NES Professor, Academic Director of the "Economics and Data Science" NES and Yandex Joint Program Daria Dzyabura took part in the largest in Russia, CIS and Eastern Europe conference on marketing analytics Matemarketing'21. She presented her joint with Renana Perez (The Hebrew University) and NES student Irina Yakovetskaya (BAE'2022) research “Data Driven Graphic Design: The Role of Color Composition in Brand Image”.
In this work, the authors apply quantitative methods to analyze brand images and how brands are represented by graphic design, the part of marketing that is still largely based on qualitative analysis. Professor Dzyabura with co-authors try to answer the questions of how brand strategy can be visualized and what images, colors and shapes are most relevant to a particular brand.
Such studies are limited by the lack of unambiguous and labeled data, and by the fact that image processing methods based on neural networks and artificial intelligence provide a rather abstract analysis with biased parameter estimates.
To overcome these limitations and find a link between graphic design elements and brand attributes, the authors conducted a comprehensive survey of brands with 1,851 respondents, allowing them to collect 4,743 collages of images associated with brands. Then, using topic modeling (an instrument used mainly for text analysis), they extracted from these collages color palettes representing a certain brand’s image.
Along with making collages, participants of the survey also tagged brands according to their quality characteristics (for example, “cheerful”, “confident”, “feminine”, “trendy”, etc.). This allowed for data-driven and interpretable relationships between colors and brand characteristics (see table here). For example, the authors of the study found that for the visuals of a family oriented brand different shades of green and orange are relevant (colors used by brands such as Kellog's and Palmolive), while shades of purple and blue are not (for example, brands Porsche and Victoria's Secret).
“As the next step, we are interested in verifying the results obtained on the data. For example, if you ask a graphic designer to create an ad image for a brand, will this image exhibit the brand attribute and match the characteristics with which the brand tries to associate itself? Another task is to teach algorithms to analyze images and carry out such an analysis using machine learning tools,” Professor Dzyabura concluded.