Abstract
E-commerce app users exhibit behaviors that are inherently logically consistent. A series of multi-scenario user behaviors interconnect to form the scene-level all-domain user moveline, which ultimately reveals the user's true intention. Traditional CTR prediction methods typically focus on the item-level interaction between the target item and the historically interacted items. However, the scene-level interaction between the target item and the user moveline remains underexplored. There are two challenges when modeling the interaction with preceding all-domain user moveline: (i) Heterogeneity between items and scenes: Unlike traditional user behavior sequences that utilize items as carriers, the user moveline utilizes scenes as carriers. The heterogeneity between items and scenes complicates the process of aligning interactions within a unified representation space. (ii) Temporal misalignment of linked scene-level and item-level behaviors: In the preceding user moveline with a fixed sampling length, certain critical scene-level behaviors are closely linked to subsequent item-level behaviors. However, it is impossible to establish a complete temporal alignment that clearly identifies which specific scene-level behaviors correspond to which item-level behaviors. To address these challenges and pioneer modeling user intent from the perspective of the all-domain moveline, we propose All-domain Moveline Evolution Network (AMEN). AMEN not only transfers interactions between items and scenes to homogeneous representation spaces, but also introduces a Temporal Sequential Pairwise (TSP) mechanism to understand the nuanced associations between scene-level and item-level behaviors, ensuring that the all-domain user moveline differentially influences CTR predictions for user's favored and unfavored items. Online A/B testing demonstrates that our method achieves a +11.6% increase in CTCVR.
Introduction
In the rapidly evolving landscape of e-commerce, understanding user intent is paramount for effective Click-Through Rate (CTR) prediction. Traditionally, models have focused on item-level interaction sequences—tracking which products a user has clicked or purchased. However, these "item-centric" approaches often overlook the broader context of a user's journey within an application. A user's path, or "moveline," encompasses a variety of scene-level behaviors such as browsing the homepage, searching for specific categories, visiting brand stores, or hunting for coupons. These multi-scenario behaviors are logically consistent and reveal deep-seated intentions that item-level data alone might miss.
The Challenges of Scene-Level Modeling
Modeling the all-domain user moveline presents two significant technical hurdles:
- Heterogeneity between Items and Scenes: Items (e.g., a specific "Blue Cup") and scenes (e.g., the "Coupon Center") exist in different data modalities. Aligning these into a unified representation space for interaction modeling is non-trivial.
- Temporal Misalignment: User behaviors at the scene level (like entering a marketing channel) do not always align perfectly in time with specific item-level clicks. This makes it difficult to determine which scene-level action influenced a particular click within a fixed-length sampling window.
The AMEN Architecture
To overcome these challenges, we propose the All-domain Moveline Evolution Network (AMEN). AMEN shifts the focus from purely item-based history to a dual-pathway architecture that integrates scene-level intent.
The architecture employs two primary pathways:
- Item-to-Scene Pathway (AISeq): This pathway constructs an "Active Intention Sequence" by extracting items exposed within specific scenes. It uses Multi-head Target Attention (MHTA) to model the relationship between the target item and these scene-specific interactions.
- Scene-to-Item Pathway (Prompting): This pathway utilizes a prompting methodology to bridge the heterogeneity gap. By transforming the target item and its context into a "virtual scene-level prompt," the model can query the user’s moveline directly in a homogeneous representation space.
Temporal Sequential Pairwise (TSP) Mechanism
A key innovation in AMEN is the Temporal Sequential Pairwise (TSP) mechanism. Rather than treating CTR prediction as a series of isolated points, TSP introduces contrastive learning. It compares the "moveline reward" of a target item against a "diff item"—an item with a different feedback label sampled from a different point in the user's timeline. This ensures the model learns how shifts in the moveline differentially impact user preferences, effectively calibrating the final prediction based on real-time intent.
Experimental Evaluation
AMEN was rigorously tested on a massive industrial dataset from Taobao, comprising 6.7 billion training instances from over 180 million users.
As shown in the table above, AMEN significantly outperforms established models like Wide&Deep, DIN, and SIM in terms of both AUC and GAUC. Specifically, it achieved an offline AUC of 0.7694, surpassing the strong internal baseline DBA (0.7621).
Online A/B testing further validated the model's efficacy:
- Feeds Scenario: A +11.6% increase in post-view click-through conversion rate (CTCVR).
- Floors Scenario: A +20.6% increase in CTR and a +20.1% increase in Gross Merchandise Volume (GMV).
Conclusion
By pioneering the modeling of user intent through the all-domain moveline, AMEN provides a more holistic view of user behavior in complex e-commerce environments. The combination of dual-pathway alignment and the TSP mechanism allows for a more nuanced understanding of how scene-level context drives item-level decisions, leading to substantial improvements in both recommendation accuracy and business metrics.
Summary updated