Mutual Recommendation Embeddings for Low-Latency Cold-Start Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing matching service systems face challenges in making efficient session-based mutual recommendations due to high computational costs and complex user behavior modeling, particularly for cold-start users, which can lead to increased latency and negative user experiences.
Innovation Solution
A method that updates user and session feature embeddings asynchronously, using a two-step learning process to efficiently model user behavior and reduce latency, involving a user feature embedding layer and a session feature embedding layer, with a conversation time prediction layer trained using a loss function.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If session-based recommendation is used to model user behavior, then recommendation accuracy for cold-start users is improved, but computational latency increases
Solution Approach 1:
The patent divides the recommendation system into two independent modules: a session-based recommendation module that captures short-term user intent, and a mutual recommendation module that leverages long-term user profiles and interaction history. This segmentation allows each module to operate with optimized computational resources, reducing overall latency while maintaining recommendation accuracy for cold-start users.
Solution Approach 2:
The patent pre-computes and stores user profiles, item embeddings, and interaction histories in advance. When a recommendation request arrives, the system quickly retrieves pre-processed data rather than computing everything from scratch. This preliminary action significantly reduces real-time computational latency while preserving the ability to provide accurate session-based recommendations.
2Measurement precision
If dynamic user behavior modeling is implemented, then mutual recommendation accuracy is improved, but system complexity increases
Solution Approach 1:
The patent separates dynamic user behavior modeling into distinct components: session-based intent modeling for short-term behavior, and profile-based modeling for long-term preferences. Each component uses simplified algorithms appropriate to its timescale, reducing overall system complexity while maintaining high mutual recommendation accuracy through their combination.
Solution Approach 2:
The patent introduces an intermediary layer that bridges session-based recommendations and mutual recommendations. This intermediary processes and harmonizes the outputs of both modules, managing the complexity of integrating dynamic user behavior data without requiring complete system redesign. The intermediary layer simplifies the integration process while preserving recommendation accuracy.
3Loss of time
If traditional recommendation filtering is used, then latency is reduced, but recommendation quality for cold-start users deteriorates
Solution Approach 1:
The patent applies different recommendation strategies to different user segments: traditional filtering methods are used for users with extensive history where they perform well and maintain low latency, while session-based methods are applied to cold-start users who need the flexibility of short-term behavior modeling. This segmentation allows the system to maintain low overall latency while improving recommendation quality for cold-start users.
Solution Approach 2:
The patent implements a hybrid approach where session-based recommendation features are selectively applied only when beneficial (e.g., for cold-start users or when recent behavior indicates strong intent), rather than universally. This partial application of the more computationally intensive session-based method maintains acceptable latency while improving recommendation quality where it matters most.
Data Source
Figure 1a
Figure 1b
Figure 2
AI summary
Disclosed are an apparatus that mutually recommends users and a method for same. The method of mutually recommending users of a matching service according to an embodiment comprises: a step wherein a user feature embedding of a first user is updated based on user information of the first user; a step wherein the session feature embedding of the first user is asynchronously updated with the user feature embedding of the first user, based on the session information of the first user; a step wherein a match request is received from the first user; a step wherein a final embedding of the first user is generated based on the user feature embedding of the first user and the session feature embedding of the first user; and a step wherein a second user to be matched with the first user is decided on from among the plurality of users, based on the final embedding of the first user and the respective final embeddings of the plurality of users included in the matching pool with the first user.