How Alibaba Inserts Marketing Cards to Boost Recommendation Revenue

Introduction

Alibaba's "For You Recommendation" module is evolving from a simple product recommendation channel to a multi‑role platform that can embed marketing scenario cards such as rankings, must‑buy lists, themed markets, and discovery sections. The goal is to increase overall slot exposure revenue while distributing traffic to these marketing scenes.

Problem and Method

Previously, cards were attached to recommended products at random probabilities, which reduced exposure revenue because card relevance and user preference were ignored.

Weak Personalization

A card quality score and user preference score are computed, and cards are selected using the formula shown in the original Figure 2 rather than random placement. The system filters the business‑provided product‑card mappings, scores multiple candidate cards per product, and applies offline‑computed user‑card preference scores. An interval control ensures a minimum number of products between cards, preventing card clustering.

Compared with the prior random approach, this method improves exposure revenue by 3.23%, though it still lags behind a baseline without cards.

Machine‑Learning Model

The task of choosing the best card for a product is framed as a CTR (click‑through‑rate) prediction problem. Samples are extracted from exposure‑click data of products eligible for card attachment. Features are divided into user, item (trigger product), and card attributes, totaling 85 features (62 numeric, 19 categorical, 4 cross features). Categorical features are embedded before model input.

Recall

From the final product recommendation list, a candidate set of product‑card pairs (item2item2card) is recalled. To capture richer relationships, an additional item2theme (item‑card) recall path is built using a SWING algorithm on multi‑day exposure‑click data, yielding a 0.79% increase in exposure revenue.

Ranking

Both Wide & Deep (WDL) and Deep & Cross Network (DCN) models were evaluated. Online A/B tests showed negligible difference, with DCN gaining a modest 0.03% in exposure revenue, so DCN was deployed globally. Models are trained daily with XTensorflow and served via RTP.

Results

The combined effect of precise distribution and improved card performance raises exposure revenue by 6.77% and average product clicks per user by 18.60% compared to the weak‑personalization baseline, and by 1.58% revenue and 0.01% clicks compared to pure product recommendation.

System Flow

The online scheduler determines the overall product order, while the card ranking model decides which cards attach to which products, respecting interval rules. Figure 4 illustrates the end‑to‑end flow, integrating both weak personalization and machine‑learning components.

Card Fallback and Cold Start

If a request lacks certain card types, a probabilistic fallback inserts at most one card per missing type, respecting interval constraints; otherwise no card is shown.

Future Work

Current models act as card selectors without jointly optimizing product and card ordering. Future plans include modeling internal card clicks, using product recommendation results as an additional recall path, and training a mixed ranking model to perform holistic ordering of products and cards.