ContrasGAN: Unsupervised domain adaptation in Human Activity Recognition via adversarial and contrastive learning

May combine two claims.

In B-GAN, sitting and standing activities are difficult to separate (accuracy 0.54 and 0.71), while ContrasGAN recognizes them well (accuracy 0.95 and 0.99).

“as presented in confusion matrices in Figure 4a and 4b, it is difficult to separate sitting and standing activities in B-GAN (with accuracy of 0.54 and 0.71), while both are well recognised in ContrasGAN (with accuracy of 0.95 and 0.99).”

May combine two claims.

On cross-user experiments, the difference between source and target data is small, and ContrasGAN has achieved the highest macro-F1 scores.

“On cross-user experiments, the difference between source and target data is small, Con-trasGAN has achieved the highest macro-F1 scores.”

May combine two claims.

GFK cannot cope with high heterogeneity well and only predicts one activity.

“Thirdly, GFK cannot cope with high heterogeneity well and only predicts one activity.”

May combine two claims.

6 out of 9 techniques perform better on tasks of same body part and different sides, achieving higher F1 scores than the other two types of tasks.

“The results show that 6 out of 9 techniques perform better on the tasks of same body part and different sides, and have achieved higher F1 scores than the other two types of tasks.”

The quote only partly states this claim.

DeepCORAL has obtained the best median macro-F1 scores.

“DeepCORAL that under-performed on the other experiments has obtained the best median macro-F1 scores.”

May combine two claims.

The training time of DANN on the W-C task is 1784 minutes, nearly 50 times the training time of ContrasGAN and significantly higher than all other techniques.

“The training time of DANN on the W-C task is 1784 mins, nearly 50 times of the training time on ContrasGAN and significantly higher than all the other techniques.”

May combine two claims.

ContrasGAN achieves the best macro-F1 scores on 11 out of 15 tasks without any negative transfer.

“ContrasGAN achieves the best macro-F1 scores on 11 out of 15 tasks without any negative transfer.”

May combine two claims.

ContrasGAN outperforms all these techniques and achieves averaged micro- and macro-F1 scores as 0.81 and 0.80, which is higher than the second best technique (ADADM) by 0.04 and 0.07 respectively.

“ContrasGAN outperforms all these techniques and achieves the averaged micro-and macro-F1 scores as 0.81 and 0.80, which is higher than the secondbest technique (ADADM) by 0.04 and 0.07 respectively.”

May combine two claims.

ShiftGAN, DAN, JAN (a variation of DAN), and ADADM achieved better overall performance.

“ShiftGAN, DAN, JAN (a variation of DAN), and ADADM have achieved better overall performance.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

ADADM seems not stable and the performance might be affected by the mixup ratio.

“Secondly, ADADM seems not stable and the performance might be affected by the mixup ratio.”

May combine two claims.

ContrasGAN outperforms ShiftGAN by 0.12 in micro-F1 and 0.15 in macro-F1 scores.

“In addition, ContrasGAN outperforms ShiftGAN with 0.12 and 0.15 in micro-and macro-F1 scores”

May combine two claims.

Among the comparison techniques, DeepCORAL and JAN do not work well on the accelerometer data and produced worse accuracy than the classic techniques TCA and GFK.

“Among the comparison techniques, DeepCORAL and JAN do not work well on the accelerometer data and have produced worse accuracy than the classic techniques TCA and GFK.”

May combine two claims.

Most techniques can produce high accuracy as the transfer learning challenge is less significant, and ContrasGAN has consistently gained better accuracy on individual classes.

“Most of the techniques can produce high accuracy as the transfer learning challenge is less significant, and ContrasGAN has consistently gained better accuracy on individual classes.”

May combine two claims.

On cross-sensor experiments, when dealing with the most challenging tasks in heterogeneous feature spaces and imbalanced class distribution, ContrasGAN can still produce consistently good performance, while some existing techniques have less stable performance and are more sensitive to the difficulty level of tasks.

“On cross-sensor experiments, when dealing with the most challenging tasks in heterogeneous feature spaces and imbalanced class distribution, ContrasGAN can still produce consistently good performance, while some of the existing techniques have less stable performance and are more sensitive to the difficulty level of tasks.”

May combine two claims.

On DSADS, DANN suffers high fluctuation in consistent accuracy, but slightly lower than ContrasGAN.

“On DSADS, DANN suffers high fluctuation in consistent accuracy, but slightly lower than ContrasGAN.”

May combine two claims.

Both ContrasGAN and ShiftGAN are better than the other methods, including the single GAN based approach ADADM.

“Both ContrasGAN and ShiftGAN are better than the others, including a single GAN based approach -ADADM”

May combine two claims.

ContrasGAN obtains the highest averaged micro- and macro-F1 scores of 0.83 and 0.81, which are 0.05 and 0.07 higher than DAN, the second best performing technique.

“It obtains the highest averaged micro-and macro-F1 scores: 0.83 and 0.81, which is 0.05 and 0.07 higher than DAN, the second best performing technique.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Both generators and discriminators converge around 2000 epochs.

“Figure 13a presents the losses of Bi-GAN, which shows that both generators and discriminators converge around 2000 epochs.”

May combine two claims.

DANN has achieved the highest micro-F1 score 0.72, which is 0.08 higher than ContrasGAN.

“DANN has achieved the highest micro-F1 score 0.72, which is 0.08 higher than ContrasGAN.”

May combine two claims.

ContrasGAN with contrastive learning demonstrated superior performance on separating similar activity groups, though it still faces difficulty differentiating subtle activities like different cycling modes.

“ContrasGAN with contrastive learning has demonstrated superior performance on separating these groups of activities, even though it also faces the challenge of differentiating subtle activities like different cycling modes.”

May combine two claims.

DeepCORAL performs the best among all comparison techniques and is very comparable to ContrasGAN.

“Overall, DeepCORAL performs the best among all the comparison techniques and is very comparable to ContrasGAN.”

Classified as probably not a substantial claim.

Shift-GAN applies a bi-directional GAN with Kernel Mean Matching to transform feature spaces between source and target domains.

“In our previous work, we have proposed an unsupervised domain adaptation technique, Shift-GAN [38], that applies a bi-directional GAN with Kernel Mean Matching to transform the feature spaces between the source and target domains.”

• cites [38] Xinlei Chen, Kaiming He, Exploring simple siamese representation learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 15750–15758. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Bi-GAN has achieved success in computer vision applications such as image-to-image translation and promising results in domain adaptation in HAR.

“This type of technique has achieved success in many computer vision applications such as image-to-image translation [56,57], and promising results in domain adaptation in HAR [38].”

• cites [56] Roggen, 2010, Collecting complex activity datasets in highly rich networked sensor environmentsNot checked yet

  Not checked in the free run.

• cites [57] Stisen, 2015, Smart devices are different: Assessing and MitigatingMobile sensing heterogeneities for activity recognitionNot checked yet

  Not checked in the free run.

• cites [38] Xinlei Chen, Kaiming He, Exploring simple siamese representation learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 15750–15758. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Maximising inter-class domain discrepancy pushes representations further away from the decision boundary so samples from different classes are well separated.

“The inter-class domain discrepancy is maximised to push the representations of each other further away from the decision boundary so that the samples from difference classes are well separated.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Bi-GAN systems enable more robust adaptation by enforcing stronger, bi-directional constraints between two GANs.

“Bi-GAN systems enable more robust adaptation by enforcing stronger, bi-directional constraints between two GANs.”

May combine two claims.

ContrasGAN uses bi-directional generative adversarial networks for heterogeneous feature transfer and contrastive learning to capture distinctive features between classes.

“ContrasGAN uses bi-directional generative adversarial networks for heterogeneous feature transfer and contrastive learning to capture distinctive features between classes.”

Probably not a substantial claim: may not pass S2 (checkable).

To address high heterogeneity between domains, an UDA technique that does not make strong assumptions about feature spaces or probability distributions is needed.

“To address the high heterogeneity between domains, we need an UDA technique that does not make strong assumptions about feature spaces or probability distributions.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Most HAR systems have different sensor deployments and host users with diverse lifestyles, leading to heterogeneous feature spaces and disparate prediction functions.

“This might not be sufficient as most of HAR systems have different sensor deployments (resulting in highly heterogeneous feature spaces between the source and target domains) and host users with diverse lifestyles (leading to disparate prediction functions).”

May combine two claims.

The cyclic constraints in Equations (1) and (2) ensure original input samples can be recovered by the two generators, leading to more robust domain adaptation.

“The cyclic constraints in Equation ( 1) and ( 2) make sure the original input samples can be recovered by the two generators: this is why Bi-GAN will lead to more robust domain adaptation.”

May combine two claims.

Shift-GAN achieved promising results on several HAR datasets and outperformed state-of-the-art transfer learning techniques.

“The technique has achieved promising results on several HAR datasets and outperformed the state-of-the-art transfer learning techniques.”

• cites [38] Xinlei Chen, Kaiming He, Exploring simple siamese representation learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 15750–15758. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Adapting contrastive loss with batch normalisation balances the magnitude of different loss components, which improves convergence of training and recognition accuracy.

“Also we have adapted contrastive loss with batch normalisation to balance the magnitude of different loss components, which improves the convergence of training and recognition accuracy.”

Classified as probably not a substantial claim. May combine two claims.

Adding contrastive learning during adaptation minimises intra-class discrepancy and maximises inter-class margin.

“To tackle these problems, we add contrastive learning during the adaptation with the goal to minimise the intra-class discrepancy and maximise the inter-class margin.”

Probably not a substantial claim: may not pass S4 (based).

Minimising intra-class domain discrepancy draws closer feature representations of samples within a class.

“The intra-class domain discrepancy is minimised to draw closer the feature representations of samples within a class so that the samples of the same class are close.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Micro-F1 and macro-F1 are appropriate metrics for domain adaptation and HAR, with macro-F1 for imbalanced datasets.

“These two metrics are often used in domain adaptation and HAR in general [4,38], with macro-F1 score is for imbalanced datasets [18,31].”

• cites [4] Bicocchi, 2010, Detecting activities from body-worn accelerometers via instance-based algorithms, Pervasive Mob. Comput. sourceNot checked yet

  Not checked in the free run.

• cites [38] Xinlei Chen, Kaiming He, Exploring simple siamese representation learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 15750–15758. sourceNot checked yet

  Not checked in the free run.

• cites [18] Pan, 2011, Domain adaptation via transfer component analysisNot checked yet

  Not checked in the free run.

• cites [31] Zou, 2013, Contrastive learning using spectral methodsNot checked yet

  Not checked in the free run.

May combine two claims.

Contrastive loss helps learn discriminative features and improve class-level recognition accuracy.

“This indicates that the contrastive loss help learn discriminative features and improve class-level recognition accuracy.”

Classified as probably not a substantial claim.

At the end of training Bi-GAN, two generators can transform samples from one domain to the other.

“At the end of training Bi-GAN, we will have two generators that can transform samples from one domain to the other.”

May combine two claims.

Contrastive learning is needed to better align the samples, which is the key contribution of the work.

“That's why we need to use contrastive learning to better align the samples, which is the key contribution of our work.”

May combine two claims.

DANN is the most expensive method because it performs adaptation with feature learning and employs an end-to-end training regime, which takes longer to converge.

“DANN is the most expensive as it performs adaptation with feature learning and employs an end-to-end training regime, which takes longer to converge.”

May combine two claims.

The difficulty of transferring activity models depends on different body parts more than different sides between the same body part: body parts, even though on different sides, have similar motion patterns.

“This suggests that the difficulty of transferring activity models depends on different body parts more than different sides between the same body part: body parts, even though on different sides, have similar motion patterns.”

May combine two claims.

ContrasGAN requires only one well-annotated domain and can transfer the activity model from this domain to many other unlabelled domains.

“Our technique requires only one well-annotated domain and can transfer the activity model from this domain to many other unlabelled domains.”

Classified as probably not a substantial claim.

We follow state-of-the-art UDA evaluation methodology by performing stratified split on target domain data into 60% training, 20% validation, and 20% testing.

“For each run, we follow the state-of-the-art UDA evaluation methodology [9,44,54] by performing stratified split on the target domain data into 60% for training, 20% for validation, and 20% for testing.”

• cites [9] Cook, 2013, Transfer learning for activity recognition: A survey, Knowl. Inf. Syst. sourceNot checked yet

  Not checked in the free run.

• cites [44] Bashmal, 2018, Siamese-GAN: Learning invariant representations for aerial vehicle image categorization, Remote Sens. sourceNot checked yet

  Not checked in the free run.

• cites [54] Plötz, 2021Not checked yet

  Not checked in the free run.

May combine two claims.

With class-level alignment, ContrasGAN can achieve more balanced accuracy on each class, leading to much higher macro-F1 scores than all comparison techniques.

“It is also worth noting that with class-level alignment ContrasGAN can achieve more balanced accuracy on each class, leading to much higher macro-F1 scores than all the comparison techniques.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A user pair with high lower-bound accuracy suggests that these users' data have similar distributions, and thus the difficulty of transfer learning will be lower.

“For example, a user pair with a high lower-bound accuracy suggests that these users' data have similar distributions, and thus that the difficulty of transfer learning will be lower.”

May combine two claims.

The prediction discrepancy between two classifiers guides the generator to better map target samples onto the source domain and guides classifier update to adapt to transformed target samples.

“This discrepancy will guide the generator to better map target samples onto the source domain and thus minimise the discrepancy, and guide the classifier update to adapt to the transformed target samples.”

May combine two claims.

In the end, the improved target generator can accurately transform target samples to the source domain and the classifier can predict labels for transformed target samples.

“In the end, we will have improved target generator that can accurately transform target samples to the source domain and the classifier that can predict labels for transformed target samples.”

Classified as probably not a substantial claim.

Combining bi-directional GAN and contrastive learning has potential in learning distinctive feature transfer and thus leading to effective activity model transfer.

“Our work suggests the potential of combining bi-directional GAN and contrastive learning in learning distinctive feature transfer and thus leading to effective activity model transfer.”

May combine two claims.

The lower-bound accuracy indicates the difficulty level of adaptation tasks: the easiest is same body part between different sides, while tasks between different body parts are more difficult.

“The lower-bound accuracy indicates the difficulty level of these adaptation tasks: the easiest one is the same body part between different sides such as left arm to right arm, while the tasks between different body parts are more difficult.”

Probably not a substantial claim: may not pass S4 (based).

Contrastive learning has strength in learning intra-class compactness and inter-class separability.

“To do so, we will first look at contrastive learning that has the strength in learning intra-class compactness (i.e., grouping samples that share the same class labels) and inter-class separability (i.e., pushing apart samples with different class labels).”

Probably not a substantial claim: may not pass S3 (load-bearing).

We run each experiment 10 times and report the mean of macro and micro F1-scores as the final result.

“For each experiment, we run 10 times and report the mean of macro and micro F1-scores as the final result.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The approach has only been validated on three datasets.

“One limitation of our work is that we have only validated our approach on three datasets”