Building a Personalized Live‑Streaming Recommendation System: From Basics to Advanced Models at Huajiao Live

The rapid growth of mobile‑internet entertainment has raised user expectations for live‑streaming and short‑video quality, making personalized recommendation a key growth factor; this talk shares how Huajiao Live built its recommendation system from zero to production.

Recommendation System Overview

Recommendation systems help users discover relevant content by modeling interests from historical behavior; early Huajiao recommendations relied on simple popularity metrics, which favored head streamers and ignored personalization.

Modern pipelines consist of a recall stage (using low‑cost models such as collaborative filtering to retrieve thousands of candidates) followed by a feature‑based ranking stage (using richer features and more complex models to select the final few items).

Recall Algorithms

1. Domain‑based collaborative filtering – item‑based similarity is computed for streamers because the number of streamers is far smaller than users, leading to low computational cost and good interpretability.

2. Latent‑vector collaborative filtering – two main approaches: matrix factorization (MF) and deep‑learning‑based embedding generation. MF decomposes a rating matrix into user and item latent vectors; implicit feedback (e.g., watch time) is often used in live‑streaming scenarios.

Matrix factorization is simple, fast online (dot‑product of user/item vectors) and storage‑efficient, but suffers from limited expressiveness and lack of side‑information.

To overcome these limits, neural collaborative filtering (NCF) replaces the inner product with a deep neural network that learns the interaction function from data.

NeuMF combines a generalized matrix factorization (GMF) layer with a multilayer perceptron (MLP) to capture both linear and non‑linear feature interactions.

More recent models such as DeepFM, DIN, and various two‑tower architectures further enrich the feature space by incorporating user/item attributes, multi‑modal signals, and attention mechanisms.

Ranking Algorithms

Feature engineering is critical; typical live‑streaming features include user demographics, historical interaction counts, real‑time metrics (viewers, gifts, comments), and contextual signals (whether the streamer is singing, dancing, etc.).

Early ranking models used linear classifiers (LR) and factorization machines (FM). FM adds automatic second‑order feature crossing, while GBDT + LR pipelines generate high‑order crossed features via tree leaves.

Deep models (DNN, Wide & Deep, DeepFM) now handle high‑order interactions directly, often combined with attention mechanisms (e.g., DIN) for better personalization.

Multi‑Objective Optimization

Live‑streaming platforms optimize several metrics simultaneously (click‑through, watch time, gifts, comments, follows, shares). Multi‑task models share embeddings across tasks while allowing task‑specific gates (e.g., MMoE) to balance objectives.

Examples include ESMM (modeling click → conversion) and advanced CVR models that address sample selection bias and data sparsity.

System Deployment

Training data are stored in HDFS, with daily or weekly batch updates; incremental samples are processed via Flink streams. Offline training produces embedding matrices that are served with TensorFlow Serving, while the online inference service is implemented in Go to transform user requests into feature vectors.

Conclusion

The Huajiao Live recommendation journey illustrates that there is no universally best model; the optimal solution depends on the specific characteristics of the live‑streaming scenario, such as multimodal content, real‑time dynamics, and hotspot effects. Understanding the domain, extracting representative features, and selecting appropriate algorithms are essential for achieving the best performance.