The data given in museum collections and exhibition catalogues is an invaluable source of information for art historians. A large and growing number of museums have online collections, many contemporary exhibitions give online access to their catalogues, and scholars have created datasets through the digitisation of catalogues of historical exhibitions. The use of quantitative analysis to answer art historical questions utilising such data sources is an expanding field. A critical barrier to entry for art historical research involving data drawn from different sources where the researcher wants to look at the level of the individual artist is that of matching artists’ names across the collections or catalogues in which they appear. With large datasets or small teams of researchers it is often the case that manual matching of names in not practical.
Name matching across data sources is a generic problem faced in both academic and commercial environments. Variations in presentation or spelling means the problem is one of finding close rather than exact matches and of distinguishing variants of the same name from different names. Researchers have investigated this issue of ‘fuzzy’ name matching, developing measures of the similarity of names or complete solutions in practical applications, often using methods from machine learning.
In this paper I investigate the use of two machine learning classification algorithms, logistic regression and random forest, for matching artists’ names with three composite datasets. Two consist of the entries for paintings by male and female artists taken from the catalogues for 106 exhibitions in nineteenth-century Paris included in the Museé D’Orsay’s ‘Salons’ dataset. The third is of modern and contemporary art, constructed by the author, and taken from the metadata available in the online collections of 135 art museums in 32 countries. The features utilised in these models are a number of name similarity measures. One challenge faced in developing the models is that the problem of matching names is highly imbalanced, in that the number of name pairs in each dataset that are variant names for an artist are several orders of magnitude smaller than the number of name pairs of different artists. There are also no pre-existent training and testing datasets.
For art historical research, name matching must identify the substantial majority of names for an artist with very few false identifications, or, in machine learning terminology, have very high levels of ‘recall’ and ‘precision’. In machine learning there is a trade-off between recall and precision, and it applies in this case. Neither logistic regression nor random forest meet the requirements. Rather, the best approach is using them in combination to identify candidate variant names that are scored manually. To complete the matching, these links are propagated and manual cross-checks are made.
I will present some Initial analyses of the name-matched datasets to illustrate their art historical value and show how they support future avenues of research. The approach developed is one scholars working with composite datasets could consider following or adapting.