Exploration of Alibaba's Feizhu Recommendation Algorithms and Full‑Space CVR Estimation Models (ESMM, ESM², HM³)

Guest: Wen Hong, Senior Algorithm Expert at Alibaba

Editor: Sun Bin, Taiyuan University of Science and Technology

Platform: DataFunSummit

Introduction: The theme is Alibaba Feizhu recommendation algorithm exploration. First, we introduce mainstream e‑commerce recommendation technologies, such as the evolution of full‑space CVR estimation (ESMM / ESM² / HM³). Then we focus on travel‑industry specifics and discuss the current status and future directions of Feizhu recommendation algorithms.

01 – E‑commerce Recommendation Algorithm Overview

Recommendation technology in e‑commerce can be divided into four parts:

① Basic Capabilities : data, samples, features, and the machine‑learning platform.

② Algorithm Models : recall, coarse ranking, fine ranking, and re‑ranking.

③ Online Services : unified TPP recommendation solution, with components such as ABFS (feature processing), BE (recall), IGraph (storage), etc.

④ Business Scenarios : "Guess you like" on Taobao, shop, merchant private domain, browsing, subscription, post‑purchase, etc.

2. Full‑Space CVR Estimation – ESMM Model Review

Traditional CVR faces three challenges: Sample Selection Bias (SSB), Data Sparsity (DS), and lack of positive purchase samples.

ESMM addresses these by introducing two auxiliary tasks (CTR and CTCVR) that are defined over the full space, sharing embeddings between them, and thus mitigating SSB and DS.

3. Full‑Space CVR Estimation – ESM² Model

ESM² improves upon ESMM by incorporating post‑click behaviors (e.g., add‑to‑cart, favorite) to alleviate the scarcity of purchase samples and to model more complex user decision paths.

Key challenges include defining suitable post‑click actions, abstracting complex purchase decisions, organizing multiple post‑click behaviors, and modeling dependencies between post‑click actions and purchase.

The modeling idea splits the click‑to‑purchase path into two disjoint branches (DAction and OAction) and uses three full‑space auxiliary tasks (exposure→click, exposure→DAction, exposure→purchase) to estimate CVR via the law of total probability.

4. Full‑Space CVR Estimation – HM³ Model

HM³ adds a micro‑behavior layer before the macro behavior layer, further refining the modeling of user actions such as clicking, adding to cart, or collecting, and integrates them into the CVR estimation.

02 – Feizhu Recommendation Exploration

Travel Recommendation Background

The user lifecycle in travel is divided into four stages: demand stimulation, pre‑travel, in‑travel, and post‑travel.

During pre‑travel, users search for destinations, hotels, and transportation; during in‑travel, they look for local attractions and activities; post‑travel involves feedback and potential new travel cycles.

Feizhu Application in Travel Cycle

When users open the Feizhu app, explicit needs (e.g., booking tickets) trigger direct clicks on specific modules, while vague needs lead to scrolling and exposure to the infinite‑feed "Guess you like" module.

Characteristics of Travel Recommendation

1) Low frequency – travel demand is sparse, causing cold‑start issues.

2) Spatio‑temporal attributes – recommendations must consider current trends and seasonal factors.

3) Periodicity – recurring patterns such as holidays influence user behavior.

Travel‑Specific Algorithm Techniques

The overall architecture mirrors e‑commerce but adds travel‑specific user understanding, recall that accounts for spatio‑temporal and lifecycle signals, and multi‑scenario ranking.

RTUS (User Center Service) aggregates full‑link logs (browse, click, search, add‑to‑cart, favorite, purchase), performs statistical analysis, and feeds user intent and preference predictions into downstream models.

User‑Journey‑Aware Recall

Features from the fused user state and target item are concatenated and passed through MLPs to obtain a recall score, showing significant offline and online gains.

Period‑Aware Sequence Modeling

Sequences are organized horizontally (yearly slices) and vertically (seasonal slices) to capture long‑term interests and periodic habits, improving fine‑ranking performance.

03 – Summary & Outlook

Future work will continue to integrate travel‑specific characteristics such as user travel cycles, spatio‑temporal attributes, and industry traits into recommendation research.

Q&A Highlights

• Post‑click actions (e.g., add‑to‑cart) have higher occurrence rates than direct purchase and are strongly correlated with purchase, helping CVR modeling.

• Adding a new post‑click action does not require extra towers; actions are split into DAction (major) and OAction (the complement).

• Labels for CTR, CTCVR are defined over the full exposure space; CVR is inferred from auxiliary tasks.

• The ESM² model is deployed in product detail page recommendation and can be adapted to other scenarios such as travel.

• Travel lifecycle stages are manually defined based on behavior analysis.

• Periodic events like Double‑11 are modeled by aligning yearly occurrences vertically.

• The presentation concludes with thanks and a call for likes, shares, and follows.