acceptodds
Under review as a conference paper at ICLR 2027

Amortizing Structural Interventions in Online Opinion Dynamics

Abstract

Minimizing opinion polarization and disagreement on social platforms is constrained by incomplete information, as individual innate opinions are private and unobservable. Recent approaches formulate this task as a low rank matrix bandit, employing an initial stage to estimate the latent opinion subspace. We show that the Friedkin-Johnsen objective admits an exact closed form representation in known, opinion free features, eliminating the need for subspace estimation. This structure allows us to delineate the amortization boundary, where offline pretraining replaces online search at a measured price. Under social homophily, a pretrained policy achieves regret governed by the rank of the shared opinion subspace plus its imitation error, with no polynomial dependence on network size and only a logarithmic penalty for enlarging the candidate intervention pool. Guided by this amortizable regime, we introduce the Learned Dynamics Tracker (LDT), a compact filter trained offline to denoise parameter estimates and deployed without updates across drifting streams. In streaming deployments, LDT achieves the lowest dynamic regret across all six horizons with statistically significant and widening margins, reducing regret by 21 percent over the strongest baseline at the longest horizon. Across twenty transfer settings, seven of them on real network topology, LDT leads the prior state of the art in every one.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Opinion Dynamics, Network Interventions, Amortized Learning, Dynamic Linear Bandits

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

This paper studies how to reduce opinion polarization and disagreement by modifying social-network connections when individuals’ innate opinions are unobservable and only noisy aggregate feedback is available. The authors show that, under the Friedkin–Johnsen model, intervention outcomes can be represented exactly as a linear relationship between known network-structural features and unknown opinion parameters. This eliminates the need to estimate the opinion subspace before learning. Based on this structure, the paper proposes Lift-TS for stationary settings and the Learned Dynamics Tracker (LDT) for settings with continuously drifting opinions. LDT combines parameter estimates and uncertainty information from multiple temporal windows in a low-dimensional spectral space and uses an offline-trained recurrent model to correct the online estimates. Experiments show that LDT achieves lower dynamic regret across different horizons, candidate-pool sizes, and network topologies.

Critical strengths
  1. The core structural observation is simple and important, eliminating the per-instance subspace-estimation stage required by prior methods.
  2. LDT’s low-dimensional parameter-tracking design is well aligned with the problem structure and is computationally efficient.
  3. The theory, algorithms, and experiments are generally well connected.
  4. The experiments include strong baselines, ablations, statistical tests, runtime analysis, and cross-network transfer evaluations.
  5. The authors clearly disclose that the real-network experiments use real topologies but synthetic opinion dynamics.
Critical weaknesses
  1. Theorem 2 evaluates posterior sampling on histories generated by the pretrained policy, but directly applies a regret bound established for histories generated by posterior sampling itself. A distribution-shift argument is missing.
  2. The drift-error bound in Theorem 3 may omit a dependence on the conditioning of the design matrix, potentially affecting the claimed dynamic-regret rate.
  3. The theoretically optimal window varies with the horizon and variation budget, whereas the deployed LDT is projected around a fixed reference window. The asymptotic guarantee for the deployed algorithm is therefore not fully established.
  4. Cross-graph transfer appears to require shared dynamics in spectral coordinates in addition to low-rank homophily, but this stronger assumption is not formalized.
  5. LDT still requires online feedback, parameter updates, and forced exploration, so the claim that pretraining “replaces online search” is overstated.
Questions for authors
  1. How can the history-distribution mismatch in Theorem 2 be addressed rigorously?
  2. Can the authors prove that the sliding-window drift error does not contain an additional design-conditioning or effective-sample-size factor?
  3. How does the deployed LDT with a fixed reference window achieve the claimed dynamic-regret rate?
  4. Does cross-graph transfer require an explicit assumption that parameter dynamics are shared in spectral coordinates?
  5. When innate opinions are unobservable, how can the gap-dependent pretraining error be estimated before deployment to determine whether amortization is appropriate?
5 comments
AuthorsOpenReview ↗

Response to Reviewer QCqN (1/3)

We sincerely thank the reviewer for the careful and constructive questions, which helped us make the theory more precise.

Our revised PDF colors every change: green for this review, orange for changes that also answer other reviews. In the revision, Theorem 2 is stated on the pretrained policy's own histories, with its proof in Appendix B; the drift-bias constant is explicit in Lemma 4 and in the proof of Theorem 3; the anytime certificate, Theorem 4′, is stated in §4.3; Assumption 2 is stated in §5.2; the phrase "replaces online search" is corrected; and Appendix F adds the pre-deployment estimate of RpreR_{\mathrm{pre}}.

W1 & Q1. Theorem 2 holds under the policy's own history distribution

No additional distribution-shift argument is needed. Theorem 2 holds under the pretrained policy's own history distribution, with the same form and one additional measurable term DTD_T; the rate is unchanged whenever DT=O(log⁡K)D_T=\mathcal O(\log K), and in our measurements DTD_T is below log⁡K\log K. The revised Theorem 2 states this form, and Appendix B gives the full proof.

Rigorous proof for Q1. The core inequality of Theorem 5 (the information-ratio bound, Russo and Van Roy, 2016, Proposition 5) holds round by round under any posterior, so it applies directly on the histories of the pretrained policy π\pi; only the final summation of the per-round information involves the history distribution, and we carry it out along the rollout of π\pi itself. With ptp_t exact posterior sampling evaluated on the histories of π\pi, this gives

E[RegretT(π)]≤max⁡{σ,Δmax⁡/2}2 deff T (log⁡K+DT)+Rpre,DT=∑tEπ[∑a(pt(a)−πt(a))+It(A⋆;Yt,a)],\mathbb E[\mathrm{Regret}_T(\pi)]\le\max\{\sigma,\Delta_{\max}/2\}\sqrt{2\,d_{\mathrm{eff}}\,T\,(\log K+D_T)}+R_{\mathrm{pre}},\qquad D_T=\sum_t\mathbb E_\pi\Big[\sum_a\big(p_t(a)-\pi_t(a)\big)_+I_t(A^\star;Y_{t,a})\Big],

with expectations under the history distribution of π\pi; DTD_T measures the information posterior sampling would have collected and π\pi did not. The proof has four steps:

  1. Proposition 4 at every history.
  2. The one-step regret of posterior sampling satisfies gt≤Γˉ Itpg_t\le\sqrt{\bar\Gamma\,I^{p}_t}, with Γˉ\bar\Gamma the information-ratio constant of Theorem 5.
  3. Itp≤Itπ+∑a(pt−πt)+It(A⋆;Yt,a)I^{p}_t\le I^{\pi}_t+\sum_a(p_t-\pi_t)_+I_t(A^\star;Y_{t,a}).
  4. Since π\pi randomizes independently of the parameter given the history, ∑tEπItπ=I(A⋆;HT+1)≤log⁡K\sum_t\mathbb E_\pi I^{\pi}_t=I(A^\star;H_{T+1})\le\log K along the rollout of π\pi; Cauchy-Schwarz concludes.

Direct comparison. Against the regret of posterior sampling on its own histories, a hybrid argument gives E[RegretT(π)]≤Ep[RegretT]+12Δmax⁡T(T+1)εˉ\mathbb E[\mathrm{Regret}_T(\pi)]\le\mathbb E_p[\mathrm{Regret}_T]+\tfrac12\Delta_{\max}T(T+1)\bar\varepsilon, where Ep\mathbb E_p denotes that rollout and εˉ\bar\varepsilon is the imitation error on the deployed policy's own histories.

Measurement. On the 40 instances of Table 36, DT=2.71D_T=2.71 nats, below log⁡K=4.61\log K=4.61, so the constant of the bound grows by at most a factor 1.26. The posterior-sampling term is small both on the policy's histories (0.39) and on posterior sampling's own histories (0.26), and both are two orders of magnitude below Rpre=28.06R_{\mathrm{pre}}=28.06, so the history mismatch has a negligible effect in practice.

AuthorsOpenReview ↗

Response to Reviewer QCqN (2/3)

W2 & Q2. No additional conditioning factor in the drift error

Yes: for the window estimate of Theorem 3(ii), which the revised proof states is computed on the scheduled design draws, the drift error contains no additional conditioning or effective-sample-size factor, so the rate T7/9T^{7/9} is unchanged; the conditioning only affects a constant independent of tt and WW. The lemma below is added as Lemma 4 in Appendix B, and the proof of Theorem 3 now states this estimate and carries its constant.

Reason. The scheduled design draws of Lemma 5 come from a GG-optimal design, so with high probability the mm draws in a window give λmin⁡(V)≥12mλ⋆\lambda_{\min}(V)\ge\tfrac12m\lambda^\star; the effective sample size is already the denominator of the variance term (Equation (21)), and it is what sets the T7/9T^{7/9} order.

Proof for Q2. For the ridge estimate on a window W\mathcal W with Gram matrix VV, the drift bias is bt=V−1∑i∈Wφiφi⊤(θi−θt)b_t=V^{-1}\sum_{i\in\mathcal W}\varphi_i\varphi_i^\top(\theta_i-\theta_t). With Ψt=max⁡i∈W∥θi−θt∥2\Psi_t=\max_{i\in\mathcal W}\|\theta_i-\theta_t\|_2,

(i)    ∥bt∥2≤Φmax⁡∣W∣/λmin⁡(V)  Ψt;(ii)    ∥bt∥2≤(2Φmax⁡2/λ⋆) Ψt    for the estimate on the scheduled draws of the window.\text{(i)}\;\;\|b_t\|_2\le\Phi_{\max}\sqrt{|\mathcal W|/\lambda_{\min}(V)}\;\Psi_t;\qquad\text{(ii)}\;\;\|b_t\|_2\le\big(2\Phi_{\max}^2/\lambda^\star\big)\,\Psi_t\;\;\text{for the estimate on the scheduled draws of the window.}

By (ii), the rate of Theorem 3(ii) is unchanged, with an extra factor (Φmax⁡2/λ⋆)1/3(\Phi_{\max}^2/\lambda^\star)^{1/3} in the constant. For estimates that use every round of the window (as in LDT), Lemma 5 gives ∣W∣/λmin⁡(V)≤4tα/(cλ⋆)|\mathcal W|/\lambda_{\min}(V)\le4t^{\alpha}/(c\lambda^\star) after a short burn-in under ρt=ct−α\rho_t=ct^{-\alpha}, and the rate is again T7/9T^{7/9} at α=2/9\alpha=2/9, the schedule of Theorem 4′ (W3 & Q3).

Measurement. Replaying the deployed LDT on five episodes of each of the 20 streams and computing btb_t and Ψt\Psi_t exactly at every decision round, the realized drift bias always stays below the factor allowed by (i), and its median is more than two orders of magnitude below it:

TTWindowRealized ∥bt∥/Ψt\lVert b_t\rVert/\Psi_t, medianFactor allowed by (i), median
20060/150/full0.017/0.030/0.0365.2/7.1/7.2
80060/150/full0.023/0.021/0.0405.3/8.3/13.9

W3 & Q3. Rate guarantees for the deployed LDT

The deployed LDT has two guarantees. First, with a fixed reference window, Proposition 2 bounds its regret along every realized trajectory, and the leading term of this bound grows as T0.77T^{0.77} on the deployed streams (Table 25), which is sublinear. Second, letting the reference window grow with the round index, with matching exploration and safeguard schedules and the trained filter unchanged, gives the claimed a-priori rate T7/9T^{7/9} (Theorem 4′), and this configuration is never significantly worse than the original one.

Reason. The theoretically optimal window grows with the horizon because a window of fixed length keeps a fixed number of observations (and, under a decaying exploration schedule, an ever smaller number of design draws), so its a-priori error bound does not decrease over time; this holds for any fixed-window method and is not specific to LDT. Theorem 4′ therefore lets the reference window of LDT grow with the round index, without knowing the horizon.

Theorem 4′. Let ρt=ct−2/9\rho_t=ct^{-2/9}, εt=ε0t−2/9\varepsilon_t=\varepsilon_0t^{-2/9} and Wt=⌈w0t2/3⌉W_t=\lceil w_0t^{2/3}\rceil. Under the budgeted path condition of Theorem 3, for filter weights ω\omega fixed before deployment, every TT and every δ∈(0,1)\delta\in(0,1), with probability at least 1−δ1-\delta,

RegretT≤O~([D(σdeffc w0 λ⋆+Φmax⁡ w0 B1c λ⋆+ε0)+c Δmax⁡] T7/9).\mathrm{Regret}_T\le\tilde{\mathcal O}\Big(\Big[D\Big(\sigma\sqrt{\tfrac{d_{\mathrm{eff}}}{c\,w_0\,\lambda^\star}}+\Phi_{\max}\,w_0\,\mathcal B\sqrt{\tfrac{1}{c\,\lambda^\star}}+\varepsilon_0\Big)+c\,\Delta_{\max}\Big]\,T^{7/9}\Big).

Every constant is explicit and independent of TT. The proof splits the regret into scheduled rounds, safeguard radius, anchor noise and drift bias; the anchor window holds Θ(t4/9)\Theta(t^{4/9}) design draws, so λmin⁡(Vt)≥14cw0λ⋆t4/9\lambda_{\min}(V_t)\ge\tfrac14cw_0\lambda^\star t^{4/9} (Lemma 5), the drift bias is controlled by (i) of W2 & Q2, and each of the four parts sums to O~(T7/9)\tilde{\mathcal O}(T^{7/9}). Theorem 4′ is now stated in §4.3, with its proof in Appendix B.

Performance. With (c,w0,ε0)=(0.05,1,1)(c,w_0,\varepsilon_0)=(0.05,1,1) chosen on the validation streams, the configuration of Theorem 4′ is also significantly better than the strongest baseline at every horizon (differences in regret):

Theorem 4′ minusT=100T=100T=200T=200T=400T=400T=800T=800T=1600T=1600
submitted LDT-1.1 [-1.6,-0.7]+1.1 [-0.6,+2.6]+2.8 [-1.4,+7.1]+6.3 [-2.9,+15.8]+7.0 [-9.7,+23.8]
SW-Ridge-4.6 [-7.4,-1.9]-12.0 [-19.5,-4.6]-28.4 [-42.9,-13.9]-66.0 [-96.2,-36.5]-131.4 [-193.9,-70.4]
AuthorsOpenReview ↗

Response to Reviewer QCqN (3/3)

W4 & Q4. Cross-graph transfer needs no shared-dynamics assumption

No. None of the guarantees for cross-graph transfer relies on an assumption of shared dynamics; such an assumption only determines how large the improvement from learning is, and the revision states it explicitly.

Reason. Proposition 2, Corollary 1 and Theorem 4′ hold for every filter weight, every drift trajectory and every graph (Theorem 4′ under Assumption 1 and the variation budget), so the performance after transfer to a new graph is never worse than these guarantees. Only the size of the improvement of LDT over its anchor (Corollary 1) depends on whether the deployment drift resembles the pretraining drift, and the revision states this as Assumption 2 in §5.2, used only for that gain statement:

Assumption 2 (shared spectral drift law). In the canonicalized spectral coordinates of Appendix E, the episode trajectories of a deployment are drawn from a law in the support of the pretraining distribution.

Experiments. Under drift laws outside the pretraining family, LDT keeps its advantage: on a persistent oscillator with no stable point (Table 2), it leads SW-Ridge significantly at T=1600T=1600, by 84.5 [37.8,133.9]; on a smooth random wander over the whole sphere, it has the lowest mean regret at every horizon. Across graphs, LDT is significantly below SW-Ridge on six of the seven real networks of Table 33 and lower in mean on the seventh, including networks 25 and 40 times larger than any pretraining graph.

W5. More precise wording

We have changed this phrase to "replaces per instance estimation and tuning" in the abstract, introduction, related work and conclusion. During deployment, LDT does not update its model parameters.

The online work of LDT consists of updating the window statistics and one forward step of the frozen recurrent network (0.11 ms per round, independent of ∣V∣|V|), with scheduled exploration on at most 2% of rounds (Table 10); LDT does use the observed feedback, as any online learner must.

By online search we mean the per-instance work of the two-stage approach on each new graph and episode, namely an exploration stage to estimate a representation followed by tuning and fitting of the learner; pretraining replaces exactly this work.

Q5. Estimation before deployment without observing opinions

It can be estimated without observing the true opinions: the pretraining term is an expectation over the prior distribution of opinions, so it depends only on the deployment graph, the candidate pool, the noise level and the prior, and can be computed before deployment from opinions simulated from the prior. On five networks, this estimate agrees with the value at deployment.

Explanation. The prior is the distribution the platform specifies in its model (the same family used in pretraining), not the true opinions; the computation follows the procedure of Table 36, and the true opinions are never read.

Experiment. The pre-deployment estimate uses 20 simulated opinion draws only, and the deployment value is computed on 20 further, disjoint draws (Appendix F):

Network (∣V∣\lvert V\rvert)RpreR_{\mathrm{pre}}, pre-deployment estimateRpreR_{\mathrm{pre}}, at deploymentRealized regret at deployment
Erdős-Rényi (32)50.0 [27.6,78.0]47.5 [31.3,65.9]47.4 [31.2,65.8]
Florentine (15)57.3 [32.2,86.1]64.3 [38.7,93.0]64.0 [38.5,92.6]
Davis (32)31.8 [19.8,51.7]39.9 [27.8,55.4]39.7 [27.7,55.2]
Karate (34)24.5 [19.0,30.6]25.0 [19.9,30.5]25.0 [19.9,30.6]
Les Misérables (77)216.7 [151.1,285.5]257.5 [150.4,388.1]256.0 [148.7,386.4]

On every network the pre-deployment estimate falls inside the interval of the deployment value and ranks the networks correctly, so whether amortization is appropriate can be judged before deployment. The same principle gives a pre-deployment rule for choosing between LDT and Lift-TS; on all 29 transfer settings of the paper, the learner it selects is never significantly worse than the better of the two.

Reviewer_dH49OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies online structural interventions for reducing polarization and disagreement under Friedkin–Johnsen opinion dynamics, where innate opinions are private and only noisy scalar feedback is available. The authors use a resolvent-based linear representation of the objective and introduce Lift-TS. For drifting opinions, they propose the Learned Dynamics Tracker (LDT), combining multiscale ridge estimates, uncertainty summaries, and recurrent residual corrections in a spectral parameter space. Experiments report improvements across multiple horizons and transfer settings.

Critical strengths

S1. This paper propose a well-designed parameter-space filtering. Estimating a shared parameter allows observations from different interventions to inform a common model and provides a coherent basis for transfer across candidate pools.

S2. The evaluation includes oracle-window trackers, multiple OPD-Min-ESTR variants, additional filtering baselines, component ablations, generator shifts, and paired statistical tests.

Critical weaknesses

W1. The central representation is already studied in the closest prior work. Cinus et al. (2026), Section 3, already define the resolvent action space and trace-inner-product objective used here. Their introduction also discusses direct linear-bandit reduction and its dimensional cost. Consequently, attributing their subspace-estimation stage to overlooking diffusion structure is misleading.

W2. The amortized regret proof does not adequately account for policy-induced history distributions.** Theorem 2 and Proposition 4 compare the learned policy with exact posterior sampling at the same history. After averaging over histories generated by the learned policy, however, the posterior-sampling term remains evaluated under that policy’s history distribution. It cannot automatically be replaced by the regret of posterior sampling executing its own actions and collecting its own observations. The proof does not establish the connection needed to apply Theorem 5. The additional guarantee under arbitrary test-distribution shift also requires assumptions connecting training-distribution imitation error to errors on deployment histories.

W3. The advertised dynamic regret rate is not established for the implemented fixed-anchor LDT. For a fixed positive variation budget, Theorem 3 obtains its rate using a window length growing as (W^\star=\Theta(T^{7/9})). The deployed algorithm instead fixes (W_{\mathrm{med}}=60) and projects its prediction into a shrinking neighborhood of that estimate. Equation (23) contains a leading (T^{7/6}/\sqrt W) term, which does not establish sublinear regret for fixed (W).

W4. The main drift generator limits the interpretation of the horizon-scaling results.** Appendix E explains that the nonlinear oscillator has a stable zero that many trajectories approach, producing sublinear path-length growth, empirically (B\propto T^{0.39}). Longer episodes can therefore become progressively more stable. Moreover, Table 15 reports approximately the same regret exponent, (0.67), for LDT, training-free Lift-TS, and Vel-Extrap. These results do not isolate how much of the observed scaling comes from learned tracking and how much comes from the generator’s stabilization. Although additional generator-shift experiments are provided, the headline scaling claims need explicit validation under persistent drift without stable attractors.

W5. The benefit of pretraining varies substantially across deployment conditions, without a validated selection criterion.** On Political blogs in Table 29, LDT incurs regret of 206.8 versus 161.5 for training-free Lift-TS. Table 28 also shows intervals including zero against the strongest alternative when interventions modify only four edges or are restricted to existing edges. This limits the operational value of the claimed amortization boundary and leaves the conditions justifying pretraining insufficiently specified.

Questions for authors
  1. What is the new technical contribution of the exact resolvent/trace representation relative to §3 of Cinus et al. (2026)?
  2. How can Theorem 2’s regret comparison be justified when the learned policy and exact posterior sampling induce different history distributions?
  3. For the implemented algorithm with fixed W_med, can the authors establish a sublinear regret guarantee without hiding horizon-dependent factors?
  4. Does LDT retain its advantage over training-free structural learners under persistent drift that does not approach a stable point?
  5. What observable criterion can determine when pretraining LDT is preferable to using Lift-TS?
4 comments
AuthorsOpenReview ↗

Response to Reviewer dH49 (1/3)

We sincerely thank the reviewer for the careful and rigorous review, which helped us sharpen both the positioning and the theory.

Our revised PDF colors every change: magenta for this review, orange for changes that also answer other reviews. In the revision, the introduction credits the trace-inner-product objective to Cinus et al. (2026) and drops the phrase on diffusion physics; Theorem 2 is stated on the pretrained policy's own histories, with its proof in Appendix B; the anytime certificate, Theorem 4′, is stated in §4.3; §6 reports the persistent-drift experiment (Table 2), the comparison of the two representations and the reward-range criterion, with the details and the validation rule in Appendix E.

W1 & Q1. Contributions beyond Cinus et al. (2026, §3)

Our new contribution is to show that the subspace-estimation stage is unnecessary: the dimension that governs learning is at most the number of candidates KK, and only 21 under homophily, independent of ∣V∣|V|; this yields the guarantees, the lower bound and the cross-graph transfer summarized below. In the paper's benchmark, even our training-free Lift-TS has only about 37% of the regret of their method (327.9 against 879.0 at T=800T=800).

On the direct linear-bandit reduction. Cinus et al. regard this reduction as unsatisfactory because the feature dimension and the regret both grow as ∣V∣2|V|^2, so they first estimate a subspace and then learn in a Θ(∣V∣)\Theta(|V|)-dimensional representation. We show that this is not the dimension that matters: the regret of the direct reduction is governed by the dimension of the span of the candidate features, at most KK (Theorem 5), and under homophily the graph spectrum gives an exact 21-dimensional representation at every network size (Proposition 5).

On our introduction. Our sentence asks whether the estimation stage is necessary, and the result above shows that it is not; it does not claim that they overlook the diffusion process. To avoid this reading, the revised introduction drops the phrase on diffusion physics and credits the trace-inner-product objective to Cinus et al. (2026) where it first appears.

Q1: the new technical contributions.

AspectCinus et al. (2026)This paper
Unknown parameterestimate the ∣V∣×∣V∣\lvert V\rvert{\times}\lvert V\rvert matrix Θ⋆\Theta^\star by T1T_1 rounds of uniform exploration and a nuclear-norm estimate, under restricted strong convexityno estimation stage: the reward-relevant part of Θ⋆\Theta^\star is identified in the span of features known before interaction (Theorems 1, 5)
Coordinates2∣V∣−12\lvert V\rvert-1, built from the estimated eigenvector of each instance21 at every network size, from the graph alone; exact under homophily and under drift within the homophily subspace (Proposition 5)
GuaranteesO~(max⁡{1/κ,∣V∣}∣V∣T)\tilde{\mathcal{O}}(\max\{1/\kappa,\sqrt{\lvert V\rvert}\}\sqrt{\lvert V\rvert{T}}), fixed opinionsO~(d⋆Tlog⁡K)\tilde{\mathcal{O}}(\sqrt{d_{\star}T\log{K}}) for any opinion prior (Theorem 5); O~(rT)+Rpre\tilde{\mathcal{O}}(r\sqrt{T})+R_{\mathrm{pre}}, free of ∣V∣\lvert V\rvert (Theorem 2); Ω(min⁡{K,T})\Omega(\min\{K,T\}) for structure-blind policies (Corollary 2); dynamic regret under drift (Theorem 4′)
Structure of the objective—exact multi-pair updates (Proposition 1); convexity and non-submodularity (Proposition 3); dimension gap of the attainable set (Lemma 1)
Settingone instance, fixed opinionsdrifting opinions; one frozen model across graphs (up to 4039 nodes) and pools
Cost per roundgrows as ∣V∣2.15\lvert V\rvert^{2.15}independent of ∣V∣\lvert V\rvert; 143×143\times faster at ∣V∣=1024\lvert V\rvert{=}1024
Regret at T=800T=800 (Table 1)879.0327.9 (Lift-TS), 270.9 (LDT)

Experiment: the two representations compared directly. We attached the same learner (SW-Ridge at its best window) to the reduced representation of Cinus et al. (2∣V∣−12|V|-1 coordinates), given the exact initial opinion direction so that it carries no estimation error, and to our spectral window (21 coordinates), so that any difference in regret comes from the representation alone (§6 and Appendix E):

Representation∣V∣=32,T=200\lvert V\rvert{=}32,T{=}200∣V∣=32,T=800\lvert V\rvert{=}32,T{=}800∣V∣=64,T=200\lvert V\rvert{=}64,T{=}200∣V∣=128,T=200\lvert V\rvert{=}128,T{=}200
Reduced, Cinus et al.167.3425.765.423.2
Spectral window (ours)113.6343.234.09.5

The spectral window is lower in every configuration, each paired 95% interval excluding zero, and roughly halves the regret at ∣V∣=64|V|=64 and 128: the reduced representation is exact only for the opinions it was built from and loses accuracy once they drift, while the spectral window stays exact under drift within the homophily subspace.

AuthorsOpenReview ↗

Response to Reviewer dH49 (2/3)

W2 & Q2. Theorem 2 along the policy's own histories

The difference in history distributions does not invalidate Theorem 2: its key inequality holds under any history distribution, so the posterior-sampling term is bounded directly on the pretrained policy's own histories, at the cost of one measurable term DTD_T, below log⁡K\log K in our measurements; the guarantee under arbitrary distribution shift needs no additional assumption. The revised Theorem 2 states this form, with the full proof in Appendix B.

Proof for Q2. The information-ratio bound used in Theorem 5 (Russo and Van Roy, 2016, Proposition 5) holds for a single round under any posterior. With ptp_t exact posterior sampling evaluated on the histories of the pretrained policy π\pi, applying it with the chain rule of mutual information along the rollout of π\pi gives

E[RegretT(π)]≤max⁡{σ,Δmax⁡/2}2 deff T (log⁡K+DT)+Rpre,DT=∑tEπ[∑a(pt(a)−πt(a))+It(A⋆;Yt,a)],\mathbb E[\mathrm{Regret}_T(\pi)]\le\max\{\sigma,\Delta_{\max}/2\}\sqrt{2\,d_{\mathrm{eff}}\,T\,(\log K+D_T)}+R_{\mathrm{pre}},\qquad D_T=\sum_t\mathbb E_\pi\Big[\sum_a\big(p_t(a)-\pi_t(a)\big)_+I_t(A^\star;Y_{t,a})\Big],

with expectations under the history distribution of π\pi itself; DTD_T is the information posterior sampling would have collected and π\pi did not, and the rate of Theorem 2 is unchanged whenever DT=O(log⁡K)D_T=\mathcal O(\log K). Four steps give the connection needed to apply Theorem 5:

  1. Proposition 4 at every history.
  2. The one-step regret of posterior sampling satisfies gt≤Γˉ Itpg_t\le\sqrt{\bar\Gamma\,I^{p}_t}, with Γˉ\bar\Gamma the information-ratio constant of Theorem 5.
  3. Itp≤Itπ+∑a(pt−πt)+It(A⋆;Yt,a)I^{p}_t\le I^{\pi}_t+\sum_a(p_t-\pi_t)_+I_t(A^\star;Y_{t,a}).
  4. Since π\pi randomizes independently of the parameter given the history, ∑tEπItπ=I(A⋆;HT+1)≤log⁡K\sum_t\mathbb E_\pi I^{\pi}_t=I(A^\star;H_{T+1})\le\log K along the rollout of π\pi; Cauchy-Schwarz concludes.

Arbitrary distribution shift. A hybrid argument gives E[RegretT(π)]≤Ep[RegretT]+12Δmax⁡T(T+1)εˉ\mathbb E[\mathrm{Regret}_T(\pi)]\le\mathbb E_p[\mathrm{Regret}_T]+\tfrac12\Delta_{\max}T(T+1)\bar\varepsilon, where Ep\mathbb E_p runs posterior sampling on its own histories and εˉ\bar\varepsilon is the imitation error on the deployed policy's own histories and can be estimated on the deployment graph with the procedure of Table 36, so no assumption linking the training distribution to deployment histories is needed.

Measurement. On the 40 instances of Table 36, DT=2.71D_T=2.71 nats, below log⁡K=4.61\log K=4.61, so the constant grows by at most a factor 1.26; the posterior-sampling term is 0.39 on the policy's histories and 0.26 on posterior sampling's own, both two orders of magnitude below Rpre=28.06R_{\mathrm{pre}}=28.06.

W3 & Q3. An anytime guarantee for the deployed tracker

For the implemented LDT with fixed WmedW_{\mathrm{med}}, Proposition 2 bounds the regret along every realized trajectory with no horizon-dependent factor, and its leading term grows sublinearly, as T0.77T^{0.77} [.73,.81] on the deployed streams (Table 25). Theorem 4′ gives an a-priori sublinear rate T7/9T^{7/9}, also free of horizon-dependent factors, by letting the anchor grow with the round index, with matching exploration and safeguard schedules; the trained filter is unchanged, and this configuration is never significantly worse than the submitted one.

On Equation (23). Its T7/6/WT^{7/6}/\sqrt W term reflects every fixed-length window, not LDT: a fixed window keeps a fixed number of observations, and under the decaying schedule ever fewer design draws, so its a-priori error bound cannot decrease over time. Theorem 3 therefore lets the window grow with TT, and Theorem 4′ does the same for the anchor of LDT.

Theorem 4′. Let ρt=ct−2/9\rho_t=ct^{-2/9}, εt=ε0t−2/9\varepsilon_t=\varepsilon_0t^{-2/9} and Wt=⌈w0t2/3⌉W_t=\lceil w_0t^{2/3}\rceil. Under the budgeted path condition of Theorem 3, for filter weights ω\omega fixed before deployment, every TT and every δ∈(0,1)\delta\in(0,1), with probability at least 1−δ1-\delta,

RegretT≤O~([D(σdeffc w0 λ⋆+Φmax⁡ w0 B1c λ⋆+ε0)+c Δmax⁡] T7/9).\mathrm{Regret}_T\le\tilde{\mathcal O}\Big(\Big[D\Big(\sigma\sqrt{\tfrac{d_{\mathrm{eff}}}{c\,w_0\,\lambda^\star}}+\Phi_{\max}\,w_0\,\mathcal B\sqrt{\tfrac{1}{c\,\lambda^\star}}+\varepsilon_0\Big)+c\,\Delta_{\max}\Big]\,T^{7/9}\Big).

Every constant is explicit and independent of TT, no horizon is input, and no factor depends on the ratio of the anchor to the optimal window. Theorem 4′ is now stated in §4.3, with its proof and a remark on Equation (23) after Corollary 6 in Appendix B.

Performance. With (c,w0,ε0)=(0.05,1,1)(c,w_0,\varepsilon_0)=(0.05,1,1) chosen on the validation streams, it is also significantly better than the strongest baseline at every horizon (regret differences):

Theorem 4′ minusT=100T=100T=200T=200T=400T=400T=800T=800T=1600T=1600
submitted LDT-1.1 [-1.6,-0.7]+1.1 [-0.6,+2.6]+2.8 [-1.4,+7.1]+6.3 [-2.9,+15.8]+7.0 [-9.7,+23.8]
SW-Ridge-4.6 [-7.4,-1.9]-12.0 [-19.5,-4.6]-28.4 [-42.9,-13.9]-66.0 [-96.2,-36.5]-131.4 [-193.9,-70.4]
AuthorsOpenReview ↗

Response to Reviewer dH49 (3/3)

W4 & Q4. LDT's lead grows under persistent drift

Yes. Under persistent drift that does not approach a stable point, LDT keeps the lowest regret and index at every horizon, and from T=200T=200 on its lead over the training-free structural learners grows with the horizon. Moreover, similar exponents in Table 17 do not mean similar performance.

On the exponents of Table 17 (Table 15 of the submission). An exponent describes how fast regret grows, not how large it is. On the paper's generator, LDT has the lowest regret at every horizon, and its lead over the strongest baseline, SW-Ridge, grows from 3.5 at T=100T=100 to 72.3 at T=800T=800 (Tables 1 and 18).

Persistent-drift experiment. To remove any effect of the generator settling, we built a persistent oscillator by capping the modulation of the paper's oscillator so that it has no stable point; its variation budget grows as T0.98T^{0.98} (against T0.39T^{0.39}). Regret of the frozen LDT, trained only on the paper's oscillator, and of the strongest training-free trackers, SW-Ridge and Vel-Extrap at their oracle windows (Table 2):

MethodT=100T=100T=200T=200T=400T=400T=800T=800T=1600T=1600
LDT (ours)69.8116.0203.4365.0679.6
SW-Ridge70.6116.3210.9395.1764.1
Vel-Extrap94.2139.9231.6414.6781.0

At T=1600T=1600, LDT leads SW-Ridge by 84.5 [37.8,133.9], Vel-Extrap by 101.4 [60.9,144.2] and Lift-TS by 190.7 [155.4,223.4].

W5 & Q5. An observable pre-deployment criterion

The observable criterion is the mean reward range over the candidate pool under the prior, which is computed from the graph and the pool before deployment: in all 24 of the paper's 29 transfer settings where this range is narrow (at most 0.2), LDT is better than Lift-TS, and only when the range is wide can Lift-TS be preferable.

The advantage of LDT. Across these 29 settings, LDT has an average regret of 114.3, against 128.0 for the strongest baseline. Political blogs, which the reviewer mentions, has the widest reward range of all settings (0.43): exactly the case in which the criterion points to Lift-TS, which is also our method, and even there LDT is significantly better than the strongest baseline, SW-Ridge.

The intervals including zero in Table 32 (Table 28 of the submission) arise only against SW-Ridge at its oracle window, and in those two settings LDT still has the lowest mean regret of all methods.

Role of the range. A wide range keeps the ranking of interventions stable even under a delayed estimate, which favours full-history averaging, while a narrow range calls for tracking.

Validation of a pre-deployment rule. The choice can also be made directly on the deployment graph: the platform simulates 10 validation episodes on its own graph and candidate pool under the nominal prior, runs both learners and keeps the one with lower validation regret. Average regret over all 29 settings (Appendix E):

SettingsSelected learnerAlways LDTAlways Lift-TS
One-factor variations (13 settings, Table 30)108.4107.4139.5
Generative scope (5 settings, Table 31)105.4104.2131.7
Candidate pool (4 settings, Table 32)89.288.9120.1
Real networks (7 settings, Table 33)142.4148.9155.5
All 29 settings113.4114.3139.3

In no setting is the selected learner significantly worse than the better of the two.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, I wanted to ensure we have addressed all your concerns satisfactorily.

Please let us know if any point needs further clarification or if you have additional questions. Your insights are invaluable to us, and we will respond promptly.

Thank you again for your time and thoughtful feedback.

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

The paper proposes to address the issue of minimizing opinion polarization and disagreement in social platforms where innate opinions are unknown. The authors first shows that expected intervention reward can be expressed as linear function of pairing innate opinions from the Friedkin-Johnsen(FJ) model with intervened Laplacian resolvent. They study how offline pretraining can reduce online search cost under an amortized regime. Followed by this, they propose their method LDT which combines observation stream, temporal windows and uncertainty summaries with a causal GRU to track dynamic beliefs in a FJ environment.

Critical strengths
  1. The paper addresses the issue of pretraining distribution in the offline setting and regrets in solving the opinion polarization and disagreement objective without knowing innate opinions and subsequent deployment of using the method after training in offline mode compared to online-based training complexity.
  2. The paper solves this by framing the intervention expected reward to be a linear combination of the intervened Laplacian and innate opinions.
  3. A good motivational figure explaining the objective well and gives an intuition of how the proposed method helps in tracking interventions in real deployment. .
  4. The paper is supported by a high volume of proofs- theorems and propositions, to say a few- and algorithms used in the paper but this has its own weakness as stated in Weakness section.
  5. The code provided looks mostly true to the algorithms provided.
  6. The experimental results are done against competitive baselines and perform better in every time horizon and network setting as shown in Table 1.
Critical weaknesses

The paper contains technical jargons which are difficult to associate in the first read. Current introduction makes it difficult to understand for the people with less familiarity with the domain. The paper starts with mathematical formulations at first. What is ss^T? Also, the terminology starts from problem definition. Why not explain the problem first and make the abstract and introduction easier to comprehend? The paper is mathematically dense to the point that the objective and the proposed solution are difficult to understand. A lot of times Lift TS has been talked about but not explained thoroughly what it does apart from being a sampling method? How different it is compared to LDT? Is Lift TS a part of the main LDT algorithm? It merely looks like a combination of [1] and [2] and contributes nothing to the LDT. The use of LDT with episode streams unclear. The fact that whether LDT is trained in offline mode with synthetic episodes or otherwise. [1]Russo, Daniel, and Benjamin Van Roy. "An information-theoretic analysis of thompson sampling." Journal of Machine Learning Research 17.68 (2016): 1-30. [2]Shipra Agrawal and Navin Goyal. Thompson sampling for contextual bandits with linear payoffs. In ICML, pp. 127–135, 2013

Questions for authors

What is Lift-TS apart from a mere Thomson sampling method? For the Lift TS objective how is phase 3 in the algorithm is associated with the code? It is difficult to comprehend from the compact code written? How is Theorem 2 different from Proposition 4? It seems like E[Regret_T(p)] is basically O(r sqrt(T)). It could have been explained in Theorem 2 itself. Observing the trend of polarization-disagreement index, it seems that at one point the index will be stagnated. Are there any experiments in an approximately very long horizon where this observation can be stated False and indeed LDT does work as the minimization objective is central to the method proposed. Also the baselines are tried on multiple seeds. What is mean and std of the regret and index? An ablation would be good on real/synthetic network where there is a dynamic opinion change and LDT fares well with their current methodology against the baselines.

5 comments
AuthorsOpenReview ↗

Response to Reviewer dJQU (1/3)

We sincerely thank the reviewer for the careful reading and the concrete suggestions, which help us improve the readability of the paper; we have also run the three experiments the reviewer asked for, and all three confirm the advantage of LDT.

Our revised PDF colors every change: blue for this review, orange for changes that also answer other reviews. The introduction now describes the problem in plain words and defines ss⊤ss^\top at its first use, Theorem 2 contains the decomposition of Proposition 4, §6 reports the T=1600T=1600 results and the persistent-drift experiment (Table 2), and Appendix F adds the comparison of Lift-TS with Thompson sampling without our features.

W1. Problem, objective and method

We thank the reviewer and apologize that the mathematical presentation at the start of the paper made it hard to follow.

The meaning of ss⊤ss^\top. The vector ss collects the private opinions of all users, and ss⊤ss^\top is the matrix whose entry (i,j)(i,j) is the product sisjs_is_j of the opinions of users ii and jj. The polarization-disagreement index after intervention aa is s⊤Xas=⟨Xa,ss⊤⟩s^\top X_as=\langle X_a, ss^\top\rangle, where the weights Xa=(I+La)−1X_a=(I+L_a)^{-1} depend only on the network and the intervention and are known in advance; only ss⊤ss^\top is unknown.

The problem. Following the reviewer's suggestion, the first paragraph of the revised introduction now describes the problem in plain words, and ss⊤ss^\top is defined where it first appears.

In short, each user holds a private opinion that the platform cannot see, and the opinion a user expresses is shaped by neighbours in the network. The platform wants to reduce how divided these opinions are; it cannot change opinions, but it can adjust connections, for example by strengthening or weakening recommendations. After each intervention it observes only a noisy measurement of this division, so it has to find, through repeated trials, the interventions that keep it lowest as the opinions slowly change.

Objective and method. The objective is to choose, round after round, interventions that make the polarization-disagreement index as low as possible, i.e. to keep the regret, the extra index compared with always choosing the best intervention, small. We propose two learners for this: Lift-TS, which needs no training and suits a single network with fixed or slowly changing opinions; and LDT, a small recurrent network trained once on synthetic data, which follows opinions that keep drifting and is used directly on new networks.

AuthorsOpenReview ↗

Response to Reviewer dJQU (2/3)

W2 & Q1. Lift-TS and its relation to LDT

Lift-TS is our training-free learner and is separate from LDT. Beyond [1] and [2], it rests on features that make Thompson sampling directly applicable to this problem, without an estimation stage, and on a guarantee for the sampler we actually run; without these features, the same type of sampler pays 4 to 49 times more regret.

Lift-TS in brief. While the opinions remain unknown, it finds online the interventions with the lowest index: it keeps a Gaussian estimate of the unknown part, and in each round draws a sample, plays the intervention that is best under it, and updates the estimate.

Contributions beyond [1] and [2].

  1. Features that make Thompson sampling directly usable. [1] and [2] assume the action features are given; here the unknown is the whole opinion matrix ss⊤ss^\top, which the existing approach (Cinus et al., 2026) first estimates in an exploration stage. We show that learning needs only a space of dimension at most the number of candidates (Theorem 5), and that under homophily a few features computed from the network represent every intervention exactly (Proposition 5).
  2. A guarantee for the sampler we actually use. The theory of [1] requires the true parameter to follow the Gaussian distribution used by the algorithm, which the opinion matrix ss⊤ss^\top does not; building on [2], Theorem 1 gives a regret guarantee for our sampler.
  3. Cost that does not grow with the network. Apart from a one-time feature computation, from 128 to 1024 users the per-round time of Lift-TS and LDT changes by at most 1.0%, while that of OPD-Min-ESTR grows 89-fold (Table 4).

New experiment, against the same sampler without our features, which treats every intervention as an unrelated arm:

Candidates KKLift-TS (ours)Same sampler without our featuresOPD-Min-ESTR
10060.4242.7903.2
40060.11147.91209.5
1600102.65017.51219.2

Every difference is significant; when KK grows 16-fold, the regret of Lift-TS grows 1.7-fold and that of the sampler without our features 20.7-fold (Appendix F).

Relation to LDT. Both learners are built on the resolvents XaX_a of Section 3 and serve different situations:

AspectLift-TSLDT
Intended useone network whose opinions are fixed or change slowlyepisode streams with drifting opinions
Trainingnonetrained once offline, frozen, reused on every network
Memoryall past roundsestimates over several window lengths and the whole episode, combined by the learned network
Work per roundone Gaussian update and one random drawrank-one window updates and one recurrent step

W3. Offline training of LDT and its use on episode streams

LDT is trained once, offline, on simulated episodes; on an episode stream, the same frozen model processes the episodes one after another, restarting its estimates at the beginning of each episode.

Use on episode streams. A stream is a sequence of episodes on the same network, and in each episode the opinions drift along a new path. In each round, LDT estimates the current opinions from its observations so far, lets the trained recurrent network correct the error caused by drift, and plays the intervention that currently looks best. All episodes and networks share the same model.

Offline training. Yes: before deployment, on 1500 simulated episodes (random networks with 16 to 48 users, a range of drift speeds and noise levels). Its parameters are then fixed; nothing is trained or tuned during deployment.

Q2. Phase 3 of Algorithm 1 in the code

StepAlgorithm 1, Phase 3Code
Draw a sampleϵt∼N(0,Id⋆),  θ~t←μt−1+c Ct−1ϵt\epsilon_t\sim\mathcal{N}(\mathbf{0},I_{d_\star}),\;\tilde\theta_t\gets\mu_{t-1}+c\,C_{t-1}\epsilon_tL=np.linalg.cholesky(Sig+1e-9*np.eye(d));th=mu+c*([email protected]_normal(d))
Choose the best intervention for itat←arg⁡min⁡a∈Aψa⊤θ~ta_t\gets\arg\min_{a\in\mathcal{A}}\psi_a^\top\tilde\theta_ta=int(np.argmin(af@th))
Observe the noisy indexyt=ψat⊤θP⋆+ζty_t=\psi_{a_t}^\top\theta^\star_P+\zeta_tr=means[t,a]+sigma*rng.standard_normal()
Predictive variancevt←Σt−1ψat,  κt←σ2+ψat⊤vtv_t\gets\Sigma_{t-1}\psi_{a_t},\;\kappa_t\gets\sigma^2+\psi_{a_t}^\top v_tp=af[a];Sp=Sig@p;den=sigma**2+p@Sp
Update the meanμt←μt−1+κt−1(yt−ψat⊤μt−1)vt\mu_t\gets\mu_{t-1}+\kappa_t^{-1}(y_t-\psi_{a_t}^\top\mu_{t-1})v_tmu=mu+(r-p@mu)*Sp/den
Update the covarianceΣt←Σt−1−κt−1vtvt⊤\Sigma_t\gets\Sigma_{t-1}-\kappa_t^{-1}v_tv_t^\topSig=Sig-np.outer(Sp,Sp)/den
Cholesky factorCtC_t from Σt\Sigma_t by rank-one downdaterecomputed in full at the next draw (first row), giving the same CtC_t
AuthorsOpenReview ↗

Response to Reviewer dJQU (3/3)

Q3. Theorem 2 and Proposition 4

  • Proposition 4 is a decomposition valid for any policy: the regret of the pretrained policy is at most a posterior-sampling term plus RpreR_{\mathrm{pre}}, the cost of deviating from posterior sampling. It needs no structural assumption and gives no rate.
  • Theorem 2 bounds both terms under the low-rank structure of Assumption 1: the posterior-sampling term E[RegretT(p)]\mathbb{E}[\mathrm{Regret}_T(p)] is of order O~(rT)\tilde{\mathcal{O}}(r\sqrt{T}), up to a small, measured information term, and RpreR_{\mathrm{pre}} is controlled by the pretraining error, which gives the rate.

Proposition 4 is thus the first step of the proof of Theorem 2; following the reviewer's suggestion, the revised Theorem 2 states this decomposition directly.

Q4. LDT keeps improving at longer horizons

The index does not stagnate: it approaches a floor that no method can go below, and the distance of LDT to this floor keeps shrinking. At twice the longest horizon of the main benchmark, LDT still has the lowest index and significantly lower regret than every baseline evaluated.

We extended the main benchmark to T=1600T=1600 (§6). The floor is the index reached by choosing the best intervention in every round (about 0.199), and the flattening in Table 1 is the approach to it:

TT1002004008001600
Floor0.19960.19910.19900.19890.1989
LDT index0.23310.22420.21910.21580.2141
LDT distance to the floor0.03350.02510.02010.01690.0152
SW-Ridge index0.23480.22750.22300.22030.2184

Regret at T=1600T=1600:

MethodRegretLead of LDT [95% CI]
LDT (ours)486.6—
SW-Ridge625.1138.5 [81.4,197.5]
Vel-Extrap664.0177.4 [117.1,240.1]
D-LinUCB889.2402.6 [348.0,457.2]
OPD-Min-ESTR1718.21231.6 [1082.9,1398.1]

Persistent drift. When the opinions keep drifting and never settle (Table 2), LDT, without retraining, still has the lowest regret and index at every horizon up to T=1600T=1600, where it leads SW-Ridge by 84.5 [37.8,133.9] and Vel-Extrap by 101.4 [60.9,144.2].

Q5. Mean and standard deviation of regret and index

LDT has the lowest mean regret and the lowest mean index at every horizon. Mean±standard deviation over the 20 streams of Table 1:

Regret:

MethodT=100T=100T=200T=200T=400T=400T=800T=800
LDT (ours)66.9±16.5100.5±27.0160.5±47.5270.9±88.8
SW-Ridge70.4±17.3113.6±30.6191.7±56.8343.2±116.3
Vel-Extrap94.5±22.8140.3±34.9223.3±59.3377.3±112.0
D-LinUCB94.3±16.9165.5±28.1292.0±47.0511.8±74.4
BOB121.6±23.6223.3±41.1439.4±84.7849.0±162.1
OPD-Min-ESTR120.9±28.6231.7±56.1449.8±111.0879.0±213.0
LowPopArt136.2±32.5261.0±60.8501.3±114.2962.5±211.3

Index:

MethodT=100T=100T=200T=200T=400T=400T=800T=800
LDT (ours)0.233±0.0230.224±0.0210.219±0.0210.216±0.020
SW-Ridge0.235±0.0230.228±0.0220.223±0.0220.220±0.021
Vel-Extrap0.247±0.0250.234±0.0230.227±0.0220.222±0.021
D-LinUCB0.247±0.0220.240±0.0210.235±0.0200.231±0.019
BOB0.260±0.0250.255±0.0240.254±0.0240.252±0.023
OPD-Min-ESTR0.260±0.0270.257±0.0270.255±0.0260.254±0.026
LowPopArt0.268±0.0290.264±0.0280.262±0.0270.259±0.026

The standard deviations mainly reflect differences between streams, each with its own network, candidates and opinion paths; all 48 paired comparisons are significant after correction for multiple testing (Table 18).

Q6. Ablation on real networks with drifting opinions

Every component of LDT contributes: removing any one of them raises the regret on all three networks, and removing the learned dynamics costs the most.

On three real networks (drifting opinions, 200 rounds, 20 streams per network), we removed one component of LDT at a time: the learned dynamics (the recurrent network that corrects for drift), the multiscale grid (estimates from several window lengths) and the uncertainty summaries (the reliability of each window estimate). Without the learned dynamics, LDT reduces to sliding-window ridge regression, so that column reports SW-Ridge at its best window, the strongest baseline.

Network (users)Full LDTWithout uncertainty summariesWithout multiscale gridWithout learned dynamics (SW-Ridge)
Karate (34)100.1106.1107.2112.0
Les Misérables (77)167.6172.6168.1210.2
Hamsterster (2000)132.4135.2145.0169.0

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.