Hybrid Recommendation Systems using Probabilistic Graphical Models
summary
Hybrid system builds a Bayesian-style graph that combines LDA topics with genres and video IDs to capture both content and co‑occurrence signals, training a PGM with pomegranate on 5,500 items in ~27s and operating on a news corpus of ~25k videos and 1.8 million view events. It selects an 11‑topic LDA by perplexity, creates an unweighted dependency graph (UDG), returns probability-ranked recommendations (handles cold start via segment‑level PGMs and a population default), and addresses CPT explosion by aggregating topics; pomegranate scales comfortably to ~50k nodes but needs TensorFlow/Edward for real‑time horizontal scaling.
AI-written summary · about
chapters (7)
related talks
25mBuilding a recommendation system for EPAM systems
23mML-Powered Search & Recommendation 101
30mGraph Machine Learning in All Its Glory!
18mFast Inference for Probabilistic Graphical Models
01h 44mpomegranate: fast and flexible probabilistic modeling in python
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
35mHybrid Recommender Systems in Python