A Transfer Machine Learning Matching Algorithm for Source and Target (TL-MAST)
| Autor: | Robin Hirt, Florin Tim Peters |
|---|---|
| Rok vydání: | 2020 |
| Předmět: | |
| Zdroj: | Machine Learning, Optimization, and Data Science ISBN: 9783030645793 LOD (2) |
| Popis: | Sequentially transferring machine learning (ML) models in a row over several data sets can result in an improvement of performance. Considering an increasing number of data sets, the possible number “transfer paths” of a transfer grows exponentially. Thus, in this paper, we present TL-MAST, a matching algorithm for identifying the optimal transfer of ML models to reduce the computational effort across a number of data sets—determining their transferability. This is achieved by suggesting suitable source data sets for pairing with a target data set at hand. The approach is based on a layer-wise, metric-supported comparison of individually trained base neural networks through meta machine learning as a proxy for their performance in a crosswise transfer scenario. We evaluate TL-MAST on two real-world data sets: a unique sales data set composed of two restaurant chains and a publicly available stock data set. We are able to identify the best performing transfer paths and therefore, drastically decrease computational time to find the optimal transfer. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|