Akshat Gupta

Highlights

• Co-published a paper - Quantifying and Leveraging User Fatigue for Interventions in Recommender Systems at SIGIR’23.

• Awarded two times –

“Game Changer” and “Ace of Initiative” by the Director of Product.

Recommendation Systems

• Content Curation Revamp: Identified & fixed gaps in content corpus selection affecting content life cycle; mitigated issues in training dataset and underlying data streams leading to an avg jump of abs 2% in model ROC AUC; set up true recall and NDCG measurements to test relevance and ranking abilities of recall layer; added video-watch-time based signals to increase recall for highly watch-worthy posts.

Result – Deployed the best test variants upping retention by abs 0.4% (150k+ extra users per day).

• Recent events-based Candidate Generator: Researched, proposed & implemented a post-post cosine similarity-based algo to generate

personalized content; replaced a very costly In-Session Personalization Layer, supervising a team of 1 MLE, 1 SDE and an analyst, conducting 4

A/B test iterations.

Result – Achieved savings of INR 8M per month, 14 seconds extra time-spent per user and 1% more video-full-watches.

• Learnt Weight Tuning: Established a working method to fine tune the weights for combining multiple signal scores - planned & enabled data

collection with bounded randomization of weights for users; A/B tested weights based on multiple models like Linear & Logistic Regression,

GBDTs and Statistical Weight Tuning functions.

Result – Made two production releases to gain 0.1% abs retention, 2% more post views and 10 seconds extra time-spent per user.

Ad-tech

• User Fatigue Model: Modelling and leveraging user fatigue from ads to intervene in ad-load policies - looked after feature selection, dataset

preparation and analytics to build a churn probability prediction model (Xgboost) with ad-load as primary feature; leveraged fatigue score to

tweak ad-load on the Landing-Feed – high/low fatigue -> less/more ads.

Result – Reaped 0.1% retention gains on the cost of 0.2% revenue loss (INR 300k p.m.).

• Ad policy Intervention: Intervening in ad-policy in real-time using a Contextual Bandits Model – drove the entire project, selected model features,

formulated the optimal reward function and ensured timely delivery while maximizing user satisfaction and ad-revenue concurrently.

Result – Generated additional revenue of INR 2M per month with satisfaction metrics neutral to positive.

Search

• Search Relevance Improvement: Upgraded simple token match based algo to an n-gram based logic in Elastic search to retrieve the most relevant

documents for a searched query.

Result – Increased interacted searches by 16% and time-spent for searching-users by 20 seconds.

• Search Ranking Improvement: Introduced a 2-stage ranking process for Search using counter features and engagement scores to increase the

mean reciprocal rank (MRR) by 45%.

• Found out better communication windows for drip marketing via pattern recognition, bringing 3k extra customer leads per day
• Led a 3-member team to figure out and fix the bugs in analytics data pipeline, eliminating a 12% error in the estimated demand
• Automated and managed the entire reporting for marketing team single-handedly - handled reporting for 20+ agencies.

• Identified user cohorts dropping off due to Delivery Fee (DF), driving DF pricing revision for them, reducing the drop off by 15%
• RCA on B2B Cart Abandonment - Identified user purchase behaviours, helping in precise re-targeting of customers leading to an uplift in the cart conversion rates by 12% in 3 months
• Built a one-stop dashboard for the SAAS business unit, in collaboration with Singapore and Indonesia business stakeholders.