Pith. sign in

REVIEW 5 major objections 6 minor 54 references

MLego: Interactive and Scalable Topic Exploration Through Model Reuse

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Merging saved models answers LDA topic queries at interactive speed

desk verdict A well-motivated system for reusing materialized LDA models, but the current draft is missing its experiments and the plan-search optimality proof doesn't hold up. read the letter →

arxiv 2508.07654 v1 pith:RI6KHCOA submitted 2025-08-11 cs.DB cs.IR

Fei Ye , Jiapan Liu , Yinan Jing , Zhenying He , Weirao Wang , X. Sean Wang This is my paper · ORCID
classification cs.DBcs.IR
keywords LatentDirichletAllocationtopicmodelingmodelreusematerializationqueryoptimizationinteractiveanalyticsmergingtop-kalgorithm
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

MLego addresses the problem that LDA topic models are too slow to retrain for interactive, region-based queries over large text corpora. Its central claim is that a query can be answered by merging already-trained topic models, and that the best merge plan can be found without enumerating the exponentially many candidate plans. The paper proposes a score sc = αlp + (1−α)ct balancing performance loss and time cost, and a hierarchical search over "relatively longest" plans that prunes candidates via a threshold/top-k procedure. If the cost model is monotonic in the number of merged models, the authors argue the search returns the optimal plan. A batch-query optimizer reorders queries to further cut total cost.

What carries the argument

Plan score sc(p) = α lp + (1−α) ct(p), where lp = 1 − P(x) with P(x) a monotone performance-loss function of the number of merged models x, and ct(p) = training cost + merging cost. The search exploits "relatively longest" (RL) plans as tree roots, orders candidate plans by lp, ct_merge, and ct_train, and uses the threshold/top-k algorithm with a "push down" rule to prune layers; Theorems 2–4 give conditions under which merge cost can be ignored without changing plan order.

What would settle it

Run Algorithm 3 on a corpus with materialized models m1..mk, compare its returned plan against the score of every candidate plan for the same query. If any unexamined plan has a strictly lower score, the monotonicity or list-ordering assumption fails; a direct measurement of performance loss versus number of merged models (P(x)) on real data would reveal whether the assumption holds.

Watch

Extended reading notes

Core claim

The paper's discovery is that LDA analytic queries are amenable to a materialize-and-reuse paradigm: instead of training from scratch on the query region, MLego combines pre-trained models using Bayesian updating, either through streaming variational Bayes or collapsed Gibbs sampling, to form an approximate model whose perplexity is close to the full model. The plan used to combine models is chosen by minimizing a linearized score of performance loss and time cost. The paper proves that all candidate plans can be generated from the set of "relatively longest" plans, and that a threshold-based hierarchical search over three ordered lists can stop early while still returning the optimal plan,

Load-bearing premise

The optimality of the plan search rests on the assumption that plan cost is monotonic — merging more models never lowers the cost, all else being equal — and on the training-complexity ordering that keeps the cost lists sorted.

Editorial extensions

If this is right

  • Users can pose ad-hoc region queries over large text corpora and get approximate LDA results in interactive time rather than waiting for full retraining.
  • The plan-search strategy applies to any cost model that is monotonic in the number of merged models, so the framework is not tied to the specific lp/ct linearization.
  • Model merging via VB or CGS is order-independent, so materialized models can be maintained incrementally and combined in any order without changing the outcome.
  • Batch query reordering reduces total cost under a monotone cost model, making multi-region exploration faster.
  • The approach complements existing visual analytics systems by supplying a query-driven engine beneath user-facing topic visualizations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The materialize-and-merge recipe is generic for Bayesian models whose posterior stays in the same exponential family as the prior, so the LDA instantiation likely transfers to other conjugate models, though the paper does not claim this.
  • The claimed training complexity O(Mi N^2 K) is not obviously consistent with standard Gibbs sampling complexity O(Mi N K) per iteration; if the true complexity is linear in N, the ordering of ct_train layers and the threshold x* would change, narrowing the proven optimality range.
  • The monotonicity assumption is verified only by a figure described in the text; a deployment would need to measure P(x) on the target corpus and model type directly, or estimate it online, rather than take it as given.
  • The user-supplied weight α encodes the accuracy-versus-latency tradeoff; an adaptive α that tunes to a response-time budget could make the interactive tradeoff automatic.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes MLego, a framework for interactive topic-modeling analytics that answers LDA-based queries by reusing and merging materialized topic models instead of training from scratch. It defines a score function sc = α·lp + (1−α)·ct balancing performance loss and time cost, gives merging primitives for variational Bayes and collapsed Gibbs sampling, and presents a hierarchical plan search (Algorithm 3) intended to find the optimal plan p* = argmin_p sc(p) without enumerating the exponential candidate set, under a monotonicity assumption on the cost model. It also sketches a batch-query reordering heuristic. The abstract claims extensive experiments and a visual analytics prototype, but the submitted manuscript contains no experimental section and the text breaks off during Section V.C, before the batch-query details, with the remainder being references.

Significance. If the proposed framework worked as claimed, it would address a real gap: interactive exploration of large text corpora with topic models, where retraining from scratch is too slow. The core idea of materializing and merging LDA models, borrowed from streaming Bayesian updating, is plausible and the two merging algorithms are concretely specified. The paper also makes a fair conceptual contribution by framing topic-model reuse as a query-planning problem and importing top-k threshold techniques. However, the significance is highly conditional: the central plan-search optimality guarantee rests on an unverified monotonicity assumption and an incompletely specified algorithm, and the absence of any experimental validation leaves the headline claims of interactivity and quality unsupported. The manuscript therefore is not currently publishable, though the direction may be viable after major revision.

major comments (5)
  1. [Abstract / Section VI (missing)] The abstract and Introduction promise 'Extensive experiments' demonstrating reduced computation cost and maintained quality, and a visual analytics prototype. The submitted manuscript contains no experimental section, no dataset descriptions, and no runtime/quality results; Figure 3, explicitly referenced for monotonicity verification, is also absent. The paper also breaks off in Section V.C before presenting Algorithm 4 or the promised NP-hardness proof. These are not presentation issues: without experiments, the core claims of interactivity and scalability are unverified.
  2. [V.B.2, Time Cost] The training complexity is stated as O(Mi N^2 K) with 'The proof directly follows from [37].' This is not supported: standard LDA collapsed Gibbs sampling has O(Mi N K) cost per token (or per iteration), and the original LDA paper [37] does not contain an N^2 term. Since this complexity is used to motivate monotonicity and to compute the threshold x* in Theorems 3–4, the incorrect exponent propagates into the pruning logic. Please derive the complexity from the actual sampling update or correct it to a defensible form.
  3. [V.B.4, Algorithm 3 and Theorem 2] The optimality guarantee of Algorithm 3 is not established. The threshold algorithm requires each list to be sorted by the aggregate score and a valid stopping bound, but the pseudocode (lines 7–16) leaves the threshold update and termination unspecified, and the list-generation order is ambiguous because ΔL is overwritten in different branches. Theorem 2 is not a proof: it assumes 'if pi2 should be pushed down' and then derives a condition, but gives no argument that all violated orders are captured; the proof ends by acknowledging order is 'not guaranteed.' If a layer containing the true optimum is pruned, the claimed p* = argmin sc(p) fails even under monotone costs. A complete correctness proof, or a counterexample, is required.
  4. [V.B.5, Theorems 3 and 4] The proofs contain algebraic slips. Equation (13) has ct(train)_pij on both sides; the correct inequality is ct(merge)_pi − ct(merge)_pij < ct(train)_pi − ct(train)_pij. Equation (14) incorrectly treats the merge cost difference as proportional to the difference in the number of models without defining tm consistently. Theorem 4's proof uses min(ct(train)_pi − ct(train)_pj) ≤ ct(train)_mmin, which is not generally true and is not justified. Since these theorems justify ignoring merge cost in the search, the pruning guarantee is unsupported.
  5. [V.B.2, Monotonicity assumption] The entire optimality argument rests on the asserted monotonic cost model: P(x) with P(x_i) > P(x_j) iff x_i > x_j. The text claims Figure 3 verifies this, but the figure is missing and the verbal description is confusing: it says performance loss increases with the number of merges, yet the interval [301,400] is said to satisfy monotonicity when loss 'remains unchanged or monotonically decreases.' A concrete experimental measurement of P(x) on real datasets, or at least a clearly described figure with error bars, must be provided. Without it, the search's optimality is conditional on an unvalidated postulate.
minor comments (6)
  1. [IV.C] There are two 'Definition 1' entries: 'Analytic Query on ML' in Section IV.A and 'Score Function' in Section IV.C. Renumber to avoid confusion.
  2. [III.B / V.B.2] N is used inconsistently: 'amount of data' in Section III.B versus 'number of words' in Section V.B.2. The relation between D, V, K in Section III.A and N is unclear; define all terms.
  3. [V.B.3, Theorem 1] The proof of Theorem 1 is only a sketch: 'The proof is obvious' and 'conflicts with RL plans' are not defined. Please provide a formal statement of RL plans and a rigorous argument that every plan is a subset of some RL plan.
  4. [V.B.2] Reference [20] is cited for the monotonicity experiments, but [20] is an earlier claim about users accepting approximate models, not a model-merging study. Please cite the actual source for the empirical relation between merge count and performance loss.
  5. [V.B.4] Figure 4 is referenced but does not appear in the submitted text. If figures are omitted in this version, please ensure all referenced figures are included.
  6. [V.C] The section promises a proof of NP-hardness and a heuristic Algorithm 4, but the text stops after 'heuristic algorithm 4.' Complete this section in a revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the plan-search optimizes a user-defined score and imports model-merging from external prior work; no prediction is equal-by-construction to its inputs.

full rationale

The central claim is that Algorithm 3 returns p* = argmin_p sc(p) for the user-defined score sc = αlp + (1−α)ct (Eq. 2). This is an optimization of an explicitly defined cost model, not a prediction derived from fitted inputs. The performance-loss simplification lp = 1 − P(x) (Section V.B.2) is stated as a monotonicity assumption, not as a fitted parameter that is later renamed as a prediction. The LDA merging operators (Algorithms 1–2) are imported from external work [31], and the top-k threshold machinery from [23], [66]; there is no evidence that co-authored reference [44] carries a load-bearing argument. The paper's main rigor risks are non-circular: Figure 3 verifying monotonicity is absent from the reviewed text, the proof of Theorem 2 is logically question-begging, and Eq. (13) contains an algebraic slip. These are correctness concerns, not definitional or constructional circularity. The derivation chain does not reduce to its own inputs by construction, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central framework rests on standard Bayesian updating and on the prior merging work of [30] and [31]. The main new burdens are the monotonicity assumption, the questionable LDA complexity expression, and the weak justification of RL-plan completeness. No new physical or conceptual entities are introduced.

free parameters (2)
  • decay factor λ (DSGS) = not specified in reviewed text
    Used in Algorithm 2 for Gibbs-based model merging; controls the weighting of past updates. Its value affects merged model quality and is not stated in the text we have.
  • user weight α = user-specified in [0,1]
    Trades off performance loss against time cost in the score function; not fitted by the authors, but the plan search output depends on it.
assumptions (5)
  • standard math Posterior updating via Eq. (4) allows sequential model merging without revisiting old data
    Foundation for both the VB and Gibbs merging methods; a standard property of Bayesian posteriors.
  • domain assumption Plan cost is monotonic in the number of merged models
    Section V.B.2 states 'The only constraint that must be satisfied by the cost model is that it is monotonic.' The optimality of the search depends on this; verification is claimed via Figure 3, which is not included in the reviewed text.
  • ad hoc to paper LDA training complexity is O(Mi N^2 K)
    Section V.B.2 says 'The proof directly follows from [37]', but standard LDA inference is O(Mi N K) per iteration. The N^2 term is not derived and affects the ct(train) ordering used in the plan search.
  • ad hoc to paper All candidate plans can be generated from RL plans (Theorem 1)
    Section V.B.3. The proof is a single paragraph relying on the definition of RL plans rather than a formal argument, so it is effectively an unproved structural assumption.
  • domain assumption Model merging via SDA-Bayes and DSGS preserves topic quality sufficiently
    Imported from cited prior work [30], [31]; the paper does not re-derive or evaluate this assumption in the reviewed text, yet it underlies the entire approach.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MLego: Interactive and Scalable Topic Exploration Through Model Reuse." pith.science (2026). https://pith.science/paper/RI6KHCOA

@misc{pith2026250807654,
  author       = {Pith},
  title        = {Pith review of: MLego: Interactive and Scalable Topic Exploration Through Model Reuse},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RI6KHCOA}},
  note         = {Machine review of arXiv:2508.07654}
}
read the original abstract

With massive texts on social media, users and analysts often rely on topic modeling techniques to quickly extract key themes and gain insights. Traditional topic modeling techniques, such as Latent Dirichlet Allocation (LDA), provide valuable insights but are computationally expensive, making them impractical for real-time data analysis. Although recent advances in distributed training and fast sampling methods have improved efficiency, real-time topic exploration remains a significant challenge. In this paper, we present MLego, an interactive query framework designed to support real-time topic modeling analysis by leveraging model materialization and reuse. Instead of retraining models from scratch, MLego efficiently merges materialized topic models to construct approximate results at interactive speeds. To further enhance efficiency, we introduce a hierarchical plan search strategy for single queries and an optimized query reordering technique for batch queries. We integrate MLego into a visual analytics prototype system, enabling users to explore large-scale textual datasets through interactive queries. Extensive experiments demonstrate that MLego significantly reduces computation costs while maintaining high-quality topic modeling results. MLego enhances existing visual analytics approaches, which primarily focus on user-driven topic modeling, by enabling real-time, query-driven exploration. This complements traditional methods and bridges the gap between scalable topic modeling and interactive data analysis.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 53 canonical work pages

  1. [37]

    Latent dirichlet allocation,

    D. M. Blei, A. Y . Ng, and M. I. Jordan, “Latent dirichlet allocation,” J. Mach. Learn. Res. , vol. 3, p. 993–1022, 2003

    work page 2003

  2. [1]

    KY k=1 qD (βk | λk) # ·

    Model Merging for LDA: Given a query q, our objec- tive is to efficiently output the topic parameter eβ such that perplexity for eβ is close to the perplexity of β where β is the topic parameter obtained by running LDA algorithm from scratch on the entire C. We seek to do this by only using < o, N,Θ > of model mi. Different posterior approximation algorit...

    work page

  3. [2]

    Merge conducts model merging, while Train handles online training for uncovered data

    Cost Analysis: Based on the plan cost in Section 4.B, we further analyze each cost for two actions: (1) Merge and (2) Train. Merge conducts model merging, while Train handles online training for uncovered data. Performance Loss. Since we train each materialized model completely, the performance loss is mainly caused by insuffi- cient model retraining afte...

    work page

  4. [3]

    The main challenge of such a ”generate- and-rank” method is that there are an exponential number of candidate plans, and the plan generation is time-consuming

    A Basic Approach and its Challenge: One approach is to generate all candidate plans based on M and D, followed by scoring and ranking. The main challenge of such a ”generate- and-rank” method is that there are an exponential number of candidate plans, and the plan generation is time-consuming. To solve this problem, we present Algorithm 3, which accelerat...

    work page

  5. [4]

    push down

    Hierarchical Plan Searching: We present a general plan searching algorithm based on the top-k algorithm (line 5 to 12). Its main idea is to generate candidate plans hierarchically based on RL plans and score function, then utilize the top-k algorithm to avoid plan enumeration. Top-k Algorithm. As described in Section 4.C, the score function consists of we...

    work page

  6. [5]

    Each point in Figure 3 represents the average value of all points within that interval

    The experiment shows that the performance loss of the merged model increases as the number of merging increases compared with the model trained directly from scratch (i.e., #models = 1). Each point in Figure 3 represents the average value of all points within that interval. Taking the interval [301, 400] as an example, the above assumption is satisfied wh...

    work page

  7. [6]

    For a fixed plan p, the training time cost is quadratic to the data uncovered by the models in p

    Candidate Plans Generation: We generate RL plans for the full plan list in order of training time cost for each query (line 1). For a fixed plan p, the training time cost is quadratic to the data uncovered by the models in p. As the data distribution in the query range is unknown, it is difficult to quickly estimate the training time cost. But when there ...

    work page

  8. [7]

    Theorem 1

    All model non-overlapping relations in each query will be preserved in RL plans. Theorem 1. All possible candidate plans for each query q can be generated by the RL plans. Proof. The proof of the Theorem 1 is obvious. We assume pi can not be generated by the RL plans, i.e., for ∀p in RL plans, there is no way for pi to be obtained by removing several mode...

    work page

Show all 54 references
  1. [9]

    To further speed up plan searching, we consider reducing the number of lists (line 8 to 9)

    Improvement Computing Correlation of Costs: The hier- archical plan searching discussed above involves multiple lists, which will weaken the searching efficiency if the generation directions of these lists are inconsistent. To further speed up plan searching, we consider reduc...

    2013

  2. [18]

    Streaming gibbs sampling for LDA model,

    Y . Gao, J. Chen, and J. Zhu, “Streaming gibbs sampling for LDA model,” CoRR, vol. abs/1601.01142, 2016

    2016 arXiv

  3. [23]

    A survey of top-k query processing techniques in relational database systems,

    I. F. Ilyas, G. Beskales, and M. A. Soliman, “A survey of top-k query processing techniques in relational database systems,” ACM Comput. Surv., vol. 40, no. 4, 2008

    2008

  4. [24]

    Culda: Solving large-scale lda problems on gpus,

    X. Xie, Y . Liang, X. Li, and W. Tan, “Culda: Solving large-scale lda problems on gpus,” in Proceedings of the 28th International Symposium on High-Performance Parallel and Distributed Computing , 2019, p. 195–205

    2019

  5. [25]

    Lda*: A robust and large-scale topic modeling system,

    L. Yut, C. Zhang, Y . Shao, and B. Cui, “Lda*: A robust and large-scale topic modeling system,” Proc. VLDB Endowment , vol. 10, no. 11, p. 1406–1417, 2017

    2017

  6. [26]

    Warplda: A cache efficient o(1) algorithm for latent dirichlet allocation,

    J. Chen, K. Li, J. Zhu, and W. Chen, “Warplda: A cache efficient o(1) algorithm for latent dirichlet allocation,” Proc. VLDB Endowment, vol. 9, no. 10, p. 744–755, 2016

    2016

  7. [27]

    Reducing the sampling complexity of topic models,

    A. Q. Li, A. Ahmed, S. Ravi, and A. J. Smola, “Reducing the sampling complexity of topic models,” in Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Mining, 2014, p. 891–900

    2014

  8. [28]

    A scalable asynchronous distributed algorithm for topic modeling,

    H.-F. Yu, C.-J. Hsieh, H. Yun, S. Vishwanathan, and I. S. Dhillon, “A scalable asynchronous distributed algorithm for topic modeling,” in Proceedings of the 24th International Conference on World Wide Web , 2015, p. 1340–1350

    2015

  9. [29]

    Lightlda: Big topic models on modest computer clusters,

    J. Yuan, F. Gao, Q. Ho, W. Dai, J. Wei, X. Zheng, E. P. Xing, T.-Y . Liu, and W.-Y . Ma, “Lightlda: Big topic models on modest computer clusters,” in Proceedings of the 24th International Conference on World Wide Web, 2015, p. 1351–1361

    2015

  10. [30]

    Streaming variational bayes,

    T. Broderick, N. Boyd, A. Wibisono, A. C. Wilson, and M. I. Jordan, “Streaming variational bayes,” in Advances in Neural Information Pro- cessing Systems, C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Weinberger, Eds., vol. 26, 2013

    2013

  11. [31]

    BY b=1 A (Cb, p(Θ)) p(Θ)−1 # p(Θ) ∝ exp (

    to achieve LDA merging. In a nutshell, the SDA-Bayes framework makes Streaming, Distributed, and Asynchronous Bayes updates to the estimated posterior according to a user- specified approximation batch primitive: p(Θ | C1, . . . , CB) ≈ q(Θ) ∝ " BY b=1 A (Cb, p(Θ)) p(Θ)−1 # p(...

  12. [32]

    Topic modeling using latent dirichlet allo- cation: A survey,

    U. Chauhan and A. Shah, “Topic modeling using latent dirichlet allo- cation: A survey,” ACM Comput. Surv., vol. 54, no. 7, 2021

    2021

  13. [33]

    A survey of topic modeling in text mining,

    R. Alghamdi and K. Alfalqi, “A survey of topic modeling in text mining,” International Journal of Advanced Computer Science and Applications, vol. 6, no. 1, 2015

    2015

  14. [34]

    Streaming, dis- tributed variational inference for bayesian nonparametrics,

    T. Campbell, J. Straub, J. W. Fisher III, and J. P. How, “Streaming, dis- tributed variational inference for bayesian nonparametrics,” in Advances in Neural Information Processing Systems , C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, Eds., vol. 28, 2015

    2015

  15. [35]

    Sharing clusters among related groups: Hierarchical dirichlet processes,

    Y . Teh, M. Jordan, M. Beal, and D. Blei, “Sharing clusters among related groups: Hierarchical dirichlet processes,” in Advances in Neural Information Processing Systems, L. Saul, Y . Weiss, and L. Bottou, Eds., vol. 17, 2004

    2004

  16. [36]

    Online learning for latent dirichlet allocation,

    M. D. Hoffman, D. M. Blei, and F. Bach, “Online learning for latent dirichlet allocation,” in Advances in Neural Information Processing Systems, vol. 23, 2010

    2010

  17. [38]

    Online but accurate inference for latent variable models with local gibbs sampling,

    C. Dupuy and F. Bach, “Online but accurate inference for latent variable models with local gibbs sampling,” J. Mach. Learn. Res., vol. 18, no. 1, p. 4581–4625, 2017

    2017

  18. [39]

    News category dataset,

    R. Misra, “News category dataset,” arXiv preprint arXiv:2209.11429 , 2022

    2022 arXiv

  19. [40]

    Misra and J

    R. Misra and J. Grover, Sculpting Data for ML: The first act of Machine Learning. Independently published, 2021

    2021

  20. [41]

    Defending against neural fake news,

    R. Zellers, A. Holtzman, H. Rashkin, Y . Bisk, A. Farhadi, F. Roesner, and Y . Choi, “Defending against neural fake news,” in Advances in Neural Information Processing Systems , 2019

    2019

  21. [42]

    Efficient computation of top-k frequent terms over spatio-temporal ranges,

    P. Ahmed, M. Hasan, A. Kashyap, V . Hristidis, and V . J. Tsotras, “Efficient computation of top-k frequent terms over spatio-temporal ranges,” in Proc. ACM SIGMOD Int. Conf. Manage. Data , 2017, p. 1227–1241

    2017

  22. [43]

    Auto- mated phrase mining from massive text corpora,

    J. Shang, J. Liu, M. Jiang, X. Ren, C. R. V oss, and J. Han, “Auto- mated phrase mining from massive text corpora,” IEEE Transactions on Knowledge and Data Engineering , vol. 30, no. 10, pp. 1825–1837, 2018

    2018

  23. [44]

    Efficiently answering top-k frequent term queries in temporal-categorical range,

    Z. He, L. Wang, C. Lu, Y . Jing, K. Zhang, W. Han, J. Li, C. Liu, and X. S. Wang, “Efficiently answering top-k frequent term queries in temporal-categorical range,” Information Sciences , vol. 574, pp. 238– 258, 2021

    2021

  24. [45]

    Scikit-learn: Machine learning in python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Van- derplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and ´Edouard Duchesnay, “Scikit-learn: Machine learning in python,” Journal of Mach...

    2011

  25. [46]

    Efficient methods for topic model inference on streaming document collections,

    L. Yao, D. Mimno, and A. McCallum, “Efficient methods for topic model inference on streaming document collections,” in Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Mining , 2009, p. 937–946

    2009

  26. [47]

    Fast collapsed gibbs sampling for latent dirichlet allo- cation,

    I. Porteous, D. Newman, A. Ihler, A. Asuncion, P. Smyth, and M. Welling, “Fast collapsed gibbs sampling for latent dirichlet allo- cation,” in Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Mining , 2008, p. 569–577

    2008

  27. [48]

    Distributed infer- ence for latent dirichlet allocation,

    D. Newman, P. Smyth, M. Welling, and A. Asuncion, “Distributed infer- ence for latent dirichlet allocation,” in Advances in Neural Information Processing Systems, J. Platt, D. Koller, Y . Singer, and S. Roweis, Eds., vol. 20, 2007

    2007

  28. [49]

    An architecture for parallel topic models,

    A. Smola and S. Narayanamurthy, “An architecture for parallel topic models,” Proc. VLDB Endowment , vol. 3, no. 1–2, p. 703–710, 2010

    2010

  29. [50]

    Plda: Paral- lel latent dirichlet allocation for large-scale applications,

    Y . Wang, H. Bai, M. Stanton, W.-Y . Chen, and E. Y . Chang, “Plda: Paral- lel latent dirichlet allocation for large-scale applications,” in Algorithmic Applications in Management , 2009

    2009

  30. [51]

    Elda: Lda made efficient via algorithm-system codesign submission,

    S. Wang, D. Li, H. Yu, and H. Liu, “Elda: Lda made efficient via algorithm-system codesign submission,” in Proceedings of the 25th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, 2020, p. 407–408

    2020

  31. [52]

    Approximate query processing: Taming the terabytes,

    M. N. Garofalakis and P. B. Gibbon, “Approximate query processing: Taming the terabytes,” in Proc. VLDB Endowment , 2001, p. 725

    2001

  32. [53]

    Data cube: a relational aggregation operator generalizing group-by, cross-tab, and sub-totals,

    J. Gray, A. Bosworth, A. Lyaman, and H. Pirahesh, “Data cube: a relational aggregation operator generalizing group-by, cross-tab, and sub-totals,” in Proceedings of the Twelfth International Conference on Data Engineering, 1996, pp. 152–159

    1996

  33. [54]

    Efficient scalable accurate re- gression queries in in-dbms analytics,

    C. Anagnostopoulos and P. Triantafillou, “Efficient scalable accurate re- gression queries in in-dbms analytics,” in 2017 IEEE 33rd International Conference on Data Engineering (ICDE) , 2017, pp. 559–570

    2017

  34. [55]

    Prediction cubes,

    B.-C. Chen, L. Chen, Y . Lin, and R. Ramakrishnan, “Prediction cubes,” in Proc. VLDB Endowment , 2005, p. 982–993

    2005

  35. [56]

    Scalable k -means clustering via lightweight coresets,

    O. Bachem, M. Lucic, and A. Krause, “Scalable k -means clustering via lightweight coresets,” in Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Mining, 2018, p. 1119–1127

    2018

  36. [57]

    Turning big data into tiny data: Constant-size coresets for k-means, pca and projective clustering,

    D. Feldman, M. Schmidt, and C. Sohler, “Turning big data into tiny data: Constant-size coresets for k-means, pca and projective clustering,” in Proceedings of the Twenty-Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, 2013, p. 1434–1453

    2013

  37. [58]

    On coresets for k-means and k-median clustering,

    S. Har-Peled and S. Mazumdar, “On coresets for k-means and k-median clustering,” in Proceedings of the Thirty-Sixth Annual ACM Symposium on Theory of Computing , 2004, p. 291–300

    2004

  38. [59]

    Fast and accurate k-means for large datasets,

    M. Shindler, A. Wong, and A. Meyerson, “Fast and accurate k-means for large datasets,” in Advances in Neural Information Processing Systems , J. Shawe-Taylor, R. Zemel, P. Bartlett, F. Pereira, and K. Weinberger, Eds., vol. 24, 2011

    2011

  39. [60]

    Modelhub: Deep learning lifecycle management,

    H. Miao, A. Li, L. S. Davis, and A. Deshpande, “Modelhub: Deep learning lifecycle management,” in IEEE 33rd International Conference on Data Engineering (ICDE) , 2017, pp. 1393–1394

    2017

  40. [61]

    Towards unified data and lifecycle management for deep learning,

    H. Miao, A. Li, L. S. Davis, and A. Deshpande, “Towards unified data and lifecycle management for deep learning,” inIEEE 33rd International Conference on Data Engineering (ICDE) , 2017, pp. 571–582

    2017

  41. [62]

    Mistique: A system to store and query model intermediates for model diagnosis,

    M. Vartak, J. M. F. da Trindade, S. Madden, and M. Zaharia, “Mistique: A system to store and query model intermediates for model diagnosis,” in Proc. ACM SIGMOD Int. Conf. Manage. Data , 2018, p. 1285–1300

    2018

  42. [63]

    Modeldb: A system for machine learning model management,

    M. Vartak, H. Subramanyam, W.-E. Lee, S. Viswanathan, S. Husnoo, S. Madden, and M. Zaharia, “Modeldb: A system for machine learning model management,” in Proceedings of the Workshop on Human-In-the- Loop Data Analytics , 2016

    2016

  43. [64]

    Applying data mining techniques for descriptive phrase extraction in digital document collections,

    H. Ahonen, O. Heinonen, M. Klemettinen, and A. Verkamo, “Applying data mining techniques for descriptive phrase extraction in digital document collections,” in Proceedings IEEE International Forum on Research and Technology Advances in Digital Libraries, 1998, pp. 2–11

    1998

  44. [65]

    Location-aware top-k term publish/subscribe,

    L. Chen, S. Shang, Z. Zhang, X. Cao, C. S. Jensen, and P. Kalnis, “Location-aware top-k term publish/subscribe,” in IEEE 34th Interna- tional Conference on Data Engineering (ICDE) , 2018, pp. 749–760

    2018

  45. [66]

    Optimal aggregation algorithms for middleware,

    R. Fagin, A. Lotem, and M. Naor, “Optimal aggregation algorithms for middleware,” in Proceedings of the Twentieth ACM SIGMOD- SIGACT-SIGART Symposium on Principles of Database Systems , 2001, p. 102–113

    2001

  46. [67]

    Combining fuzzy information from multiple systems (ex- tended abstract),

    R. Fagin, “Combining fuzzy information from multiple systems (ex- tended abstract),” in Proceedings of the Fifteenth ACM SIGACT- SIGMOD-SIGART Symposium on Principles of Database Systems, 1996, p. 216–226

    1996

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.