Feature Incremental Clustering with Generalization Bounds

The problem formulation is novel and well-motivated by real-world applications such as sensor-based activity recognition. The theoretical analysis for FIC-MR is particularly strong: when the pre-trained centers $\mathbf{U}^0$ align with the current distribution (i.e., $\Gamma(\mathbf{U}^0) \to 0$), the generalization bound tightens to a fast rate of $\tilde{\mathcal{O}}(1/n_2)$ instead of the standard $\tilde{\mathcal{O}}(\sqrt{k/n_2})$. This provides a rigorous justification for model reuse in incremental learning. The experimental evaluation across nine diverse datasets consistently validates the theoretical predictions, showing FIC-MR outperforms alternatives especially when current-stage data is scarce.

The reconstruction mechanism in FIC-DR relies on a heuristic k-means-based completion objective (Eq. 5) that lacks theoretical guarantees for accurate feature recovery; consequently, the $\mathcal{Y}$-discrepancy term in Theorem 3 may be uncontrollably large in practice. The generalization bounds contain unspecified problem-dependent constants (e.g., $V$, $\gamma$) that limit their quantitative utility for hyperparameter selection. Additionally, the analysis assumes bounded features $\|\mathbf{x}\| \leq \gamma$ and relies on correlations between old and new features for reconstruction, assumptions that may fail in high-dimensional or adversarial settings. The paper also omits comparison to modern deep clustering baselines, restricting evaluation to k-means variants.

The empirical results generally support the theoretical hierarchy of methods, with FIC-MR > FIC-DA > FIC-DR across most datasets, validating that model reuse is preferable to distribution-sensitive reconstruction when historical data is unavailable. However, the evaluation relies on external validity metrics (ACC, NMI, F-score) that require ground-truth labels, which creates a methodological tension since the proposed algorithms optimize an unsupervised squared-norm criterion and the generalization bounds are derived for expected clustering risk. The comparison to baselines (KM-P1 and KM-C1) is fair for isolating the value of incremental features, but the absence of streaming clustering or deep incremental clustering baselines limits the assessment of practical competitiveness.

The experimental protocol is described in sufficient detail to permit replication: nine public datasets are employed with a standardized split (50% previous, 20% test, remainder current), and the single hyperparameter $\theta$ in FIC-MR is searched over a disclosed grid $\{10^0, 10^1, 10^2, 10^3\}$. However, no code or implementation has been made publicly available, and the FIC-DR reconstruction procedure lacks algorithmic specifics (e.g., convergence criteria, initialization of missing features). Furthermore, the experimental results report means and standard deviations over only ten runs, which may be insufficient for establishing statistical significance given the variance observed in Tables II and III.