acceptodds
Under review as a conference paper at ICLR 2027

PlaceGuard: Record-Placement Verification for Multimodal Graph OOD Detection

Abstract

Open-world catalog admission must determine whether a submitted text-image record belongs at a claimed graph identity before assigning a known class. Existing multimodal OOD detectors evaluate record familiarity, and graph detectors evaluate a node with its topology, yet neither view directly verifies the record-position pairing. We study record-placement OOD, where a record and graph position are individually in-distribution but their assignment is unsupported. An in-distribution record can therefore be out of place. A position-controlled study reassigns intact records while keeping recipients, relations, and target-excluded contexts constant. One-to-one reassignment makes every deterministic record-only score exactly nondiscriminative and leaves position-only scores invariant, whereas record-position compatibility separates cross-class and within-class reassignment. PlaceGuard turns that observation into three necessary checks for within-record validity, record-position fit, and context familiarity. It multiplies the first two supports so that both must be high and uses a one-way context check that increases unknownness only beyond the clean validation tail. Across six public graphs and three class splits, PlaceGuard reaches conventional semantic OOD AUROC 0.913 and macro weakest-locus AUROC 0.780, compared with 0.504 for the strongest method under the shared protocol. The results show that assignment verification complements conventional semantic detection when a familiar record is submitted to an unsupported relational position.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Multimodal Attributed Graphs, Out-of-Distribution Detection, Record-Placement Verification, Graph OOD Detection, Open-World Recognition

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fEGVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper attempts to solve the problem that existing graph anomaly detection algorithms often fail to recognize records with normal content that are misplaced in the graph. The authors confirm this detection blind spot through controlled experiments and propose the PlaceGuard method, which identifies misplacement anomalies through three-layer verification and achieves much better detection results than existing methods.

Critical strengths
  1. A new OOD detection problem scenario has been proposed, where the record content is normal but does not match the graph position.
  2. Through controlled experiments and theoretical analysis, it has been proven that relying solely on records or graph positions cannot identify such mismatch anomalies.
  3. The authors proposed a non-compensatory verification approach that requires both internal consistency and positional matching to be satisfied, avoiding high confidence in one dimension masking anomalies in another dimension.
Critical weaknesses
  1. All cases of record misplacement are artificially constructed and lack validation in real-world scenarios.
  2. The baseline methods do not use the same targeted training strategy for misplacement, which raises concerns about the fairness of the experimental comparison.
  3. The core design of the method combines multiple existing techniques, with limited methodological novelty.
  4. Detection performance remains unsatisfactory for within-class record misplacement cases.
Questions for authors
  1. Can artificially constructed record misplacement samples represent anomalies in real scenarios? Is there any real-world data to validate the method’s effectiveness?
  2. If the baseline methods are also trained with the same misplacement samples, can PlaceGuard still maintain a significant advantage?
  3. Compared to simple feature fusion or score weighting, does PlaceGuard’s non-compensatory verification have significant advantages?
  4. The detection performance of within-class record misplacement is limited and the false positive rate is high. Can this method meet practical application requirements?
1 comment

We thank the reviewer for the questions on realistic errors, training fairness and practical use. AUROCs/rates are percentages. We average six graphs within each split; ± is sample SD over the three split means.

W1 & Q1 — Real errors and controlled validation. Yes. We recovered public Wikimedia Commons corrections: photographs of the U.S. Navy Commodores ensemble had been assigned to the same-named R&B band category. Editors corrected the category association while leaving the photos and descriptions unchanged (correction discussion).

These corrections document real occurrence; the reported AUROCs come from controlled reassignments on six real graphs. The controlled test changes record assignments while holding recipient contexts and the record multiset fixed (Section 5.1). It isolates whether a detector checks correspondence: record-only scores cannot detect an intact-record permutation (Proposition 1). BOND [1] likewise evaluates injected and organic outliers separately. Our control targets the distinction illustrated by the real corrections: normal content can still have an incorrect association.

W2 & Q2 — The actual baselines with the same misplaced samples. After training all seven Table 2 baselines on the same four replacement types, with identical replacement examples and donors, the best native baseline reaches 50.5±0.1 weakest-locus AUROC, versus 79.5±0.7 for revised PlaceGuard. TIDE leads on conventional semantic OOD in the same comparison (92.4 versus 91.4). We retain each native architecture and inference score, using two fixed exposure strengths, at most 120 additional epochs and patience 20. These rows select epochs on clean known validation, then exposure strength on structured validation (Appendix E.10, Table 17).

We separately align the strongest exposed baseline, TNT-OOD, with the MLP/PlaceGuard structured-validation criterion at every epoch. Its weakest AUROC rises to 51.9±0.1, leaving a paired gap of 27.6±0.7 points to revised PlaceGuard. At that same TNT checkpoint, removing propagation gives 61.4 weakest AUROC, while conventional AUROC falls from 90.7 to 67.8. Thus selection and propagation each affect the result, but neither change closes the gap.

The MLP directly learns to reject a record–recipient pairing from their separate feature vectors. The native baselines retain their own representations, objectives and scores; identical negative samples do not make them the same detector. The PlaceGuard–MLP comparison below additionally matches input features, parameter count and selection rules.

W3 & Q3 — Technical contribution, fusion and score weighting. PlaceGuard trains two distinct decisions: whether the record is internally valid and whether it fits this recipient. Clean records receive (validity, fit)=(1,1), intact misplaced records (1,0), and one-modality replacements (0,0). The validity head sees the record alone; the fit head also sees its interactions with known neighbors. These labels distinguish normal-but-misassigned content from corrupt content (Section 4.3).

For supports in [0,1], the submitted product is bounded by each support, so a strong check cannot cancel a weak one. With the same heads, clean calibration records and context check, it exceeds equal weighting by 5.6 weakest-locus and 2.0 conventional AUROC points. Appendix E.9 reports this fixed-head comparison.

Table 15 tests the full feature-fusion MLP separately. Revised PlaceGuard improves weakest/conventional means over it by 1.1/7.2 points (79.5/91.4 versus 78.4/84.2). Even without context rejection, the revised verifier gives 79.7/88.6. An intact-record permutation preserves validity scores as a multiset, but multiplying them can reorder fit scores. The revision leaves fit unchanged for ordinary validity values; validity adds rejection only in the lowest 2.5% of clean validation validity scores. The heads stay fixed. The rule and cutoff were fixed after the original analysis and before evaluation on the existing reporting splits (Appendix E.9). The submitted product is retained as the scoring control in Table 16.

W4 & Q4 — Same-class errors and practical thresholds. Same-class swaps preserve class and text–image coherence, leaving recipient-specific fit as the detection signal. With a threshold set for 95% clean validation acceptance, revised scoring achieves 37.4±0.8% same-class recall and 6.2±0.7% clean-test rejection (Appendix B.4, Table 8). At 95% same-class error recall, clean FPR is 69.0%. For review-assisted admission, users set the threshold from a desired clean-validation acceptance rate and apply it to incoming records.

References

[1] K. Liu et al. BOND: Benchmarking Unsupervised Outlier Node Detection on Static Attributed Graphs. NeurIPS Datasets and Benchmarks, 2022.

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

The paper studies record-placement OOD in multimodal graphs. A familiar text-image record can be assigned to an unsuitable graph position. The controlled protocol reassigns records while keeping recipient positions and contexts fixed. PlaceGuard multiplies within-record validity and record-position fit scores. A separate context check can only increase rejection. Training uses structured negatives from known classes. Experiments cover six graphs and three class splits. The paper reports 91.3% conventional semantic AUROC and 78.0% weakest-locus AUROC under fixed recipients, compared with at most 50.4% for the other methods.

Critical strengths
  1. The controlled protocol makes the information limit clear. Proposition 1 shows why deterministic record-only scores cannot detect a one-to-one record permutation. Figure 1 shows that record-context compatibility can still detect cross-class and same-class reassignment. This gives a useful diagnostic for placement errors that a record-only detector cannot identify in this setting.

  2. The stated selection protocol is careful. Development seeds 53/54 are separate from final seeds 42-44. Donor pools are disjoint across training, validation, and evaluation. Thresholds use clean known records. Table 7 also reports clean acceptance and weakest-shift recall. These details help assess both the evaluation boundary and the practical rejection trade-off.

Critical weaknesses
  1. The benefit of factorization is not isolated. PlaceGuard learns from four types of structured negatives (Sec. 4.3). A plain joint record-context classifier trained on the same negatives is not reported. Table 3 compares product-based verification variants, so it does not provide this control. Appendix E.7 also describes cosine and CoOD-style compatibility controls, but their five-shift results are not reported alongside Table 2. This comparision matters because chance placement AUROC for record-only methods already follows from Proposition 1. The large gap in Table 2 does not establish the added value of the proposed factorization.

  2. Table 1 has inconsistent averages. The six displayed TIDE scores average to 91.58, not 90.5. GRASP averages to 89.55, not 89.9, and TNT-OOD to 90.67, not 90.9. These gaps exceed what one-decimal rounding can explain. With equal graph weights, TIDE would rank above PlaceGuard's 91.3 in this table. Section 5.2 also gives an upper macro AUROC of 92.5, which does not appear in the Avg. column. The aggregation rule and reported values need correction or clarification. This affects confidence in the conventional semantic OOD comparison.

Questions for authors
  1. Clarification: How is the Avg. column in Table 1 computed? Please resolve the discrepancies above and provide the aggregration rule and corrected table from the saved results.

  2. Validation request: How does a single MLP or bilinear record-context scorer perform with the same features, four negative types, donor pools, and selection budget? This would test the contribution of factorization.

  3. Clarification: Can you report the Appendix E.7 controls under the Table 2 protocol for all six graphs, including the record-context cosine and CoOD-style compatibility scores?

  4. Clarification: TIDE and TNT-OOD obtain only 52.4 and 51.5 semantic-record AUROC under fixed recipients. Can you check their sensitivity to the target record before and after support-to-query propagation? This would help assess whether the adapted scores retain the intended content signal.

2 comments

We thank the reviewer for the precise control requests and numerical check. AUROCs are percentages. We average six graphs within each split; ± is sample SD over the three split means.

W1 & Q2 — A single MLP with matched training. Revised PlaceGuard improves weakest-locus AUROC over the MLP by 1.1 points (79.5±0.7 versus 78.4±0.6) and conventional semantic AUROC by 7.2 points (91.4±0.9 versus 84.2±1.3; Appendix E.9, Table 15).

Both models receive identical frozen text/image features, target-excluded text/image contexts, prediction disagreement, four types of replacement examples and donor pools. The MLP is Linear–ReLU–Linear with one binary admission target and a negative-logit score. Parameters differ by at most 0.063%. Both use two optimizer configurations, at most 120 epochs and patience 20. Epoch/config selection maximizes the lowest recall across the four replacement types at 95% clean validation acceptance, then the lowest-type AUROC.

PlaceGuard explicitly separates two questions: is the record itself valid, and does it fit this recipient? Intact misplaced records receive (validity, fit)=(1,0), one-modality replacements (0,0), and clean records (1,1). Thus training distinguishes a valid record at the wrong position from a record whose content is corrupted. The submitted product gives a 7.1-point conventional AUROC advantage over the MLP, with weakest-locus AUROC 0.5 points lower (Tables 15–16).

Multiplication can reorder fit scores even when a permutation preserves intact records. We revised only the score combination, keeping the heads fixed: ordinary validity scores leave fit unchanged; validity adds rejection only in the lowest 2.5% of clean validation validity scores. SIRC [1] provides a related rationale for preserving a primary score while adding auxiliary confidence. The rule and inherited 0.975 cutoff were fixed after the original analysis and before evaluation on the previously observed reporting splits; E.9 defines the revision, and Tables 1–2 use the submitted scoring rule, with Table 1 averages corrected below.

Even without context rejection, the revised verifier reaches 79.7±0.6 weakest-locus / 88.6±0.5 conventional AUROC, both above the MLP; context checking adds 2.8 conventional points. PlaceGuard thus uses the same training examples and capacity to detect familiar records at incompatible positions while retaining stronger rejection of semantic unknowns.

W2 & Q1 — Table 1 averages. Thank you for catching the reporting inconsistencies. We recalculated Avg. as the equally weighted mean of the six displayed graph AUROCs. Before rounding, averaging over graphs and splits in either order gives the same mean. The ± values are the previously reported sample SDs across the three split-level macro means. In table order, the corrected means are 84.0, 89.6, 86.7, 88.6, 90.7, 74.1, 91.6 and 91.3. Section 5.2 now reports 74.1–91.6. TIDE ranks first (91.6), and PlaceGuard ranks second (91.3). The submitted PlaceGuard also reaches 78.0 weakest-locus AUROC on fixed recipients (Table 2). With identical replacement examples and donors, the strongest of the seven retrained baselines reaches 51.9 after checkpoint alignment, versus 79.5 for revised PlaceGuard (Table 17).

Q3 — Both E.7 controls on all six graphs. Table 14 now reports all five shifts for each of the six graphs under exactly the Table 2 recipients and donor assignments. The CoOD-style score is the component-role compatibility control specified in E.7. Weakest-locus AUROCs are 71.8±0.6 for record–context cosine and 67.1±0.3 for that compatibility score, versus original PlaceGuard's 78.0±0.8 (Table 13). Weakest locus takes the minimum over five shifts within each graph/split before averaging.

Q4 — Target sensitivity before/after propagation. Appendix E.8 reports this checkpoint-fixed control. With the same checkpoints and clean/replacement pairs, disabling only propagation raises semantic-record AUROC from 52.4±0.1 to 60.3±0.4 for TIDE and from 51.5±0.1 to 60.6±0.7 for TNT-OOD. After propagation, only 25.0% and 12.5% of their mean absolute paired score changes remain. These are averages of the after/before ratios over graphs and splits.

The additional TNT exposure/selection check in E.10 also gives 81.3 semantic-record AUROC before propagation versus 56.6 after it. Thus a trained model can retain substantial target-content evidence before propagation, which the full score then attenuates. Table 2 retains the original checkpoints and their full propagated scores; E.10 reports the additional-training check separately.

References

[1] G. Xia and C.-S. Bouganis. Augmenting the Softmax with Additional Confidence Scores for Improved Selective Classification with Out-of-Distribution Data. IJCV, 2024.

Thanks for fixing Table 1 and adding the MLP control, but the gain of the factorized design over the MLP is small (about 1 point on weakest-locus, and lower with the submitted rule), so I keep my score.

Reviewer_kUdKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies record-placement OOD detection, where a familiar text–image record is assigned to an incompatible graph position. PlaceGuard checks within-record validity and record–context compatibility using two learned heads, combines their support scores multiplicatively, and applies an additional context-based rejection rule. Training uses synthetic text, image, and complete-record replacements constructed from known-class data, while calibration uses clean known records. Experiments on product and social graphs evaluate semantic unknowns and several controlled reassignment conditions, reporting improvements over adapted OOD baselines.

Critical strengths

1.The paper studies multimodal graph OOD detection, which is a relatively new and interesting problem setting.

2.The experimental comparisons mainly involve new baselines, making the results more informative for assessing the current competitiveness of the proposed approach.

3.The paper considers two different types of distribution shifts, including Semantic OOD and the proposed Record-placement OOD, which broadens the scope of the evaluation.

Critical weaknesses

1.The definition of Record-placement OOD is somewhat difficult to interpret. As the authors note, some OOD samples are not intrinsically unusual records but are instead records placed in unusual locations, which appears closely related to anomaly detection; therefore, the motivation for formulating this setting specifically as an OOD detection problem, rather than an anomaly detection problem, needs to be clarified.

2.A “record” essentially corresponds to node features containing both textual and visual attributes, while “placement” mainly reflects the connectivity of the node, making Record-placement OOD resemble feature- or structure-level anomaly detection. This setting may potentially be addressed by relatively simple classical methods, but such baselines are not evaluated; moreover, it remains unclear what specific challenges multimodality introduces in this setting and why the proposed method is particularly necessary for addressing them.

3.The experimental analysis is relatively limited in explaining why the proposed method works and why it can effectively handle both Semantic OOD and Record-placement OOD, leaving the underlying effectiveness of the method insufficiently validated.

Questions for authors

See Weaknesses.

2 comments

We thank the reviewer for the questions on the task, multimodality and mechanism. AUROCs are percentages. We average six graphs within each split; ± is sample SD over the three split means.

W1 — Relation to anomaly detection. Record misplacement is a contextual anomaly. For example, a normal product title and image can be attached to another product's identifier and purchase history. The record is familiar, but it does not belong at that position. We call this record-placement OOD because the record–context pairing departs from known training examples. Known-class training and clean-known calibration let the same detector reject both unseen classes and misplaced familiar records.

“Position” means the existing identifier and relations; no edge needs to change. Our controlled test keeps these relations and the target-excluded context fixed. An intact permutation preserves every record, so any deterministic record-only detector has identical clean/shifted score multisets (Proposition 1). A joint attribute–structure anomaly detector can use the correspondence that a record-only score misses. Section 2.2 now makes this overlap explicit.

W2 — Simple methods and the multimodal challenge. Weakest-locus AUROC takes the minimum over five shifts within each graph/split before averaging. We evaluated the simple record–context cosine score, the E.7 component-role compatibility score, and DOMINANT [1], which reconstructs attributes and structure. Their weakest-locus AUROCs are 71.8±0.6, 67.1±0.3 and 50.9±0.2, versus original PlaceGuard's 78.0±0.8 (Appendix E.7, Tables 13–14). DOMINANT uses the authors' network adapted to known-support training and query–support reconstruction. These results concern that inductive adaptation.

Multimodality creates two distinct errors: replacing only the image can break text–image agreement, whereas moving an intact pair preserves that agreement but can break its fit to the recipient. A text-only score gives 50.0 AUROC on image replacement, and an image-only score gives 50.0 on text replacement. Original PlaceGuard gives 91.2/94.9 on text/image replacement. Detecting both errors requires checking the record itself and its relation to the claimed node; simple similarity does not perform as well as learning these checks.

W3 — Why the two checks handle different shifts. The validity head sees the text–image record alone. The fit head also sees its interactions with connected known records. Clean examples receive (validity, fit)=(1,1); an intact misplaced pair receives (1,0); a one-modality replacement receives (0,0). This teaches the model that a normal record can still be wrongly assigned, instead of treating every negative as internally invalid (Section 4.3).

With the heads fixed (Table 16), letting validity add rejection only in its lowest 2.5% of clean validation scores raises text/image AUROC from 85.11/92.88 to 88.92/94.62, while weakest-locus AUROC remains close (79.51 versus 79.52). Validity helps detect corrupt content; fit carries the intact-placement signal. The context check uses known-class predictions from the neighborhood, excluding the target. It adds rejection only when their uncertainty exceeds the 97.5th clean-validation percentile. With the same revised assignment score, it raises conventional AUROC from 88.58 to 91.38 (Table 15). An intact misplaced record can fail fit while remaining internally valid; an unfamiliar semantic neighborhood can trigger the context check. Appendix E.9 defines the revised combination and compares it with the submitted product.

References

[1] K. Ding et al. Deep Anomaly Detection on Attributed Networks. SDM, 2019.

Thank you for the rebuttal. I find multimodal graph OOD detection an interesting research problem. However, after considering the authors' response, I remain concerned that the proposed problem setting is more closely aligned with anomaly detection than OOD detection. Since this distinction is fundamental to the validity of the problem formulation, I have decided to maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.