Time series data, prevalent in fields like medical, e-commerce, finance, etc., is used for forecasting, such as predicting next quarter’s product demand based on past trends. However, some problems necessitate causal models to answer questions like “What the product demand would have been without a specific intervention (e.g., products with slower delivery time suppressed from the search results)?” Such questions require causal models to estimate unobserved counterfactual outcome. In this paper, we propose a novel Graph Causal Forecasting (GCF) model, that predicts the unobserved demand leveraging the relationship of a product with other similar products in the marketplace (spatial aspect), along with change in demand over time for each product (temporal aspect). The core idea is to estimate the counterfactual outcome using a synthetic control unaffected by the treatment. Our approach uses RGCN-dilated CNN based network, which leverages domain knowledge to automatically design a synthetic control during training. Using GCF for our demand forecasting problem, we achieve 75.3% lower MAPE compared to baseline. We use the forecasted values to recommend high demand products, in terms of our business metric (discussed later) which tracks the quality of these recommendations, we achieve a significant jump of 61.2%. Moreover, it adds 67.8% more high demand products to the marketplace, compared to existing model in production. Deployment of GCF in 2023, led to +1399 bps improvement in number of products with a view from customers, and +310 bps improvement in number of products with a sale. We also compare GCF with state of the art forecasting methods on a semi-synthetic data, created by simulating a treatment on open source traffic data METR-LA. We achieve 30% lower MSE against TGCN, a time series forecasting approach and 30% lower MSE against CRN and 25% lower MSE against Google Causal Impact model, both of which are causal forecasting approaches.

Click-through Rate (CTR) module is the foundation block of recommendation system and used for search, content selection, advertising, video streaming etc. CTR is modelled as a classification problem and extensive research is done to improve the CTR models. However, uncertainty method for these models are still an unexplored area. In this work we analyse popular uncertainty methods in the context of recommendation system. We found that popular uncertainty models fails to capture the predictive uncertainty of the CTR model that exist unique to the recommendation models and is not prevalent in the traditional classification models. We empirical show why a different uncertainty measure is required for the recommendation system CTR prediction models. We propose PRU (Predictive Relevance Uncertainty), a single forward pass uncertainty approach for a sample as a distance from the predictive relevance samples of the training data. We show the efficacy of the proposed predictive relevance uncertainty (PRU) on selective prediction. Further, we demonstrate the utility of the proposed framework on the downstream task of OOD detection and active learning while maintaining the latency of a single pass deterministic model.

Internet users actively search for trending products on various social media services like Instagram and YouTube which serve as popular hubs for discovering and exploring fashionable and popular items. It is imperative for e-commerce giants to have the capability to accurately identify, predict and subsequently showcase these trending products to the customers. E-commerce stores can effectively cater to the evolving demands of the customer base and enhance the overall shopping experience by offering recent and most sought-after products in a timely manner. In this work we propose a framework for predicting and surfacing trending products in e-commerce stores, the first of its kind to the best of our knowledge. We begin by defining what constitutes a trending product using sound statistical tests. We then introduce a machine learning-based early trend prediction system called TrendSpotter to help users identify upcoming product trends. TrendSpotter is a unique adaptation of the state-of-the-art InceptionTime model\citeInceptionTime that predicts the future popularity of a product based on its current customer engagement, such as clicks, purchases, and other relevant product attributes. The effectiveness of our approach is demonstrated through A/B tests, where we first showcase the effectiveness of our statistical test based labeling strategy, resulting in an incremental sales lift of 59 bps\footnotebps or basis points are a measure of percentages. 1 bps = 0.01% across two experiments on home page and search page. Subsequently, we conduct a comparison between our machine learning model and the statistical labeling baseline and observe an additional sales gain of 14 bps, reflecting the importance of early identification of trending products.

Change Point Detection (CPD) models are used to identify abrupt changes in the distribution of a data stream and have a widespread practical use. CPD methods generally compare the distribution of data sequences before and after a given time step to infer if there is a shift in distribution at the said time step. Numerous divergence measures, which measure distance between data distributions of sequence pairs, have been proposed for CPD \citeMStatisticNIPS, BergCPD and often the choice of divergence measure depends on the data used. Density Ratio Estimation (DRE) \citeRelDivCPD,BergCPD can be used to estimate a broad family of f-divergences, which includes widely used CPD divergences like Kullback-Leibler (KL) and Pearson, and thus DRE is a popular approach for CPD. In this work, we improve upon the existing DRE techniques for CPD, by proposing a novel objective that combines DRE seamlessly with adversarial sample generation. The adversarial samples allows for a robust CPD with DRE to track subtle changes in distribution, leading to a reduction in false negatives. We experiment on a wide variety of real-world, public benchmark datasets to show that our approach improves upon existing state-of-the-art (SoTA) methods, including DRE based CPD methods, by demonstrating an \sim 5% increase in F-score.

The temporal difference (TD) learning framework is a major paradigm for understanding value-based decision making and related neural activities (e.g., dopamine activity). The representation of time in neural processes modeled by a TD framework, however, is poorly understood. To address this issue, we propose a TD formulation that separates the time of the operator (neural valuation processes), which we refer to as internal time, from the time of the observer (experiment), which we refer to as conventional time. We provide the formulation and theoretical characteristics of this TD model based on internal time, called internal-time TD, and explore the possible consequences of the use of this model in neural value-based decision making. Due to the separation of the two times, internal-time TD computations, such as TD error, are expressed differently, depending on both the time frame and time unit. We examine this operator-observer problem in relation to the time representation used in previous TD models. An internal time TD value function exhibits the co-appearance of exponential and hyperbolic discounting at different delays in intertemporal choice tasks. We further examine the effects of internal time noise on TD error, the dynamic construction of internal time, and the modulation of internal time with the internal time hypothesis of serotonin function. We also relate the internal TD formulation to research on interval timing and subjective time.