Transfer Learning in Body Sensor Networks using Ensembles of Randomised Trees Pierluigi Casale SPS Group-TU Eindhoven and Holst Center/IMEC Eindhoven, The Netherlands Email: pierluigi.casale@imec-nl.nl Marco Altini Holst Center/IMEC Eindhoven, The Netherlands Email: marco.altini@imec-nl.nl Oliver Amft University of Passau Passau, Germany olver.amft@uni-passau.de Abstract—In this work we investigate the process of trans- ferring the activity recognition models of the nodes of a Body Sensor Network and we proposed a methodology that supports and makes the transferring possible. The methodology, based on a collaborative training strategy, makes use of classifier ensembles of randomised trees that allow to generate activity recognition models able to be successfully transferred through the nodes of the network. Experimental results evaluated on 17 subjects with a network of 5 wearable nodes with 5 everyday life activities show that the recognition models can be transferred to a new untrained node replacing a node previously present in the network without a significant loss in the recognition performance. Moreover, the models achieve good recognition performance in nodes located in previously unknown positions. I. I NTRODUCTION Supervised Machine Learning techniques are widely used in the recognition process of physical activities using Wearable Sensors and Body Sensor Networks (BSN). Nevertheless, these data-driven methodologies can represent a limitation when the retraining of the activity recognition models of the network is required. For the aim of example, consider the situation where users need to replace a broken node or they want to relocate a node previously used on a specific position in a new position, e.g., from the chest to the waist. In both cases, the possibility to avoid a retraining still maintaining the recognition func- tionality of the node represents an appealing capability from the application point of view. In this work, we investigate the process of directly transfer the recognition models, i.e., the classifiers running in the nodes of a BSN, to a new node previously untrained in order to avoid the training phase after the deployment of the BSN. This transferring process, known as Transfer Learning [1], is accomplished in this work through the combination of a collaborative training strategy and the use of classifiers ensemble based on randomised trees. Using the collaborative training strategy, a limited amount of data shared between all the nodes of the BSN is used in combination to the data of the node for training a classifier ensemble before the deployment of the network. This ensemble, while still able to provide high recognition performance for the node, contains a degree of redundancy helpful during the transferring process. When algorithms based on randomized trees like Bagging [2], Random Forest [3] or Rotation Forest [4] are considered, the proposed strategy allows to learn ensembles that can be transferred through the nodes of the network and are able to recognize the activities sensed by nodes positioned at different locations. The random transformations that these algorithms apply to the training set are beneficial in the transferring process of the recognition models through the nodes placed at different position. Since the performance of the recognition models depend on the training data, the amount of data shared at training time is a quantity that needs to be taken into account in order to find a good trade-off between the performance of the node and the performance of the transferred classifiers. The methodology has been applied in situations where a broken node needs to be replaced by a new node located in the same position (replacement scenario) and a node already present in the network is relocated to a previously unknown position (relocation scenario). A dataset collected using a BSN of 5 wearable nodes has been used for evaluating the methodology in both scenarios, using sensor data from 17 subjects. In both scenarios considered, experimental results show that the recognition model of a node can be transferred and high recognition performance can be obtained in the replacement scenario and good recognition performance are achieved in the relocation scenario. Results have been validated using several K-folds cross-validation protocols in order to test the performance of the methodology when different amount of data are shared between nodes. This paper makes the following contributions: 1) We define a collaborative training strategy that, by sharing a limited amount of data between nodes, allows to generate classifier ensembles that, besides achieving high recognition performance in the node, contain a degree of redundancy useful in the trans- ferring of the recognition models through the nodes of the network. 2) We use ensemble learning algorithms based on ran- domised trees that, using random transformations on the training set generated with the collaborative training strategy, allows the direct transferring of the recognition models through the nodes of the network. II. TRANSFER LEARNING AND ITS APPLICATION IN BSN Classifier ensembles have been already considered as a mechanism for Transfer Learning ([5], [6], [7], [8]). Of par- ticular interest, the work of Kamishima et al. [9] applied a bagging approach for transfer the learning capabilities of a model through different domains. In their work, an high number of trees was learned on data from both source and target domains and a pruned version of the final ensemble was used to predict examples of the target domain. The pruning step was used in order to avoid the decreasing of performance due