Sequential Transfer Machine Learning in Networks: Measuring the Impact of Data and Neural Net Similarity on Transferability

In networks of independent entities that face similar predictive tasks,\ntransfer machine learning enables to re-use and improve neural nets using\ndistributed data sets without the exposure of raw data. As the number of data\nsets in business networks grows and not every neural net transfer is\nsuccessful, indicators are needed for its impact on the target performance-its\ntransferability. We perform an empirical study on a unique real-world use case\ncomprised of sales data from six different restaurants. We train and transfer\nneural nets across these restaurant sales data and measure their\ntransferability. Moreover, we calculate potential indicators for\ntransferability based on divergences of data, data projections and a novel\nmetric for neural net similarity. We obtain significant negative correlations\nbetween the transferability and the tested indicators. Our findings allow to\nchoose the transfer path based on these indicators, which improves model\nperformance whilst simultaneously requiring fewer model transfers.\n