Pith. sign in

REVIEW 3 major objections 5 minor 15 references

A Hierarchical Integer Linear Programming Approach for Optimizing Team Formation in Education

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read An integer linear program that partitions students into teams by their expressed preferences can satisfy team-size and skill-coverage constraints while realizing more student preferences than teacher-assigned teams do.

desk verdict A solid formalization of educational team formation with a real, fixable bug in the O2 ILP: reflexive pairs cap the max-min objective at zero, so the O2-based strategies don't solve the stated problem. read the letter →

arxiv 2506.02756 v1 pith:C2J36IKO submitted 2025-06-03 math.OC cs.CY

Aaron Kessler , Tim Scheiber , Heinz Schmitz , Ioanna Lykourentzou This is my paper
classification math.OCcs.CY MSC 90C1090C2968Q17
keywords teamformationintegerlinearprogrammingmulti-objectiveoptimizationstudentpreferenceseducationalproblemNP-completenesstaxonomyclassroomassignment
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

The paper's aim is to show that team formation in education can be cast as a partitioning optimization problem and solved well enough for real classrooms. It defines EDUCATIONAL TEAM FORMATION (EDU-TF), where every student must be placed in exactly one team, team sizes are bounded, each team must cover a minimum number of skills, and students carry mutual preference values; it also places this problem inside a taxonomy of team formation families. It proves that even the feasibility version is NP-complete, and it proposes a modular hierarchical integer linear program whose objectives can be chosen and ordered by the teacher. On nine datasets from eight university courses, ten strategies derived from the model mostly realized more student preferences than the manual teacher-assigned teams, with runtimes usually from seconds to about an hour and many strategies reaching good solutions within 15 minutes. The broader claim is that preference-aware, constraint-respecting algorithmic team formation is practical enough to use in routine course planning.

What carries the argument

The carrying mechanism is a hierarchical integer linear program: a base feasibility ILP with assignment variables $x_{a,j}$, skill-count variables $y_{j,i}$, and coverage indicators $z_{j,i}$, extended by preference variables $t_{a,b,j}$ and $q_{a,b}$. Constraints (3)-(8) assign every student to exactly one team, keep team sizes within $[k_{min},k_{max}]$, and require at least $c$ covered skills per team. The objectives O1, O2, and O3 are added as modules and solved in sequence, fixing the value of each earlier objective before optimizing the next, so the teacher can prioritize avoidance of bad pairs over maximization of good ones. The NP-completeness proof is carried by all-rounder students who cover every skill: they make all but one team trivially feasible, so the remaining team must cover the full skill set exactly when the underlying SET COVER instance has a solution.

What would settle it

Remove that asymmetry: give the instructors the same full preference matrices the algorithm used, let them form teams, and compare the sum of realized preferences. If their teams reach or exceed the ILP strategies' scores, the outperformance claim would not replicate.

Watch

Extended reading notes

Core claim

The central result is a formal problem definition with a matching solver. EDU-TF asks for an $n$-partition of students satisfying team-size bounds and minimum skill coverage; Theorem 1 proves this decision problem is NP-complete by reduction from SET COVER, adding all-rounder students so that all but one team are trivially feasible. The paper then adds three preference objectives as plug-in modules: maximizing the sum of realized preferences, maximizing the smallest realized preference, and maximizing or minimizing the number of realized preferences of a given value. Objectives are solved lexicographically in teacher-chosen order, producing the ten strategies tested. In the experiments, most strategies on most datasets yielded a higher sum of realized preferences than the manual teacher solution, with the pure preference-sum strategy S2.1 always highest and strategies based on objective O2 being the slowest.

Load-bearing premise

The load-bearing premise is that the teacher-assigned teams used as the baseline are comparable to the algorithmic ones, but in most datasets the instructors had not seen the complete list of student-to-student preferences that the algorithm optimized, so they were solving an easier or different version of the problem.

Editorial extensions

If this is right

  • If the paper is right, instructors can hand team formation to an optimizer that honors student preferences while still enforcing team-size and skill-coverage constraints, at least for courses up to roughly 80 students.
  • The NP-completeness result justifies the use of ILP and heuristic solvers: no polynomial-time exact algorithm should be expected for the general feasibility problem, even with a trivial lower size bound and full skill coverage.
  • Because the objectives are modular and ordered hierarchically, a course can switch from a 'maximize satisfaction' strategy to a 'minimize misery' strategy by reordering objectives, without redesigning the model.
  • On large cohorts, using sparse preference matrices (explicit teammate requests rather than dense profile-derived similarities) made all but the O2-based strategies solve optimally in about a minute, so practical deployments can trade preference richness for speed.
  • Strategies that put objective O2 first were consistently slower and sometimes timed out before later objectives ran; the paper reads this as a warning to prefer O3 variants when minimizing dissatisfaction under time pressure.

Reading between the lines

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

  • The paper leaves implicit that the same modular ILP could apply beyond classrooms to any partition-formation setting where pairwise compatibility scores are available, such as hackathons, project staffing, or cohort grouping in organizations.
  • A natural extension the paper does not test is to cap the number of skills a single student can count toward a team's coverage; the all-rounder example in the NP-completeness proof shows how skill coverage can be satisfied by one member, so a per-student skill cap would change the feasible region materially.
  • The sensitivity of the results to the preference-calibration pipeline is unexamined: mapping profile distances into buckets and rescaling Likert ratings to [-2,2] or [-1,1] are design choices, and a systematic perturbation of those mappings would show how stable the strategy rankings are.
  • One could also test whether a re-run of the optimizer on updated preferences mid-course would improve team-level outcomes; the model's hierarchical structure makes such dynamic regrouping a straightforward extension, though the paper only lists it as future work.
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

3 major / 5 minor

Summary. The paper introduces EDUCATIONAL TEAM FORMATION (EDU-TF), a partitioning problem that combines teacher-imposed team-size and skill-coverage constraints with student preference objectives. The authors propose a taxonomy of team formation problem families, prove that the feasibility version of EDU-TF is NP-complete via a Set Cover reduction, and present a modular hierarchical ILP that can combine objectives O1 (maximize sum of realized preferences), O2 (maximize the smallest realized preference), and O3 (maximize/minimize the number of realized preferences at a given value). Ten strategies derived from this ILP are evaluated on nine real-world course datasets against manual teacher-assigned teams, using the sum of realized preferences and runtime as metrics.

Significance. If corrected, the paper would provide a useful bridge between educational practice and multi-objective optimization: the taxonomy helps organize a scattered literature, the NP-completeness proof for the feasibility version is a self-contained formal contribution, and the modular hierarchical ILP is a practical improvement over single-objective formulations. The Set Cover reduction in Theorem 1 is sound as presented. The reported experiments, however, rest on two load-bearing issues: the O2 ILP extension does not encode the formally defined objective, and the manual baselines for most datasets were created without access to the full preference matrix used by the algorithm. These issues undermine the strength of the main empirical and algorithmic claims as currently stated, though both appear fixable within the scope of the paper.

major comments (3)
  1. [Section 4.2, constraints (12)–(14)] The O2 extension is written for all a,b∈[m], including reflexive pairs. For a=b, constraint (13) cannot hold with q_{a,a}=0 because x_{a,j}=1 for exactly one team j, so every feasible solution sets q_{a,a}=1; constraint (14) then implies r≤p_{a,a}=0. Consequently the printed ILP cannot represent a positive optimal value for EDU-TF(O2), whereas the definition in Section 3.3 maximizes the minimum over the non-reflexive realized pairs M. This is a load-bearing defect: strategies S1.1, S3.1, and S3.2 rely on O2, and Table 3 reports O2 as 'solved optimally' for them (e.g., D8.2 S3.1 and S3.2), which the printed model cannot actually certify. The fix is to quantify (12)–(14) over a≠b and to handle the M=∅ singleton case separately; alternatively, the paper must present the implemented model if it differs from the printed one.
  2. [Section 5.3 and Section 6.1.1] The comparison against manual solutions is not made under equal information. Table 3 labels six of the nine manual baselines (D1, D2, D4, D5, D7, D8.1) with an asterisk, and Section 5.3 explains that those instructors did not have access to the full preference matrix used by the optimizer. Since the evaluation metric is the sum of all realized preferences, including the profile-derived entries hidden from the manual baseline, the broad conclusion that 'algorithmic solutions outperform manual ones' is partly an artifact of asymmetric information. The acknowledgment that teachers saw the raw profiles does not remove the bias. Please report the comparison on the three datasets where the manual solution used the same information (D3, D6, D8.2), or recompute all scores using only explicitly stated preferences, and adjust the wording of the main claim to match that subset.
  3. [Section 1, Section 5.3, and Section 6.1.3] The statement that the approach achieves 'optimal solutions and good-enough solutions within 15 minutes' overstates the experimental evidence. In Table 3 several one-hour runs time out without a proof of optimality (D5 S2.1, D7 S2.1, D8.1 S2.1 all at 60:01), and the time to the last solution improvement is often close to the limit (D7 S2.1 T*=52:47; D8.1 S4.1 T*=57:23). The 15-minute runs can be substantially worse than the one-hour result, for example D8.1 S3.1 drops from 173 to 43 and D7 S2.2 drops from 178 to 156; D8.1 S3.1 at 15 minutes is even below the manual baseline. The runtime conclusions should be limited to exact optimality on the smaller instances and to 'good-enough' solutions on the larger ones.
minor comments (5)
  1. [Section 3.2 and Section 4.2] The headings 'Feasiblity' and 'objecitves' contain typos; please correct them.
  2. [Table 3] The meaning of T and T* is explained in the caption, but the header row 'T T ∗' is easy to misread; consider renaming the columns to 'Total time' and 'Time to best solution'.
  3. [Section 3.1] The notation [x] = [0, x−1] for positive integers is nonstandard and may confuse readers; a brief example or a more standard indexing convention would help.
  4. [Section 7] The conclusion says 'four strategies (comprising ten sub-strategies)' while the rest of the paper refers to ten strategies; please align the terminology.
  5. [Reproducibility] The paper does not state whether the datasets, preference matrices, and solver scripts are available; sharing them would support replication of Table 3.

Circularity Check

1 steps flagged · score 2.0 of 10

Central NP-completeness proof and ILP design are self-contained; the only partial circularity is the headline comparison against manual teacher assignments, where the evaluation metric is the same preference objective the model optimizes and the baselines did not have the full preference matrix.

  1. fitted input called prediction [Abstract; Section 3.3 (O1); Section 5.3 (evaluation metrics and manual-baseline caveat)]
    "Results from evaluating ten strategies derived from our model on real-world university datasets indicate that our approach outperforms heuristic teacher-assigned teams by better accommodating student preferences. ... We evaluate each of the ten proposed strategies on every dataset using the following metrics: 1) the quality of the resulting solution as defined by the sum of all realized preferences ... not all manual solutions were created using the identical problem instances as the ones the strategies were applied to. ..."

    The reported 'outperforms' is measured by the sum of realized preferences, which is exactly objective O1 from Section 3.3 and is the quantity that the evaluated strategies maximize (or lexicographically optimize through the O3 stages). Since the manual teacher-assigned baselines were created without access to the full preference matrix, they were not solving the same optimization problem on the same input. The comparison therefore largely verifies that an optimizer scores higher on its own objective than non-optimizers who did not see the full objective input. This makes the qualitative conclusion 'better accommodating student preferences' partly forced by the evaluation design.

full rationale

The paper's central theoretical result, Theorem 1, is a direct and checkable polynomial-time reduction from SET COVER, an external NP-complete benchmark, and the construction does not invoke any self-referential or fitted quantity. The hierarchical ILP is presented as an explicit modular formulation, and the experimental section reports runtimes and solution quality against time limits without fitting parameters to the target outcomes. The authors' self-citations (e.g., Lykourentzou et al. 2016, 2017; Glaßer et al. 2010) are background or pointers and are not load-bearing for the uniqueness or correctness of any derived result. The one genuinely circular element is the manual-baseline comparison: the evaluation metric is the same O1 sum-of-realized-preferences objective that the strategies optimize, and the manual solutions were created without the full preference matrix, as the paper itself concedes. This partially reduces the headline claim of outperformance to a tautology. Separately, the O2 ILP constraints in Section 4.2 are written for all a,b in [m]; combined with p_{a,a}=0 and the assignment constraints, they force q_{a,a}=1 and r <= 0, so the printed model does not realize the max-min objective over non-reflexive pairs when the optimal minimum is nonnegative. That is a correctness defect in the printed formulation, but it is not a circular derivation, so it is not counted in the circularity score. Overall, the formal contributions are self-contained, and only the comparative evaluation claim suffers from partial circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central model has no fitted parameters; the hierarchical objectives and NP-completeness proof do not depend on tuning. The free parameters listed are experimental preprocessing choices that shape the datasets, not theoretical constants. Axioms are standard complexity results and domain assumptions about preference data and solver correctness. No new physical or conceptual entities are introduced.

free parameters (3)
  • Per-skill coverage thresholds
    Set by each instructor to decide which students count as covering each skill; changes feasibility and solution quality in the evaluation (Section 5.1).
  • Profile similarity bucket count
    A 'fixed number of evenly sized buckets' used to map profile distances to preference values; not numerically reported and changes the preference matrix (Section 5.1.2).
  • Preference scale mappings
    Likert ratings remapped to [-2,2], strong preferences to plus/minus 4, and profile similarities to [-1,1] or [-2,2]; these choices shape the input preferences (Section 5.1).
assumptions (4)
  • standard math SET COVER is NP-complete
    Used as the source problem in the reduction in Theorem 1.
  • domain assumption Pairwise preference sums approximate team interaction quality
    The objectives O1 and O2 rest on this assumption; the paper itself flags it as a limitation in Section 6.2.2.
  • domain assumption Instructor-defined skill thresholds and profile preprocessing yield valid input data
    The experimental results depend on the thresholds and similarity bucketing being appropriate for each course (Section 5.1).
  • domain assumption Gurobi correctly solves the hierarchical ILP and indicator constraints
    The implementation relies on Gurobi 10.0.1 for exact and time-limited solutions (Section 4.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Hierarchical Integer Linear Programming Approach for Optimizing Team Formation in Education." pith.science (2026). https://pith.science/paper/C2J36IKO

@misc{pith2026250602756,
  author       = {Pith},
  title        = {Pith review of: A Hierarchical Integer Linear Programming Approach for Optimizing Team Formation in Education},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C2J36IKO}},
  note         = {Machine review of arXiv:2506.02756}
}
read the original abstract

Teamwork is integral to higher education, fostering students' interpersonal skills, improving learning outcomes, and preparing them for professional collaboration later in their careers. While team formation has traditionally been managed by humans, either instructors or students, algorithmic approaches have recently emerged to optimize this process. However, existing algorithmic team formation methods often focus on expert teams, overlook agency in choosing one's teammates, and are limited to a single team formation setting. These limitations make them less suitable for education, where no student can be left out, student agency is crucial for motivation, and team formation needs vary across courses and programs. In this paper, we introduce the EDUCATIONAL TEAM FORMATION problem (EDU-TF), a partitioning optimization problem model tailored to the unique needs of education, integrating both teacher and student requirements. To solve EDU-TF, we propose a modular optimization approach, one of the first to allow the flexible adjustment of objectives according to educational needs, enhancing the method's applicability across various classroom settings rather than just research environments. Results from evaluating ten strategies derived from our model on real-world university datasets indicate that our approach outperforms heuristic teacher-assigned teams by better accommodating student preferences. Our study contributes a new modular approach to partition-based algorithmic team formation and provides valuable insights for future research on team formation in educational settings.

Figures

Figures reproduced from arXiv: 2506.02756 by the authors.

Figure 1
Figure 1. Taxonomy of the General Team Formation Problem (General TFP) families, defined by the number of teams [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Evolution of the solution quality expressed as the sum of all realized preferences during the calculation of [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 13 canonical work pages

  1. [5]

    A comparative study of team formation in social networks

    Xinyu Wang, Zhou Zhao, and Wilfred Ng. A comparative study of team formation in social networks. InDatabase Systems for Advanced Applications: 20th International Conference, DASFAA 2015, Hanoi, Vietnam, April 20-23, 2015, Proceedings, Part I 20, pages 389–404. Springer,

    work page 2015

  2. [11]

    Lift: integrating stakeholder voices into algorithmic team formation

    Emily M Hastings, Albatool Alamri, Andrew Kuznetsov, Christine Pisarczyk, Karrie Karahalios, Darko Marinov, and Brian P Bailey. Lift: integrating stakeholder voices into algorithmic team formation. InProceedings of the 2020 CHI conference on human factors in computing systems, pages 1–13,

    work page 2020

  3. [12]

    Shaping Collaborations with Algorithms: How Agency and Heterogeneity Criteria Influence Team Formation and Outcomes

    Diego Gomez-Zara, Victoria Kam, Charles Chiang, Leslie DeChurch, and Noshir Contractor. Augmenting team diversity and performance by enabling agency and fairness criteria in recommendation algorithms.arXiv preprint arXiv:2410.00346,

    work page Pith review arXiv

  4. [15]

    Team dating leads to better online ad hoc collaborations

    Ioanna Lykourentzou, Robert E Kraut, and Steven P Dow. Team dating leads to better online ad hoc collaborations. In Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing, pages 2330–2343,

    work page 2017

  5. [1972]

    doi:10.1007/978-1-4684-2001-2_9

    ISBN 978-1-4684-2001-2. doi:10.1007/978-1-4684-2001-2_9. URL https://doi.org/10.1007/ 978-1-4684-2001-2_9. G. Ausiello, P. Crescenzi, G. Gambosi, V . Kann, A. Marchetti-Spaccamela, and M. Protasi.Complexity and Approx- imation – Combinatorial Optimization Problems and Their Approximability Properties. Teubner,

    doi:10.1007/978-1-4684-2001-2_9 2001

  6. [2001]

    You want me to work with who? stakeholder perceptions of automated team formation in project-based courses

    Farnaz Jahanbakhsh, Wai-Tat Fu, Karrie Karahalios, Darko Marinov, and Brian Bailey. You want me to work with who? stakeholder perceptions of automated team formation in project-based courses. InProceedings of the 2017 CHI conference on human factors in computing systems, pages 3201–3212,

    work page 2017

  7. [2010]

    Lianying Zhang and Xiang Zhang

    doi:10.1109/SocialCom.2010.12. Lianying Zhang and Xiang Zhang. Multi-objective team formation optimization for new product development. Computers & Industrial Engineering, 64(3):804–811,

    work page doi:10.1109/socialcom.2010.12 2010

  8. [2012]

    An eth-tight algorithm for multi-team formation

    Daniel Lokshtanov, Saket Saurabh, Subhash Suri, and Jie Xue. An eth-tight algorithm for multi-team formation. In 41st IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science (FSTTCS 2021). Schloss-Dagstuhl-Leibniz Zentrum für Informatik,

    work page 2021

Show all 15 references
  1. [2016]

    Designing a multi-disciplinary software engineering project

    28 A Hierarchical Integer Linear Programming Approach for Optimizing Team Formation in Education Patricia Lago, Joost Schalken, and Hans van Vliet. Designing a multi-disciplinary software engineering project. In 2009 22nd Conference on Software Engineering Education and Traini...

    2009

  2. [2017]

    Megan Hammond, Joan Martinez, and Joseph B. Herzog. Work in progress: An optimization model for assigning students to multidisciplinary teams by considering preferences and skills. In2023 ASEE Annual Conference & Exposi- tion, number 10.18260/1-2–44448, Baltimore , Maryland, June

    doi

  3. [2019]

    Submodularity in team formation problem

    Avradeep Bhowmik, Vivek Borkar, Dinesh Garg, and Madhavan Pallan. Submodularity in team formation problem. In Proceedings of the 2014 SIAM international conference on data mining, pages 893–901. SIAM,

    2014

  4. [2020]

    Multi-skill collaborative teams based on densest subgraphs

    Amita Gajewar and Atish Das Sarma. Multi-skill collaborative teams based on densest subgraphs. InProceedings of the 2012 SIAM international conference on data mining, pages 165–176. SIAM,

    2012

  5. [2021]

    doi:https://doi.org/10.1016/j.eswa.2021.114886

    ISSN 0957-4174. doi:https://doi.org/10.1016/j.eswa.2021.114886. Abhijeet Lele. Formation of an efficient team by improvising employee selection process using ahp-lp for a software company in india.Management and Labour Studies, 40(1-2):22–33,

    2021

  6. [2024]

    Assigning or recommending you a team? the algorithmic effects on team formation and performance

    Diego Gomez-Zara. Assigning or recommending you a team? the algorithmic effects on team formation and performance. InAcademy of Management Proceedings, volume 2024, page 12540. Academy of Management Valhalla, NY 10595,

    2024

  7. [2025]

    doi:https://doi.org/10.1016/j.eswa.2024.125289

    ISSN 0957-4174. doi:https://doi.org/10.1016/j.eswa.2024.125289. Manoel Campêlo and Tatiane Fernandes Figueiredo. Integer programming approaches to the multiple team formation problem.Computers & Operations Research, 133:105354,

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.