[Paper Review] MaTrust: An Effective Multi-Aspect Trust Inference Model

Yuan Yao, Hanghang Tong+3 more|arXiv (Cornell University)|Nov 9, 2012

Access Control and Trust34 references3 citations

TL;DR

MaTrust proposes a multi-aspect trust inference model that learns latent factors from local trust ratings and incorporates prior knowledge like trust bias as specified factors. It outperforms state-of-the-art models in both accuracy (up to 16.5% better MAE on PGP) and efficiency (up to 3.5 million times faster than MoleTrust), scaling linearly with data size.

ABSTRACT

Trust is a fundamental concept in many real-world applications such as e-commerce and peer-to-peer networks. In these applications, users can generate local opinions about the counterparts based on direct experiences, and these opinions can then be aggregated to build trust among unknown users. The mechanism to build new trust relationships based on existing ones is referred to as trust inference. State-of-the-art trust inference approaches employ the transitivity property of trust by propagating trust along connected users. In this paper, we propose a novel trust inference model (MaTrust) by exploring an equally important property of trust, i.e., the multi-aspect property. MaTrust directly characterizes multiple latent factors for each trustor and trustee from the locally-generated trust relationships. Furthermore, it can naturally incorporate prior knowledge as specified factors. These factors in turn serve as the basis to infer the unseen trustworthiness scores. Experimental evaluations on real data sets show that the proposed MaTrust significantly outperforms several benchmark trust inference models in both effectiveness and efficiency.

Motivation & Objective

• Address the limitations of existing trust inference models that rely solely on transitivity and ignore multi-aspect trust properties.
• Enable trust inference without requiring additional input data beyond local trust ratings, unlike prior multi-aspect models.
• Incorporate prior knowledge such as global, trustor, and trustee biases directly into the inference process.
• Develop a scalable and efficient model suitable for large-scale trust networks.

Proposed method

• Model trust as a combination of latent factors derived from observed trust ratings using matrix factorization.
• Introduce three types of specified bias factors: global bias, trustor bias, and trustee bias, to represent prior knowledge.
• Formulate the trust inference problem as an optimization task that jointly learns latent factors and bias coefficients from the trust matrix.
• Use regularization to prevent overfitting and ensure generalization, with a closed-form solution for efficient inference.
• Decompose the trust matrix into trustor and trustee factor matrices, plus bias components, enabling constant-time prediction after pre-computation.
• Apply alternating least squares (ALS) for optimization, with complexity linear in the number of users and observed trust pairs.

Experimental results

Research questions

• RQ1 Can modeling trust as multiple latent aspects improve inference accuracy compared to transitivity-only approaches?
• RQ2 Can prior knowledge such as trust bias be effectively integrated into trust inference without requiring additional input data?
• RQ3 How does the proposed model perform in terms of accuracy and efficiency compared to state-of-the-art trust inference models?
• RQ4 How sensitive is the model to hyperparameters like the number of latent factors and regularization coefficient?
• RQ5 Does the model scale efficiently with increasing data size in terms of users and observed trust pairs?

Key findings

• MaTrust significantly outperforms SVD and KBV in both RMSE and MAE, improving MAE by 16.5% on the PGP dataset.
• The model achieves up to 3.5 million times faster inference than MoleTrust, with wall-clock time measured in microseconds.
• MaTrust scales linearly with the number of users and observed trust pairs, making it suitable for large-scale applications.
• The model is robust to hyperparameter changes: performance remains stable across different values of latent factor size (r) and regularization (λ).
• Fixing bias coefficients to 1 (as in KBV) results in worse performance than learning them jointly, confirming the benefit of end-to-end optimization.
• The pre-computation stage of MaTrust requires only O(|K| + nr + r²) space, far less than O(n²) space used by models like Guha that fill all missing entries.