Unsupervised Kernel Weighting by Agreement

What Does “Agreement” Mean?

When several omics views are available, their similarity matrices need not describe patients in the same way. An unsupervised weighting criterion can favor a fused kernel that agrees with a chosen consensus similarity pattern, without using a clinical outcome. The choice of target matters: optimizing agreement is meaningful only when the target represents a defensible reference rather than a pattern the candidate weights can reproduce by construction.

Let \(n\) be the number of patients and \(M\) the number of views. For view \(m\), let \(K^{(m)}\) be its \(n\times n\) kernel matrix, with the same patients in the same order in every view. A weighted additive kernel is \(K_w=\sum_{m=1}^M w_m K^{(m)}\). The earlier material on weighted kernel combinations establishes why nonnegative weights summing to one yield a convex combination and preserve positive semidefiniteness when the component kernels are positive semidefinite.

For agreement scoring, first center each kernel across patients, and then scale it to unit Frobenius norm. Write \(H=I_n-\mathbf{1}\mathbf{1}^{T}/n\), and define \(A_m=HK^{(m)}H/\|HK^{(m)}H\|_F\), provided the centered matrix is not zero. Let \(C\) be a fixed, nonzero, centered consensus similarity matrix on the same patients, scaled so that \(\|C\|_F=1\). The consensus may come from a prespecified external source or a separately constructed consensus network. It must be held fixed while the candidate weights are optimized.

Objective and Constraints

A common unsupervised choice maximizes \(J(w)\) over the simplex:

The nonnegativity constraint prevents a view from being subtracted from the others; the sum-to-one constraint fixes the scale of the weights and makes them interpretable as relative contributions. The denominator must be positive for every evaluated candidate. If the weighted centered kernel is zero, its direction and hence its cosine are undefined. Upper or lower bounds on individual weights can be added when required by the analysis plan, but they change the feasible set and should be reported.

This objective is not the same as maximizing an unnormalized inner product. Without normalization, the score can reward matrix magnitude. Nor is it generally a linear optimization: the norm of the fused matrix depends on the weights. A numerical optimizer can search the simplex, but the result should be checked against boundary solutions and alternative initial values when the optimization procedure is iterative.

Why the Score Has This Form

The Frobenius inner product is an ordinary dot product between the entries of two matrices viewed as vectors. Define \(a_m=\langle A_m,C\rangle_F\) and \(G_{mr}=\langle A_m,A_r\rangle_F\). By linearity of the inner product, the numerator is \(w^Ta\). Expanding the squared denominator gives:

Thus the objective can also be written \(J(w)=w^Ta/\sqrt{w^TGw}\), when \(w^TGw>0\). The matrix \(G\) is a Gram matrix, so it is positive semidefinite. This representation clarifies two features: a view’s score against the target is not the whole story, because similarities among the views affect the fused norm; and redundant views can supply nearly the same direction rather than distinct evidence.

Interpreting Constrained Choices

The simplex prevents negative contributions but does not require every view to receive positive weight. An optimum can lie on a face or vertex of the simplex. Conversely, high agreement from a mixture can reflect complementary patterns, rather than one view being individually closest to the target. The weights are criterion-dependent: they summarize how the selected objective favors the views under the chosen preprocessing and constraints.

In practice, report the optimized score, the weight vector, the constraints, and the way the target was formed. A large weight does not establish that the corresponding assay is biologically more important. In particular, if two kernels are strongly correlated, many weight allocations can yield very similar fused matrices and scores.

Scaling, Diagnostics, and Failure Modes

Normalization is essential because kernel magnitude can otherwise dominate a weighted sum. For a simple calculation, suppose two raw kernels have Frobenius norms 10 and 1, and their individual cosines with a unit-norm target are 0.80 and 0.60. Assume their unit directions are orthogonal. An equal-weight combination of the raw kernels has target inner product \((8+0.6)/2=4.3\) and norm \(\sqrt{(10^2+1^2)/4}=\sqrt{25.25}=5.025\), giving cosine \(4.3/5.025=0.856\). If instead each kernel is first normalized to unit norm, the equal-weight combination has cosine \((0.80+0.60)/\sqrt{2}=0.990\). The change arises from scale, not new information. The normalized result is also mathematically feasible here: the squared projections \(0.80^2+0.60^2=1\) account for the target direction.

A consensus made from the same candidate views may still be useful as a descriptive reference, but it creates a circularity risk: a view can score well partly because it helped define the target. Where appropriate, build the target without the view being evaluated, or use a consensus specified independently of the candidate fusion. Clearly distinguish a fixed target from a target recomputed for each candidate weight vector; the latter defines a different optimization problem.

Useful diagnostics include checking that all matrices use the same patient order, verifying symmetry and centering, inspecting kernel norms before normalization, and confirming that the optimized fused norm is not near zero. Examine pairwise kernel similarities and the objective across the simplex or near the optimum. Similar scores for substantially different weights indicate weak identification of the weights. A boundary solution can be legitimate, but it should prompt checks for an uninformative view, inconsistent preprocessing, and a consensus dominated by one source.

• Scale confounding: applying the agreement criterion to unnormalized kernels can make weights respond to magnitude rather than similarity pattern. Center and normalize using a stated rule.
• Target circularity: reusing candidate kernels to construct the target can make their apparent agreement optimistic. Use a defensible fixed target and acknowledge shared information.
• Overinterpretation: an unsupervised agreement weight is not an estimated causal effect, a measure of assay quality in general, or evidence of prediction of a clinical endpoint.
• Degeneracy and redundancy: a zero-norm fused matrix makes the cosine undefined; highly similar kernels can also make weights unstable while leaving the fused matrix nearly unchanged.

How This Criterion Differs From Alternatives

Equal weighting is simple and avoids choosing weights to fit a target, but it does not respond to differences in agreement. Kernel alignment between two views, as covered in the earlier material on kernel alignment, describes their pairwise similarity; it need not identify which combination matches a consensus. SNF iteratively exchanges information among patient networks, whereas this criterion selects weights for a weighted kernel using a fixed agreement target. Outcome-guided weighting instead uses an observed response and answers a supervised question; it is not interchangeable with this unsupervised procedure. Product kernels encode joint similarity in a different way and do not provide the same additive weight interpretation.

The method is limited by the target, preprocessing, and objective. Different consensus constructions can produce different weights. Centering and norm scaling can alter which structures count as agreement, while cosine similarity is insensitive to overall scale and does not guarantee that local neighborhoods or biological subgroups are preserved. No outcome labels are used, so agreement alone cannot establish predictive performance. If the scientific aim is prediction, validate against held-out outcomes; if it is network structure, inspect that structure directly rather than relying only on a scalar score.