TLDR: The research paper introduces Personalized Landing Page Modeling (PLPM), a new task in recommender systems focused on selecting the most suitable landing page for users upon app entry. To address this, the authors propose KLAN (Kuaishou Landing-page Adaptive Navigator), a hierarchical framework. KLAN comprises three modules: KLAN-ISP for inter-day static page preferences, KLAN-IIT for intra-day dynamic interest transitions using reinforcement learning, and KLAN-AM for adaptively combining these two signals. Online experiments on the Kuaishou platform show KLAN significantly improves user engagement, retention, and reduces page drop-off, demonstrating its effectiveness and scalability.
Modern online platforms, like popular social media and e-commerce apps, often feature a variety of pages designed to cater to diverse user needs. Think of a video app with a ‘Featured’ page, an ‘Explore’ page, or a ‘Following’ page. This multi-page structure creates a two-step interaction: first, users navigate to a specific page, and then they engage with content within that page. While much research has focused on improving the content users see once they are on a page, there has been less attention paid to the initial step: how to best guide users to the most suitable landing page when they first open the app.
To address this gap, researchers from Kuaishou Technology have formally defined a new task called Personalized Landing Page Modeling (PLPM). The goal of PLPM is to proactively select the most appropriate landing page for a user upon app entry. This selection aims to optimize both short-term metrics, like reducing the chance of a user immediately leaving the page (Page Drop-off Ratio), and long-term metrics, such as overall user engagement and satisfaction.
The challenge in PLPM lies in understanding complex user behavior. Users have stable, long-term preferences for certain pages (inter-day static preferences), but their interests can also change rapidly within a single day based on immediate context (intra-day dynamic interests). For instance, a user might typically prefer the ‘Featured’ page, but if they receive a notification about a live stream from a creator they follow, their immediate interest might shift to the ‘Following’ page.
To tackle these complexities, the Kuaishou team proposes a hierarchical solution framework called KLAN (Kuaishou Landing-page Adaptive Navigator). KLAN is designed with three key components working together:
KLAN-ISP: Capturing Long-Term Preferences
This module focuses on understanding a user’s stable, long-term preferences for different pages. It uses a technique called uplift modeling, which helps estimate how assigning a user to a specific page will influence their subsequent behavior. KLAN-ISP is trained on daily user data and is designed to handle issues like exposure bias (where some pages are seen more often than others) and the inherent differences between various page types.
KLAN-IIT: Modeling Dynamic Interests
This component is all about capturing a user’s rapidly changing interests within a single day. It leverages reinforcement learning, a type of AI that learns by trial and error to make optimal decisions. KLAN-IIT is trained on hourly data and is particularly adept at adapting to real-time contextual triggers. The researchers also introduced a ‘Dynamic Conservative Coefficient’ within this module to ensure stable learning, especially when user activity fluctuates throughout the day, a phenomenon they refer to as the ‘Tidal Phenomenon’.
Also Read:
- Large Language Models Transform Recommender Systems: A Deep Dive into New Capabilities and Remaining Hurdles
- Boosting Personalized Learning: A New Framework for Smarter Question Recommendations
KLAN-AM: Adaptive Integration
The final piece of KLAN is the Adaptive Modulation module. This component acts as a smart balancer, dynamically weighing the importance of a user’s long-term static preferences (from KLAN-ISP) and their short-term dynamic interests (from KLAN-IIT). It uses real-time streaming data to predict a user’s immediate behavioral tendency, ensuring that the final landing page decision is always the most relevant for the current moment.
The effectiveness of KLAN was rigorously tested through extensive online experiments on the Kuaishou platform, which serves hundreds of millions of users daily. The results were highly positive, showing significant improvements in key metrics. For example, KLAN led to a +0.205% increase in Daily Active Users (DAU) and a +0.192% improvement in user Lifetime (LT), a crucial indicator of long-term retention. Other metrics like app usage time, watch time, and video views also saw substantial growth, while the Page Drop-off Ratio decreased significantly by -28.158%, indicating that users were more satisfied with their assigned landing pages. The system also maintained real-time efficiency with negligible latency.
A notable finding from the experiments was KLAN’s ability to encourage users to explore more pages. The proportion of single-page users decreased, while multi-page users increased. This is significant because multi-page users tend to have higher DAU and LT, contributing to sustained platform growth. The full KLAN model consistently outperformed its individual components, validating the hierarchical design and the adaptive fusion of static and dynamic signals.
This work marks a significant step in personalized app navigation, filling a critical research gap. The KLAN framework is now fully deployed on the Kuaishou platform, demonstrating its real-world impact and scalability. The researchers also plan to release their dataset and code to encourage further innovation in this important area. You can find more details in the research paper: KLAN: Kuaishou Landing-page Adaptive Navigator.
