How AI Powers Heterogeneous Content Ranking in E‑Commerce Search

Research Background

Search engines play a crucial role in e‑commerce by guiding users toward potential purchases. Traditional product search returns a list of items, but the rise of user‑generated content such as articles, reviews, and videos creates a new "content stream" that can aid shopping decisions. Ranking heterogeneous data—both product listings and content streams—poses challenges, requiring cross‑domain knowledge and support for multimedia ordering.

Proposed Algorithms

The paper introduces two algorithms for content‑stream type ranking:

Independent Multi‑Armed Bandit (iMAB) : Computes a ratio θ from IPV and PV using a Beta prior for each content type (post, list, video). The posterior is updated in real time.

Personalized Markov Deep Neural Network (pMDNN) : Learns low‑dimensional embeddings for users and queries via a graph‑based node2vec model, then predicts the next slot type using a Markov‑style DNN that incorporates the previous slot’s embedding.

The pMDNN model consists of three fully‑connected layers with ReLU activations and a Softmax output, trained with cross‑entropy loss. For the first slot, cross‑domain knowledge is used by mapping recent product browsing behavior into content‑slot features.

Experiment Results

The models were deployed in an A/B testing environment and evaluated on five key metrics: PV (impressions), PV‑click, UV (unique users), UV‑click, and UV‑CTR. The pMDNN consistently outperformed iMAB, especially in UV‑click and UV‑CTR, indicating higher user engagement and relevance of the recommended content streams.

References

Róbert Busa‑Fekete and Eyke Hü​llermeier. A survey of preference‑based online learning with bandit algorithms.

Ali Mamdouh Elkahky, Yang Song, and Xiaodong He. 2015. A multi‑view deep learning approach for cross‑domain user modeling in recommendation systems. In Proceedings of the 24th International Conference on World Wide Web, 278–288.