💻 computer science

An Integrated Hybrid Recommendation and Trust-Aware Review Intelligence Framework for Reliable Service Provider Matching

This paper proposes an integrated hybrid framework that combines TF-IDF, semantic similarity, and collaborative filtering for candidate provider selection with a trust-aware review intelligence system using Multi-Task BiLSTM and Isolation Forest to analyze aspect-specific sentiment and review credibility, thereby achieving more reliable service provider matching through a unified re-ranking pipeline.

Original authors: W. A.M. Tharushika, N. T. Hewapathirana, Dhammika De Silva, Samantha Rajapaksha

Published 2026-08-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: W. A.M. Tharushika, N. T. Hewapathirana, Dhammika De Silva, Samantha Rajapaksha

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Finding a skilled professional to fix a leaky pipe, install a ceiling, or repair a car is something almost everyone does, yet the digital marketplaces designed to connect customers with these workers often leave people guessing. When a user types a request into a search box, the system must decide which of thousands of registered providers is the best fit. Traditional systems rely on simple keyword matching, looking for exact word overlaps between what a customer asks and what a provider offers, or they lean heavily on a single numerical star rating. However, these methods often miss the mark. A customer asking to "fix a dripping faucet" and a provider describing themselves as a "plumber who repairs leaks" are talking about the same need, but a simple word-counting system might fail to see the connection. Furthermore, a single number cannot capture the nuance of a service experience, which might be excellent in terms of work quality but poor regarding punctuality or communication. Even more troubling is that not all reviews are trustworthy; some are fake, manipulated, or written by people with unusual behavior patterns, and treating every review as equally valid can distort a provider's true reputation. Researchers at the Sri Lanka Institute of Information Technology have developed a new framework to solve these problems by combining two distinct approaches: one that finds relevant providers and another that verifies the reliability of the feedback they receive.

The researchers built a two-stage system designed to first identify a shortlist of likely candidates and then deeply analyze the feedback associated with them to ensure trustworthiness. In the first stage, the system acts as a broad filter, scanning a database of ten thousand providers against twenty thousand customer requests. Instead of relying on just one method, it uses three different ways to measure relevance simultaneously. It checks for direct keyword matches, uses advanced language models to understand the meaning behind different ways of phrasing a request, and looks at historical data to see how often a provider has successfully completed jobs for users similar to the one searching. These three signals are blended together to create a single score, narrowing the massive list of providers down to a manageable group of the most relevant candidates. This approach proved more effective than using any single method alone, successfully identifying the right providers with high precision.

Once the system has a shortlist of candidates, it moves to the second stage, which functions as a quality control and trust verification unit. Here, the system does not just look at the average star rating; it reads the actual text of twenty-five thousand customer reviews across fourteen different service categories, such as masonry, electrical work, and vehicle repair. It breaks down each review to understand specific aspects of the service: the quality of the work, whether the provider was punctual, how well they communicated, and their level of professionalism. Using a sophisticated neural network, the system predicts the sentiment for each of these four areas separately, recognizing that a provider might be praised for their skill but criticized for being late. At the same time, a separate module examines the reviews for signs of unreliability. It looks for linguistic patterns and behavioral clues, such as unusual word choices or repetitive phrasing, to flag reviews that might be suspicious or anomalous. Reviews that appear credible are given more weight, while those flagged as potentially fake or manipulated are downplayed, ensuring that the final ranking reflects genuine customer experiences.

The results of this integrated approach demonstrate that combining relevance with trust-aware intelligence significantly improves service matching. The sentiment analysis model successfully predicted the specific feelings behind reviews with an average accuracy of nearly eighty-nine percent, correctly identifying whether a customer was happy or unhappy about quality, punctuality, communication, or professionalism. The system that detected fake or unreliable reviews was even more precise, achieving ninety-six percent overall accuracy in distinguishing between genuine feedback and suspicious anomalies. By fusing these insights, the framework can re-rank the initial list of candidates, promoting those with strong, credible, and specific praise while demoting those with unreliable or mixed feedback. The researchers found that this two-step process, which separates the task of finding relevant providers from the task of verifying their reputation, creates a more reliable way to match customers with professionals. While the study acknowledges that the specific weights used to combine these different signals were manually selected and that further testing on real-world data is needed, the current findings show that treating relevance and trust as complementary, rather than separate, concerns offers a clear path forward for online service marketplaces.

Technical Summary: An Integrated Hybrid Recommendation and Trust-Aware Review Intelligence Framework for Reliable Service Provider Matching

Problem Statement
Online service marketplaces face a dual challenge: identifying service providers who are semantically relevant to a customer's specific request and ensuring those providers are reliable based on trustworthy feedback. Existing mechanisms often rely on simple keyword matching, which fails to capture semantic equivalence (e.g., "repair leaking pipe" vs. "fix plumbing issue"), or on numerical ratings that lack granularity. Furthermore, standard approaches frequently ignore the reliability of reviews, leaving platforms vulnerable to opinion spam and manipulated ratings. Consequently, customers struggle to distinguish between providers who are merely popular and those who are genuinely suitable and trustworthy for specific service aspects like quality, punctuality, communication, and professionalism.

Methodology
The study proposes a two-stage integrated framework that separates provider selection into relevance matching and trust-aware quality evaluation.

Stage 1: Hybrid Candidate Recommendation
This stage filters a large pool of providers to generate a relevant candidate set using a weighted fusion of three distinct signals:

Lexical Relevance: A TF-IDF model calculates keyword overlap between customer requests and provider descriptions.
Semantic Similarity: A multilingual Sentence Transformer (paraphrase-multilingual-MiniLM-L12-v2) generates dense embeddings to capture contextual meaning beyond exact keyword matches.
Credibility-Based Collaborative Filtering (CF): A custom CF module incorporates provider credibility attributes (normalized average rating, booking success rate, and interaction frequency) and personalizes scores based on a user's interaction history.
These three scores are normalized and fused using a weighted sum (30% TF-IDF, 35% Semantic, 35% CF) to rank and select the top- $K$ candidate providers.

Stage 2: Trust-Aware Review Intelligence
The candidate providers from Stage 1 undergo a deeper evaluation using their associated customer reviews:

Multi-Task Aspect-Based Sentiment Analysis (ABSA): A Multi-Task Bidirectional Long Short-Term Memory (Bi-LSTM) network with Multi-Head Attention simultaneously predicts sentiment for four service aspects: Quality, Punctuality, Communication, and Professionalism. This allows for granular evaluation rather than a single aggregate score.
Review Credibility Detection: An unsupervised Isolation Forest module analyzes linguistic and behavioral features (e.g., review length, word diversity, punctuation usage, reviewer activity) to identify anomalous or suspicious reviews.
Trust-Aware Fusion: The sentiment scores for each aspect are weighted and aggregated. Crucially, these sentiment scores are multiplied by the credibility score of the review. This ensures that feedback from suspicious sources has a diminished impact on the final provider ranking.
Re-ranking: The final trust-aware scores are aggregated at the provider level to re-rank the candidates, returning the most reliable providers.

Key Contributions
The paper outlines six principal contributions:

An integrated two-stage architecture for service-provider recommendation that decouples relevance from trust evaluation.
A hybrid candidate generation mechanism combining TF-IDF, multilingual semantic matching, and a custom credibility-based collaborative signal.
A credibility-based collaborative formulation that explicitly integrates rating, booking success, interaction frequency, and history.
A Multi-Task Bi-LSTM with Multi-Head Attention designed for simultaneous sentiment prediction across four distinct service-quality aspects.
An Isolation Forest-based module for detecting review credibility using linguistic and behavioral features.
A Trust-Aware Sentiment Fusion mechanism that adjusts aspect-level sentiment by review credibility for provider re-ranking.

Results
The framework was evaluated using 10,000 providers, 20,000 user requests, 100,000 interaction records, and 25,000 customer reviews across fourteen service categories.

Hybrid Recommendation: The hybrid method outperformed its individual components (TF-IDF only, BERT only, CF only) in Precision@K and NDCG@K metrics. The hybrid approach achieved a Precision@5 of 0.9490 and NDCG@5 of 0.9496.
Sentiment Analysis: The Multi-Task Bi-LSTM achieved an average testing accuracy of 88.60%. Performance varied by aspect, with "Quality" achieving the highest accuracy (96.10%) and "Communication" and "Professionalism" achieving 85.48%.
Credibility Detection: The Isolation Forest module achieved 96% overall accuracy in distinguishing between real and suspicious reviews, with a weighted F1-score of 0.96.

Significance and Claims
The authors claim that the framework demonstrates the complementary value of combining relevance, personalization, aspect-level service evaluation, and review reliability in a unified pipeline. The study posits that while textual relevance is a strong retrieval signal, historical interaction data alone is insufficient for primary retrieval but serves as a vital complementary signal for personalization. Similarly, the results suggest that service quality cannot be accurately represented by a single overall sentiment; aspect-level analysis is necessary.

The paper modestly notes that the two components were originally evaluated independently. It acknowledges that a controlled end-to-end ablation study comparing the full framework against its sub-components using common ranking metrics was not performed in this specific integration. However, the authors argue that the two-stage design (candidate generation followed by detailed re-ranking) is computationally practical and effectively separates the problem of "who is relevant?" from "who is trustworthy?"

Limitations and Future Work
The authors identify several limitations, including the use of domain-specific datasets, manually selected fusion weights, and the use of engineered features for credibility detection rather than deep learning-based feature extraction. Future work is suggested to include adaptive fusion weights, multilingual transformer models for ABSA, more sophisticated credibility features (e.g., temporal bursts, reviewer-provider graphs), and a rigorous end-to-end ablation study.

Technical Summary: An Integrated Hybrid Recommendation and Trust-Aware Review Intelligence Framework for Reliable Service Provider Matching

More like this