acceptodds
Under review as a conference paper at ICLR 2027

OFIGM: Online Feature-Instance Graph Memory for Open-World Class-Incremental Learning

Abstract

Real-world data streams are inherently open and nonstationary: previously unseen classes may emerge, labels can be sparse or delayed, and the observable feature space itself may evolve over time. However, existing continual, open-world, and streaming learning methods typically address these challenges in isolation. We propose OFIGM, an Online Feature-Instance Graph Memory framework for open-world class-incremental learning under simultaneous class and feature evolution. OFIGM maintains a bounded dynamic bipartite graph comprising transient instance nodes and persistent feature nodes, performs sparse graph message passing to learn schema-adaptive representations, and preserves previously acquired class knowledge through slowly updated prototype memory. A graph-aware novelty score integrates prototype distance, predictive uncertainty, and structural residuals to identify unknown samples, while a persistence-aware unknown buffer promotes coherent novel groups to new classes without global retraining. Across Digits, Wine, and Iris, OFIGM achieves mean Macro-F1 of , novel-class F1 of , and novelty AUROC of . Relative to a matched no-graph ablation, graph memory improves novelty AUROC from to on Wine and from to on Iris. We further evaluate generalization on an external Louisiana shellfish outbreak dataset containing 11 documented episodes from 2002-2025, where OFIGM achieves Macro-F1 of and outbreak recall of . These results demonstrate the feasibility of persistent feature-instance graph memory as a unified mechanism for novel-class detection, incremental class creation, and continual adaptation in evolving open-world streams.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Online Feature-Instance Dynamics, Class-Incremental Learning, Data Streams, Online Learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_iE6BOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses open-world class-incremental learning where both classes and features evolve over time, with limited labels and no explicit task boundaries.

OFIGM maintains a bounded bipartite graph of instances and features, using message passing to maintain representations as new features arrive. It combines prototype-based classification, anchoring for historical knowledge retention, and a novelty detection mechanism that uses prototype distance, prediction uncertainty, and neighborhood consistency. New classes are created when accumulated candidates satisfy persistence and coherence criteria.

Section 5 outlines an extensive evaluation protocol covering SENC, OCIL, distribution shifts, and feature growth. However, Section 6 reports only preliminary experiments on Digits, Wine, and Iris against two simple baselines and a no-graph ablation, along with an external outbreak detection experiment.

Critical strengths

(1) Relevant problem formulation. Jointly handling emerging classes and evolving feature spaces is an interesting setting that deserves more attention. The bipartite instance-feature representation is a reasonable design choice because new features can be introduced without resizing the model's input layer. I also appreciate that Section 2.4 distinguishes this application from prior work on instance-feature graphs.

(2) Transparent reporting of negative results. Section 6.2 acknowledges that removing graph memory improves several results, including all reported metrics on Digits, while Section 6.3 reports weaker early-warning performance for the graph variant. Although these results weaken the contribution, reporting them explicitly is appreciated.

Critical weaknesses

(1) The experimental evaluation falls substantially short of what the paper proposes. Section 5 describes multiple evaluation suites, including SENC, OCIL, and feature-growth benchmarks, with several established baselines. However, none of the direct SENC competitors listed in Table 2 are evaluated in Section 6. Instead, Table 3 reports results on Digits, Wine, and Iris against a prototype baseline, windowed kNN, and a no-graph variant.

The statistical protocol is also inconsistent. Section 5.3 specifies ten seeds for tabular experiments, significance testing, and confidence intervals, whereas Section 6 uses only three seeds without these analyses. Section 5.1 proposes label budgets of {1%, 5%, 10%, 25%}, but Section 6.1 reports only 20%. These are not minor omissions; the experiments needed to support the main contributions have largely not been conducted.

(2) The reported results do not convincingly establish the benefit of graph memory. On Digits, the no-graph ablation outperforms the full model across the reported metrics. On Wine, the simple prototype baseline achieves better Macro-F1 (0.889 vs. 0.872), AUROC (0.962 vs. 0.946), and forgetting. The AUROC improvements highlighted against OFIGM-noFG are small relative to the reported variability over three seeds.

This raises a basic question about what the persistent feature graph actually contributes. The class-creation thresholds, prototype anchoring, pseudo-labeling, and graph loss are not independently ablated. More importantly, feature growth is a central contribution, yet there is no convincing experiment isolating performance under evolving features or comparing against relevant methods such as SAFC.

(3) The external evaluation and presentation further weaken the submission. Section 6.3 reports Macro-F1 of 0.474 ± 0.186 and early AUROC of 0.577 over only 11 outbreak episodes. The graph model also performs worse on the pre-confirmation early-warning metrics. These results are too limited to support claims of generalization to real-world outbreak detection.

There are also several issues with the manuscript itself. Equation (1) and multiple mathematical symbols in Sections 3–4 are incorrectly rendered. Algorithm 1 and Appendices A, D, and E are referenced but absent from the supplied manuscript. These omissions make it difficult to assess parts of the method and the external evaluation.

Questions for authors

(1) Could the authors provide results against at least some of the direct baselines listed in Table 2, such as SEEN, KNNENS, HANOL, or SAFC, on the SENC datasets? Without these comparisons, it is difficult to assess the contribution relative to existing methods.

(2) On Wine, the prototype baseline outperforms full OFIGM on several metrics. How do the authors reconcile this with the claim that persistent graph memory improves performance?

(3) Could the authors repeat the experiments with the ten seeds specified in Section 5.3 and report statistical significance? The AUROC differences against OFIGM-noFG are relatively small compared with the reported standard deviations.

(4) Where is the evaluation of the proposed feature-growth capability? A comparison against a fixed-schema variant or SAFC under controlled feature arrival would help establish this contribution.

(5) Could the authors ablate the persistence-based class-creation mechanism, particularly the minimum component size, coherence threshold, and age requirement? How do these parameters affect detection delay and false-novel rates?

(6) Section 5.1 specifies label budgets of {1%, 5%, 10%, 25%}, while Section 6.1 uses 20%. Which settings were actually evaluated, and are results for the proposed budgets available?

Minor Comments

(1) Several mathematical symbols in Sections 3–4 are incorrectly rendered, making some definitions difficult to follow.

(2) Algorithm 1 and Appendices A, D, and E are referenced but missing from the supplied manuscript.

(3) Section 6.2 contains an incomplete concluding sentence.

(4) The statement in Section 1 regarding the stability–plasticity trade-off needs grammatical revision.

(5) Section 5's proposed evaluation protocol should be clearly distinguished from experiments actually completed.

Reviewer_UX5pOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes OFIGM, an online feature-instance graph memory framework for open-world class-incremental learning with evolving feature spaces. It maintains persistent feature nodes and bounded instance memory, combines graph-based representations with class prototypes for knowledge retention, and employs graph-aware novelty detection and persistence-based class creation. Experiments on three public datasets and an external outbreak dataset investigate its effectiveness in online learning settings.

Critical strengths
  1. The paper considers an interesting open-world learning scenario where new classes and features may emerge simultaneously, with sparse or delayed supervision.

  2. The proposed method integrates dynamic feature-instance graph memory, prototype-based retention, novelty detection, and incremental class creation within a single online learning framework.

Critical weaknesses
  1. Significant mismatch between the proposed experimental design and the actual evaluation.

Section 5 introduces an extensive evaluation plan covering several benchmark suites, including streaming emerging-new-class learning, online class-incremental learning, and feature-space growth. The paper also identifies several relevant methods, such as SENCForest, SEEN, KNNENS, HANOL, MLG, and SAFC, as direct baselines. However, the actual experiments are limited to Digits, Wine, and Iris, with only prototype-based learning, kNN, and an internal no-graph ablation as comparisons.

Why are the proposed benchmark suites and state-of-the-art baselines not included in the actual experimental results? Without direct comparisons against representative methods under matched protocols, it is difficult to evaluate the competitiveness of OFIGM or support the claimed contributions to open-world class-incremental learning.

  1. Limited evidence supporting the effectiveness of graph memory.

The central contribution of OFIGM is the persistent feature-instance graph memory. However, its empirical benefits appear limited and inconsistent. On Digits, removing graph memory slightly improves Macro-F1, Novel-F1, and AUROC. Although OFIGM improves AUROC relative to its no-graph variant on Wine and Iris, simpler baselines still achieve higher AUROC on these datasets. Furthermore, the no-graph ablation removes both the graph memory and the graph-residual novelty signal, making it difficult to identify which component contributes to the observed improvements.

What evidence demonstrates that the feature-instance graph memory is necessary and consistently beneficial? Could the authors provide separate ablations for graph-based representation learning, persistent feature memory, and graph-aware novelty scoring? Stronger comparisons would be needed to establish the contribution of the proposed graph mechanism.

  1. Unclear memory scalability and class creation behavior.

Although the number of transient instance nodes is bounded, feature nodes are persistent, and new class prototypes are added as new classes emerge. Thus, the overall memory requirement may continue to grow with the number of features and classes.

What is the actual total memory complexity of OFIGM as the feature and class spaces expand? Additionally, how reliably does the persistence-aware unknown buffer create new classes? Could the authors report class discovery delay, false class creation rates, and class assignment quality under different novelty thresholds and delayed-label conditions?


The figures appear AI-generated and lack the clarity and precision expected of a top-tier conference submission. The authors should treat the submission more seriously.

Questions for authors

See weakness

Reviewer_bQpBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes OFIGM, an Online Feature-Instance Graph Memory framework for open-world class-incremental learning under simultaneous class and feature evolution. OFIGM maintains a bounded dynamic bipartite graph to learn schema-adaptive representations, preserves previously acquired class knowledge through slowly updated prototype memory, and combines a graph-aware novelty score with a persistence-aware unknown buffer to identify and register new classes online. Experiments on Digits, Wine, Iris, and an external Louisiana shellfish outbreak dataset show that graph memory improves novelty detection on some datasets.

Critical strengths
  1. This paper focuses on an online learning scenario where new classes and new features occur simultaneously. This is a more common and more challenging scenario in real-world applications.
Critical weaknesses
  1. First of all, many details are unclear, and there are inconsistencies in the use of symbols. It’s difficult to list them all because there are so many. For example, the message functions φF\varphi_F and φI\varphi_I in Eqs. (5) and (7) lack explicit computational definitions, while qtq_t denotes stream metadata in Eq. (8) but pseudo-label confidence in Eq. (18).

  2. Although samples satisfying st>θts_t>\theta_t are marked as unknown and buffered, the paper does not specify whether they are handled for model training. This leaves the role of novelty detection in the online learning process insufficiently defined. Besides, the neighborhood in Eq. (15) is induced by shared feature nodes. If instances observe the same feature set, Eq.(15) will reduce to the distance between the current embedding and the window embeddings.

  3. Reducing the prototype update rate merely slows down the rate at which the prototype changes; it does not guarantee that it remains aligned with the current embedding space. Furthermore, the method for initializing the prototype is not specified. The nature of zˉc\bar{z}_c in Eq. (11) is also unknown, making it difficult to analyze the specific role played by LstabL_stab.

  4. The paper contains a large number of manually defined metrics, which makes it read more like an engineering-driven product. For example, Eq.(16) uses the size, internal consistency, and duration of candidate clusters to detect the identification of new classes caused by known class drift. The generalizability of these manually defined criteria is difficult to assess, and I believe these three metrics do not adequately address the problem, since known classes can also drift during model training and may similarly satisfy these conditions.

Questions for authors
  1. Could the authors explicitly define φF\varphi_F and φI\varphi_I in Eqs. (5) and (7), and clarify the two different meanings of qtq_t in Eqs. (8) and (18)?

  2. How are samples satisfying st>θts_t>\theta_t used during training? Are they excluded from the known-class pseudo-label loss, and do they still update the feature memory? When all instances share the same feature set, does Eq. (15) reduce to the average cosine distance from the instance window? If so, what additional structural information does it capture?

  3. How are class prototypes initialized and kept compatible with the evolving embedding space? What exactly is zˉc\bar z_c in Eq. (11), and how is it computed when class cc is absent from the current instance window?

  4. What assumptions allow the cluster-size, coherence, and persistence criteria in Eq. (16) to distinguish genuinely new classes from drifted known classes? How are their thresholds selected, and how does the method correct a mistaken class registration?

  5. The experimental results are inconsistent, and the descriptions vary from one section to another.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.