Abstract
Traditional online rating systems, which rely on global averages (O-RATING), are highly susceptible to bribery. This paper explores an alternative: personalized, network-based scores (P-RATING), where a user's rating for a service is derived from their trusted peers. Using a game-theoretic framework, we demonstrate that P-RATING is significantly more resilient to bribery, particularly when the briber lacks perfect knowledge of the social network's structure or the identities of specific users. Under certain cost constraints, P-RATING can even be made fully bribery-proof.
Introduction
Online platforms like TripAdvisor and Yelp are essential for consumer decision-making, yet they are notoriously vulnerable to manipulation. Service providers often offer incentives—such as free meals or cash—in exchange for positive reviews. Most current systems utilize an "Objective Rating" (O-RATING), a single global average that provides a predictable and high return on investment for bribers, especially when many potential customers have not yet left reviews.
This research proposes a shift toward "Personalised Rating" (P-RATING). In this system, a customer’s score for a business is calculated only from the reviews of their trusted peers within a social network. The core objective is to determine whether this locality of influence makes bribery fundamentally more difficult for a malicious actor.
Methods
The study employs a game-theoretic model involving a service provider (the briber) and a set of customers (the voters). Customers are represented as nodes in a social network.
Influence Modeling
In O-RATING, every customer is influenced by the global average. In P-RATING, influence is local: a customer is only affected by the average of their neighbors' scores. A critical metric introduced is the influence weight ($w_c$), which quantifies how a single customer's opinion propagates through the network. In P-RATING, nodes connected to "information-starved" peers (those with few other connections) possess higher influence weights.
The Bribery Game and Knowledge Constraints
The briber allocates a budget to increase the ratings of specific customers. The success of this strategy is analyzed under three knowledge scenarios:
- Perfect Knowledge: The briber knows the network structure and the identity of every node.
- Structural Knowledge only: The briber knows the graph topology but not which customer occupies which node.
- Zero Knowledge: The network structure is completely unknown.
Key Results
The analysis reveals several critical advantages of personalized systems over global ones.
The Vulnerability of Global Averages
O-RATING is proven to be non-bribery-proof if even a single person has not yet voted. Bribing "non-voters" is the most profitable strategy because they add new positive data without displacing existing information, monotonically increasing revenue for the briber.
Resilience through Locality and Anonymity
P-RATING's primary defense is its locality; a bribe only affects the target's immediate neighbors. Furthermore, if the briber knows the network structure but not the specific positions of individuals (as illustrated in Figure 1), P-RATING becomes "bribery-proof in expectation." On average, a random bribe in such a system yields zero net gain for the briber.
Quantitative Limits and Efficiency
Proposition 13 establishes that the gain from bribing an individual in a personalized system can never exceed the number of that individual's neighbors. While an optimal bribery strategy can be calculated in polynomial time (Algorithm 1) given perfect information, P-RATING remains "fully bribery-proof" if the cost of a bribe exceeds the maximum possible influence of a single node.
Discussion and Limitations
While P-RATING offers superior resistance, the model relies on several assumptions. It assumes a linear relationship between bribes and rating increases, and between ratings and utility. Real-world behavior is often non-linear (e.g., the difference between a 1-star and 2-star review might matter more than 4-star vs 5-star). Additionally, the social network is treated as static, whereas real trust networks evolve dynamically. Future research should investigate competitive scenarios between multiple service providers and cases involving partial knowledge of customer evaluations.
Conclusion
Moving from global to personalized rating systems significantly increases the cost and complexity of bribery. By leveraging the local nature of social networks and the inherent anonymity of user positions, P-RATING provides a robust framework for maintaining the integrity of online reviews.
Summary updated