[Paper Review] Model-based Pricing for Machine Learning in a Data Marketplace

Lingjiao Chen, Paraschos Koutris+1 more|arXiv (Cornell University)|May 26, 2018

Privacy-Preserving Technologies in Data20 references4 citations

TL;DR

This paper proposes a model-based pricing (MBP) framework that directly prices machine learning model instances instead of raw data, using a noise-injection mechanism to control model accuracy and prevent arbitrage. The approach enables real-time, low-runtime pricing with high revenue and affordability, validated through extensive experiments on synthetic and real datasets.

ABSTRACT

Data analytics using machine learning (ML) has become ubiquitous in science, business intelligence, journalism and many other domains. While a lot of work focuses on reducing the training cost, inference runtime and storage cost of ML models, little work studies how to reduce the cost of data acquisition, which potentially leads to a loss of sellers' revenue and buyers' affordability and efficiency. In this paper, we propose a model-based pricing (MBP) framework, which instead of pricing the data, directly prices ML model instances. We first formally describe the desired properties of the MBP framework, with a focus on avoiding arbitrage. Next, we show a concrete realization of the MBP framework via a noise injection approach, which provably satisfies the desired formal properties. Based on the proposed framework, we then provide algorithmic solutions on how the seller can assign prices to models under different market scenarios (such as to maximize revenue). Finally, we conduct extensive experiments, which validate that the MBP framework can provide high revenue to the seller, high affordability to the buyer, and also operate on low runtime cost.

Motivation & Objective

• To address the inefficiencies in existing data market pricing, where high data acquisition costs limit buyer affordability and seller revenue.
• To eliminate arbitrage opportunities in ML model markets by ensuring higher-accuracy models are never cheaper than lower-accuracy ones.
• To provide a practical, efficient pricing mechanism that allows sellers to set prices based on model accuracy, not data volume.
• To enable buyers to select models within their budget while maintaining high model utility and affordability.
• To design a framework that supports revenue maximization and dynamic pricing under varying market conditions.

Proposed method

• The MBP framework prices ML model instances based on the variance of Gaussian noise injected into a pre-trained optimal model.
• Noise injection ensures that model error increases monotonically with noise variance, enabling precise control over accuracy and price.
• The broker uses a price-error curve to map noise variance to price, allowing buyers to choose models based on budget and accuracy needs.
• A dynamic revenue optimization algorithm computes optimal price-error curves to maximize seller revenue, with polynomial-time complexity.
• The framework is formally proven to prevent arbitrage by ensuring monotonicity of error with respect to noise variance.
• Experiments compare MBP against MILP, linear, and heuristic pricing methods, showing MBP achieves near-optimal revenue with significantly lower runtime.

Experimental results

Research questions

• RQ1 How can a pricing framework for ML models prevent arbitrage while enabling affordable access for budget-constrained buyers?
• RQ2 What mechanism can enable real-time, low-runtime pricing of ML model instances based on accuracy and price?
• RQ3 How can a seller maximize revenue when offering multiple model instances with varying accuracy and price?
• RQ4 To what extent does the noise-injection approach maintain formal guarantees on model error and pricing monotonicity?
• RQ5 How does the MBP framework compare to existing pricing methods in terms of revenue, affordability, and runtime efficiency?

Key findings

• The MBP framework achieves near-optimal revenue—within 50% of the MILP-based optimal solution—across all tested datasets and buyer distributions.
• Runtime performance shows MBP is significantly faster than MILP, scaling quadratically, while MILP grows exponentially with the number of price points.
• MBP maintains high affordability, with affordability ratios consistently above 0.8 across all experiments, outperforming or matching other methods.
• The noise-injection mechanism provably ensures monotonic error increase with noise variance, eliminating arbitrage opportunities.
• Despite its simplicity, MBP outperforms naive linear and median-based pricing in both revenue and affordability, even though they are faster.
• The framework remains robust across diverse buyer valuation curves, with performance close to the theoretical optimum in most realistic scenarios.