Automated Information Flow Selection for Multi-scenario Multi-task Recommendation

Multi-scenario multi-task recommendation (MSMTR) systems must address recommendation demands across diverse scenarios while simultaneously optimizing multiple objectives, such as click-through rate and conversion rate. Existing MSMTR models typically consist of four information units: scenario-shared, scenario-specific, task-shared, and task-specific networks. These units interact to generate four types of relationship information flows, directed from scenario-shared or scenario-specific networks to task-shared or task-specific networks. However, these models face two main limitations: 1) They often rely on complex architectures, such as mixture-of-experts (MoE) networks, which increase the complexity of information fusion, model size, and training cost. 2) They extract all available information flows without filtering out irrelevant or even harmful content, introducing potential noise. Regarding these challenges, we propose a lightweight Automated Information Flow Selection (AutoIFS) framework for MSMTR. To tackle the first issue, AutoIFS incorporates low-rank adaptation (LoRA) to decouple the four information units, enabling more flexible and efficient information fusion with minimal parameter overhead. To address the second issue, AutoIFS introduces an information flow selection network that automatically filters out invalid scenario-task information flows based on model performance feedback. It employs a simple yet effective pruning function to eliminate useless information flows, thereby enhancing the impact of key relationships and improving model performance. Finally, we evaluate AutoIFS and confirm its effectiveness through extensive experiments on two public benchmark datasets and an online A/B test.

💡 Research Summary

The paper “Automated Information Flow Selection for Multi-scenario Multi-task Recommendation” addresses key challenges in modern industrial recommendation systems, which must operate across multiple business scenarios (e.g., homepage, product detail page) while optimizing for multiple objectives (e.g., click-through rate, conversion rate) simultaneously. Existing Multi-scenario Multi-task Recommendation (MSMTR) models typically deconstruct the problem into four informational components: scenario-shared, scenario-specific, task-shared, and task-specific networks. The interactions between these components generate four distinct types of information flows. However, current approaches suffer from two major drawbacks: 1) They often rely on complex, parameter-heavy architectures like Mixture-of-Experts (MoE), leading to high model complexity and training costs. 2) They indiscriminately fuse all possible information flows, which can introduce redundant or even harmful noise, as not all flows are equally relevant for every scenario-task combination.

To overcome these limitations, the authors propose a novel lightweight framework named Automated Information Flow Selection (AutoIFS). The framework is built on two core innovations. First, to achieve efficient and flexible modeling, AutoIFS incorporates Low-Rank Adaptation (LoRA). Instead of using separate bulky networks for each informational component, AutoIFS maintains a base shared network and attaches small, trainable low-rank matrices (LoRA adapters) for each specific scenario and task. This design elegantly decouples the four information units with minimal parameter overhead, allowing for simple linear combinations to represent the various information flows.

Second, and most crucially, AutoIFS introduces an Information Flow Selection Network. This network dynamically learns to select or prune the four information flows on a per-instance basis. It takes the input instance features and the current model’s performance feedback as input and outputs a set of binary masks (one for each flow type). A pruning function then eliminates the flows masked as zero. This mechanism allows the model to automatically filter out irrelevant or detrimental information flows, thereby amplifying the influence of the most critical relationships for a given prediction. The selection network is trained end-to-end alongside the main recommendation model, guided by the overall multi-task loss function.

The proposed AutoIFS framework was extensively evaluated on two public benchmark datasets and via an online A/B test in a real-world financial technology platform. Experimental results demonstrate that AutoIFS consistently outperforms state-of-the-art MSMTR baselines in terms of prediction accuracy (measured by GAUC). Furthermore, it achieves this superior performance with significantly fewer parameters and lower computational cost, validating its efficiency and practicality. The work highlights a paradigm shift from exhaustively modeling all complex interactions to adaptively selecting the most salient ones, offering a powerful and efficient solution for industrial-scale multi-scenario multi-task recommendation.