Boosting Spring Festival Activity: Alibaba’s Full‑Link Intelligent Delivery Framework

Background: Spring Festival Activity Decline

During the Chinese New Year period, logistics delays and store closures cause a seasonal dip in e‑commerce activity, leading to a noticeable drop in Hand‑Taobao DAU. Predictive algorithms identify users likely to become inactive and apply targeted interventions—pre‑holiday personalized virtual rights and in‑festival push notifications—to maintain engagement.

Full‑Link Intelligent Delivery Framework

The framework, integrated with the Pagani operation platform, consists of two main modules: intent recognition to select target users and rights recommendation to personalize the benefits delivered. It enables end‑to‑end, data‑driven user activation.

Intent Recognition

Answers “whether to send” by selecting the right audience for rights distribution.

Uses a visit‑intent model to predict users who will not visit during the festival and intervenes via in‑app pages or push messages.

Rights Recommendation

Answers “what to send” by personalizing the type of rights each user receives.

Implements a “one‑size‑one‑face” approach for individualized rights allocation.

Algorithm Modules

3.1 Intent Recognition Model

Based on user profiles and historical behavior, the model predicts the probability of future actions such as visit, click, favorite, add‑to‑cart, or purchase. Multiple objectives (visit rate, retention, conversion, ROI) are balanced through multi‑intent training.

User purchase intent: AUC = 0.83, F1 = 0.76

User visit intent: AUC = 0.86, F1 = 0.78

User click intent: AUC = 0.76, F1 = 0.88

User add‑to‑cart intent: AUC = 0.80, F1 = 0.64

User favorite intent: AUC = 0.88, F1 = 0.58

3.2 Rights Recommendation

The rights recommendation module adds a rights entity to the feature space, requiring user‑rights cross features. It balances user preference, rights sensitivity, inventory, and platform subsidy cost to achieve a three‑party win.

Cold‑Start Strategy

Two techniques are used: (1) an ε‑Greedy algorithm that randomly explores a small ε proportion of traffic while exploiting prior knowledge for the rest; (2) a decision‑tree built on prior user profile and early behavior when rights have no exposure data.

CTR Model (XFTRL)

XFTRL, an extension of the FTRL optimizer built on Alibaba’s eXtreme Parameter Server, supports billions of features and continuous incremental training, providing high‑performance click‑through‑rate estimation for rights.

E&E Optimization (Thompson Sampling)

To address the long‑tail item problem, Thompson Sampling is employed for exploration‑exploitation, modeling each item’s reward probability with a Beta distribution and selecting items based on sampled values, offering robustness against delayed or batch feedback.

Experimental Results

Intervention bucket increased daily visit rate by 1.2%.

Proportion of users whose activity did not decline rose by 1.42%.

Rights personalization bucket improved redemption rate by 40% and redemption conversion by 100% compared with a random bucket.

Future Directions

Build a more complete data pipeline to close gaps in event tracking and data retrieval.

Introduce a supply‑demand balanced allocation mechanism using game‑theoretic design for multi‑rights mixing.

Develop lifecycle‑aware algorithmic intervention strategies to continuously refine user growth rights delivery.