Regions coloured by what their kitchens use, borders drawn where the spice repertoire changes fastest, and a route finder that lists what drops out and what enters as one travels.
Choose two kitchens and read what drops out of the pot and what enters, border by border, along the shortest overland path.
The map is the idea of Dr Malabika Biswas, who proposed reading the world's cuisines as a continuous field of spice transitions, with the map organised around what enters and what drops out as one travels from one region to the next, and the regions defined by what their kitchens use, whatever they happen to grow. This page is a first prototype of that idea. The coding of the regional matrix, the pipeline and any errors in either are Anirban Ghatak's.
Three layers are offered. The expert-coded layer scores 59 spices and aromatic herbs on a 0–3 prominence scale for 117 culinary regions, coded by the author from culinary reference knowledge. It is a structured prior, not a measurement, and it is the only layer with sub-national resolution for South Asia, China, Southeast Asia, Africa and the Americas. The CulinaryDB layer estimates, for 23 cuisines, the share of recipes in which each spice appears, from about 45,000 recipes in CulinaryDB (Bagler group, IIIT Delhi). The Ahn et al. (2011) layer does the same for the 11 macro-regions of the 56,000-recipe corpus behind the flavour network paper (Ahn, Ahnert, Bagrow and Barabási, Scientific Reports). Named blends in recipe text (garam masala, curry powder, five-spice, berbere, and so on) are decomposed into their usual constituents.
Each region is a composition vector over the lexicon. Colours come from a principal component analysis of Hellinger-transformed compositions, with the first two components mapped to the CIELAB chroma axes, so that hue distance approximates compositional distance. Lightness is reserved for spice intensity, measured as the sum of a region's prominence scores in the coded layer and as the expected number of lexicon spices per recipe in the corpus layers. Lightly spiced kitchens, the Arctic, the Congo Basin, Mongolia, Polynesia, therefore appear pale, and the map distinguishes using few spices from using different ones. Each region's territory is drawn within the countries it is allowed to cover. Where several regions share a country, the country is divided among them by nearness to their representative points (a Voronoi split, computed on the Natural Earth projection and clipped to the country's boundary), so a cell never crosses a border into a country it does not cover. Two regions are neighbours when their territories share a border, and a small number of short sea crossings are added by hand and drawn as dashed lines. Three frontier scores are computed for each border. The composition score is the Jensen–Shannon distance between the two normalised profiles and ignores how much spice is used, which makes any sparsely coded region look far from every neighbour. The combined score, used by default, appends to each profile an unspiced share equal to the region's shortfall from the most intensely spiced region in the layer before normalising, so that two lightly spiced kitchens are close to each other and both are far from a heavily spiced one. The intensity score is the absolute gap in the intensity index. Routes minimise great-circle distance over the neighbour graph.
Antecedents. Rozin (1973) on flavour principles; Sherman and Billing (1999, BioScience) on climate and spice use; Cavalli-Sforza, Menozzi and Piazza (1994) on synthetic gene-frequency maps; Monmonier (1973, Geographical Analysis) on maximum-difference barriers; Zhu et al. (2013, PLoS ONE) on geography and similarity of Chinese regional cuisines; Jain, Rakhi and Bagler (2015, PLoS ONE) on Indian regional cuisines. Recipe corpora are English-language and tilted toward restaurant and diaspora cooking, and their region tags are national at best, which is why the fine structure of the map rests on the coded layer for now.