todo:
requires bridging the gap between fundamental algorithms and modern machine learning applications.
✔️Recent research in top AI conferences has focused on overcoming the NP-hard nature of learning optimal decision trees, often using techniques that echo the dynamic programming and search-based approaches from the BST literature.
1.Rigor of Proofs: The proof sketch for Theorem 5.1 (Correctness) and the 'Refinement for Practical Cases' in Theorem 5.2 (Time Complexity) need to be expanded with more formal details and rigorous arguments. Clarify the apparent contradiction in the time complexity analysis regarding the cost per subproblem. If the cost per subproblem is indeed O(log n), then the overall complexity should be derived consistently.
2.Adaptive Parameter Selection: Provide more theoretical justification or empirical analysis for the adaptive parameter selection (Algorithm 4.3). How robust is it to different weight distributions? Are there cases where it might perform sub-optimally?
3.Worst-Case Analysis: While adversarial datasets are mentioned in experiments, a more in-depth theoretical analysis of the worst-case behavior of HWP would be beneficial. Under what specific conditions does the algorithm degrade, and what is the theoretical upper bound in such cases?
4.Detailed MLST Description: Provide a more detailed description of the Multi-Level Segment Tree (MLST) data structure, including its operations (query_profile, suggest_splits) and how they achieve the claimed logarithmic time complexities. Pseudocode for these specific MLST functions would be helpful.
5.Clarity on 'Practical Cases' vs. 'Theoretical': Clearly delineate when the O(n² log² n) complexity holds (i.e., under 'practical weight distributions' or 'γ-clustering') versus the more general O(n³ log² n) derived from the initial steps of Theorem 5.2. This distinction is crucial for reviewers.
6.Broader Impact/Societal Implications: While the paper touches on applications, a brief discussion on potential broader impacts, including any ethical considerations or societal implications of highly optimized search algorithms, could be added, especially given the increasing emphasis on responsible AI.
7.Minor Typos/Formatting: A thorough proofreading for minor typos and formatting inconsistencies would be beneficial. For example, in Section 10.1, the Python code for AKKL and CGMY algorithms has some minor indentation issues and missing imports if run directly.
This paper presents a highly significant and novel contribution to the field of algorithmic Track. Recommendation: Accept with minor revisions.
If possible, the authors might briefly mention any modern AI/machine-learning contexts (e.g. static decision trees for ensembles, though those usually optimize a different objective).
@chatgpt review
I recommend acceptance with minor revisions. The result is exciting and the work appears solid, but a few areas could be improved or clarified before final publication:
Proof Details: Ensure all proofs are complete and clearly explained. In particular, the new structural lemma (RMLK) and the analysis of HWP should be double-checked. If any steps are lengthy, consider moving them to an appendix with a roadmap in the main text.
Explain HWP Clearly: Since “Hierarchical Weight Partitioning” is a new idea here, the authors should give intuition (perhaps with a small example) on how weights are partitioned. Diagrams or pseudocode for this part would help.
Complexity Rigor: The $O(n^2\log^2 n)$ bound should be derived step-by-step in the paper. If any data-structure operations have non-trivial costs, they should be accounted for. Maybe add a remark on space complexity as well.
Experiments Details: Fully document the datasets and experimental setup: where the real-world data comes from, how weights were generated, hardware specs, compiler flags, etc. Provide the code or a link to it if possible, or at least pseudocode for reproducibility. If the baseline algorithm had to stop early due to time, mention that limitation.
Comparison with Approximations: The Chrobak et al. paper [18] also mentions an $O(n\log n)$ additive-3 approximation for this problem (their Theorem (ii)). Although the focus here is exact optimality, a brief mention of approximate alternatives and their practical cost/benefit could be informative (even if not the main topic).
Writing polish: Review for clarity, especially in sections describing the multi-level segment tree. It might be helpful to add transitions summarizing why each major idea is needed (e.g. “We next show how to use segment trees to speed up this DP step.”).