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

The talk, presented by senior algorithm expert Wen Hong from Alibaba, introduces the background of e‑commerce recommendation technologies and then focuses on the specific characteristics and current status of the Feizhu travel recommendation algorithm.

1. E‑commerce Recommendation Overview

The recommendation pipeline consists of four major parts: basic capabilities (data, samples, features, ML platform), algorithm models (recall, coarse ranking, fine ranking, re‑ranking), online services (feature processing ABFS, recall engine BE, storage IGraph, etc.) and business scenarios (Taobao "You May Like", shop, private domain, browsing, subscription, post‑purchase, etc.).

2. Full‑Space CVR Estimation – ESMM Model

Traditional CVR suffers from Sample Selection Bias, data sparsity, and lack of purchase positives. ESMM introduces two auxiliary tasks—CTR and CTCVR—trained on the full exposure space, sharing embeddings to alleviate sparsity and bias. The model predicts CTR (exposure→click) and CTCVR (exposure→click→purchase) and derives CVR as CTCVR/CTR.

3. Limitations of ESMM

Only models the linear path exposure→click→purchase.

Ignores complex post‑click behaviors (e.g., add‑to‑cart, wishlist).

Does not address the scarcity of purchase positives.

4. ESM² Model

ESM² adds a post‑click stage to capture richer user actions, defining DAction (e.g., add‑to‑cart/collect) and its complement OAction. Three auxiliary tasks are introduced: exposure→click (CTR), exposure→DAction, and exposure→OAction. The final purchase probability is modeled as:

pPurchase = y1 * [y2 * y3 + (1‑y2) * y4]

where y1 is CTR, y2 is the probability of DAction, y3 and y4 are the probabilities of purchase given DAction and OAction respectively. This formulation resolves Sample Selection Bias, data sparsity, and the lack of purchase positives.

5. Challenges of ESM²

Defining appropriate post‑click actions.

Abstracting complex purchase decision processes.

Organizing multiple post‑click actions efficiently.

Modeling dependencies between post‑click actions and final purchase.

6. HM³ Model

HM³ further refines the hierarchy by inserting a micro‑behavior layer before macro actions, allowing the model to capture fine‑grained interactions (e.g., button clicks inside a product page) and then aggregate them into macro behaviors for purchase prediction.

7. Travel Recommendation – Feizhu

The user lifecycle in travel is divided into four stages: demand stimulation, pre‑travel, in‑travel, and post‑travel. Each stage exhibits distinct user intents and behaviors, which the recommendation system must adapt to.

Key characteristics of travel recommendation are low frequency, strong spatio‑temporal attributes, and periodicity (e.g., holidays). The architecture mirrors the e‑commerce pipeline but incorporates travel‑specific user understanding, time‑aware recall, and multi‑scenario ranking.

8. RTUS (User Center Service)

RTUS aggregates full‑link logs (browse, click, search, add‑to‑cart, collect, purchase) and provides real‑time user state, preferences, and travel intent, which are fed into downstream CTR/CVR models for better personalization.

9. User‑Journey‑Aware Recall

The recall module incorporates the user's current travel stage (pre‑travel, in‑travel, post‑travel) into the user profile, extracts stage‑aware sequence features, concatenates them with item features, and computes recall scores via a dot‑product. Offline and online experiments show significant improvements over generic e‑commerce recall.

10. Periodicity‑Aware Sequence Modeling

Users' historical behaviors are organized horizontally (year‑by‑year) and vertically (seasonal slices such as Spring Festival). This structure enables the model to capture both long‑term interest evolution and periodic habits, improving fine‑ranking performance.

11. Summary & Outlook

The presented models (ESMM, ESM², HM³) and travel‑specific enhancements demonstrate the importance of full‑space auxiliary tasks, post‑click behavior modeling, and spatio‑temporal awareness. Future work will further integrate travel cycle, location attributes, and industry‑specific signals into recommendation research.

Q&A

The session concludes with a series of questions about post‑click behavior frequencies, model scalability, label definition for CTR/CVR/CTCVR, handling of periodic events, and practical deployment scenarios.