons we learned while designing TEEN to work in an industrial setting.

The universe is the vast expanse of cosmic space consisting of billions of galaxies. Galaxies are made of billions of stars that revolve around a gravitation center of the black hole. Quasars are quasi-stellar object which emits electromagnetic radiation more potent than the luminosities of the galaxies, combined. In this paper, the fourth version, the Sloan Digital Sky Survey (SDSS-4), Data Release 16 dataset, was used to classify the SDSS dataset into galaxies, stars, and quasars using machine learning and deep learning architectures. We efficiently utilize both image and metadata in tabular format to build a novel multi-modal architecture and achieve state-of-the- art results. For the tabular data, we compared classical machine learning algorithms (Logistic Regression, Random Forest, Decision Trees, Adaboost, LightGBM, etc.) with artificial neural networks. Deep learning architecture such as Resnet50, VGG16, EfficientNetB2, Xception, and Densenet121 have been used for images. Our works shed new light on multi-modal deep learning with their ability to handle imbalanced class datasets. The multi- modal architecture further resulted in higher metrics (accuracy, precision, recall, F1 score) than models using only images or tabular data.

Leaks have undoubtedly been one of the biggest problems plaguing piping and cabling systems across industries like electricity and power, building and smart cities, oil and gas, etc. Addressing these leaks in time becomes paramount as failure leads to a complete standstill of the transportation chain. Most AI based leak detection systems have failed to reach the deployment state as these systems are prone to output false positives. It is imperative to observe that these leaks don’t occur every day or in other words they are rare events. But when they do occur, these leaks more often than not go unnoticed. Due to the insufficient number of identified leak points, it becomes difficult to build an AI based model for the same. In an attempt to aid/replace rule-based and physics-based leak detection systems, this paper proposes a novel AI based leak detection solution using reinforcement learning which not only reduces false positives but also extends itself to multi armed bandit-based leak localization. By using this methodology, we model the latent behavior of any piping or cabling systems and provide a Q-learning based shortest path recommendation in order to help the maintenance team reach the leak node in a short amount of time.

With the ever-increasing use of complex machine learning models in critical applications, explaining the model’s decisions has become a necessity. With applications spanning from credit scoring to healthcare, the impact of these models is undeniable. Among the multiple ways in which one can explain the decisions of these complicated models, local post hoc model agnostic explanations have gained massive adoption. These methods allow one to explain each prediction independent of the modelling technique that was used while training. As explanations, they either give individual feature attributions or provide sufficient rules that represent conditions for a prediction to be made. Current state of the art methods use rudimentary techniques to generate explanations using a fairly simple surrogate model. The data structures they employ restrict them to local explanations which fail to scale up to big data. In this paper, we come up with a novel pipeline for building explanation. A Generative Adversarial Network is employed for generating synthetic data, while a piecewise linear model in the form of Linear Model Trees is used as the surrogate model. The combination of these two techniques provides a powerful yet intuitive data structure to explain complex machine learning models. The novelty of this data structure is that it provides explanation in the form of both decision rules and feature attributions. The data structure also enables us to build global explanation model in a computationally efficient manner such that it scales with large amounts of data.

Digital advertising enables advertisers to promote their products on various online and digital channels. Real Time Bidding is an advanced advertising method which allows advertisers to target potential buyers and acquire ad space on websites, in the form of programmatic auctions. Hundreds of billions of such auctions take place every day. Predicting future performance and developing customized targeting strategies enables advertisers to make better use of their budgets and ultimately improve their KPI measures. With Google's recent policy on phasing out third-party cookies, organizations around the world are shifting more towards data-protected contextual targeting strategies. This paper proposes a machine learning-based approach to predicting future ad-campaign performance by focusing on contextual features such as browser, operating system, device type, and so on. First, a custom metric encompassing cost, performance and campaign delivery is developed. This metric's predicted value is used to score and recommend targeting strategies. To generate new features, feature engineering techniques such (The CMO's Guide to Programmatic Buying, n.d.)as Lag feature generation, statistical encodings, Graph-based embeddings for websites and cyclical feature encodings are prepared for downstream tasks such as CTR prediction This paper then compares and chooses the best performing linear, tree-based, and deep learning models for our proposed hypotheses. Finally, we develop heuristic criteria for offline testing of the recommended strategies and calculate theoretical performance uplift for comparison with older ranker score methodologies, which serve as our baseline results. Our proposed solution outperforms the baseline solution and proposes a novel way of recommending strategies.

Predictive model design for accurately predicting future stock prices has always been considered an interesting and challenging research problem. The task becomes complex due to the volatile and stochastic nature of the stock prices in the real world which is affected by numerous controllable and uncontrollable variables. This paper presents an optimized predictive model built on long- and-short-term memory (LSTM) architecture for automatically extracting past stock prices from the web over a specified time interval and predicting their future prices for a specified forecast horizon, and forecasts the future stock prices. The model is deployed for making buy and sell transactions based on its predicted results for 70 important stocks from seven different sectors listed in the National Stock Exchange (NSE) of India. The profitability of each sector is derived based on the total profit yielded by the stocks in that sector over a period from Jan 1, 2010 to Aug 26, 2021. The sectors are compared based on their profitability values. The prediction accuracy of the model is also evaluated for each sector. The results indicate that the model is highly accurate in predicting future stock prices.