Abstract
One of the major challenges in Human Activity Recognition (HAR) based on machine learning is the scarcity of labeled data. Indeed, collecting a sufficient amount of training data to build a reliable recognition problem is often prohibitive. Among the many solutions in the literature to mitigate this issue, collaborative learning is emerging as a promising direction to distribute the annotation burden over multiple users that cooperate to build a shared recognition model. One of the major issues of existing methods is that they assume a static activity model with a fixed set of target activities. In this paper, we propose a novel approach that is based on Growing When Required (GWR) neural networks. A GWR network continuously adapts itself according to the input training data, and hence it is particularly suited when the users share heterogeneous sets of activities. Like in federated learning, for the sake of privacy preservation, each user contributes to the global activity classifier by sharing personal model parameters, and not by directly sharing data. In order to further mitigate privacy threats, we implement a strategy to avoid releasing model parameters that may indirectly reveal information about activities that the user specifically marked as private. Our results on two well-known publicly available datasets show the effectiveness and the flexibility of our approach.
Introduction
Human activity recognition (HAR) is a well-established research topic [9]. The objective of HAR is to infer human daily activities ranging from low-level activities like walking or sitting, to high-level activities like cooking or watering flowers. HAR has been a key enabler for many real-world applications, including healthcare and well-being.
Sensor-based HAR performs the recognition task by analyzing data generated by wearable sensors and/or environmental sensors [9]. Recently, some efforts also focused on activity recognition based on radio frequency signals [13, 18]. The majority of sensor-based approaches in the literature are based on supervised machine learning [28]. Those methods require a large amount of training data to build a robust activity classifier. Unfortunately, the acquisition of labeled sensor data is costly, time-consuming and intrusive [17]. Indeed, one of the major challenges in HAR is the scarcity of labeled data for realistic scenarios [6].
In order to mitigate this problem, solutions based on collaborative learning have been proposed [15, 53]. Those approaches assume that each participating user is able to collect a few labeled data that is willing to share. The data collected from the participating subjects is then aggregated to build a shared activity classifier. The results of those works are promising, since the shared activity model can reliably recognize activities while requiring limited annotated data from each user. The major drawback of those approaches is that data about human activities can be considered as sensitive by many individuals. Hence, relevant privacy concerns arise when activity data is shared to untrusted third parties [7].
Recently, the federated learning paradigm has been proposed to mitigate the privacy problem in general collaborative machine learning [52]. Indeed, federated learning is a particular collaborative learning approach, where each participant only shares the parameters of a locally trained model instead of sharing data. While federated learning significantly reduces the amount of outsourced sensitive information, it relies on the assumption that the model architecture is static, and hence that the set of output classes (i.e., activities in our domain) is fixed. Moreover, it has been proven that it is still possible to infer sensitive information even from model parameters [46]. Despite some privacy preserving techniques have been proposed to mitigate the indirect leak of private information, it is not trivial to consider personal privacy preferences in this framework. In the specific case of activity recognition, users may have different preferences on which activities they wish to keep private.
In this paper, we present a novel collaborative learning approach, called
In
In order to mitigate privacy risks, the participating users only share a portion of their local model parameters (i.e., GWR neurons). Moreover,
The key contributions of the paper are summarized as follows:
We propose a novel distributed collaborative learning approach for human activity recognition. Thanks to the underlying machine learning algorithm,
In order to mitigate privacy issues, in
We performed an extensive empirical evaluation considering two publicly available datasets. The results show the effectiveness and flexibility of
The rest of the paper is organized as follows. Section 2 reviews the relevant work in HAR with focus on semi-supervised learning, transfer learning and collaborative learning. Section 3 describes
Related work
Human activity recognition (HAR) has attracted a lot of attention in the last decades and various techniques have been proposed [4, 28, 49]. In this section, we review the most relevant techniques in the literature to tackle the labeled data scarcity problem for HAR, including knowledge-based reasoning, semi-supervised learning, transfer learning, and collaborative learning. We also review relevant works that tackled the privacy problems that arise in collaborative learning settings.
Knowledge-based reasoning
In order to completely avoid the acquisition of labeled data, many works in the literature proposed knowledge-based approaches for activity recognition, especially considering smart-home settings [10, 34]. Those techniques rely on logic formalism (e.g., ontologies) that explicitly model the complex relationships between sensor events and activities based on common-sense knowledge. In particular, complex activities are defined in terms of their simpler components. The sequences of simple actions, recognized based on firing of specific sensor events, are then matched in real-time to activity definitions in order to infer the current activity performed by the monitored subject. Knowledge-based reasoning has been also proposed in combination with probabilistic reasoning [16, 21]. Semantic reasoning can also be used for the dynamic segmentation of sensor events [36, 54]. The main issue of knowledge-based approaches is their rigidity. Moreover, domain modeling is a time consuming and manual effort, and there is no guarantee that the resulting semantic model can comprehensively cover all the possible context conditions. Indeed, domain experts and knowledge-engineers are required to create the formal model. Finally, knowledge-based approaches cannot be directly applied on inertial sensors data obtained from mobile and wearable devices.
Semi-supervised learning
A few works proposed alternative machine learning approaches for activity recognition to overcome the issues of supervised learning. For instance, there exists some research efforts on unsupervised learning techniques [27, 29, 48]. However, those approaches require a large pool of data to discover significant patterns. Moreover, a certain amount of labeled data is still required to reliably associate each discovered cluster with its corresponding activity class.
In order to combine the strengths of supervised and unsupervised approaches, semi-supervised learning methods for activity recognition have been proposed [2]. These approaches only rely on a small labeled training sets to initialize the recognition model, that is continuously improved using unlabeled data. In the literature, the most common semi-supervised strategies used for activity recognition are self-learning [32], co-learning [22, 30], and active learning [1, 23, 24, 35].
In particular, active learning proved to be particularly effective for activity recognition. This approach requires an explicit feedback from the users to obtain labels for the most informative unlabeled data points: when the classifier is uncertain about the current prediction, a query is triggered to the user in order to obtain the actual activity that he/she was performing. The feedback is used to update the recognition model with new labeled examples.
Transfer learning
Transfer learning is another approach that has been explored to tackle the data scarcity challenge for activity recognition [17]. The objective of transfer learning is to transfer knowledge learned from a source domain (with labeled data) to a target domain (with unlabeled data). Transfer learning in activity recognition has been mainly applied to take advantage of labeled data acquired in a specific sensing setting to recognize activities in different sensing settings.
Wang et al. [12] proposed a stratified transfer learning to improve the recognition rate on cross-domain activity recognition. In this work cross-domain means different body positions where wearable sensors are worn. Recently, Change et al. [8] designed several domain adaptation techniques and performed more systematic evaluation on sensor wearing diversity; that is, transferring the activity model from one wearing position (e.g., chest) to another (e.g., thigh).
Collaborative learning
Activity recognition based on collaborative learning is becoming increasingly popular. Collaborative learning aims at creating a shared model by aggregating smaller models that are built by periodically updates from multiple participants [51]. Moreover, the participants also receive periodic updates from the shared model to further improve their local model thanks to collaboration.
Civitarese et al. [15] proposed a knowledge-based approach based on active learning to collaboratively refine a probabilistic semantic model in charge of recognizing activities in real-time. While this method does not require training data, the human engineering effort that is required to build a comprehensive semantic model is significant. Ye et al. [53] proposed a cross learning technique to train each personal model on its own data and complement each other’s labeling by querying the uncertain data to other personal models. This technique works on heterogeneous environments that have different sensor deployments with diverse sets of activities. While the above mentioned methods mitigate the data labeling problem thanks to collaboration, they did not take into account the privacy issues that emerge when sharing personal data.
On the other hand, federated learning is a more generic collaborative learning framework that explicitly considers privacy aspects [52]. In particular, only the parameters of local models are shared and aggregated in a privacy-preserving manner to build a collaborative model, thus avoiding the sharing of private data. Federated learning has been recently applied to activity recognition [11, 47, 55]. The major drawback of federated learning is that the collaborative model takes into consideration a predetermined set of activities that is fixed and does not change with time.
Privacy preservation in collaborative learning
Several works have considered the privacy issues arising by sharing data or model parameters as part of collaborative learning. Among them, Liu et al. propose a collaborative learning system that train a cloud model by sharing users’ data in a privacy-preserving manner [31]. The system perturbs the training data on the mobile device using lightweight transformations to preserve the privacy of the individual training samples. The idea is to approximately reconstruct the association between encrypted feature vectors and labels using regression, without compromising the privacy of the original feature vectors.
Osia et al. have proposed a deep private feature extraction approach that extract features that maximize the mutual data with the primary information and minimize the correlation with the sensitive information [37]. In this way, the features extracted will not disclose private or identifiable information but still result in high accuracy in classification tasks. They apply the technique over face image datasets and remove identity information, but extract other features that would be effective for classification such as age, gender and facial traits.
Shokri et al. have designed a model sharing technique based on deep learning [46]. The idea is to build a personal model using each user’s own data and then, during this personal network training, the system performs a selective sharing of model parameters (gradient). This parameter sharing, interleaved with local parameter updates during stochastic gradient descent, allows users to benefit from other users’ models without explicit sharing of training data inputs. They also extend their system with a differential privacy technique. The overall objective is to collaboratively train a neural network that can be used privately and independently by each participant. Each user then downloads a subset of the parameters from the server and uses them to update the local model.
Our collaborative learning approach takes inspiration from the above model sharing approach [46], with three main differences: a) by using GWR as a learning model, what our users share are neurons from their private GWR, instead of gradient based parameters, b) the choice of neurons to be shared is influenced both by the user’s privacy preferences, and by a randomized process, c) there is no assumption on the complete homogeneity of the set of activities considered by the participating users.
CollAR
Overall architecture
The main objective of
In
The shared model in the cloud needs an initialization phase. During this phase, a limited set of users is required to acquire a small labelled dataset to create personal activity models. Once the personal models are built, these users share some of the personal parameters to initialize the collaborative model.
Once the shared activity model is ready, new users can start using
Since personal models incrementally evolve over time thanks to active learning, each user can periodically contribute in updating the collaborative activity model by sharing small portions of the updated personal model.
The combination of collaborative and active learning allows our system to significantly reduce the reliance on training data and, at the same time, to reach satisfactory recognition rates.

Workflow of
The data workflow of
In the following, we describe each component of the system in details. We introduce the background of GWR neural network in Section 3.2. We show how we use GWR to train personal models in Section 3.3. Section 3.4 describes how to integrate personal models to train the shared model. Section 3.5 shows our semi-supervised approach to continuously update personal models. Finally, Section 3.6 presents our privacy-preserving parameter sharing strategy.
A Growing When Required (GWR) network is an incremental neural network capable of dynamically adding and removing neurons whenever the current state of the model does not sufficiently match the input data [33]. A GWR network is represented by a collection of neurons and edges. Similarly to self-organising maps [26], each neuron j is associated with a weight vector
During the training phase, when a new feature vector is available, the GWR algorithm analyzes if it is sufficiently matched by the current network. First, given the current input
In the literature [14] this set is often considered a singleton since it is very unlikely in practice that multiple neurons with the exactly the same distance exist.
The network activation at the current time t is computed as the distance d between the current input and the weight vector of the best matching neuron:
If d exceeds the activation threshold ACT_T, a new neuron is added to the network. The new neuron is associated with a weight vector that is the average of the input feature vector and the weight vector
On the other hand, if the distance does not exceed the threshold, the weight vector
The GWR algorithm also creates an edge between the best and the second best matching neuron based on input data. The second best is computed as shown before but excluding b from the considered neurons. Each edge is associated with an age factor. Whenever the algorithm considers a new input, the age of all edges is incremented; the exception is that the age of both the edge between the best and second best neurons and of the edge between the new neuron (if created) and the best is assigned to zero. Depending on their age, neurons and edges can be dynamically added and/or removed, thus reflecting the changes of input data. Hence, the network is growing as required. More details about the GWR algorithm can be found in [33].
The algorithm that we propose to train and update personal GWR models from labelled data in
As we explained in Section 3.2, a new neuron n is added to the network when a new input data labelled as l is far away from the best matching neuron. In this case, the network will store
The frequencies of each neuron are continuously updated with incoming labelled input data. Suppose that a new input data labelled as

Updating the personal GWR network with labelled data points
Once the network is trained, it can be used to classify unlabelled data points. For each unlabelled input data we derive a posterior probability distribution over the possible activities. The most likely activity in the distribution is the output of the classification.
The standard strategy proposed in the literature is to compute the probability distribution from the frequencies of the best-matching neuron [41]. However, during preliminary experiments we observed that the best matching neuron alone is not completely representative for the input data. Hence, differently from [41], we average the probability distribution obtained from the frequencies of the best matching neuron with two other probability distributions:
top-k: we select the top-k best matching neurons to the input. For each one of these neurons, we obtain a probability distribution from the frequencies. We compute the average of the probability distributions weighted by the distance.
best matching and its neighbours: we compute the average of the probability distributions of the best-matching neuron and its neighbours (i.e., the neurons directly linked to the best-matching neuron).
The participating users in
The shared GWR model requires a bootstrap phase in order to be initalized. A restricted number of volunteers acquire a small amount of labelled data to initialize their personal GWR model using Algorithm 1. Subsequently, these users share a portion of their neurons to the cloud model using a sharing approach that will be described in Section 3.6.
New users of
Periodically, the participating user may contribute with a personal update to further improve the shared model by sharing a portion of the neurons from their personal GWRs. The sharing mechanism is the same one used for model initialization and it will be presented in Section 3.6.
The shared model is initialized/updated with a subset of neurons from the personal GWRs of multiple users. Each shared neuron is associated with its weight vector and frequencies. The shared GWR model is initialized/updated using Algorithm 2. This algorithm is very similar to Algorithm 1. The main difference is that the input is a set of shared neurons, instead of labelled data. For each shared neuron j, we compute the distance between j and its best matching neuron b. If the distance is below the activation threshold, we combine the frequencies of b with the frequencies of j mitigated by the distance between b and j. Otherwise, we directly include j in the GWR network and we link it to b. In this latter case, the corresponding weight vector is the average between
Note that, to reliably update the shared model, the cloud service triggers an update process only when a sufficient number of shared neurons are available. At the end of the training process, the cloud service notifies the participating users that a new version of the shared model is available. When participating users are notified about a new available update of the shared model, they may decide to download it to improve their personal model. The local integration of the shared model update is performed using Algorithm 2.

Update GWR by neuron sharing
Each user can further personalize the personal GWR model adopting a semi-supervised strategy based on active learning. For each unlabeled feature vector, the current personal GWR model is used to obtain a probability distribution over the possible activities. If the probability of the most likely activity is lower than a threshold ACTIVE_LEARNING_T, we assume that there is uncertainty in the prediction and hence we ask the user to provide a feedback about the activity that he/she is actually performing. The feedback is associated with the feature vector. When a sufficient number of new labeled feature vectors is collected through active learning (according to the ACTIVE_LEARNING_BATCH_SIZE parameter), they are provided as new labeled examples to the personal GWR (using Algorithm 1).
Since personal models evolve over time, the participating users may periodically share with the cloud a portion of their neurons to collaboratively improve the shared GWR model, using the strategy presented in Section 3.6. The update can be performed periodically; e.g., daily, weekly, monthly, etc.
Privacy-preserving neuron sharing
Users may have privacy concerns in releasing to untrusted parties information about certain activities that they perform. In our architecture, the cloud service provider, as well as the other participating users, are considered honest but curious adversaries: they follow the protocol but they may try to infer from the data that they observe information about the activities performed by a given user (e.g., type of activity, when it was performed, and possibly even how it was performed).
As opposed to a data sharing model that would release directly data or feature vectors from which the cloud provider may infer the activity and possibly even finer grain information,
Despite neurons are only model parameters, they have associated data on the position in the feature space (i.e., the weight vector) and on the frequency of activation for the different activities. Hence, an adversary may indeed exploit this data to infer sensitive information.
Several privacy protection techniques may be applied to mitigate this privacy violation risk, including for example, perturbing weight vectors and frequencies in order to achieve differential privacy following an approach similar to the one proposed in [46]. In practice, this means that for hiding activity
We take a simpler approach in favor of utility, exploiting the observed robustness of the system even when sharing only a small portion of the neurons, and the property of
Based on user privacy preferences, we select the neurons that should be shared from a copy of the personal GWR network that has never seen any data about private activities. We call this network the private twin of the personal GWR. More precisely,
Each user provides a set
When a personal GWR is created, its private twin
During the supervised initialization (Algorithm 1) the two networks are trained in parallel except that feature vectors labeled with
When the user shares model parameters with the cloud, the neurons are taken from the private twin network that has never been exposed to private data. Sharing all the neurons of
Each shared neuron is only associated with its original weight vector and frequencies.
When a user receives an update from the cloud, the neurons are integrated both in the personal GWR and in its private twin in the same way (using Algorithm 2).
While this technique avoids sharing of information directly connected with a private activity, some attacks based on background knowledge on the correlation between private and non-private activities may still be performed. These attacks are naturally mitigated by the absence of precise temporal information at the cloud side (the sharing is performed in batches) and may also be contrasted by using the same knowledge at the user side, considering as private also the correlated activities.
Clearly, if all the participating users consider activity A as private the globally initialized network will not be able to recognize A and the parameters integrated by the cloud at each update step will not help anyone in improving their ability to recognize A. This is anyway expected if nobody wants to share information about that activity, but privacy preferences often differ among users and this problem may be mitigated if the system has many participants. The effect of this privacy-preserving strategy on performance will be illustrated in Section 4.5.
Experiments and evaluation
In this section, we evaluate the effectiveness of online collaborative learning – How effective is heterogeneous sets of activities – How much utility (i.e., recognition rate)
Datasets
In order to validate the effectiveness of
PAMAP2 contains data collected from 3 Colibri wireless inertial measurement units (IMU) from 9 users performing 12 activities. Each user wore the sensors on their wrist of the dominant arm, chest, and ankle on the dominant side. From this dataset we excluded a user (identified as 9 in the dataset) that only performed one activity. Also, we excluded 4 activities that were insufficiently represented in the dataset. Hence, in this work we consider the following 8 activities from PAMAP2: lying, sitting, standing, walking, ascending stairs, descending stairs, vacuum clearning, and ironing. Each of those activities was performed approximately for 3 minutes by each user, except for ascending/descending stairs due to the limitations of the building where the activities were carried out. More information about this dataset can be found in [44].
DSADS contains data collected from 8 users wearing accelerometer units in 5 positions: torso, right arm, left arm, right leg, and left leg. This dataset includes the following 19 activities: sitting, standing, lying on the back, lying on the right side, ascending stairs, descending stairs, standing in an elevator still, moving around in an elevator, walking, walking on a treadmill, walking on a treadmill in inclined position, running on a treadmill, exercising on a stepper, exercising on a cross trainer, using an exercise bike (horizontal), using an exercise bike (vertical), rowing, jumping, and playing basketball. Each user in the dataset performed each activity for 5 minutes. Hence, we did not exclude any user or activity in our experiments when using DSADS. More information about this dataset can be found in [5].
Implementation details and hyperparameter tuning
We implemented
Table 1 summarizes the many hyperparameters of
Hyperparameters in CollAR
Hyperparameters in
In order to evaluate the effectiveness of GWR as activity classifier, we compared it with other classifiers. In particular, we evaluated if the supervised approach proposed in Algorithm 1 reaches recognition rates that are comparable with other classifiers commonly used for activity recognition. We also evaluate if GWR reaches acceptable results even if trained with a limited amount of data. Note that in this experiment we evaluated GWR as a supervised classifier for activity recognition, while the experiments on collaborative learning are presented in the following subsections.
For each user, we split data into training (80%) and testing (20%). Then we use p% of training data to train the classifier, varying p from 10 to 100 with step size 10. We iterated this process for each user and averaged the resulting macro F1-scores. We selected a large number of supervised learning techniques for comparison, including SVM (with RBF kernel), Random Forest, and Multi-Layer Perceptron (MLP).2
We use the implementation of these techniques from the Python Scikit-learn library [43].
Figure 2 shows the results on both datasets. GWR achieves comparable results to the state-of-the-art techniques, leading to F1_Scores slightly lower than the best performing techniques. However, with respect to the compared classifiers, the advantage of GWR is the ability to adapt to new samples and learn new activities, which is desirable for long-term sustainable collaborative activity learning. Hence, the small drop in F1 score is acceptable considering the advantages of GWR.

F1-scores of
The key contribution of
We adopted the same evaluation methodology for all the following experiments. We have experimented with different neuron sharing percentages on p% (see Section 3.6) ranging from 10% and 90% to obtain the lowest value that provided robust results. Indeed, it is preferable that each user only shares as little neurons as possible while achieving satisfying recognition rates. In the end, we selected
In order to demonstrate the advantage of
personal model – This baseline is useful to compare the recognition rate of a personal GWR model (i.e., only trained with data from the left out user) with the
data sharing model – This baseline is useful to compare the recognition rate of a collaborative system that is based on sharing data instead of model parameters with the recognition rate of
Comparison of F1-scores between CollAR , personal training and sharing data. AL (Active learning), Q (Query rate)
Comparison of F1-scores between
Table 2 compares the recognition accuracy of
While the recognition rates reported in Table 2 may seem not completely satisfactory, it is important to note that the considered datasets only include a rather low number of users (both PAMAP2 and DSADS involve data from 8 users). We expect that our collaborative approach would show its benefits (in terms of recognition rate) by considering a significantly larger number of participating users.
In the following, we provide insights about the recognition rate of

PAMAP: confusion matrix. ac1 = lying, ac2 = sitting, ac3 = standing, ac4 = walking, ac5 = ascending stairs, ac6 = descending stairs, ac7 = vacum cleaning, ac8 = ironing.

DSADS: confusion matrix. ac1 = sitting, ac2 = standing, ac3 = lying on the back, ac4 = lying on the right side, ac5 = ascending stairs, ac6 = descending stairs, ac7 = standing in an elevator still, ac8 = moving around in an elevator, ac9 = walking, ac10 = walking on a treadmill, ac11 = walking on a treadmill in inclined position, ac12 = running on a treadmill, ac13 = exercising on a stepper, ac14 = exercising on a cross trainer, ac15 = exercise bike (horizontal), ac16 = exercise bike (vertical), ac17 = rowing, ac18 = jumping, ac19 = playing basketball.
One of the key reasons that we advocate GWR in
This flexibility is useful both in the case of heterogeneous sets of activities due to different user routines and in the case of heterogeneous sets of activities due to privacy preservation following the strategy described in detail in Section 3.6. In both cases the neurons shared by each user may not contain any information about some activities.
In order to estimate the drop in performance observed in these cases, we perform a leave-one-subject-out cross-validation and, at each fold, we randomly assign to each user a subset of hidden activities (either private and hence not considered in the private twin network, or not even performed and hence not considered in both of the personal GWR and its twin). For each of the N-1 users, we train a personal model with all of his/her available data. Note that in both cases of activities hidden because private or hidden because not performed, the resulting private twin network will be the same, hence the same neurons and associated data are shared to the cloud model. Based on these sets of shared neurons,
In order to understand the impact, we compare the results with the ideal condition in which each user contributes to the whole set of activities

Comparison of performance with different numbers of hidden activities.
As we can see, there is no significant drop in F1-scores on
In Fig. 5(b), we witness a performance drop on DSADS dataset, when only half of the activities are being used. After inspecting the results, we find that the drop may be due to the fact that DSADS has more classes to learn; i.e., 19 activities vs 8 in PAMAP2. In principle, the less classes each user contributes, the more difficult the learning task; that is, we need to complement with a large number of classes from the other users. So when each user only contributes half of the activities, the cloud model needs to find another half from the other users. In DSADS, we need to find 10 new activities while in PAMAP2 we only need to find other 4. Considering the limited number of users, and consequently the limited number of personal models covering the missing activities, the above observation explains the drop of performance on the dataset with the larger number of activities.
In the following, we discuss the strengths and the limitations of
Flexibility
The results presented in Section 4.5, and especially the ones reported in Fig. 5, indicate that
While these results are encouraging, they still need to be confirmed on datasets with a larger number of users. Another limitation of our experiments is that we randomly sampled the shared activities for each user based on a uniform distribution, while in real-world scenarios the actual distribution may be influenced by user profiles and personal habits.
Privacy properties of CollAR
However, we did not evaluate if the collaborative model can still reveal some sensitive information about participating users. For instance, an adversary may use domain specific background knowledge on the correlation between private and non-private activities to infer hidden activities given the ones that are shared. An adversary could also try to analyze the properties of shared neurons (e.g., the weight vectors) to reconstruct users’ raw sensor data. In order to mitigate these potential problems,
We also want to mention that the privacy properties of
Impact of active learning
In order to fine-tune the shared model on each user,
Heterogeneous sensing setups
In this work, we make the assumption that participating users that contribute to the shared model are actually monitored with the homogeneous sensing setups. However, different users may use different types of wearable sensors positioned in different parts of the body (e.g., a user may use his/her smartphone in the pocket while another one the smart-watch on the wrist). In order to make another example, when considering smart-home activity recognition, different homes may have very different configurations.
As we mentioned in Section 2.3, transfer learning tackles the orthogonal problem of adapting a recognition model trained on a specific sensor settings to recognize activities monitored with different sensing solutions. In future work we will investigate how to include transfer learning in
Need for evaluation on a larger scale
We evaluated
Conclusion and future work
We presented
Our results indicate that
Besides the many possible future research directions that we previously presented in Section 5, we will also explore if and how alternative self-organizing neural networks (as e.g., [50]) can be effectively adopted in our collaborative framework.
Conflict of interest
The authors have no conflict of interest to report.
