Pith. sign in

REVIEW 5 major objections 5 minor 155 references

PRL claims that any deep recommender can be improved by a plug-and-play Bayesian layer that models the prediction residual and debiases domain confounders via do-calculus.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 09:07 UTC pith:V5HIYYNX

load-bearing objection Worth a look for the plug-and-play residual+causal idea, but the printed training algorithm is mathematically undefined for vector confounders; major revision plus code/data release is needed before the claims can be trusted. the 5 major comments →

arxiv 2607.20863 v1 pith:V5HIYYNX submitted 2026-07-23 cs.IR cs.AIcs.LG

Probabilistic Residual Learning for Online Recommendations

Wenyuan Wang , Yusong Zhao , Zihao Xu , Hengyi Wang , Qi Xu , Zhigang Hua , Yan Xie , Yi Wang
show 6 more authors
Zihao Zhao Bo Long Chengzhi Mao Shuang Yang Hengguan Huang Hao Wang
This is my paper
classification cs.IR cs.AIcs.LG
keywords Bayesian Deep LearningProbabilistic Graphical ModelRecommender SystemsCross-Domain RecommendationCold-StartResidual LearningCausal InferenceDo-Calculus
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that any existing deep recommender can be upgraded by adding a lightweight Bayesian layer that learns the difference between the base model's predictions and the ground truth—the residual rating—and then corrects for it, without retraining the base model. It argues this matters most in cold-start cross-domain settings, where base models are misled by domain-specific spurious patterns such as country-level exposure or popularity effects. To that end, PRL models the residual with a hierarchical Bayesian model that clusters users into latent groups and uses do-calculus to average over a domain-level confounder, aiming to remove the confounder's influence. The experiments report consistent gains when PRL is added to five different base recommenders on two datasets, with the full causal version outperforming its non-causal counterpart.

Core claim

The central claim is that cross-domain cold-start recommendation can be reformulated as a plug-and-play probabilistic residual-correction problem. Given any fixed base recommender, PRL learns the residual eR = R - \hat{R} with a hierarchical Bayesian model in which a domain-level confounder s_m affects both user and item latent vectors and the residual, and users are assigned to latent clusters. At inference, for the user's inferred cluster, the residual is predicted by the do-calculus marginalization of Eq. 28, p^{(k)}(eR | do(u_i), do(v_j)) = \sum_m p^{(k)}(eR | u_i, v_j, s_m) p(s_m), which the paper argues cuts the spurious effect of the domain factor. Experiments on XMRec (18 countries)

What carries the argument

The load-bearing mechanism is the residual rating eR_{ij} = R_{ij} - \hat{R}_{ij}, combined with the do-calculus debiasing formula of Eq. 28. Users are probabilistically clustered using a Gaussian-mixture prior on the user latent u_i, whose cluster assignment π_i selects a cluster-specific sub-model. A domain-level confounder s_m (a 'country' embedding in XMRec, a release-year embedding in MovieLens) is modeled as influencing u_i, v_j, and eR, and is inferred with a variational domain-index approximation. At prediction time, Eq. 28 replaces the raw conditional p(eR | u_i, v_j, s_m) with the interventional average \sum_m p(eR | u_i, v_j, s_m) p(s_m), which removes the confounder's direct infl

Load-bearing premise

The causal graph assumed in Fig. 2 — that a single domain-level confounder s_m is the only common cause influencing user and item latents and the residual — is load-bearing; if unobserved confounders exist or s_m is not a true confounder, the do-calculus marginalization of Eq. 28 does not recover a true interventional distribution.

What would settle it

Run PRL on a synthetic dataset where the generative process includes a known observed domain variable plus a hidden confounder that also affects both user/item latents and the rating. If PRL's post-hoc correction fails to remove the hidden confounder's bias in top-k ranking (compared against a ground-truth no-confounding system), the identifiability claim behind Eq. 28 is falsified. Similarly, on real data with per-domain item popularity measurements, check whether PRL equalizes exposure across domains beyond what a non-causal version does; the paper only reports one such case study.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any frozen base recommender can be enhanced without retraining it, saving compute in online systems where the base model is already serving traffic.
  • The causal marginalization reduces country-specific recommendation bias; the paper's case study on XMRec reports a 38.5% reduction in the imbalance of camera recommendations across countries.
  • PRL automatically discovers user clusters that are interpretable, making it possible to inspect which user groups drive the residuals.
  • The full causal PRL outperforms the non-causal version across all reported base models, splits, and cutoff values, indicating the debiasing step is the main contributor to the gains.
  • The gains persist even when the base model is scaled up to roughly match PRL's parameter count, suggesting the improvement is structural rather than a capacity artifact.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The residual-correction recipe could transfer to other frozen prediction models (e.g., ranking, click-through rate, forecasting) where retraining is too expensive, using the same cluster-and-debias pattern.
  • Real markets likely have multiple confounders at once; extending PRL to a vector-valued s_m and testing whether additional debiasing helps would be a natural next step.
  • Because the clusters form around residual errors, they could double as a diagnostic tool: clusters with concentrated large residuals may reveal where the base model is systematically blind, pointing to targeted fixes.
  • A synthetic experiment with known ground-truth confounders would directly test whether Eq. 28 actually recovers the interventional distribution; the paper does not report such a validation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes Probabilistic Residual Learning (PRL), a plug-and-play framework that models the residual between a fixed base recommender's predictions and ground-truth ratings using a hierarchical Bayesian model with user clustering, domain-level confounders, and do-calculus-based causal adjustment. The authors report experiments on XMRec and MovieLens with five base recommenders (CDL, DLRM, PerK, NCF, LightGCN), showing that PRL improves the base models in cold-start cross-domain settings and that the causal component provides further gains.

Significance. The core idea—enhancing arbitrary deep-learning recommenders by learning a probabilistic residual correction with user clusters and confounder adjustment—is appealing and potentially practical. The paper includes multi-dataset, multi-base-model experiments and an ablation, which are useful. However, the technical derivation contains several load-bearing errors: the stated KL divergence is incorrect, the update rules invert a singular matrix, and the causal adjustment formula is not a valid backdoor adjustment. These issues call into question whether the reported results could have been produced by the algorithm as described. The contribution is not yet established; the manuscript requires substantial revision before it can be considered for publication.

major comments (5)
  1. [Section 2.2, Eq. (17)] The expression labeled D_KL(q(v_j) || p(v_j|x_v_j,s_m)) is not the KL divergence between a Gaussian q(v_j)=N(μ_v, Λ_v^{-1}I) and the product-of-Gaussians prior. With the prior's precision (Λ_v+λ_v)I and mean (Λ_v f_v + λ_v W_v s_m)/(Λ_v+λ_v), the quadratic term should be (Λ_v+λ_v)/2 ||μ_v − μ_pog||^2, not −λ_v/2||μ_v−W_v s_m||^2 − Λ_v/2||μ_v−f_v(x_v)||^2. Since this term is part of the ELBO and drives the update rules, the learning objective is not correctly derived.
  2. [Section 2.3, Eqs. (25)–(27)] The update rules for W_u, W_v, and w_R contain (s_m s_m^T)^{-1}. For any g>1, s_m s_m^T is a rank-one matrix and is singular; Section 2.1 allows g≪h with no restriction g=1, and Section 2.4 suggests learning country embeddings, which violates the invertibility assumption. Moreover, Algorithm 1 updates these parameters inside the per-domain loop, but Eqs. (25)–(27) sum over all users/items, so each iteration overwrites the previous domain's update instead of accumulating evidence. The algorithm as printed cannot be executed for typical g.
  3. [Section 2.4, Eq. (28)] The proposed causal adjustment p(eR|do(u),do(v)) = Σ_{m=1}^M p(eR|u,v,s_m)p(s_m) is not a well-defined backdoor formula. The generative model defines s_m as a continuous vector with prior N(0,I), so summing over a discrete domain index is meaningless; a correct adjustment would integrate over the confounder distribution. The paper also does not verify that the backdoor criterion holds for the graph in Fig. 2—e.g., it does not exclude unobserved confounders between u_i, v_j, and eR. The causal debiasing claim is therefore unsupported.
  4. [Tables 2 and S2] The 'DLRM PRL w/o Causality' row in Table S2 (MovieLens) is numerically identical to the 'DLRM PRL w/o Causality' row in Table 3 (XMRec): 0.0232, 0.0026, 0.0039, 0.0014, 0.0014. This is almost certainly an error and undermines confidence in the MovieLens results. Additionally, in the first MovieLens split, PRL (Full) slightly degrades DLRM's NDCG@20 (0.2431 vs 0.2433) and MAP@20 (0.0284 vs 0.0285), which is inconsistent with the blanket claim that PRL improves base models. No standard deviations or significance tests are reported.
  5. [Section 3.3] The experimental evaluation compares PRL only with its own base recommenders and an ablated version. Since the paper's stated scope is cross-domain recommendation, it should compare against existing cross-domain recommendation methods (e.g., CoNet, DDTCDR, or domain-adaptation baselines). Without such comparisons, and without code/data, the claim of consistent cross-domain improvement is not established. The authors should also report statistical significance across multiple runs.
minor comments (5)
  1. [Section 2.1] The notation f_x(·) appears in Algorithm 1 and Section 2.3('4)') but is not defined; the text uses f_v(·) for the item encoder. Please clarify.
  2. [Section 2.2, Eq. (10)] The term E_q(π)[p(π|θ)] is missing a log; in the standard ELBO it should be E_q(π)[log p(π|θ)]. This appears to be a typo but should be corrected.
  3. [Eq. (22)] The notation V λ_{eR(i,:)} V^T is ambiguous: λ_{eR(i,:)} is defined as a diagonal matrix over J items, but a user may have rated only a subset. The authors should specify how missing entries are handled (e.g., zero-padding or masking).
  4. [Table S11] The final row contains '.0002' instead of '0.0002', a minor formatting error.
  5. [General] The paper alternates between 'eR' and 'e_R' for the residual rating; please use a consistent notation throughout.

Circularity Check

0 steps flagged

No significant circularity: the residual target is defined independently of the base predictor, and the causal adjustment is a standard formula evaluated on held-out data.

full rationale

PRL's derivation chain is not circular. The residual is defined as eR_ij = R_ij - bR_ij with bR fixed (Sec. 2.1), so the final prediction R = bR + eR is an additive correction. Learning eR from observed residuals and evaluating on held-out cold-start users/domains (Tables 1-3, Fig. 4) is a genuine out-of-sample empirical claim, not a tautology: a residual model could in principle hurt performance. Equation 28 is the standard backdoor adjustment for the graph in Fig. 2, p(eR|do(u),do(v)) = sum_m p(eR|u,v,s_m)p(s_m); it is not fitted from the data and is compared against PRL w/o Causality on held-out metrics. The self-cited VDI [40] is used only as an estimator for the confounder embedding and is an externally published method, while the product-of-Gaussians identity is cited to Gales & Airey [5]; neither is a self-citation used to forbid alternatives. Two soundness concerns are worth noting but are not circularity: Eq. 25-27 use (s_m s_m^T)^{-1}, which is undefined for g>1 and suggests the printed update rules are not directly executable, and the backdoor/identifiability condition for s_m is assumed rather than proven. These affect correctness/verifiability, not whether the prediction reduces by construction to its inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The model introduces several fitted parameters (cluster centers/covariances, global mappings, domain embeddings, encoder weights) beyond the base recommender. The causal debiasing relies on a domain SCM that is assumed, not tested. The confounder s_m is an invented latent variable with no independent evidence.

free parameters (6)
  • Cluster parameters {μ_k, Σ_k} = K=3, values learned from data (Eq. 20-21)
    Gaussian mixture parameters for user clusters, updated via EM-like updates.
  • W_u, W_v, w_R = learned via Eq. 25-27
    Mappings from confounder s_m to user latent, item latent, and residual rating.
  • Domain confounder s_m = estimated via VDI [40]
    Domain embedding that represents country/market or release-year; used in causal adjustment.
  • K (number of user clusters) = K=3 in all experiments
    Chosen a priori, not validated; affects model complexity and cluster granularity.
  • Precision constants λ_eR, λ_v, Λ_v, λ_u, Λ_u = not specified
    Treated as constants in the paper, but values are not reported; effectively free hyperparameters.
  • Encoder f_v parameters = trained via gradient ascent
    Parameters of the deep encoder that maps item features to item latent prior.
axioms (4)
  • domain assumption The generative SCM in Fig. 2: domain-level confounder s_m affects user latent u_i and item latent v_j, and residual rating eR depends on u_i, v_j, and s_m.
    This causal model justifies the debiasing step. If s_m is not a confounder or other confounders exist, the causal claim fails.
  • domain assumption Backdoor criterion holds for Eq. 28: conditioning on s_m blocks all confounding paths between (u_i, v_j) and eR.
    The do-calculus formula marginalizing over s_m is valid only if s_m is the sole confounder. The paper does not verify this.
  • standard math Product-of-Gaussians prior and mean-field variational factorization (Eq. 7-9) are used.
    Standard Bayesian modeling assumptions; the KL equations (Eq. 11-17) rely on these.
  • domain assumption VDI [40] provides a valid posterior approximation p(s_m | eR, x_v, clusters).
    The confounder s_m is estimated by a self-cited variational domain indexing method, which is treated as a black box.
invented entities (1)
  • Domain-level causal confounder s_m no independent evidence
    purpose: Represents domain-specific bias (e.g., country, release year) that influences user and item representations and residual ratings; used for debiasing via Eq. 28.
    Introduced as a latent/observed domain variable. No external falsifiable handle is given; its validity is only assessed through downstream performance and a case study.

pith-pipeline@v1.3.0-alltime-deepseek · 25462 in / 16895 out tokens · 166912 ms · 2026-08-01T09:07:28.917735+00:00 · methodology

0 comments
read the original abstract

Modern recommender systems are typically based on deep learning (DL) models, where a dense encoder learns representations of users and items. As a result, these systems often suffer from the black-box nature and computational complexity of the underlying models, making it difficult to systematically enhance their recommendation capabilities. To address this problem, we propose Probabilistic Residual Learning (PRL), a causal Bayesian recommendation model that models the residual between ground-truth and base predictions, enabling targeted refinement of existing systems. Specifically, PRL (1) probabilistically groups users for localized residual modeling, (2) models domain-level confounders that influence user and item representations, and (3) aggregates cluster-specific residual predictions over the confounders using do-calculus. Experiments demonstrate that our plug-and-play PRL is compatible with various base deep learning recommender systems, improving their performance while automatically discovering meaningful user clusters.

Figures

Figures reproduced from arXiv: 2607.20863 by Bo Long, Chengzhi Mao, Hao Wang, Hengguan Huang, Hengyi Wang, Qi Xu, Shuang Yang, Wenyuan Wang, Yan Xie, Yi Wang, Yusong Zhao, Zhigang Hua, Zihao Xu, Zihao Zhao.

Figure 1
Figure 1. Figure 1: Overview of PRL. During training, PRL models the residuals of a fixed base recommender, probabilistically clusters [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Probabilistic graphical model of PRL. The cluster [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Causal inference in PRL is equivalent to removing [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Recall@N on all three XMRec splits, comparing each base model with PRL without causality and full PRL. fine-grained results for different clusters discovered by PRL in Ap￾pendix. Results show that our PRL can usually improve performance in most clusters. Case Study on Debiasing. To explore PRL’s debiasing capability and cross-domain generalization, we conducted a detailed analysis using the first domain pa… view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE projections of PRL user latents u𝑖 on XMRec. Colors denote cluster assignments 𝜋𝑖 inferred by Eqn. 24, rather than ground-truth labels. 0 2000 4000 6000 8000 10000 Item Index 0.2 0.4 0.6 Score Item Scores by Cluster Clusters Cluster 1 Cluster 2 Cluster 3 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: User clusters based on users’ highest rated items, us [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

155 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Michael Downes and Barbara Beeton , organization =. The

  2. [2]

    Cristophe Fiorio , year = 2015, month = oct, note =

    2015

  3. [3]

    The algorithms bundle , author =

  4. [4]

    The Listings Package , author =

  5. [5]

    Publication quality tables in

    Simon Fear , year = 2005, month = apr, note =. Publication quality tables in

    2005

  6. [6]

    The subcaption package , author =

  7. [7]

    A package to create a nomenclature , author =

  8. [8]

    User Manual for glossaries.sty v4.44 , author =

  9. [9]

    David Carlisle , month = oct, year = 2004, note =. The

    2004

  10. [10]

    TKDE , year=

    Domain adaptation on graphs by learning aligned graph bases , author=. TKDE , year=

  11. [11]

    CVPR , pages=

    Adagraph: Unifying predictive and continuous domain adaptation through graphs , author=. CVPR , pages=

  12. [12]

    KDD , pages=

    Collaborative knowledge base embedding for recommender systems , author=. KDD , pages=. 2016 , organization=

    2016

  13. [13]

    ECCV , volume =

    Zhengming Ding and Sheng Li and Ming Shao and Yun Fu , title =. ECCV , volume =

  14. [14]

    Joty and Muhammad Imran , editor =

    Firoj Alam and Shafiq R. Joty and Muhammad Imran , editor =. Domain Adaptation with Adversarial Training and Graph Embeddings , booktitle =

  15. [15]

    SIU , pages=

    Domain Adaptation on Graphs via Frequency Analysis , author=. SIU , pages=

  16. [16]

    arXiv preprint arXiv:2004.10240 , year=

    Neural forecasting: Introduction and literature overview , author=. arXiv preprint arXiv:2004.10240 , year=

    Pith/arXiv arXiv 2004

  17. [17]

    arXiv preprint arXiv:1906.05264 , year=

    Gluonts: Probabilistic time series models in python , author=. arXiv preprint arXiv:1906.05264 , year=

    Pith/arXiv arXiv 1906

  18. [18]

    ICML , year=

    Continuously Indexed Domain Adaptation , author=. ICML , year=

  19. [19]

    RecSys , pages=

    Deep neural networks for youtube recommendations , author=. RecSys , pages=

  20. [20]

    NIPS , pages=

    Deep learning for precipitation nowcasting: A benchmark and a new model , author=. NIPS , pages=

  21. [21]

    Convolutional

    Shi, Xingjian and Chen, Zhourong and Wang, Hao and Yeung, Dit-Yan and Wong, Wai-Kin and Woo, Wang-chun , booktitle=. Convolutional

  22. [22]

    KDD , pages=

    Collaborative variational autoencoder for recommender systems , author=. KDD , pages=

  23. [23]

    NIPS , pages=

    Collaborative recurrent autoencoder: Recommend while learning to fill in the blanks , author=. NIPS , pages=

  24. [24]

    TDKE , volume=

    Towards Bayesian deep learning: A framework and some existing methods , author=. TDKE , volume=

  25. [25]

    CSUR , volume=

    A Survey on Bayesian Deep Learning , author=. CSUR , volume=

  26. [26]

    AAAI StarAI Workshop , year=

    Causal discovery from incomplete data: A deep learning approach , author=. AAAI StarAI Workshop , year=

  27. [27]

    Frontiers in genetics , volume=

    Review of causal discovery methods based on graphical models , author=. Frontiers in genetics , volume=. 2019 , publisher=

    2019

  28. [28]

    ICML , year =

    Jingquan Yan and Hao Wang , title =. ICML , year =

  29. [29]

    arXiv preprint arXiv:2203.11014 , year=

    DHEN: A deep and hierarchical ensemble network for large-scale click-through rate prediction , author=. arXiv preprint arXiv:2203.11014 , year=

    Pith/arXiv arXiv

  30. [30]

    2017 , school=

    Bayesian Deep Learning for Integrated Intelligence: Bridging the Gap between Perception and Inference , author=. 2017 , school=

    2017

  31. [31]

    2022 , booktitle=

    OrphicX: A Causality-Inspired Latent Variable Model for Interpreting Graph Neural Networks , author=. 2022 , booktitle=

    2022

  32. [32]

    Nature medicine , year=

    Assessment of Medication Self-Administration using Artificial Intelligence , author=. Nature medicine , year=

  33. [33]

    Nature medicine , volume=

    Artificial intelligence-enabled detection and assessment of Parkinson’s disease using nocturnal breathing signals , author=. Nature medicine , volume=. 2022 , publisher=

    2022

  34. [34]

    arXiv preprint arXiv:1111.0352 , year=

    Revisiting k-means: New algorithms via Bayesian nonparametrics , author=. arXiv preprint arXiv:1111.0352 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:1906.00091 , year=

    Deep learning recommendation model for personalization and recommendation systems , author=. arXiv preprint arXiv:1906.00091 , year=

    Pith/arXiv arXiv 1906

  36. [36]

    Acm transactions on interactive intelligent systems (tiis) , volume=

    The movielens datasets: History and context , author=. Acm transactions on interactive intelligent systems (tiis) , volume=. 2015 , publisher=

    2015

  37. [37]

    Proceedings of the 26th international conference on world wide web , pages=

    Neural collaborative filtering , author=. Proceedings of the 26th international conference on world wide web , pages=

  38. [38]

    Computer Speech & Language , volume=

    Product of Gaussians for speech recognition , author=. Computer Speech & Language , volume=. 2006 , publisher=

    2006

  39. [39]

    Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

    Lightgcn: Simplifying and powering graph convolution network for recommendation , author=. Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

  40. [40]

    Relational stacked denoising autoencoder for tag recommendation , booktitle =

    Hao Wang and Xingjian Shi and Dit. Relational stacked denoising autoencoder for tag recommendation , booktitle =

  41. [41]

    TKDE , volume=

    Relational collaborative topic regression for recommender systems , author=. TKDE , volume=

  42. [42]

    Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

    Collaborative topic modeling for recommending scientific articles , author=. Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

  43. [43]

    Proceedings of the 30th ACM International Conference on Information & Knowledge Management , pages=

    Cross-market product recommendation , author=. Proceedings of the 30th ACM International Conference on Information & Knowledge Management , pages=

  44. [44]

    IJCAI , pages =

    Collaborative topic regression with social regularization for tag recommendation , author=. IJCAI , pages =

  45. [45]

    Bell and Chris Volinsky , title =

    Yehuda Koren and Robert M. Bell and Chris Volinsky , title =. IEEE Computer , volume =. 2009 , pages =

    2009

  46. [46]

    Session-based Recommendations with Recurrent Neural Networks , booktitle =

    Bal. Session-based Recommendations with Recurrent Neural Networks , booktitle =

  47. [47]

    KDD , year =

    Yifei Ma and Murali Balakrishnan Narayanaswamy and Haibin Lin and Hao Ding , title =. KDD , year =

  48. [48]

    Zico Kolter and Vladlen Koltun , title =

    Shaojie Bai and J. Zico Kolter and Vladlen Koltun , title =. CoRR , volume =

  49. [49]

    Proceedings of the VLDB Endowment , volume=

    Probabilistic demand forecasting at scale , author=. Proceedings of the VLDB Endowment , volume=. 2017 , publisher=

    2017

  50. [50]

    NIPS , pages=

    Gaussian process priors with uncertain inputs application to multiple-step ahead time series forecasting , author=. NIPS , pages=

  51. [51]

    Maddix and Jan Gasthaus and Dean Foster and Tim Januschowski , title =

    Yuyang Wang and Alex Smola and Danielle C. Maddix and Jan Gasthaus and Dean Foster and Tim Januschowski , title =. ICML , pages =

  52. [52]

    CoRR , volume =

    Valentin Flunkert and David Salinas and Jan Gasthaus , title =. CoRR , volume =

  53. [53]

    IEEE transactions on power systems , volume=

    ARIMA models to predict next-day electricity prices , author=. IEEE transactions on power systems , volume=. 2003 , publisher=

    2003

  54. [54]

    international conference on machine learning , pages=

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning , author=. international conference on machine learning , pages=

  55. [55]

    Advances in neural information processing systems , pages=

    Spatial transformer networks , author=. Advances in neural information processing systems , pages=

  56. [56]

    Natural-Parameter Networks:

    Hao Wang and Xingjian Shi and Dit. Natural-Parameter Networks:. NIPS , pages =

  57. [57]

    arXiv preprint arXiv:1312.6114 , year=

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    Pith/arXiv arXiv

  58. [58]

    2013 , publisher=

    The skew-normal and related families , author=. 2013 , publisher=

    2013

  59. [59]

    Kingma and Jimmy Ba , title =

    Diederik P. Kingma and Jimmy Ba , title =. ICLR , year =

  60. [60]

    Xudong Mao and Qing Li and Haoran Xie and Raymond Y. K. Lau and Zhen Wang and Stephen Paul Smolley , title =. ICCV , pages =

  61. [61]

    Gordon , title =

    Han Zhao and Remi Tachet des Combes and Kun Zhang and Geoffrey J. Gordon , title =. ICML , pages =

  62. [62]

    VLDB , volume=

    Detecting change in data streams , author=. VLDB , volume=

  63. [63]

    AAAI , pages =

    Seiichi Kuroki and Nontawat Charoenphakdee and Han Bao and Junya Honda and Issei Sato and Masashi Sugiyama , title =. AAAI , pages =

  64. [64]

    Adversarial Multiple Source Domain Adaptation , booktitle =

    Han Zhao and Shanghang Zhang and Guanhang Wu and Jos. Adversarial Multiple Source Domain Adaptation , booktitle =

  65. [65]

    Castillo and Rama Chellappa , title =

    Swami Sankaranarayanan and Yogesh Balaji and Carlos D. Castillo and Rama Chellappa , title =. CVPR , pages =

  66. [66]

    CVPR , pages =

    Kuniaki Saito and Kohei Watanabe and Yoshitaka Ushiku and Tatsuya Harada , title =. CVPR , pages =

  67. [67]

    Jordan , title =

    Mingsheng Long and Zhangjie Cao and Jianmin Wang and Michael I. Jordan , title =. NIPS , pages =

  68. [68]

    TKDE , volume=

    A survey on transfer learning , author=. TKDE , volume=

  69. [69]

    JAMA , volume=

    The National Sleep Research Resource: towards a sleep data commons , author=. JAMA , volume=

  70. [70]

    JAMA , volume=

    Appendicular bone density and age predict hip fracture in women , author=. JAMA , volume=

  71. [71]

    Sleep , volume=

    The sleep heart health study: design, rationale, and methods , author=. Sleep , volume=

  72. [72]

    NIPS , pages=

    PyTorch: An imperative style, high-performance deep learning library , author=. NIPS , pages=

  73. [73]

    1994 , publisher=

    Mixture density networks , author=. 1994 , publisher=

    1994

  74. [74]

    NIPS , pages=

    Generative adversarial nets , author=. NIPS , pages=

  75. [75]

    Jaakkola and Matt T

    Mingmin Zhao and Shichao Yue and Dina Katabi and Tommi S. Jaakkola and Matt T. Bianchi , title =. ICML , pages =

  76. [76]

    ICCV workshop on Transferring and Adapting Source Knowledge in Computer Vision (TASK-CV) , pages =

    Baochen Sun and Kate Saenko , title =. ICCV workshop on Transferring and Adapting Source Knowledge in Computer Vision (TASK-CV) , pages =

  77. [77]

    Machine learning , volume=

    A theory of learning from different domains , author=. Machine learning , volume=

  78. [78]

    TNN , volume=

    Domain adaptation via transfer component analysis , author=. TNN , volume=

  79. [79]

    arXiv preprint arXiv:1412.3474 , year=

    Deep domain confusion: Maximizing for domain invariance , author=. arXiv preprint arXiv:1412.3474 , year=

    Pith/arXiv arXiv

  80. [80]

    CVPR , pages=

    Maximum classifier discrepancy for unsupervised domain adaptation , author=. CVPR , pages=

Showing first 80 references.