acceptodds
Under review as a conference paper at ICLR 2027

Online Multi-Source Transfer Learning under Shifting Source Reliability

Abstract

Transfer learning leverages labeled source data to improve online prediction when labeled target data are limited. However, both target and source distributions may change over time. A source that helps early in the stream may then become unhelpful, making fixed transfer weights unreliable. We propose Adaptive Multi-Source Transfer Learning (AMTL), which combines a target-only predictor with multiple source-specific transfer predictors. After each target batch, AMTL updates predictor weights based on their losses and uses a structured transition to adjust how much to transfer and which sources to trust. The theoretical analysis establishes a pathwise cumulative-loss bound relative to target-only prediction and examines the sources of error for sparse logistic models. Experiments on synthetic streams, CIFAR-10, and hospital records show that AMTL performs well in online prediction relative to existing methods.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Online multi-source transfer learning, Shifting source reliability, Adaptive weighting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aHDXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a transfer learning method that uses adaptive transfer weights since source distributions may change over time. The proposed method combines a target-only predictor and multiple source-specific transfer predictors.

Critical strengths

1.The paper is easy to follow.

  1. The problem of transfer learning unders shifting source reliability is of great importance.
Critical weaknesses
  1. It seems that the weight update does not affect the result of Corollary 1. In the proof, the bound holds for any weights as long as they sum to one. A second point to note is that the last term on the RHS OF the inequlity is somehow large. For example, when T = 20, δ = 0.05 the term is around 4*B (if you use B=−log⁡ϵ=−log⁡0.1=2.3B=-\log \epsilon=-\log0.1=2.3, then 4B is about 9.2) while cumulative loss is approximatey around 9 in Table 9.

  2. The baseline parameters are not consistent across methods. For example, Flat-FS and AMTL use different values of γ\gamma and q1q_1. So the conclusion may be affected by the those parameters. Can the authors provide any insight of how the hyperparameters were chosen? A sensitivity analysis is needed before making the conlusion the .

  3. Also, in the theory, it uses ($q_1 = 1/2, \alpha_o = 1/T, c_\lambda ≥ 2\sqrt{2}\nu), but in the experiments, it uses (0.8, 0.025, c_λ = 0.10). Can the authors explain the gap between the theory and the experiments?

  4. The real data anaysis has fixed source reliability, which is different than the paper's asscumption on shiting source reliabilities. The hospital sources are fixed. In CIFAR the sources are fixed and the target shifts. On the heart disease data, the proposed AMTL has cumulative loss 5.98 vs SafePool is 6.05. The different seems not statistically significant. In Figure 4b, it shows the transfer weight moving from 0.8 to about 0.65 after 12 batches, so the result may mostly reflect q1q_1. The results on the 123 held-out records are not reported.

  5. Related work is incomplete. What about the paper BOTL (McKay et al., 2020),Ge et al. (2013, 2014)? In addition, model-averaging transfer (Hu and Zhang, 2023) and robust transfer with unreliable sources (Fan et al., 2025) address the same source-weighting question offline. The online transfer methods cited in the introduction (Zhao et al. 2014; Wu et al. 2017; Du et al. 2019; Xie and Pratama 2022) are not compared to the proposed method. Is there a specific reason not doing so? Also, Credit-TL and SafePool have no citation when first mentioned.

Questions for authors
  1. Please address the 5 comments list in the weaknesses.

  2. How does AMTL compare with re-scoring each source on the target history whenever a snapshot is replaced?

  3. Can the authors provide any insight on fixing γ\gamma instead of letting it grow with NtN_t?

2 comments
AuthorsOpenReview ↗

Response to Reviewer aHDX (1/2)

W1. Thank you for pointing this out. We now state the general convex-weight result and the bound using AMTL's updates separately. The former is Corollary 3 in Appendix A.7. The main-text Corollary 2 combines Theorem 1 with candidate estimation bounds and includes ∑t=1TEst,t−log⁡Π(s1:T)\sum_{t=1}^T E_{s_t,t}-\log\Pi(s_{1:T}), which depends on the realized transition and comparison path.

For batch-mean loss, the fluctuation term is now

B12log⁡1δ∑t=1T1bt.B\sqrt{\frac12\log\frac1\delta\sum_{t=1}^T\frac1{b_t}}.

With equal batch sizes bb, it becomes BTlog⁡(1/δ)/(2b)B\sqrt{T\log(1/\delta)/(2b)}, reducing the former batch-level term by 1/b1/\sqrt b under the same probability allocation. Appendices A.7–A.8 give the conditional within-batch calculation and the resulting path bound.

W2. All main comparisons select parameters by cumulative log loss on shared validation data, separately for each setting. Appendices B.1 and B.3 give the synthetic grids, settings, and baseline implementations. Appendices C–D give the image and hospital protocols. Matched Flat retains AMTL's models and initialization, while independently tuned Flat uses the same full search budget as AMTL. Figure 5 gives initialization and sharing-rate profiles, and Table 2 reports the transition comparison.

W3. Section 4.1 gives the comparison for the executed initialization:

∑t=1Tℓt(pt)−ℓR,t≤−log⁡(1−q1)+(T−1)log⁡11−αo.\sum_{t=1}^T\\{\ell_t(p_t)-\ell_{\mathrm{R},t}\\}\leq-\log(1-q_1)+(T-1)\log\frac1{1-\alpha_o}.

The constant log⁡2+1\log2+1 uses q1=1/2q_1=1/2, αo=1/T\alpha_o=1/T, and T≥2T\geq2. The actual cost is approximately 12.11112.111 in Near, 11.51311.513 in six other settings, and 10−510^{-5} in Opposite. The reference is AMTL's internal Target at the same penalty scale.

Theorem 1 applies to the executed parameters. Theorem 2 requires its stated penalty, score, curvature, and localization conditions. Appendix B.1 gives the empirical scales and validation protocol. Table 16 also reports the theory-initialization control, fixing q1=1/2q_1=1/2 and αo=1/T\alpha_o=1/T, retaining the selected fits and choosing αs\alpha_s on validation data.

W4. We added experiments with changing source snapshots and retained the original target-color and cold-start tasks separately. In Section 5.4, the CIFAR-10 target rule and labels stay fixed while three source snapshots change color rules over five phases. AMTL's cumulative loss is 1.449±0.0401.449\pm0.040, compared with 1.528±0.1001.528\pm0.100 for Target. Table 10 reports all methods and held-out test losses.

The hospital experiment preserves labels while changing source composition through younger patients, older patients, the full cohort, and younger patients again. The warm-start protocol uses 60 calibration records followed by 120 scored Cleveland records. The cold-start protocol scores 180 records. Both reserve 123 records for testing, and validation uses only external hospitals with Cleveland excluded.

Section 5.5 reports warm-start online losses of 3.657±0.0723.657\pm0.072 and 3.696±0.0703.696\pm0.070, the lowest means among the 11 main methods. We report paired comparisons with the strongest baselines; these intervals include zero. Tables 14–15 provide held-out test loss, AUC, and the cold-start comparisons. Figure 3 separates total candidate mass, conditional source weights, and nominal source-objective weight.

AuthorsOpenReview ↗

Response to Reviewer aHDX (2/2)

W5. We expanded the Introduction and added HomOTLMS [1], BOTL-OLS based on [2], and TAB based on [4]. They cover online multi-source classification, weighting from target history, and robust source use. Appendix B.3 gives the binary snapshot adaptations and shared batch-feedback protocol.

HomOTLMS extends the single-source setting of Zhao et al. [3], using a passive-aggressive target learner and Hedge weights. BOTL-OLS fits combination weights from current-model predictions on target history. TAB uses its target-boundary decision rule.

The related-work discussion includes local source weighting and prediction consistency [5, 6], continuous-response model averaging [7], domain-stream ensembles [8], and unlabeled-target representation learning [9]. Credit-TL and SafePool are component baselines defined in this study. Section 5.2 identifies them as such, and Appendix B.3 gives their updates and tuning rules.

Q1. We address the five weaknesses in responses W1–W5 above.

Q2. The Source score control in Appendix B.2 scores each source-only model on the observed target history at replacement, using that score for all its strength candidates. AMTL scores each source–strength candidate by cross-fitting. Both retain the Target weight and total transfer mass and select rates on the same validation streams.

Table 8 reports paired AMTL-minus-Source-score losses of −0.564±0.018-0.564\pm0.018, 0.002±0.0020.002\pm0.002, and −1.548±0.061-1.548\pm0.061 in Flip, Rotate, and Switch under the main configuration. It also includes uniform reset. Table 2 compares refresh with inherited weights.

Q3. The revised method fits several strengths and adapts their weights. Appendix A.5 gives the sample-size dependence through Eq. (26):

1neff,j,t(γ)=1Nt+nj,t+(1Nt+1nj,t)(γ−NtNt+nj,t)2.\frac1{n_{\mathrm{eff},j,t}(\gamma)}=\frac1{N_t+n_{j,t}}+\left(\frac1{N_t}+\frac1{n_{j,t}}\right)\left(\gamma-\frac{N_t}{N_t+n_{j,t}}\right)^2.

For an aligned source, the variance-minimizing fraction Nt/(Nt+nj,t)N_t/(N_t+n_{j,t}) increases with target history. Source discrepancy adds the bias term in Eq. (25). The candidate library lets different sources and rounds favor different strengths, with the finite-grid variance effect quantified by this identity. Corollary 1 gives the aligned-source rate.

References

[1] Q. Wu, H. Wu, X. Zhou, M. Tan, Y. Xu, Y. Yan, and T. Hao. Online transfer learning with multiple homogeneous or heterogeneous sources. IEEE Transactions on Knowledge and Data Engineering, 29(7):1494–1507, 2017.

[2] H. McKay, N. Griffiths, P. Taylor, T. Damoulas, and Z. Xu. Bi-directional online transfer learning: a framework. Annals of Telecommunications, 75:523–547, 2020.

[3] P. Zhao, S. C. H. Hoi, J. Wang, and B. Li. Online transfer learning. Artificial Intelligence, 216:76–102, 2014.

[4] J. Fan, C. Gao, and J. M. Klusowski. Robust transfer learning with unreliable source data. The Annals of Statistics, 53(4):1728–1752, 2025.

[5] L. Ge, J. Gao, and A. Zhang. OMS-TL: a framework of online multiple source transfer learning. In Proceedings of the 22nd ACM International Conference on Information and Knowledge Management, pp. 2423–2428, 2013.

[6] L. Ge, J. Gao, H. Ngo, K. Li, and A. Zhang. On handling negative transfer and imbalanced distributions in multiple source transfer learning. Statistical Analysis and Data Mining, 7(4):254–271, 2014.

[7] X. Hu and X. Zhang. Optimal parameter-transfer learning by semiparametric model averaging. Journal of Machine Learning Research, 24(358):1–53, 2023.

[8] H. Du, L. L. Minku, and H. Zhou. Multi-source transfer learning for non-stationary environments. In International Joint Conference on Neural Networks, 2019.

[9] R. Xie and M. Pratama. Automatic online multi-source domain adaptation. Information Sciences, 582:480–494, 2022.

Reviewer_A4vPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes AMTL, which combines a target-only predictor with one target-source predictor per source. Predictor weights are updated by exponential loss weighting and a structured transition matrix. Separate parameters control movement between target-only and transfer prediction and movement among sources. The theory gives a pathwise log-loss comparison and finite-sample results for sparse logistic transfer estimators. Experiments use synthetic source shifts, color-modified CIFAR-10, and multi-hospital heart-disease data.

Critical strengths

Retaining a target-only predictor provides an explicit reference for transfer decisions. The cumulative-loss comparison is transparent and remains applicable when source usefulness changes.

The experiments report failures as well as gains. The AMTL-Flat comparison keeps the fitted predictors fixed, helping isolate the contribution of the transition structure.

Critical weaknesses

The aggregation mechanism and pathwise bound are direct instances of exponential weighting under a Markov prior. Generic MarkovHedge already permits transition matrices and gives the corresponding path-prior penalty. The main addition is a particular two-rate transition structure. The paper needs a stronger case that this specialization and the accompanying estimation analysis provide sufficient methodological novelty.

The additional benefit of the structured transition is setting-dependent. AMTL improves on AMTL-Flat in five of eight synthetic settings, but their hospital mean losses are nearly equal, and the hospital sources are fixed. Flat-FS has lower mean loss at 6 and 9 sources in the CIFAR source-pool study. A clear matched tuning protocol is needed to assess the advantage over a well-tuned flat transition.

Questions for authors

Table 5 gives the final parameter values. What data and criterion were used to choose these settings for AMTL and AMTL-Flat, and did both methods receive the same tuning budget?

1 comment
AuthorsOpenReview ↗

Response to Reviewer A4vP

Q1. Thank you for highlighting this connection. We build on Markov aggregation [1, 2]. Our contribution is the construction and evaluation of transfer candidates as source snapshots change.

Each source supplies separately fitted models at several strengths. At replacement, cross-fitted target-history losses evaluate the current candidates before the next batch and reallocate transfer mass. Theorem 1 applies Markov aggregation to this history-dependent transition, and Proposition 1 separates mode, candidate, and refresh costs. Theorem 2 and Corollaries 1–2 connect source sample size and transfer bias to candidate accuracy and cumulative online loss.

Table 2 separates the effects of multiple strengths and refresh from the transition design. The revised Introduction and Conclusion center the contribution on these transfer models, snapshot updates, and their analysis.

Q2. Section 5.3 gives two controlled comparisons. Matched Flat keeps AMTL's fitted models, initialization, and refresh. Independently tuned Flat uses the same validation data and search budget. In Near, AMTL reduces loss by 0.220±0.0640.220\pm0.064 and 0.138±0.0680.138\pm0.068, respectively. The changing-hospital paired difference against matched Flat is −0.0011±0.0010-0.0011\pm0.0010.

Table 11 retains the source-count comparison under both source-data budgets. Appendix C also reports the experiment with a fixed target and changing source snapshots.

Q3. Matched AMTL-Flat retains AMTL's fitted models, strength list, initialization, clipping, loss updates, and refresh. It selects only its flat rate on the same validation data. Independently tuned Flat and AMTL use the same candidate model scales, strength lists, initializations, and search budget. Both select by cumulative validation log loss. Appendix B.1 gives the search spaces, including the shared scale in Near.

Each synthetic setting uses 12 validation streams and 50 paired evaluations. Table 2 uses the component configurations in Appendix B.2. Near and Scarce share the current main evaluation records, while the other six settings use the separate component evaluation records. The two image tasks use 24 validation orderings and 30 paired evaluations. Warm-start hospitals use 36 external-hospital validation orderings, equally weighted across hospitals, and 60 paired Cleveland evaluations. All baselines use the corresponding validation data and criterion. Appendices C–D give the real-data protocols.

References

[1] N. Cesa-Bianchi and G. Lugosi. Prediction, Learning, and Games. Cambridge University Press, 2006.

[2] W. M. Koolen and S. de Rooij. Combining expert advice efficiently. In Proceedings of the 21st Annual Conference on Learning Theory, pp. 275–286, 2008.

Reviewer_eZAVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies online binary prediction on a target stream with i.i.d. target data and mm labeled source domains whose distributions may change over time. The proposed method, AMTL, refits two kinds of sparse logistic models each round: a target-only model, and one transfer model per source that minimizes a fixed γ\gamma-weighted combination of target and source risks. Their clipped probability outputs are combined with exponential weights followed by a structured fixed-share transition. One rate αo\alpha_o moves mass between the target-only and transfer states, and a second rate αs\alpha_s moves mass among sources. The theory gives (i) a pathwise log-loss bound against any state path, including an O(1)O(1) excess over target-only prediction, and (ii) a finite-sample error bound for the transfer estimators (estimation error + transfer bias + optimization error), combined into an end-to-end prequential bound. Experiments cover eight synthetic regimes, a CIFAR-10 color-shift task, and UCI multi-hospital heart disease data.

Critical strengths
  1. Relevant and well-motivated problem. Source reliability that changes over time is realistic (e.g., hospital data drift), and existing online transfer methods mostly assume fixed sources.
  2. Simple aggregation with a clean safety guarantee. The aggregation layer is modular, and its cumulative loss is at most a constant worse than target-only prediction. The two-rate transition is interpretable and reduces to standard fixed share when m=1m=1.
  3. Reasonably thorough empirical study. It includes targeted mechanism tests (Flip, Rotate, Switch), paired confidence intervals, and an ablation against a single-rate variant.
Critical weaknesses
  1. Adaptivity is limited to mixing a fixed set of predictors; the transfer estimator itself is not adaptive. The weight γ\gamma is fixed and shared across all sources and rounds (γ=0.05\gamma = 0.05 in the experiments). A degraded source therefore still dominates its own transfer model, and the method can only downweight that model's output. Mixing the predictions of a target-only model and a γ\gamma-transfer model cannot reproduce the model one would obtain with an intermediate γ\gamma. As a result, the comparator class in Theorem 1 may exclude the predictor that actually transfers well.

  2. The weights do not reflect the current source–target relationship. The only signal driving the weights is the cumulative prequential loss of each state, and the weights are tied to the source index rather than to the snapshot. When a snapshot is replaced, the new snapshot inherits the old snapshot's weight and must be corrected through several rounds of online losses. This is wasteful, because the replacement time is known and target history is already available to evaluate the new source directly. The authors themselves attribute Credit-TL's advantage in Rotate to its weight reset at replacement, which supports this concern.

  3. The presentation is hard to follow, and it is unclear how the method benefits from multiple sources. The setup introduces many interacting components (snapshots and their persistence, γ\gamma, q1q_1, αo\alpha_o, αs\alpha_s, clipping, several regularization schedules). The bounds are heavily overloaded with notation (ej,te_{j,t}, Ej,tE_{j,t}, ER,tE_{R,t}, ξϵ\xi_\epsilon, BB, neffn_{\mathrm{eff}}, τγ,j,t\tau_{\gamma,j,t}, ηopt\eta_{\mathrm{opt}}, δRSC\delta_{\mathrm{RSC}}, ...). Corollary 1 is essentially a weight-averaged sum of per-state errors. It does not state when, or by how much, AMTL improves over target-only prediction as a function of the source sample sizes, the number of sources mm, and the source–target discrepancy. The path bound even charges a log⁡m\log m cost per switch. The reader therefore cannot see at a glance whether, or how, the method benefits from aggregating multiple source samples.

Questions for authors
  1. Why not adapt γ\gamma, per source and over time? As a minimal step, why not include transfer models with several γ\gamma values as additional states? How would this compare with the current design?
  2. Have you considered re-initializing a source's weight at snapshot replacement using its out-of-sample loss on the accumulated target history (e.g., a source-only fit evaluated on D0,1:t−1D_{0,1:t-1})? Since the transition would remain predictable, a time-varying prior should keep Theorem 1 intact.
  3. Can you state a simplified corollary that makes the gain over target-only explicit? For example, assume one good source with discrepancy τ\tau and nn samples, and show the resulting rate. Can you also say how this gain scales with the number of useful sources?
  4. How sensitive are the results to (q1,αo,αs)(q_1, \alpha_o, \alpha_s), and how were these values chosen? The single-rate variant AMTL-Flat outperforms AMTL on Near, Mixed, Scarce, and the hospital data, which weakens the case for the two-rate design.
1 comment
AuthorsOpenReview ↗

Response to Reviewer eZAV

Q1. Thank you for this suggestion. Section 3.1 now fits a separate model for every source jj and chosen strength γ\gamma:

=\gamma\widehat L_{0,t-1}(\theta)+(1-\gamma)\widehat L_{j,t}(\theta) +\lambda_{j,t}(\gamma)\|\theta\|_1.$$ The common finite strength list is fixed after validation. Each source–strength pair has its own weight, updated by target feedback. Different sources and rounds can therefore favor separately fitted candidates at different strengths. Table 2 compares this library with one validation-selected strength, keeping AMTL's penalty scale and initialization. Multiple strengths reduce cumulative loss by $0.138\pm0.063$ in Near, $0.748\pm0.031$ in Flip, and $1.630\pm0.093$ in Switch. Table 16 extends the strength comparison to the image and hospital tasks. **Q2.** We added snapshot refresh in Section 3.3. At replacement, all current transfer candidates are scored on held-out folds of the observed target history, using the same partition. Each scoring fit uses the other folds and its current source snapshot. For $N_t\geq2$, Eq. (10) gives $$w_{s,t}\leftarrow q_t\frac{\exp\left(-N_t\widehat L^{\mathrm{cv}}_{0,t-1}(s)\right)}{\sum_{a\ne\mathrm{R}}\exp\left(-N_t\widehat L^{\mathrm{cv}}_{0,t-1}(a)\right)},\quad s\ne\mathrm{R}.$$ Refresh retains the Target weight and total transfer mass. The executed models use the full target history. Appendix A.2 incorporates refresh into $Q_t$, formed before the current batch, and Theorem 1 uses this transition. In Table 2, refresh reduces loss relative to inherited weights by $0.779\pm0.026$, $1.139\pm0.042$, and $3.306\pm0.139$ in Flip, Rotate, and Switch. Table 8 additionally compares uniform reset and source-only scoring under the main-comparison configuration. **Q3.** We separated fitting, aggregation, and refresh in Sections 3–4 and moved technical conditions and proofs to Appendix A. Theorem 2 separates estimation error, transfer bias, and optimization error. Its effective sample size is $$n_{\mathrm{eff},j,t}(\gamma)=\left(\frac{\gamma^2}{N_t}+\frac{(1-\gamma)^2}{n_{j,t}}\right)^{-1}.$$ For an aligned source, Corollary 1 replaces $N_t$ in the Target upper bound by $n_{\mathrm{eff},j,t}(\gamma)$. Under its conditions and zero optimization error, including $\gamma=N_t/(N_t+n_{j,t})$ reduces this upper bound by the factor $N_t/(N_t+n_{j,t})$. Appendix A.8 quantifies the contribution of several useful sources through their combined path mass. With persistent snapshots, one strength per source, and $m/2$ useful sources of equal cumulative loss, the allocation cost is $\log(2/q_1)$, compared with $\log(m/q_1)$ for one specified source, with the same persistence term. Theorem 1 gives improvement over the internal Target when a supported path's cumulative loss reduction exceeds its path cost. Appendix A.5 quantifies source discrepancy and the finite strength grid. **Q4.** Parameters are selected separately for each synthetic setting on 12 shared validation streams by cumulative log loss. Appendix B.1 gives the grids and selected settings. Figure 5 profiles $q_1$, $\alpha_o$, and $\alpha_s$ while minimizing over the other aggregation parameters. Validation selects $q_1=.99999$ in seven settings, $q_1=10^{-5}$ in Opposite, and gives $\alpha_o=\alpha_s=.003$ in Near. The other settings use $\alpha_o=0$. Near uses 1,000 targets in batches of five and source gaps $1,1.3,1.6,2$, with the original source sample budget. Section 5.3 distinguishes matched Flat from independently tuned Flat with the same full search budget. In Near, AMTL lowers loss by $0.220\pm0.064$ and $0.138\pm0.068$, respectively. AMTL ranks first in Near and has the best average rank, $2.000$, across the eight settings. Table 2 gives all paired comparisons.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.