E-CARE: An Efficient LLM-based Commonsense-Augmented Framework for E-Commerce

TL;DR

E-CARE tackles the cost and latency of leveraging LLM-based commonsense in e-commerce retrieval by offline constructing a reasoning factor graph from historical query–product interactions and training adapters to map inputs to this graph. During inference, a single LLM forward pass embeds the query, and the precomputed graph guides the retrieval process, enabling efficient, commonsense-aware decision making. The approach yields consistent improvements across search relevance and app recall without supervised fine-tuning, demonstrating practical scalability for large catalogs. By reducing graph redundancy through clustering and edge filtering, E-CARE maintains rich reasoning capabilities while keeping inference fast and cost-effective.

Abstract

Finding relevant products given a user query plays a pivotal role in an e-commerce platform, as it can spark shopping behaviors and result in revenue gains. The challenge lies in accurately predicting the correlation between queries and products. Recently, mining the cross-features between queries and products based on the commonsense reasoning capacity of Large Language Models (LLMs) has shown promising performance. However, such methods suffer from high costs due to intensive real-time LLM inference during serving, as well as human annotations and potential Supervised Fine Tuning (SFT). To boost efficiency while leveraging the commonsense reasoning capacity of LLMs for various e-commerce tasks, we propose the Efficient Commonsense-Augmented Recommendation Enhancer (E-CARE). During inference, models augmented with E-CARE can access commonsense reasoning with only a single LLM forward pass per query by utilizing a commonsense reasoning factor graph that encodes most of the reasoning schema from powerful LLMs. The experiments on 2 downstream tasks show an improvement of up to 12.1% on precision@5.

Figures (4)

• Figure 1: Diagram of reasoning factor graph generation procedure of E-CARE pipeline with 2 input query-product pairs as examples. Firstly, Large Language Models (LLMs) are prompted to perform commonsense reasoning over query-product pairs, producing product features as well as commonsense reasoning that connect queries to products through intermediate need and utility factors. The factors are then organized into a graph structure, namely, the reasoning factor graph. Secondly, a node clustering procedure is applied to merge nodes with similar semantic meaning with respect to different types of factors. Finally, LLM-based edge filtering is employed to filter out unreliable edges from the reasoning factor graph, further improving the quality.
• Figure 2: Overview diagram of two downstream applications—Search Relevance (SR) and App Recall (AR)—integrated with the reasoning factor graph. (a) Offline Training: Historical query-product interaction pairs are processed through the E-CARE pipeline (described in §\ref{['sec3']}) to construct the reasoning factor graph and train corresponding adapters. (b) Online Inference (SR): The trained adapters augment queries and products with reasoning factors before being fed into the downstream model to predict search relevance scores.(c) Offline Inference (AR): In the app recall scenario, trained product adapters are applied offline to connect new products to relevant factors, thereby reducing computation during online inference. (d) Online Inference (AR): For each input query, the predicted factors serve as key factors, and products associated with the most overlapping key factors are returned as recall results.
• Figure 3: Statistics of the reasoning factor graph for the ESCI dataset along the E-CARE pipeline. Both the graph size and the in-group similarity of factors connected to each query or product node decrease as the pipeline progresses, indicating that node clustering and edge filtering reduce the redundancy of factors within groups, potentially facilitating downstream tasks.
• Figure 4: Similarity evaluation of product adapter results regarding 4 different factor types.