LLMSARec: large language model with semantic alignment for Web service recommendation

• Date: 18 November 2024
• Pages: 37-53
• DOI: https://doi.org/10.1108/IJWIS-09-2024-0262
• Published date: 18 November 2024
• Subject Matter: Information & knowledge management,Information & communications technology,Information systems,Library & information science,Information behaviour & retrieval,Metadata,Internet
• Author: Shangjie Feng,Buqing Cao,Ziming Xie,Zhongxiang Fu,Zhenlian Peng,Guosheng Kang

LLMSARec: large language model

with semantic alignment for Web

service recommendation

Shangjie Feng,Buqing Cao,Ziming Xie,Zhongxiang Fu,

Zhenlian Peng and Guosheng Kang

School of Computer Science and Engineering,

Hunan University of Science and Technology,Xiangtan, China

Abstract

Purpose –With the continuous increase in Web services, eff‌icient identif‌ication of Web services that meet

developers’needs andunderstanding their relationships remains a challenge.Previous research has improved

recommendation effectiveness by using correlations between Web services through graph neural networks

(GNNs), while it has not fullyleveraged service descriptions, limiting the depth and diversityof learning. To

this end, a Web services recommendation method called LLMSARec, based on LargeLanguage Model and

semantic alignment, is proposed.This study aims to extract potential semantic information from services and

learn deeperrelationships between services.

Design/methodology/approach –This method consists of two core modules: prof‌ile generation and

maximizing mutual information. The prof‌ile generation module uses LLM to analyze the descriptions of

services, infer and construct service prof‌iles. Concurrently,it uses LLM as text encoders to encode inferred

service prof‌iles for enhancedservice representation learning. The maximizing mutual informationmodel aims

to align the semantic features of the services text inferred by LLM with structural semantic features of the

services captured by GNNs, thus achieving a more comprehensive representation of services. The aligned

representation serves as an inputfor the model to identify services with superior matching accuracy,thereby

enhancingthe service recommendation capability.

Findings –Experimentalcomparisons and analyses were conducted on the Programmable Webplatform data

set, and the results demonstratedthat the effectiveness of Web service recommendations can be signif‌icantly

improvedby using LLMSARec.

Originality/value –In this study, the authors propose a Web service recommendation approach based on

Large Language Model and semanticalignment. By extracting latent semantic information from servicesand

effectively aligning semantic features with structural features, new representations can be generated to

signif‌icantlyenhance recommendation accuracy.

Keywords Web services recommendation, Graph neural network, Large language model,

Services semantic alignment

Paper type Research paper

1. Introduction

With the development of services computing, the number of Web API are dramatically

increasing, posing challenges for developers in selecting and discovering Web services.

The work of this paper is supported by National Natural Science Foundation of China with Grant No.

62376062, 62177014, the National Key R&D Program of China with Grant No. 2018YFB1402800,

Hunan Provincial Natural Science Foundation of China with Grant No. 2022JJ30020 and the Science

and Technology Innovation Program of Hunan Province with Grant No. 2023sk2081.

International

Journal of Web

Information

Systems

37

Received13 September 2024

Revised14 October 2024

Accepted20 O ctober 2024

InternationalJournal of Web

InformationSystems

Vol.21 No. 1, 2025

pp. 37-53

© Emerald Publishing Limited

1744-0084

DOI 10.1108/IJWIS-09-2024-0262

The current issue and full text archive of this journal is available on Emerald Insight at:

https://www.emerald.com/insight/1744-0084.htm

Analyzing various types of information, such as service text and interaction behavior,

enables developers to quickly identify the services that best meet their needs from an

extensive library of Web API. This not only facilitates the creation of new service

compositions but alsosignif‌icantly enhances development eff‌iciency.

Traditional service recommendation methods primarily extract key features of Web

services from Web services Description Language documents and calculate the similarity

between services using metrics such as Jaccard Similarity, cosine similarity and Euclidean

distance, recommending services with the highest similarity (Hu et al., 2024;Wu et al.,

2020). However, these traditional methods usually overlook the interdependencies and

relationships between services, causing the resulting service representations to lack both

depth and diversity. To address this problem, some researchers have attempted to construct

service networks that integrate various service information and relationships (Liu et al.,

2015a;Liu et al., 2015b;Huang et al., 2014;Cao et al., 2024). Nevertheless, owing to the

sparsity and complexityof service networks, modeling service semantics ina comprehensive

and eff‌icient manner remains a challenging task. Recently, graph-based recommendation

methods have used graph neural networks (GNNs) to eff‌iciently capture complex service

relationships among service nodes. Models such as NGCF (Wang et al., 2019), LightGCN

(He et al., 2020) and GCCF (Chen et al., 2020a) typically construct a graph structure to

represent service relationships and encode feature information to further delineate these

relationships. These models train a GNN to learn high-order representations of nodes to

generate embeddings of services. The similarity or matching degree of these embeddings is

then calculated to produce services recommendation lists that offer more accurate and

personalized recommendations. Although GNN-based methods can capture service

relationships,they still encounterseveral challenges.

GNN-based methods mainly focus on extracting structural features of the service

network, heavily relying on ID-based information while neglecting the more detailed and

comprehensive latent semantic information contained within service descriptions. These

GNN-based graph methods primarily learnfrom data consisting of implicit feedback, which

may introduce biases [e.g. false clicks (Wang et al.,2021) or popularity bias (Chen et al.,

2020b)]. Thus, the representations learned by these models are heavily dependent on the

inherent quality of data. When the data contains noise, it can lead to the generation of

inaccurate representations,thereby reducing the accuracy of the recommendations.

Recently, Large Language Model (LLM) have made signif‌icant progress with their

powerful text processing and reasoningcapabilities and have started to be applied in service

recommendation. Figure 1 illustrates how GNNs and LLM are connected and deployed.

LLM can deeply analyze and encode service description texts using their strong text

understanding capabilities, but they have limitations in capturing structural information.

Kolasani (2023) suggested using LLM to analyze user interaction data to determine user

needs and preferences. Wei et al. (2024) suggested generating additional auxiliary

information for text data to gain a deeper understanding of content. Although LLM excels in

processing service descriptiontexts, it has a limitation in capturing the structural information

of services. To address this issue, Ren et al. (2024) combined the structural information

capture capability of GNNs with the text understanding capability of LLM, a more

comprehensive service network learning can be achieved, thereby enhancing the

performance of service recommendations. Inspired by this approach, this study introduced

LLMSARec into Web service recommendations. To address the f‌irst problem, an LLM-

based service reason constructionmodule was designed. This leverages the reasoningability

of LLM, which acts as a bridge to integrate the GNN-based graph model with the LLM. In

addition, LLM uses reasoning based on relevantservice descriptions to extract more detailed

IJWIS

21,1

38