Shapes as Product Differentiation: Neural Network Embedding in the Analysis of Markets for Fonts

Sukjin Han, Eric H. Schulman, Kristen Grauman, Santhosh RamakrishnanarXiv 2021 · arXiv

Introduction

In traditional economic analysis, researchers typically rely on "structured" data—quantifiable metrics such as price, quantity, and technical specifications. However, for a vast array of consumer goods ranging from fashion and furniture to software interfaces, the primary driver of value is design. Because design is inherently "unstructured" and high-dimensional, economists have historically treated it as an "unobservable" factor, often relegated to the statistical error term. This limitation makes it difficult to assess how market forces, such as competition or mergers, influence the evolution of style and creativity.

This paper introduces a novel framework that transforms visual "shapes" into measurable economic variables. By utilizing deep learning to quantify font styles, the authors demonstrate that neural network embeddings can reveal strategic shifts in product differentiation that traditional metrics completely fail to detect.

Methodology: Quantifying Visual Style

To map the aesthetic properties of fonts into a mathematically tractable format, the researchers employed a multi-step machine learning and econometric approach.

1. Neural Network Architecture and Triplet Loss

The authors adapted a state-of-the-art Convolutional Neural Network (CNN) architecture, similar to FaceNet, to learn font embeddings. Rather than performing a simple classification task, the network was trained to map images into a 128-dimensional Euclidean space.

The training process utilized Triplet Loss. In this setup, the model is presented with three inputs: an Anchor font, a Positive font (from the same family but a different style), and a Negative font (from a different family). The loss function penalizes the model unless it places the Anchor and Positive fonts closer together in the embedding space than the Anchor and Negative fonts. Over time, the network learns to identify the subtle visual features that define a font's "style."

2. Pangram Inputs

To ensure the model captured more than just individual character shapes, the researchers used images of entire pangrams—sentences containing every letter of the alphabet. This allowed the network to account for complex design elements such as kerning (spacing), ligatures, and x-height, which are critical to the overall visual impact of a typeface.

3. Validation via Natural Language Processing

To confirm that the AI’s visual clusters were economically and human-meaningful, the authors compared the embeddings against user-contributed tags (e.g., "curly," "geometric," "serif"). Using Word2vec to cluster these linguistic tags, they found a high correlation between visual and linguistic clusters. The image embeddings captured nearly double the information about user perceptions compared to traditional industry categories (NMI of 0.473 vs 0.261), proving the AI "sees" the same stylistic nuances that humans describe.

The Impact of Mergers on Creativity

The core of the study applies this framework to a major industry event: the 2014 merger between Monotype and FontFont. To isolate the effect of the merger, the authors used the Synthetic Control Method (SCM), creating a "Synthetic FontFont"—a weighted average of non-merged foundries—to serve as a counterfactual.

Key Findings

  • The Cannibalization Strategy: The most significant discovery was that the merged firm significantly increased its visual differentiation post-merger. In economics, when two competing firms merge, they often push their designs further apart to avoid "cannibalizing" their own sales. By making their products more visually distinct, the merged entity can cover a broader segment of the market without its own products competing against one another.
  • The Failure of Traditional Data: If the researchers had only analyzed the number of new fonts released or their technical specifications, the merger would have appeared to have no effect. The strategic shift was purely visual, occurring in the "unstructured" dimension of shape.
  • Statistical Significance: The treatment effect on the "Gravity Measure"—a metric calculating a font's distance from its nearest neighbors in the embedding space—was approximately 0.5 to 1 standard deviation. This effect was most pronounced in the first two years following the merger.
  • Local Monopolies: The font market behaves as a series of "local monopolies." Because professional designers often require a very specific aesthetic for a project, they are relatively price-insensitive. This explains why prices remained stable even as the firm strategically shifted its design output.

Conclusion and Future Directions

By treating design as a coordinate in a high-dimensional space, this research provides a blueprint for incorporating aesthetics into rigorous economic modeling. Beyond market analysis, this framework has significant implications for Intellectual Property (IP) law. It could potentially be used to quantify "visual similarity" in copyright disputes, establishing a mathematical threshold for design infringement.

The authors suggest that the next frontier is the integration of these embeddings into structural demand models (such as BLP estimation). This would allow firms and regulators to predict how a change in a product's physical shape might directly influence its market share and consumer welfare.

Summary updated

Abstract

Many differentiated products have key attributes that are unstructured and thus high-dimensional (e.g., design, text). Instead of treating unstructured attributes as unobservables in economic models, quantifying them can be important to answer interesting economic questions. To propose an analytical framework for these types of products, this paper considers one of the simplest design products-fonts-and investigates merger and product differentiation using an original dataset from the world's largest online marketplace for fonts. We quantify font shapes by constructing embeddings from a deep convolutional neural network. Each embedding maps a font's shape onto a low-dimensional vector. In the resulting product space, designers are assumed to engage in Hotelling-type spatial competition. From the image embeddings, we construct two alternative measures that capture the degree of design differentiation. We then study the causal effects of a merger on the merging firm's creative decisions using the constructed measures in a synthetic control method. We find that the merger causes the merging firm to increase the visual variety of font design. Notably, such effects are not captured when using traditional measures for product offerings (e.g., specifications and the number of products) constructed from structured data.
BibTeX
@misc{han2021shapes,
  author = {Sukjin Han and Eric H. Schulman and Kristen Grauman and Santhosh Ramakrishnan},
  title = {{Shapes as Product Differentiation: Neural Network Embedding in the Analysis of Markets for Fonts}},
  year = {2021},
  eprint = {2107.02739},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2107.02739},
}
Figure