acceptodds
Under review as a conference paper at ICLR 2027

On Unlearning for Time-series Forecasting

Abstract

Time-series forecasting is widely used in healthy, transportation, finance, and many other domains. Models in these settings are often trained on longitudinal user or entity level records, which may later require removal because they contain sensitive or proprietary information or have been corrupted by sensor failures. To address such deletion requests without costly retraining, machine unlearning has been widely studied as a practical mechanism for privacy protection and data governance. However, the application of machine unlearning to time series prediction has not yet been well realized, it mainly dues to the following unique challenges: Gradient-based unlearning can be unstable because a deleted observation participates in multiple causally connected forecasting windows, causing parameter updates to propagate beyond the requested interval and degrade retained forecasting utility. Label-guided updating offers a more controlled alternative, but continuous and context-dependent forecasts lack a suitable replacement target, while the exact-retrained output is unavailable during unlearning. Moreover, the remaining support for a deleted temporal pattern is highly non-uniform. Some affected windows retain structurally similar counterparts in the retained data, whereas others become underrepresented or isolated. We present RDTU, a Residual Diffusion framework of Time-series Unlearning that constructs labels without accessing an exact-retrained model at unlearning time. RDTU first uses a retained-set neural tangent kernel predictor to obtain a deletion-compatible base forecast. Then it quantifies the global and local structural support of each affected window using the volume contribution of the retained-reference data. Conditioned on the query window, retained-set prediction, and structural support, a diffusion model generates a residual correction that estimates the counterfactual forecast, yielding a pseudo-label field that guides a lightweight model update. Experiments across real-world forecasting benchmarks show that RDTU consistently produces unlearned models that most closely match exact retraining among the evaluated efficient unlearning methods. It achieves this alignment while maintaining retained forecasting utility and reducing membership-inference signals.

Primary area:
learning on time series and dynamical systems
Keywords:
machine unlearning, time series, diffusion model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wSefOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses approximate unlearning for time series forecasting. The authors propose RDTU, which combines a retained-set NTK predictor, a structural irreplaceability score, and a conditional residual diffusion model to construct counterfactual labels, which then guide a lightweight unlearning update. Experiments on four datasets demonstrate improved functional agreement with exact retraining while maintaining forecasting utility and efficiency.

Critical strengths
  1. The problem to be addressed is a real need in the time series forecasting.
  2. The novelty of the proposed method is adequate.
Critical weaknesses
  1. The base forecast model used in the experiments is a Transformer. However, I cannot find any detailed introduction, neither in the main text nor in the appendix, about the structure and scale of the Transformer. The scale of the model influences the discrepancy of NTK and the performance of the diffusion module a lot for unlearning.
  2. If the proposed method is considered 'agnostic to the original model', experiments on different types of models, including LSTM, TCN, PatchTST, etc., should be conducted.
  3. One real-world motivation for unlearning is privacy, as mentioned in the Introduction. The experimental setting should align with this claim. Take the traffic dataset (which can be treated as multi-dimension time series forecasting)as an example, if one sensor is considered to have a privacy issue, then what ought to be removed from the training data is the whole corresponding channel, not some random time windows.
  4. The scale of test data is TOO SMALL, which brings 3 major issues: (a) Consider the time consumption for retraining in Table 3. If the longest time to retrain the model is less than 40 seconds, the practical benefit of learning is questionable. (b) In larger scale datasets, the error of Nystrom can be greater. In small datasets, this problem is hidden. (c) Larger datasets may require the diffusion to have more sampling steps, which may deteriorate the usability of the proposed method.

    Based on the critical flaws w.r.t. experimental settings, I tend to give a score between 1 and 2 in the first round. However, I am open to increase my score if the above mentioned concerns can be adequately addressed.
Questions for authors
  1. Why use a diffusion model instead of some more efficient generative models like flow matching or rectified flow? I think it is a good idea to investigate whether these models could reduce the sampling steps while maintaining similar performance in RDTU.
  2. Instead of using NTK, can we simply use a lightweight NN model, like DLinear or an LSTM, to form the base forecast?
  3. Please also refer to my concerns listed in the weaknesses.
10 comments
AuthorsOpenReview ↗

Response to Reviewer wSef [part 1]

Reviewer wSef W1

The base forecast model used in the experiments is a Transformer. However, I cannot find any detailed introduction, neither in the main text nor in the appendix, about the structure and scale of the Transformer. The scale of the model influences the discrepancy of NTK and the performance of the diffusion module a lot for unlearning.

Transformer: 44 layers, hidden dimension 128128, 88 heads, feed-forward dimension 256256, dropout 0.10.1, and GELU activation. It is trained with AdamW for at most 80 epochs using batch size 512512, learning rate 3×10−43\times10^{-4}, weight decay 10−410^{-4}.

LSTM: 22 layers, hidden dimension 256256, dropout 0.10.1, unidirectional recurrence, and a linear prediction head. It is trained with AdamW for at most 200200 epochs using batch size 512512, learning rate 10−310^{-3}, and weight decay 10−410^{-4}.

TCN: 44 residual blocks, 128128 channels, 22 causal convolutions per block, kernel size 33, dilation factors {1,2,4,8}\{1,2,4,8\}, dropout 0.10.1, weight normalization, and ReLU activation. It is trained with AdamW for at most 200200 epochs using batch size 512512, learning rate 10−310^{-3}, and weight decay 10−410^{-4}.

PatchTST: 22 encoder layers, hidden dimension 128128, 88 heads, feed-forward dimension 256256, patch length 44, stride 22, dropout 00, and GELU activation, with channel independence, RevIN, and a flatten-and-linear prediction head. It is trained with AdamW for at most 200200 epochs using batch size 512512, learning rate 10−310^{-3}, and weight decay 10−410^{-4}.

Reviewer wSef W2

If the proposed method is considered 'agnostic to the original model', experiments on different types of models, including LSTM, TCN, PatchTST, etc., should be conducted.

Thanks for the suggestion. We conducted supplementary experiments on different base models, and the specific results of LSTM are shown in the following table, results of TCN, PatchTST are added in Appendix. LSTM:

MethodAir Quality U/NAir Quality F-P GapAir Quality NRMSEAir Quality TimeTraffic U/NTraffic F-P GapTraffic NRMSETraffic TimeEEG U/NEEG F-P GapEEG NRMSEEEG TimeHAR70 U/NHAR70 F-P GapHAR70 NRMSEHAR70 Time
Original0.57324.730.2484--0.638414.760.3933--0.55321.590.0474--0.67308.560.2968--
Retrain0.54370.000.25459.2640.54330.000.38797.5610.52010.000.04889.2290.51800.000.300732.845
SSD0.545970.400.79351.4360.533082.980.99890.6990.519876.820.77120.5720.525779.950.56691.952
NegGrad+0.533820.550.35572.7940.538135.560.51982.0970.521859.390.52321.2300.5269217.951.113312.873
Zero-label0.545050.590.59411.4160.534850.560.67650.8300.518245.160.43940.1610.523359.050.35220.665
Random-label0.545550.590.59511.4200.534550.960.69320.7860.518245.440.43990.1610.520443.740.48820.724
SCRUB-R0.544547.770.56531.1260.541137.240.51200.7590.51916.090.07530.2040.526511.150.29761.425
TS-Unlearn0.53795.520.25982.3440.548315.180.41302.4950.52774.210.05561.8410.529810.920.31671.404
Ours0.53684.800.24882.59400.533913.970.39571.5210.51892.550.04870.9470.51929.400.30813.1775
AuthorsOpenReview ↗

Response to Reviewer wSef [part 2]

Reviewer wSef W3

One real-world motivation for unlearning is privacy, as mentioned in the Introduction. The experimental setting should align with this claim. Take the traffic dataset (which can be treated as multi-dimension time series forecasting)as an example, if one sensor is considered to have a privacy issue, then what ought to be removed from the training data is the whole corresponding channel, not some random time windows.

Thanks for the question. We designed experiments for both single-variable and multi-variable scenarios. For the single-variable data, we selected the complete data of 10 users for the experiment, and randomly selected all the data of one of the users for forgetting. For multivariate data, we also selected 10 users and randomly chose the data from a certain sensor/dimension of one of the users for forgetting. The detailed results are shown in the following table. It can be observed that our method has significant advantages and performance in this scenario.

MethodAir U/NAir F-P Gap (%)Air NRMSEAir Time (s)Traffic U/NTraffic F-P Gap (%)Traffic NRMSETraffic Time (s)
Original0.5988 ± 0.0666--0.2217 ± 0.0011--0.7319 ± 0.0621--0.3153 ± 0.0100--
Retrain0.5568 ± 0.0229--0.2250 ± 0.0048192.7964 ± 58.22050.6135 ± 0.0312--0.3163 ± 0.0058119.0407 ± 30.3364
SSD0.5613 ± 0.036359.6761 ± 21.28770.6040 ± 0.16510.6852 ± 0.10420.5628 ± 0.051678.3237 ± 45.06560.6812 ± 0.22570.5832 ± 0.1281
NegGrad+0.5571 ± 0.0402859.6122 ± 91.84907.3509 ± 0.770781.3413 ± 3.13150.5695 ± 0.0452861.6575 ± 138.95722.6960 ± 0.701733.8003 ± 1.5389
Zero-label0.5519 ± 0.026966.2707 ± 8.53500.6293 ± 0.052915.4022 ± 0.84560.5832 ± 0.062579.9763 ± 38.94630.3794 ± 0.04796.6967 ± 0.6467
Random-label0.5590 ± 0.025068.4220 ± 8.02510.6503 ± 0.051217.3528 ± 2.01060.5809 ± 0.055981.2705 ± 38.66830.3959 ± 0.04806.8879 ± 0.6449
SCRUB-R0.5531 ± 0.035224.3947 ± 0.93580.3234 ± 0.001397.3652 ± 5.06730.6251 ± 0.051830.8972 ± 9.03580.3624 ± 0.010347.1269 ± 2.1898
TS-Unlearn0.5598 ± 0.030270.3468 ± 7.22470.6664 ± 0.047416.6571 ± 2.39680.6133 ± 0.061381.0927 ± 35.60630.3869 ± 0.04497.6543 ± 1.2011
RDTU0.5507 ± 0.04376.1794 ± 1.33910.2293 ± 0.00443.4558 ± 0.16740.6083 ± 0.069323.1773 ± 9.98360.3267 ± 0.01385.2726 ± 0.1996
MethodEEG U/NEEG F-P Gap (%)EEG NRMSEEEG Time (s)HAR70+ U/NHAR70+ F-P Gap (%)HAR70+ NRMSEHAR70+ Time (s)
Original0.5670 ± 0.0737--0.0622 ± 0.0078--0.5981 ± 0.0471--0.1244 ± 0.0070--
Retrain0.5597 ± 0.0510--0.0580 ± 0.0083114.1573 ± 15.63180.5847 ± 0.0712--0.1219 ± 0.0071227.1485 ± 27.6805
SSD0.5548 ± 0.041084.4921 ± 2.58160.6527 ± 0.07760.4189 ± 0.09930.5407 ± 0.028426.5559 ± 10.46530.5155 ± 0.09710.6252 ± 0.1200
NegGrad+0.5658 ± 0.05231011.6810 ± 254.69424.6332 ± 1.133226.7890 ± 2.11310.5636 ± 0.0332656.0686 ± 353.61261.3259 ± 0.686062.9475 ± 1.3050
Zero-label0.5863 ± 0.051414.4248 ± 2.19410.0891 ± 0.00904.2419 ± 0.33460.5830 ± 0.071026.6644 ± 11.60350.1225 ± 0.007012.6703 ± 3.7334
Random-label0.5884 ± 0.079914.7301 ± 4.52210.0861 ± 0.01174.8149 ± 0.97150.5807 ± 0.072954.3932 ± 16.28760.1470 ± 0.009814.4579 ± 4.7531
SCRUB-R0.5646 ± 0.054315.8047 ± 2.63050.0988 ± 0.008530.9713 ± 1.57010.5813 ± 0.058218.2323 ± 6.51660.1306 ± 0.006976.7681 ± 3.9383
TS-Unlearn0.5899 ± 0.062415.4128 ± 5.26990.0901 ± 0.01014.9452 ± 0.91330.5802 ± 0.067045.1040 ± 13.55150.1389 ± 0.008214.2351 ± 3.7874
RDTU0.5597 ± 0.049114.2764 ± 10.48580.0629 ± 0.00944.4898 ± 0.23620.5721 ± 0.073416.5749 ± 6.08400.1233 ± 0.00776.7578 ± 0.1847
AuthorsOpenReview ↗

Response to Reviewer wSef [part 3]

Reviewer wSef W4

The scale of test data is TOO SMALL, which brings 3 major issues: (a) Consider the time consumption for retraining in Table 3. If the longest time to retrain the model is less than 40 seconds, the practical benefit of learning is questionable. (b) In larger scale datasets, the error of Nystrom can be greater. In small datasets, this problem is hidden. (c) Larger datasets may require the diffusion to have more sampling steps, which may deteriorate the usability of the proposed method.

Thanks for the question. We conducted experiments of two larger common time series forecasting benchmark datasets(Ettm1, Weather), the results of the transformer base model shows below and more results of other base model have been added to Appendex.

MethodETTm1 NRMSEETTm1 F-P Gap%ETTm1 U/NETTm1 TimeWeather NRMSEWeather F-P Gap%Weather U/NWeather Time
Original0.44886613.7600.829193—0.18028821.2970.842255—
Retrain0.4537080.0000.53076247.2910.2001020.0000.56268344.892
SSD0.84901429.3560.5016171.2090.29989641.2330.5902561.297
NegGrad+1.184357217.5110.71357736.9250.34750592.7230.68086240.030
Zero Label0.50225721.6760.7382973.0740.20858733.8760.7208563.259
Random Label0.52935422.3160.7271423.0250.20409934.0470.7241213.290
SCRUB-R0.47787233.1760.6717074.5920.20049531.7070.6879035.142
TS-Unlearn0.46990728.1200.78758217.8180.28004559.4790.67372122.547
Ours0.46166217.4590.52787810.2510.18613316.8200.66020213.914

We can find that our model still performs well on larger scale datasets, although the error of Nystrom is greater, the residual capture of diffusion can effectively make up for this part. In addition, we have optimized the implementation of diffusion in terms of efficiency, such as DDPM acceleration. Since time-series data is inherently not overly complex, the optimized diffusion can effectively mitigate the time impacts caused by multi-step sampling and other issues.

Reviewer wSef Q1

Why use a diffusion model instead of some more efficient generative models like flow matching or rectified flow? I think it is a good idea to investigate whether these models could reduce the sampling steps while maintaining similar performance in RDTU.

Thanks for the question. We conducted ablation experiments by replacing diffusion with flow matching and rectified flow, and the specific results are shown in the table below. It can be observed that, without significant differences in running time, flow matching and rectified flow did not achieve better performance than diffusion. On one hand, this is because in the diffusion part, we used DDPM to accelerate and improve the running efficiency of diffusion. On the other hand, we believe that diffusion can better capture this uncertain residual relationship.

MethodAir Quality U/NAir Quality F-P GapAir Quality NRMSEAir Quality TimeTraffic U/NTraffic F-P GapTraffic NRMSETraffic TimeEEG U/NEEG F-P GapEEG NRMSEEEG TimeHAR70 U/NHAR70 F-P GapHAR70 NRMSEHAR70 Time
NTK + Diffusion (Ours)0.50014.63120.260011.7100.54209.92510.39354.2610.52443.75260.11102.3220.51789.77030.34245.744
NTK + VP Flow Matching0.53835.33200.26868.9210.542110.32640.39617.4410.52683.92020.11022.8960.521610.57700.34916.822
NTK + Rectified Flow0.53145.30660.26688.9110.543110.32870.39557.3820.52893.80730.11002.8950.523210.40070.34866.870
Reviewer_cf3jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RDTU, a residual diffusion framework for approximate machine unlearning in time-series forecasting. It combines a retained-set neural tangent kernel (NTK) predictor, a Structural Irreplaceability Score (SIS), and conditional residual diffusion to construct pseudo-labels that guide model updates toward the behavior of exact retraining. Experiments on four public datasets show promising results in unlearning effectiveness and forecasting utility. However, the key assumption that residual patterns learned from the original training set transfer reliably to the retained-data setting remains insufficiently validated.

Critical strengths
  1. Well-motivated problem formulation. The paper identifies the difficulty of handling overlapping forecasting windows during data deletion and introduces strict temporal deletion closure to address it.

  2. Promising methodological design. Combining an NTK-based predictor with SIS-conditioned residual diffusion offers an interesting approach to constructing counterfactual labels without exact retraining for every deletion request.

Critical weaknesses
  1. Insufficient validation of the residual-transfer assumption. RDTU assumes that the conditional residual relationship learned under the complete-data condition remains transferable to retained-set conditions. Although the ablations support the usefulness of residual correction, they do not directly establish the stability of this relationship after deletion. Evaluations across deletion ratios, rare-pattern removal, and sequential deletion requests are needed, particularly because strict temporal deletion closure can amplify the fraction of affected training windows.

  2. Limited experimental coverage. Experiments use four datasets, with ten series or subjects per dataset and three deletion seeds per model, and evaluate only a Transformer forecasting backbone. Although the appendix includes multi-step results for horizons of 3 and 6, coverage remains limited for longer-horizon forecasting, higher-dimensional multivariate data, alternative architectures, and more challenging deletion scenarios.

  3. Empirical evidence does not establish formal unlearning guarantees. The paper evaluates predictive agreement with exact retraining and membership indistinguishability using candidate-conditioned LiRA. These are useful empirical checks, but they do not establish distributional equivalence to retained-only retraining or provide certified bounds on residual data influence. Robustness to adaptive, white-box, and reconstruction attacks remains untested.

  4. Insufficient characterization of computational scalability. The reported runtime includes NTK computation, diffusion training and sampling, and model updating, and the paper studies the effect of diffusion steps. However, it does not characterize runtime and peak memory as the number of training windows, model parameters, variables, or deletion requests increases. Separating one-time preprocessing from per-request and amortized costs would clarify the practical scalability of RDTU.

Questions for authors

See the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer cf3j [part 1]

Reviewer cf3j W1 [part1]

Insufficient validation of the residual-transfer ... particularly because strict temporal deletion closure can amplify the fraction of affected training windows.

Thanks for the suggestion. We conducted supplementary experiments on the deletion requirements at different proportions of 10% and 20%, and the specific results are shown in the table below.

MethodAir Quality U/NAir Quality F-P GapAir Quality NRMSEAir Quality TimeTraffic U/NTraffic F-P GapTraffic NRMSETraffic TimeEEG U/NEEG F-P GapEEG NRMSEEEG TimeHAR70 U/NHAR70 F-P GapHAR70 NRMSEHAR70 Time
Retrain0.551900.266414.5830.5426500.400499.1120.5191600.137217.8580.5164600.3510777.532
SSD0.545451.42320.66660.7970.5375568.93030.877870.6340.5224156.190.643570.4450.5237793.32630.584151.425
NegGrad+0.5317104.95531.06316.1920.5344288.64100.946192.6520.5301377.080.739181.8510.52458234.37130.8589021.713
Zero-label0.541956.47020.63444.5330.5381450.48130.686782.5130.5246947.280.479680.5670.5255965.07910.408953.032
Random-label0.542555.70020.63064.5640.5383951.52920.706312.5090.5256047.420.483610.5780.5174944.20430.536443.000
SCRUB-R0.5408475.28914.74984.8830.53499178.44041.790612.6470.5283241.400.417420.7550.5184654.50950.454354.460
TS-Unlearn0.540812.61750.30059.0490.5416411.92940.395284.8900.530687.890.123871.4490.5372111.18300.353425.991
Ours0.53135.62610.261111.9440.5377210.68090.388398.6800.525525.970.110892.8970.516069.66890.343818.108
MethodAir Quality U/NAir Quality F-P GapAir Quality NRMSEAir Quality TimeTraffic U/NTraffic F-P GapTraffic NRMSETraffic TimeEEG U/NEEG F-P GapEEG NRMSEEEG TimeHAR70 U/NHAR70 F-P GapHAR70 NRMSEHAR70 Time
Retrain0.538000.26838.2390.5389500.403457.6750.5231700.137875.9810.5176600.3576169.001
SSD0.541557.17760.71820.9140.5354868.75300.872610.6540.5291156.550.658500.4630.5224988.26250.580251.528
NegGrad+0.544258.08580.64113.9510.5253363.46000.753522.1890.5204456.860.559111.2550.52179200.21440.8891318.874
Zero-label0.538660.92770.69936.2810.5358450.18670.685992.9240.5288048.840.500090.8670.5236161.61100.421065.379
Random-label0.536559.44630.68946.2860.5316350.63910.705362.9080.5280449.010.502390.9530.5202544.05110.560395.378
SCRUB-R0.5439704.32367.00446.3730.52484215.90012.170753.0960.5243063.050.628220.9840.5179299.02730.759086.527
TS-Unlearn0.547918.52190.354012.3740.5431511.85650.396035.5740.519307.460.121432.0060.5359112.71950.3537010.455
Ours0.54566.68900.263512.1730.5276310.08160.389588.9800.524575.400.111033.2010.517319.81110.346198.452

A 20% raw-timestamp deletion request affects approximately 50% of the original training windows after strict temporal closure. This value already represent substantial effective forgetting ratios. The consistently strong performance of RDTU under such demanding settings empirically supports our assumption that, within a reasonable range of deletion requests, the conditional residual relationship learned under the complete-data condition remains transferable to the retained-set condition.

AuthorsOpenReview ↗

Response to Reviewer cf3j [part 2]

Reviewer cf3j W1 [part2]

We also conducted experiments of different pattern, we designed experiments with one-shot delection of 5% whole unlearning window, and the results shows below.

MethodAir Quality U/NAir Quality F-P Gap (%)Air Quality NRMSEAir Quality Time (s)*Traffic U/NTraffic F-P Gap (%)Traffic NRMSETraffic Time (s)*EEG U/NEEG F-P Gap (%)EEG NRMSEEEG Time (s)*HAR70+ U/NHAR70+ F-P Gap (%)HAR70+ NRMSEHAR70+ Time (s)*
Retrain0.5142600.00000.25995533.2890.6065100.00000.39427811.2350.5489430.00000.1112256.6440.5295680.00000.34935320.347
SSD0.52106691.37090.9459760.6300.56080781.35571.0040720.4130.56619256.61400.6570300.3070.54159574.25310.5472410.474
NegGrad+0.514421387.37863.17773011.6450.580729188.33341.3325273.5120.551429118.08280.9687251.6610.543921367.97651.6971745.678
Zero-label0.52108361.81870.6280430.8680.57200554.90880.6078980.4660.55910554.42680.5126060.2260.54116162.89520.3889940.477
Random-label0.52084963.38350.6404120.9410.56888050.46570.6047030.4740.55959253.99810.5118410.2320.55413552.83660.5022700.483
SCRUB-R0.5100996.19390.25896112.9580.59252629.56060.4672660.7230.55654219.60160.2181310.3600.5408315.99100.3436262.697
TS-Unlearn0.51124912.99940.27026216.9240.58880214.49830.3952986.9160.55406911.34600.1304333.3940.56113922.23850.35482111.393
Ours0.5111185.96290.25970311.3530.58567712.33980.3888766.9090.5541843.72790.1116594.7100.5372245.50200.34643612.246

Additionally, we designed the experiments of sequential deletion. Based on a 5% forgetting rate, we continuously issued 5 requests for 2% forgetting, and recorded the corresponding indicators of the forgetting model at each step. The results of the EEG dataset are shown in the table below. The results of other datasets are included in the Appendix.

MethodMetricRound 1 (7%)Round 2 (9%)Round 3 (11%)Round 4 (13%)Round 5 (15%)
RetrainU/N0.55390.54380.56180.55270.5598
RetrainF-P Gap (%)00000
RetrainNRMSE0.11450.11610.11890.12590.1245
RetrainTime (s)5.966.025.875.805.20
SSDU/N0.56090.56370.55140.55890.5626
SSDF-P Gap (%)85.7393.2395.9997.6798.34
SSDNRMSE0.89980.97141.00151.01621.0271
SSDTime (s)0.300.110.110.110.11
NegGrad+U/N0.56720.56780.54100.56060.5608
NegGrad+F-P Gap (%)149.29200.46218.76227.98239.78
NegGrad+NRMSE1.34301.77842.00062.07032.2026
NegGrad+Time (s)1.441.351.351.361.14
Zero-labelU/N0.56740.56050.54310.55000.5547
Zero-labelF-P Gap (%)46.1949.1648.3749.9649.80
Zero-labelNRMSE0.44690.46020.47110.47960.4903
Zero-labelTime (s)0.240.240.240.240.24
Random-labelU/N0.56200.53780.54310.54910.5543
Random-labelF-P Gap (%)47.6348.6850.7751.2351.12
Random-labelNRMSE0.46210.45870.49350.49220.5035
Random-labelTime (s)0.250.250.250.250.24
SCRUB-RU/N0.56390.55640.55190.55190.5606
SCRUB-RF-P Gap (%)18.0522.4019.5520.2323.44
SCRUB-RNRMSE0.21120.24430.22180.21920.2520
SCRUB-RTime (s)0.430.360.360.360.34
TS-UnlearnU/N0.55870.55660.55680.55770.5694
TS-UnlearnF-P Gap (%)10.2813.9615.9820.3720.89
TS-UnlearnNRMSE0.14130.15860.17330.19710.2118
TS-UnlearnTime (s)3.092.712.712.792.35
OursU/N0.55590.55310.54340.54780.5544
OursF-P Gap (%)4.014.464.735.214.64
OursNRMSE0.11280.11300.11420.11410.1145
OursTime (s)1.861.851.861.831.83
AuthorsOpenReview ↗

Response to Reviewer cf3j [part 3]

Reviewer cf3j W2

Limited experimental coverage. Experiments use four datasets, with ten series or subjects per dataset and three deletion seeds per model, and evaluate only a Transformer forecasting backbone. Although the appendix includes multi-step results for horizons of 3 and 6, coverage remains limited for longer-horizon forecasting, higher-dimensional multivariate data, alternative architectures, and more challenging deletion scenarios.

Thanks for your suggestion. We conducted the experiments of H=12 for the four datasets of the paper, and the results shows below. We can observe that our model performs well with longer-horizon forecasting.

MethodAir Quality U/NAir Quality F-P Gap (%)Air Quality NRMSEAir Quality Time (s)Traffic U/NTraffic F-P Gap (%)Traffic NRMSETraffic Time (s)EEG U/NEEG F-P Gap (%)EEG NRMSEEEG Time (s)HAR70+ U/NHAR70+ F-P Gap (%)HAR70+ NRMSEHAR70+ Time (s)
Original0.6327415.9510.69014--0.6829815.7220.60089--0.7638520.0210.69704--0.7014412.94030.48587--
Retrain0.526880.0000.6951120.5690.527920.0000.6005114.4180.522450.0000.715397.0740.520340.00000.4887579.433
SSD0.5339654.1360.885780.6360.5279786.4931.076430.5280.5307562.7371.099210.5100.5220682.64650.669241.263
NegGrad+0.53755264.5522.516009.8030.53349108.7161.168433.6320.5211496.7181.090311.8350.52583165.64381.0085022.667
Zero-label0.5338247.4510.809831.2230.5329050.7150.784190.6050.5333150.1490.925860.2840.5219759.37790.515021.422
Random-label0.5337747.2470.810461.2580.5304750.6540.795340.5970.5302150.8510.925500.3020.5303444.98830.623131.439
SCRUB-R0.5361615.1770.6887411.1850.5291153.0370.791830.9150.5490227.6060.715640.4790.5608212.51880.4866810.034
TS-Unlearn0.5352616.9960.694612.6430.5274126.9830.663481.4230.5598918.3100.715560.7900.5457613.93030.487982.971
RDTU0.5305312.8940.686479.5180.5276716.5360.607968.7650.5463116.2530.710993.4700.5413010.68580.4810215.687

We also conducted experiments of higher-dimensional multivariate data on two larger common time series forecasting benchmark datasets(Ettm1, Weather) with H=96, the results of the transformer base model shows below and more results of other base model have been added to Appendex.

MethodETTm1 NRMSEETTm1 F-P Gap%ETTm1 U/NETTm1 TimeWeather NRMSEWeather F-P Gap%Weather U/NWeather Time
Original0.44886613.7600.829193—0.18028821.2970.842255—
Retrain0.4537080.0000.53076247.2910.2001020.0000.56268344.892
SSD0.84901429.3560.5016171.2090.29989641.2330.5902561.297
NegGrad+1.184357217.5110.71357736.9250.34750592.7230.68086240.030
Zero Label0.50225721.6760.7382973.0740.20858733.8760.7208563.259
Random Label0.52935422.3160.7271423.0250.20409934.0470.7241213.290
SCRUB-R0.47787233.1760.6717074.5920.20049531.7070.6879035.142
TS-Unlearn0.46990728.1200.78758217.8180.28004559.4790.67372122.547
Ours0.46166217.4590.52787810.2510.18613316.8200.66020213.914
Reviewer_daUWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors cope with the challenge of handling deletion requests when dealing with sensitive, proprietary, or corrupted information while performing time series forecasting, focusing on a relatively underexplored methodological area, so-called machine unlearning. The proposed RDTU method (Residual Diffusion framework of Time-series Unlearning) does not access a retrained model during unlearning to create labels, but rather employs a neural tangent kernel (NTK) with Nystrom features as a regression surrogate, followed by a conditional diffusion model that learns the residual between the trained model and its NTK predictor based on retained-set conditions. A Structural Irreplaceability Score (SIS), which combines a log-determinant volume term with a k-NN support term, is used to condition the diffusion model while summarizing both complementary global directional change and locally retained support. Conditioned on the retained-set NTK prediction and SIS, diffusion generates a residual correction that yields suitable targets for forgotten windows, by staying close to the true targets of retained ones. Experiments span 4 real-world forecasting datasets, several horizon lengths, and a transformer base model and consider representative baselines in the field. RDTU consistently approximates exact retraining more closely than the other efficient unlearning methods evaluated, while being relatively fast, especially compared to retraining.

Critical strengths

Main strengths:

  1. [Significance of Contribution] The authors very clearly recognize the conceptual gaps in formalizing machine unlearning for time series forecasting, while correctly positioning the method against related ones in the literature (e.g., gradient-based unlearning, label-guided updating). Handling forecasting windows that are ordered and extensively overlapping requires careful consideration when defining deletion to avoid information leakage between inputs and targets.
  2. [Clean Design] The residual generator can be reused for repeated deletion requests, which is useful in practice. The ablation shows the clear advantage of the diffusion residual and SIS, several metrics and alternative models for residuals (e.g., MLP, ARIMA) are tested.
Critical weaknesses

Main weaknesses:

  1. [Incremental Improvements] Performance improvements are mostly supported based on relative reductions (e.g., F-P gaps), while in absolute terms, differences are mostly small (e.g., in NRMSE). Improvements are also smaller for longer horizons. No standard deviations or significance tests are mentioned for the results, so their overall significance is questionable.
  2. [Limited Evaluation] Common time series forecasting benchmark datasets are not considered in this work (see https://github.com/thuml/Time-Series-Library, https://github.com/SalesforceAIResearch/gift-eval). Instead, very small subsets (10 series) are considered from the 4 chosen datasets, out of which half are univariate. Additionally, only a single backbone is considered. Exact details about splits, seeds, and hyperparameters are missing.
Questions for authors
  • [Q1] It is unclear whether reported times include training the diffusion model (check Appendix A.4, mentioning diffusion training happens once, vs A.5, mentioning the reported time includes it).
  • [Q2] Please give clear details on hyperparameters, splits, and standard deviations characterizing performances, mentioning statistical significance.
  • [Q3] Consider expanding experiments to common benchmarks and additional base models to support the generalizability of results.
  • [Q4] Consider making clearer the methodological novelty of the several complex components of the method throughout the intro and methods parts (i.e., NTK ensemble, Nyström features, a conditional diffusion model, SIS, and fine-tuning), while discussing any scalability issues (or other limitations).
4 comments
AuthorsOpenReview ↗

Response to Reviewer daUW [part 1]

Reviewer daUW W1

Performance improvements ...,so their overall significance is questionable.

Thanks for your suggestion. F-P Gap is both appropriate and sufficiently informative for measuring functional agreement with exact retraining in our forecasting setting. Unlike classification accuracy, forecasting errors are continuous, unbounded, and strongly dependent on the scale and variability of the underlying time series. Reporting only an unnormalized prediction difference would therefore make results across datasets difficult to interpret and potentially misleading[1]. Classification-based unlearning studies commonly report absolute accuracy differences from Retrain. For example, on Tiny-ImageNet, UAM obtains (\Delta\mathrm{Acc}=1.75), whereas the competing methods produce gaps ranging from (4.48) to (69.98)[2]. So UAM's test-accuracy difference of (1.50) percentage points relative to the retrained accuracy of (62.36%) corresponds to a relative deviation of (2.41%). Similarly, the absolute differences reported by MUDA on Tiny-ImageNet correspond to relative deviations of approximately (1.03%) to (11.11%), depending on the evaluated metric[3]. Therefore, the magnitude of our reported F-P Gaps is consistent with retraining-relative discrepancies reported in existing unlearning studies. For time-series forecasting, F-P Gap measures the prediction discrepancy relative to the intrinsic variation of each time series and provides a consistent interpretation across datasets with substantially different scales. In particular, for low-variance time series, even a numerically small absolute prediction difference may represent a considerable fraction of the meaningful temporal variation. Consequently, the small absolute magnitude of the prediction discrepancy does not invalidate the F-P Gap; rather, it demonstrates why scale-aware normalization is necessary for evaluating time-series unlearning.

Meanwhile, we also present the standard deviation of the experimental results as follows

MethodAir Quality U/NAir Quality F-P GapAir Quality NRMSEAir Quality TimeTraffic U/NTraffic F-P GapTraffic NRMSETraffic Time
Retrain0.5034 ± 0.0155--0.2622 ± 0.025026.323 ± 5.7150.5291 ± 0.0155--0.3955 ± 0.055312.549 ± 2.845
SSD0.5218 ± 0.013652.8954 ± 11.72470.6683 ± 0.10980.946 ± 0.0330.5371 ± 0.021472.6908 ± 12.99800.8928 ± 0.15500.395 ± 0.015
NegGrad+0.5090 ± 0.0065162.1498 ± 53.43441.5902 ± 0.52229.871 ± 0.1870.5412 ± 0.0269140.9170 ± 64.33451.3989 ± 0.63474.340 ± 0.159
Zero-label0.5093 ± 0.012854.0268 ± 4.59040.6009 ± 0.05054.161 ± 0.1490.5387 ± 0.016551.8856 ± 6.44560.6896 ± 0.07722.150 ± 0.112
Random-label0.5069 ± 0.009854.2146 ± 5.31390.6052 ± 0.05770.775 ± 0.1380.5374 ± 0.017252.5161 ± 6.98090.7063 ± 0.08290.786 ± 0.110
SCRUB-R0.5195 ± 0.01298.4768 ± 1.84980.2720 ± 0.02426.854 ± 0.4790.5499 ± 0.027111.9866 ± 47.54230.4142 ± 0.43624.913 ± 0.037
TS-Unlearn0.5197 ± 0.01167.5967 ± 2.43670.2698 ± 0.02948.227 ± 0.2530.5458 ± 0.020513.5214 ± 3.62360.4189 ± 0.05563.579 ± 0.034
RDTU0.5001 ± 0.00814.6312 ± 1.15200.2600 ± 0.026011.710 ± 0.1190.5420 ± 0.02299.9251 ± 2.69190.3935 ± 0.05354.261 ± 0.048
MethodEEG U/NEEG F-P GapEEG NRMSEEEG TimeHAR70 U/NHAR70 F-P GapHAR70 NRMSEHAR70 Time
Retrain0.5235 ± 0.0086--0.1148 ± 0.04838.910 ± 0.1330.5210 ± 0.0091--0.3459 ± 0.212632.371 ± 12.531
SSD0.5180 ± 0.010854.1037 ± 6.34270.5877 ± 0.13430.227 ± 0.0720.5204 ± 0.010572.0281 ± 9.87050.5763 ± 0.25430.623 ± 0.013
NegGrad+0.5261 ± 0.0115102.1933 ± 19.21470.9080 ± 0.18922.397 ± 0.0250.5372 ± 0.0126288.0736 ± 104.13611.6082 ± 0.916210.051 ± 0.170
Zero-label0.5184 ± 0.008460.1355 ± 3.71760.5854 ± 0.07190.698 ± 0.0120.5227 ± 0.010958.6257 ± 5.51220.4214 ± 0.23601.009 ± 0.020
Random-label0.5183 ± 0.008960.0174 ± 4.17990.5846 ± 0.07880.695 ± 0.0030.5220 ± 0.013145.3235 ± 5.17900.5377 ± 0.23861.007 ± 0.014
SCRUB-R0.5368 ± 0.008218.9612 ± 2.93570.2149 ± 0.03992.992 ± 0.0210.5497 ± 0.021711.6701 ± 5.45250.3565 ± 0.22424.802 ± 0.073
TS-Unlearn0.5254 ± 0.00788.5701 ± 2.51150.1353 ± 0.04861.310 ± 0.0840.5385 ± 0.023712.8356 ± 5.38920.3539 ± 0.22981.933 ± 0.095
RDTU0.5244 ± 0.01253.7526 ± 0.57070.1110 ± 0.04322.322 ± 0.0560.5176 ± 0.01009.7703 ± 4.05600.3424 ± 0.20605.744 ± 0.123
AuthorsOpenReview ↗

Response to Reviewer daUW [part 2]

Reviewer daUW W2 & Q3

Common time series forecasting benchmark datasets are not considered in this work (see https://github.com/thuml/Time-Series-Library, https://github.com/SalesforceAIResearch/gift-eval). Instead, very small subsets (10 series) are considered from the 4 chosen datasets, out of which half are univariate. Additionally, only a single backbone is considered. Exact details about splits, seeds, and hyperparameters are missing. Consider expanding experiments to common benchmarks and additional base models to support the generalizability of results.

Thanks for the suggestions. We conducted experiments of two larger common time series forecasting benchmark datasets(Ettm1, Weather), the results of the transformer base model shows below and more results of other base model have been added to Appendex.

MethodETTm1 NRMSEETTm1 F-P Gap%ETTm1 U/NETTm1 TimeWeather NRMSEWeather F-P Gap%Weather U/NWeather Time
Original0.44886613.7600.829193—0.18028821.2970.842255—
Retrain0.4537080.0000.53076247.2910.2001020.0000.56268344.892
SSD0.84901429.3560.5016171.2090.29989641.2330.5902561.297
NegGrad+1.184357217.5110.71357736.9250.34750592.7230.68086240.030
Zero Label0.50225721.6760.7382973.0740.20858733.8760.7208563.259
Random Label0.52935422.3160.7271423.0250.20409934.0470.7241213.290
SCRUB-R0.47787233.1760.6717074.5920.20049531.7070.6879035.142
TS-Unlearn0.46990728.1200.78758217.8180.28004559.4790.67372122.547
Ours0.46166217.4590.52787810.2510.18613316.8200.66020213.914

We also conducted supplementary experiments on different base models, and the specific results of LSTM are shown in the following table, more results of other base model have been added in Appendix.

LSTM:

MethodAir Quality U/NAir Quality F-P GapAir Quality NRMSEAir Quality TimeTraffic U/NTraffic F-P GapTraffic NRMSETraffic TimeEEG U/NEEG F-P GapEEG NRMSEEEG TimeHAR70 U/NHAR70 F-P GapHAR70 NRMSEHAR70 Time
Original0.57324.730.2484--0.638414.760.3933--0.55321.590.0474--0.67308.560.2968--
Retrain0.54370.000.25459.2640.54330.000.38797.5610.52010.000.04889.2290.51800.000.300732.845
SSD0.545970.400.79351.4360.533082.980.99890.6990.519876.820.77120.5720.525779.950.56691.952
NegGrad+0.533820.550.35572.7940.538135.560.51982.0970.521859.390.52321.2300.5269217.951.113312.873
Zero-label0.545050.590.59411.4160.534850.560.67650.8300.518245.160.43940.1610.523359.050.35220.665
Random-label0.545550.590.59511.4200.534550.960.69320.7860.518245.440.43990.1610.520443.740.48820.724
SCRUB-R0.544547.770.56531.1260.541137.240.51200.7590.51916.090.07530.2040.526511.150.29761.425
TS-Unlearn0.53795.520.25982.3440.548315.180.41302.4950.52774.210.05561.8410.529810.920.31671.404
Ours0.53684.800.24882.59400.533913.970.39571.5210.51892.550.04870.9470.51929.400.30813.1775
AuthorsOpenReview ↗

Response to Reviewer daUW [part 3]

Reviewer daUW Q1

It is unclear whether reported times include training the diffusion model (check Appendix A.4, mentioning diffusion training happens once, vs A.5, mentioning the reported time includes it).

Thanks for the question. The runtime reported for RDTU in the main table follows a conservative first-request end-to-end accounting protocol and includes the time required to train the diffusion model. Specifically, it covers complete-data NTK construction, diffusion training, retained-set NTK and SIS construction, diffusion sampling, and the final unlearning update. The statements in Appendices A.4 and A.5 describe two different aspects of the method. Appendix A.4 describes the deployment lifecycle: the diffusion model is trained once for a given original forecasting model and can subsequently be reused across multiple deletion requests. Appendix A.5 describes the runtime convention adopted in the main comparison, in which this one-time diffusion-training cost is included. Therefore, the reported results correspond to the more conservative cost of processing the first deletion request. For subsequent requests on the same original model, diffusion training is not repeated, resulting in a lower per-request runtime.

Reviewer daUW Q4

Consider making clearer the methodological novelty of the several complex components of the method throughout the intro and methods parts (i.e., NTK ensemble, Nyström features, a conditional diffusion model, SIS, and fine-tuning), while discussing any scalability issues (or other limitations)

Thanks very much for the suggestion. In the updated article, we have improved the relevant descriptions, making the novelty presentation of each component more clear.

AuthorsOpenReview ↗

Response to Reviewer daUW [part 4]

Reviewer daUW Q2

Please give clear details on hyperparameters, splits, and standard deviations characterizing performances, mentioning statistical significance.

Data splits and window counts

Each series is chronologically divided into 60% training, 5% validation, 15% auxiliary, and 20% test data. The auxiliary split is used only for standardization and auditing. Windows are constructed separately within each split to prevent temporal leakage. We evaluate ten users with three deletion seeds {101,202,303}, giving 30 deletion requests per dataset.

DatasetOriginal Train DDForget FFRetain RRValidationTest
Air Quality15,7645,25010,5141,7296,990
Traffic7,8702,6405,2308613,494
EEG3,3243902,9343041,234
HAR70+13,09881612,2821,1544,634
ETTm132,4501,91930,53111,42511,425
Weather34,6612,03532,6265,17510,444

Hyperparameters

Transformer: 4 layers, hidden dimension 128, 8 heads, feed-forward dimension 256, dropout 0.1, and GELU activation. It is trained with AdamW for at most 80 epochs using batch size 512, learning rate 3×10−43\times10^{-4}, weight decay 10−410^{-4}.

LSTM: 2 layers, hidden dimension 256, dropout 0.1, unidirectional recurrence, and a linear prediction head. It is trained with AdamW for at most 200 epochs using batch size 512, learning rate 10−310^{-3}, and weight decay 10−410^{-4}.

TCN: 4 residual blocks, 128 channels, 2 causal convolutions per block, kernel size 3, dilation factors {1,2,4,8}, dropout 0.10.1, weight normalization, and ReLU activation. It is trained with AdamW for at most 200 epochs using batch size 512, learning rate 10−310^{-3}, and weight decay 10−410^{-4}.

PatchTST: 2 encoder layers, hidden dimension 128, 8 heads, feed-forward dimension 256, patch length 4, stride 2, dropout 0, and GELU activation, with channel independence, RevIN, and a flatten-and-linear prediction head. It is trained with AdamW for at most 200 epochs using batch size 512, learning rate 10−310^{-3}, and weight decay 10−410^{-4}.

NTK/Nyström: one initialization, 512 anchors, Jacobian batch size 150, ridge multiplier 10−310^{-3}, and eigenvalue cutoff 10−1010^{-10}.

Diffusion: 128 steps with a linear schedule from 10−410^{-4} to 0.02, hidden dimension 256, 60 epochs, batch size 512, learning rate 2.6×10−32.6\times10^{-3}, and 32 generated residual samples. We use 20% of the complete-data residual pairs for diffusion training.

We report the mean and standard deviation of datasets Air Quality and Traffic over the 30 user–deletion-seed combinations in the tables below, and complete seed formulas and module-level configurations are provided in the Appendix.

MethodAir Quality U/NAir Quality F-P Gap (%)Air Quality NRMSEAir Quality Time (s)Traffic U/NTraffic F-P Gap (%)Traffic NRMSETraffic Time (s)
Original0.6009 ± 0.03936.5600 ± 2.04690.2546 ± 0.0250--0.6454 ± 0.040015.5700 ± 4.52800.3941 ± 0.0561--
Retrain0.5034 ± 0.0155--0.2622 ± 0.025026.323 ± 5.7150.5291 ± 0.0155--0.3955 ± 0.055312.549 ± 2.845
SSD0.5218 ± 0.013652.8954 ± 11.72470.6683 ± 0.10980.946 ± 0.0330.5371 ± 0.021472.6908 ± 12.99800.8928 ± 0.15500.395 ± 0.015
NegGrad+0.5090 ± 0.0065162.1498 ± 53.43441.5902 ± 0.52229.871 ± 0.1870.5412 ± 0.0269140.9170 ± 64.33451.3989 ± 0.63474.340 ± 0.159
Zero-label0.5093 ± 0.012854.0268 ± 4.59040.6009 ± 0.05054.161 ± 0.1490.5387 ± 0.016551.8856 ± 6.44560.6896 ± 0.07722.150 ± 0.112
Random-label0.5069 ± 0.009854.2146 ± 5.31390.6052 ± 0.05770.775 ± 0.1380.5374 ± 0.017252.5161 ± 6.98090.7063 ± 0.08290.786 ± 0.110
SCRUB-R0.5195 ± 0.01298.4768 ± 1.84980.2720 ± 0.02426.854 ± 0.4790.5499 ± 0.027111.9866 ± 47.54230.4142 ± 0.43624.913 ± 0.037
TS-Unlearn0.5197 ± 0.01167.5967 ± 2.43670.2698 ± 0.02948.227 ± 0.2530.5458 ± 0.020513.5214 ± 3.62360.4189 ± 0.05563.579 ± 0.034
RDTU0.5001 ± 0.00814.6312 ± 1.15200.2600 ± 0.026011.710 ± 0.1190.5420 ± 0.02299.9251 ± 2.69190.3935 ± 0.05354.261 ± 0.048

[1] R. J. Hyndman and A. B. Koehler, “Another Look at Measures of Forecast Accuracy,” International Journal of Forecasting, vol. 22, no. 4, pp. 679–688, 2006.

[2] H. Kim, K. Kim, S. Chae, and S. Yoon, “Unlearning-Aware Minimization,” Advances in Neural Information Processing Systems, vol. 38, 2025.

[3] S. Seo, D. Kim, and B. Han, “Revisiting Machine Unlearning with Dimensional Alignment,” in Proceedings of the Winter Conference on Applications of Computer Vision, pp. 3206–3215, 2025.

Comments on the paper

15 comments
AuthorsOpenReview ↗

Revision Change Summary

  1. [Introduction] We optimized the design and presentation of the introduction section of the model, and further clarified the relevant contributions of this paper, making the motivation and novelty of our method more explicit.
  2. [Section 4.1 & Table 1] We enhanced the data set segmentation and provided detailed data descriptions, thereby making the experimental setup more comprehensive.
  3. [Section 4.3 & Table 3] We added experiments of User-Level and User-Sensor Unlearning, making the experimental scenarios more in line with our motivation and actual application scenarios.
  4. [Section 4.4 & Table 4] Based on the original ablation experiments, we added four more ablation experiments to demonstrate the excellent performance and characteristics of the NTK and diffusion modules in our model.
  5. [Appendix A.1 & Table 5] We provided more detailed descriptions of the large-scale datasets used in the additional experiments, as well as the experimental settings.
  6. [Appendix A.2 ] We added the detailed settings for the four base models used in the experiment, thereby enhancing the reproducibility of the experiment.
  7. [Appendix B.1 ] We presented the main experimental results including the standard deviation.
  8. [Appendix B.3 ] We conducted comprehensive experiments on two large-scale datasets to demonstrate the reliability and stability of our method in complex data scenarios.
  9. [Appendix B.4 ] We designed comprehensive experiments on three different base models to demonstrate the robustness and scalability of our method.
  10. [Appendix B.5 & B.6 ] We conducted experiments under various forgetting strategies, including different forgetting ratios, special forgetting patterns, and continuous forgetting settings, in order to demonstrate the robustness and stability of our approach.
  11. [Appendix B.7 ] We added performance indicators used to measure the consistency of the distribution and the performance under reconstruction attacks, in order to further demonstrate the forgetting effect of our model.
  12. [Appendix C.2 ] We added running time and memory analysis of our model, thereby enhancing the scalability analysis of the model.
AuthorsOpenReview ↗

Thanks for All Reviewers

We would like to thank all the reviewers for their insightful comments and constructive suggestions. Below we provide our detalled response to eachreviewer. We have uploaded a revised version of our submission, with major changes highlighted in blue. New experiments are also conducted, and the results are added to the tables in blue color. Thank you and we look forward to further feedback and discussion.

AuthorsOpenReview ↗

A Gentle Reminder of Further Feedback – To Reviewer daUW

Dear Reviewer daUW,

The conclusion of the discussion period is closing, and we eagerly await your response. We greatly appreciate your time and effort in reviewing this paper and helping us improve it.

Thank you again for the detailed and constructive reviews. We hope our response is able to address your comments related to the comparision and discussion with more details of experiments, novelty clarification, and experiments of more benchmarks and base models. We take this as a great oppartunity to improve our work and shall be grateful for any additional feedbak you could give us.

Best Regards,

Authors of Paper 18599

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.