Trajectory Boosted Transformer Model and KG/LLM based μ-Genre for PV Offer/Content Type Arbitration

In the realm of video streaming services, with diverse offers and vast array of content, presenting personalized content to customers poses unique challenges for recommending thematic carousels on the homepage. We capture diverse and niche customer interests by devising micro-genre (μ-genre) carousels by leveraging knowledge graphs and Large Language Models. While traditional Next Item Recommendation (NIR) systems focus on predicting immediate customer engagement based on past behavior, recommending different carousels on homepage requires balancing diverse content, considering both long-term and short-term user interests, and exploring among different offers. To achieve this, we first analyze and derive insights from the customer journey in terms of transitions between states/segments using a semi-supervised Input Output Hidden Markov Model (IOHMM). Next, we propose Transformer based customer engagement trajectory driven propensity model that helps to capture customer’s long term interests. We influence the propensity model learning process using insights derived from customer journey analysis. Finally, we introduce an impression-discounted Upper Confidence Bound (UCB) method for recommending under-impressed offers/content types and page/neighborhood diversification modified Maximum Marginal Relevance (MMR) approach for generating a diversified page. We evaluate the efficacy of the proposed approaches in an intuitive manner that aligns well with business decision-making and enables the selection of treatments for A/B experiment.