VIVEK SINGH

Project 1: Demand Forecasting Engine using ML:

• Built an end-to-end demand forecasting engine using ML scaled to ~34 stores(pan India), responsible for replenishing ~11k unique items(~ 0.1 Mn unique time-series) from the warehouse to dark-stores, leading to significant automation of the process
• Prepared the data train, val and inference datasets for the 34 dark stores, with overall ~1 lacks unique time-series with 92 features
• Implemented a splitting logic with training incorporating last 5 months and validation and inference data incorporating the 7 days
• Created a logic to segregate the high-selling items and the low-selling items based on the mean quantity sold slot for the items
• Formulated an accuracy metric based on business logic, ie, either percent diff should be between 80-120, or value diff less than 1
• Implemented Facebook-Prophet for ~5k time-series and improved the overall accuracy by 25% and savings on dump by 5cr per week
• Tuned hyperparameters prior_scales and fourier orders) of prophet using optuna leading to further ~10% improvement in accuracy
• Experimented with growth(logistic) and trend parameters of prophet, along with log transformation and improved the model accuracy
• Implemented the LGBM and Catboost model for the ~1 lacs, low-selling time series and improved the overall model accuracy by 15%
• Improved item availability logic by adding weights for each item at the city level, which improved the availability and decreased dump

Project 2: INSIDER-Residual Analysis Tool

• Built a tool to identify time-series on which model performs poor, along with the reasons if there’s noise in data or problem in model
• Applied statistical tests and domain logics to tag the high residuals and a rolling mean-based approach to get reasons for poor forecasts
• Implemented a logic based on residual analysis to select the model which gives the best forecasts for each ~5k time series
• Automated the process along with the report generation having interactive plots and metric values, tracking models with residual analysis

Eagle-Anomaly Detection Tool:

• Built an end-to-end time-series based Anomaly detection tool, which tags the anomalous points based on several time series factors
• Implemented statistical logics to find the contribution of the feature categories leading to the anomaly on a specific day
• Formulated data distribution-based logic for deviation in feature-category and its contribution to the deviation in the target-metric
• Incorporated the weights to each feature category based on which category changes correlate maximum to the target metric
• Dashboarded the results for monitoring the metrics and displayed each anomalous point along with its reasons in the priority manner

Project 1: Product recommendation system and market basket analysis

• Built a product Recommendation system using User-item( interaction between the customer to the product) and item-feature(featuresof the items) based sparse coo matrix(using scipy) for hybrid collaborative filtering using LightFM library with an AUROC value of 85%
• Obtained frequently bought together products combination(~1000 optimum association rules) using ARM over a market basket dataset
• Designed custom dictionary to correct product names using the Levenshtein distance and uesd inflexion point analysis to get threshold
• Project 2: Binary Classification Model For Uninstallation Of The Application:
• Implemented a binary classification model to identify the retailers who are likely to uninstall the LoveLocal retail application
• Clustered the stores using k-means clustering with a haversine distance matrix to fetch the centroid data from Google-places API
• Implemented RF classifier and XGBoost on an imbalanced dataset with an AUROC value of 87%, and the model went into production

• Created Propensity model via supervised learning techniques to create target groups
• Created propensity model via supervised learning techniques, used to create target groups for cross-selling marketing campaigns
• Selected optimal features, trained the model and best performance (Recall:84 &AUROC:85) is achieved by, weighted XG-Boost algorithm
• Calculated the true churn rate per group, and the top decile contains 10% of the population which is most likely to buy SCV Cargo

Yatra support email classifier bot:

• Implemented the multi-class text classification which categorized customer’s query into 5 classes (ticket cancellation, refund etc)
• Applied LDA(genism) to cross verify the miss-classified labels for the email body and corrected the labels if not in top 5 optimum words
• Cleaned the email body followed by basic pre-processing, tokenization, stop words removal and lemmatization using spacy and genism
• Vectorized the text data using Tf-Idf, selected optimal features using Chi2, and trained the model using random forest classifier
• Achieved the AUROC value of 82% (macro average), dockerized and then deployed the text classification application using Flask