Why Fusion Parameters Need Sensitivity Analysis
A fused patient-similarity matrix is an output of both the data and a set of analytic choices. Distance bandwidths determine how quickly similarity declines with distance; neighbor counts determine which network edges are retained; diffusion iterations control how far information propagates through a network; kernel weights set the relative contribution of each omics view; and normalization changes the scale on which those views are combined. Two reasonable parameter settings can therefore produce different patient relationships from the same measurements.
Sensitivity analysis asks whether the substantive similarities are stable across a defensible range of settings. It is not a search for the one parameter setting that produces the most appealing network. In an unsupervised analysis, there is usually no known patient-level truth against which to optimize these choices. A useful analysis reports the range examined, the features that persist, and the features that change.
Comparisons require the same patients, patient ordering, input views, and preprocessing. If one setting uses a different set of patients or a different feature-processing pipeline, its matrix difference cannot be attributed to fusion parameters alone. The data and the fusion implementation should be held fixed while the parameter of interest is varied. When several parameters plausibly interact, also examine combinations rather than assuming that each parameter acts independently.
Measuring Change in Patient Similarities
A raw difference between two matrices is useful only when their scales are comparable. If matrices have been put on a common scale, one summary is the relative Frobenius change:
Here \(\theta_0\) is a reference setting. This measure summarizes changes across all matrix entries; it does not identify which patients changed. Because symmetric similarity matrices repeat each off-diagonal entry, the same convention should be used for every comparison. If a method fixes the diagonal, comparing off-diagonal entries alone can make the summary easier to interpret.
A complementary measure is the Spearman rank correlation between the upper-triangle entries of two matrices. It indicates whether pairs that are relatively similar under one setting remain relatively similar under another, even when the numerical scales differ. For local structure, define the \(k\)-nearest-neighbor set \(N_i^{(a)}\) for patient \(i\) under setting \(a\). Its overlap with the set under setting \(b\) is
The mean of \(O_i^{(a,b)}\) across patients summarizes neighborhood stability. These measures answer different questions: a matrix can retain the overall ranking of pairwise similarities while changing a few important local neighborhoods, or vice versa. Report more than one when both global and local structure matter.
Bandwidths and Neighbor Counts
For a Gaussian distance kernel with distance \(d_{ij}\) and bandwidth \(h>0\), define
This formula assumes that the distances are nonnegative and meaningfully comparable within the view. A smaller \(h\) makes similarity decline more rapidly with distance; a larger \(h\) makes similarities more alike. If distance scales differ greatly across views, a shared numerical bandwidth need not represent a shared degree of local smoothing. The bandwidth convention should therefore be fixed and stated, especially when comparing views.
A neighbor count \(k\) is a discrete alternative way to control locality in a network: each patient retains edges to a specified number of nearby patients before any chosen symmetrization or normalization. Smaller \(k\) yields a sparser graph and can isolate local structure, but it can also disconnect the network or make it sensitive to small distance changes. Larger \(k\) retains more edges, which can connect distinct patient groups and weaken local contrast. The exact graph depends on tie handling and whether directed edges are symmetrized by a union or mutual-neighbor rule; keep those choices fixed across the sensitivity comparison.
Worked Example: Changing a Gaussian Bandwidth
Suppose three hypothetical patient pairs have distances \(d=1\), \(d=2\), and \(d=3\) in one omics view. Compare the Gaussian similarities at bandwidths \(h=1\) and \(h=2\). The distance is assumed to be nonnegative and comparable across these pairs, and the same distance definition is used for both settings.
For the pair at distance 1, the similarities are \(K(1)=\exp(-1^2/(2\cdot 1^2))=\exp(-0.5)=0.6065\) and \(K(2)=\exp(-1^2/(2\cdot 2^2))=\exp(-0.125)=0.8825\). For the pair at distance 2, they are \(K(1)=\exp(-2^2/2)=\exp(-2)=0.1353\) and \(K(2)=\exp(-2^2/8)=\exp(-0.5)=0.6065\).
For distance 3, the values are \(K(1)=\exp(-9/2)=0.01111\) and \(K(2)=\exp(-9/8)=0.3247\). Thus increasing the bandwidth raises all three similarities, especially for the more distant pairs. This reduces contrast between near and far pairs; it does not establish that the distant patients are biologically similar. The relevant sensitivity question is whether conclusions about patient neighborhoods or groups persist over a defensible bandwidth range.
Worked Example: Changing a Neighbor Count
Consider four hypothetical patients A, B, C, and D with distances from A given by \(d_{AB}=0.2\), \(d_{AC}=0.4\), and \(d_{AD}=0.9\). Assume ties do not occur and use directed \(k\)-nearest neighbors followed by union symmetrization: an undirected edge is retained if either patient selects the other. For this illustration, the other pairwise distances are \(d_{BC}=0.3\), \(d_{BD}=0.7\), and \(d_{CD}=0.5\).
At \(k=1\), the nearest choices are A to B, B to A, C to B, and D to C. After union symmetrization, the retained edges are AB, BC, and CD. The graph is a chain, and A and D are not direct neighbors.
At \(k=2\), A selects B and C; B selects A and C; C selects B and A; and D selects C and B. The retained edges are AB, AC, BC, BD, and CD. The edge set has gained AC and BD, and the graph is more connected. The change follows from the stated distances and symmetrization rule. It illustrates why a result that depends on a single edge should be checked across plausible \(k\) values rather than treated as fixed evidence of a patient relationship.
Diffusion Iterations
Diffusion propagates similarity through paths in a patient network. In a simple random-walk description, let \(P\) be a row-stochastic transition matrix: \(P_{ij}\geq 0\) and \(\sum_j P_{ij}=1\). The entry \((P^t)_{ij}\) is the probability of being at patient \(j\) after \(t\) transitions starting from patient \(i\). Increasing \(t\) allows information to travel along longer paths. This interpretation assumes the transition matrix is fixed and describes the network being diffused; an implementation that updates multiple view-specific networks may use a different iteration formula. Its formula must be held fixed when assessing iteration sensitivity.
Worked Example: Two and Three Diffusion Steps
For three hypothetical patients, use the row-stochastic transition matrix \(P=\begin{pmatrix}0.5&0.5&0\\0.25&0.5&0.25\\0&0.5&0.5\end{pmatrix}\). Each row sums to 1 and all entries are nonnegative, so it satisfies the stated transition-matrix conditions.
After two steps, \(P^2=P P\). Its first row is \(0.5(0.5,0.5,0)+0.5(0.25,0.5,0.25)=(0.375,0.5,0.125)\); the second row is \((0.25,0.5,0.25)\); and the third is \((0.125,0.5,0.375)\). Thus the two-step transition probability from patient 1 to patient 3 is 0.125, despite the absence of a direct transition between them in \(P\).
Multiplying once more, the first row of \(P^3\) is \(0.5(0.375,0.5,0.125)+0.5(0.25,0.5,0.25)=(0.3125,0.5,0.1875)\). The probability of reaching patient 3 from patient 1 has risen from 0.125 at two steps to 0.1875 at three. More iterations can spread information across a network, but may also make distinctions between local neighborhoods less pronounced. These probabilities illustrate diffusion; they are not, by themselves, estimates of clinical risk.
Weights and Normalization Choices
For two kernel matrices \(K^{(1)}\) and \(K^{(2)}\), a weighted additive fusion is \(F=w_1K^{(1)}+w_2K^{(2)}\), with nonnegative weights summing to 1. This constraint makes the weights interpretable as relative contributions only when the input matrices are on comparable scales. A high-scale matrix can dominate the sum even if its weight is not larger. Normalization therefore affects not just numerical presentation but potentially which view drives patient similarities.
Worked Example: Normalization Changes the Relative Contribution
Take hypothetical kernels \(K^{(1)}=\begin{pmatrix}4&2\\2&4\end{pmatrix}\) and \(K^{(2)}=\begin{pmatrix}1&0.8\\0.8&1\end{pmatrix}\), with weights \(w_1=w_2=0.5\). Assume the matrices are used as given and compare their off-diagonal similarities after fusion.
Without normalization, the fused diagonal is \(0.5(4)+0.5(1)=2.5\), and the off-diagonal is \(0.5(2)+0.5(0.8)=1.4\). The off-diagonal relative to the diagonal is \(1.4/2.5=0.5600\).
Now normalize each matrix to unit diagonal by dividing by its diagonal value. The first matrix has off-diagonal \(2/4=0.5\), while the second has off-diagonal \(0.8/1=0.8\). Their equally weighted fusion has diagonal 1 and off-diagonal \(0.5(0.5)+0.5(0.8)=0.6500\). The relative off-diagonal has changed from 0.5600 to 0.6500. The weight labels remained equal, but normalization changed the fused similarity because it changed the scale of each view. Both normalizations are calculable; neither is automatically the scientifically correct choice without regard to the analysis goal.
Other common choices include centering, unit-trace scaling, and unit-Frobenius scaling. They can preserve different aspects of a matrix and are not interchangeable. Record which operation is applied, in what order, and whether it is applied separately to each view or to the fused matrix. If the aim is to interpret weights as relative view contributions, compare results under reasonable normalization choices and inspect each view's contribution on the resulting scale.
A Practical Sensitivity Workflow
Choose a reference configuration and ranges for bandwidths, neighbor counts, iterations, weights, and normalization. State the distance, graph, and update conventions used.
First change one parameter at a time to see its direct effect. Then examine selected combinations, because bandwidth, sparsity, and diffusion can interact.
Use relative matrix differences only for comparable scales; also compare ranks of pairwise similarities and patient-neighbor overlap.
Report which patient pairs or local neighborhoods persist, which depend on parameter choices, and the settings under which changes occur.
Diagnostics, Alternatives, and Limitations
Several checks help identify a sensitivity analysis that is misleading. Confirm that every fused matrix uses the same patient order and that the matrix is finite and has the expected dimensions. Check nonnegativity or symmetry if the chosen construction requires them, and verify stochastic row sums if the matrix is interpreted as a transition matrix. For a sparse network, inspect the number of connected components and isolated patients as \(k\) changes. For kernels, inspect per-view scales before and after normalization. A sudden structural change at a small parameter adjustment may indicate a threshold effect, a tie, a disconnected graph, or a coding error; it should be investigated rather than averaged away.
A one-at-a-time sweep is easy to interpret but can miss interactions: a bandwidth may be stable at one neighbor count and unstable at another. A factorial grid over a small set of defensible values reveals such interactions but can require many computations. Stability summaries are another alternative to a single selected setting, but they do not identify which setting is biologically correct. Choosing parameters by downstream clustering quality can be useful when an appropriate external validation target exists; without such a target, selecting the setting that gives the clearest-looking clusters risks circular reasoning.
Sensitivity is not statistical significance, and stability is not proof that a similarity is true or causal. The chosen parameter range can omit important alternatives; the distance itself may be inappropriate; and all settings may share the same preprocessing bias or technical artifact. Results should therefore describe robustness to the choices actually varied, not claim robustness in general. The introductory material on kernel matrices and patient networks supplies the relevant similarity-matrix context; the focus here is how design choices change those similarities.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- For \(d=1.5\), calculate the Gaussian similarity at bandwidths \(h=1\) and \(h=2\), using \(K(d)=\exp(-d^2/(2h^2))\). Which is larger?
- A patient has neighbors \(\{A,B,C\}\) under one setting and \(\{B,C,D\}\) under another, with \(k=3\). Calculate the neighbor-set overlap.
- Why can equal kernel weights fail to give equal influence to two views?
- What does a rise in a multi-step transition probability between two patients establish, and what does it not establish?
Solutions. (1) At \(h=1\), \(K=\exp(-1.5^2/2)=\exp(-1.125)=0.3247\). At \(h=2\), \(K=\exp(-2.25/8)=\exp(-0.28125)=0.7548\); the larger bandwidth gives the larger similarity. (2) The intersection is \(\{B,C\}\), so overlap is \(2/3=0.6667\). (3) If the matrices have different scales, the same numerical weight multiplies entries of different magnitudes; normalization can therefore change relative influence. (4) It establishes that the specified transition process permits more probability to travel between the patients in the stated number of steps. It does not establish biological causation, clinical risk, or that the patients are truly similar under an external standard.