How Interactive Recommendation Boosts User Engagement: The “Wind Vane” Framework Explained

Introduction

Recommendation systems are widely used in industry, but they often suffer from two problems: (1) a single passive experience for users, and (2) limited understanding of user preferences. Enabling effective interaction between users and the system can alleviate these issues by increasing user participation and providing richer feedback for the model.

Background

Unlike passive TV watching, children prefer interactive mobile devices, which offer higher engagement. Similarly, mainstream recommendation systems lack sufficient interactivity. Introducing interaction can both improve user experience and help the system learn better preferences.

Interactive Recommendation Framework

Based on the KDD 2018 paper Q&R: A Two‑Stage Approach toward Interactive Recommendation , the framework consists of three stages: question generation, user feedback, and item recommendation.

Question Generation Model

Generating full natural‑language questions is still difficult, so the task is transformed into a keyword recommendation problem. Keywords such as “scarf”, “coat”, “hat” are displayed as cards; clicking a keyword implies the corresponding question (e.g., “Do you want a scarf?”). Candidate keywords are sourced from search logs because they are meaningful and abundant.

User Feedback

When a user clicks a keyword, the system treats it as a positive response. Additional signals (e.g., clicks on other keywords, add‑to‑cart, collection, dwell time) can be incorporated to enrich feedback.

Item Recommendation Model

Clicked keywords are used as queries for a standard search algorithm to retrieve items. This “recommend‑then‑search” process combines the low‑cost of recommendation with the precision of search.

Product Form – “Wind Vane”

The product, named “Wind Vane”, follows three principles: interactive, demand‑focused with strong explainability, and scenario‑aware. Users click an item on the homepage, interact on the detail page, and the system may display a “Wind Vane” card with contextual copy and a set of keywords for further exploration.

Technical Solution

The overall architecture includes data generation from search logs, product information, and knowledge graphs; a recall stage; a ranking stage; a display‑control module; and endpoint intelligence. The pipeline is illustrated in the diagram below.

Recall

Recall is modeled as a meta‑path problem on a heterogeneous graph of users, items, keywords, scenes, and categories. Three meta‑path strategies are used:

u2i2q – reverse the query‑to‑item relation to link clicked items back to queries.

u2i2scene2q – map clicked items to scenes, then scenes to keywords.

u2i2c2q – incorporate category information from a knowledge graph.

Future work includes expanding beyond keywords to videos, guides, and other modalities.

Ranking

Ranking consists of a coarse “xftrl” stage and a fine “Attention_GRU” stage.

xftrl : An engineering‑friendly implementation of the FTRL algorithm, handling billions of sparse features. Offline experiments show an AUC of 0.67.

Attention_GRU : Addresses the limitations of xftrl by modeling sequential user behavior and high‑order feature interactions. Experiments demonstrate progressive AUC improvements from 0.5685 (category only) to 0.6830 (full feature set).

Further enhancements add time decay and behavior‑type weighting to the attention mechanism, raising AUC to 0.6999 and speeding up training by 40%.

Display Control

Two aspects are controlled: (1) position, timing, and intent to avoid bad cases, and (2) scenario and industry interventions to generate more contextual keywords and copy.

Endpoint Intelligence

Rich signals from the detail page (add‑to‑cart, collection, dwell time, swipe trajectories) are incorporated, significantly boosting keyword recommendation performance. Future directions include modeling the full behavior sequence with Attention_GRU and moving parts of the pipeline to the client side for real‑time personalization.

Double‑11 Results

During Alibaba’s 2018 Double‑11 event, “Wind Vane” exceeded expectations. Different time slots employed distinct strategies: early‑stage focused queries, mid‑stage scenario‑driven queries, and late‑stage convergence queries.

Conclusion and Outlook

Interactive recommendation shows great promise. The “Wind Vane” system achieved notable gains, yet many improvements remain: richer detail‑page signals, more sophisticated scene‑based query recall using knowledge graphs, enhanced feature sets and loss functions that consider second‑hop metrics, and client‑side model deployment.

References

Christakopoulou et al., “Q&R: A Two‑Stage Approach toward Interactive Recommendation,” KDD 2018.

Zhao et al., “Meta‑graph based recommendation fusion over heterogeneous information networks,” KDD 2017.

McMahan et al., “Ad click prediction: a view from the trenches,” KDD 2013.

Zhu et al., “What to do next: Modeling user behaviors by time‑LSTM,” IJCAI 2017.

Zhu et al., “A Brand‑level Ranking System with the Customized Attention_GRU Model,” IJCAI 2018.

Chorowski et al., “Attention‑based models for speech recognition,” NIPS 2015.

Mnih et al., “Recurrent models of visual attention,” NIPS 2014.