Dimensionality reduction for visualizing single-cell data using UMAP.
Summary (2 min read)
Introduction
- The optimal power flow (OPF) problem optimizes certain objective such as power loss and generation cost subject to power flow equations and operational constraints.
- Recently, convex relaxations of the OPF problem have been proposed.
- SDP/SOCP relaxation is not always exact, especially when the underlying network is not radial.
- In [14], the authors perturbed the objective function, and this technique is guaranteed to work in some cases, but in general case the authors still do not have a rank one feasible solution.
B. OPF and SDP relaxation
- The OPF problem seeks to optimize certain objective, e.g. total line loss, or generation cost, subject to power flow equations (1) and various operational constraints.
- The voltage magnitude at each load bus i ∈ N needs to be maintained within a prescribed region, i.e. V mini ≤ |Vi| ≤ V maxi .
- Notice that this optimization is non-convex because of the rank constraint rankW ≤.
III. AN ADMM HEURISTIC FOR THE OPF
- The authors apply the ADMM method to derive a heuristic for the nonconvex OPF problem (5).
- The rank constraint helps us to come up with the tractable non-convex minimization, hence the ADMM provides a sequence of convex program which approximately solves the original non-convex OPF.
- Another important property of their algorithm is that if the initial iterates Z0,Λ0 are Hermitian matrices, then all W k, Zk,Λk are Hermitian matrices: Proposition 3.
A. Feasible point
- The convergence of the ADMM heuristic for a non-convex problem is still an open question [16].
- If it converges, then the authors are guaranteed to have a rank one feasible point of the OPF.
- Finally, since W ∗ ∈ C, the authors can conclude that W ∗ is a rank one feasible point in the optimal power flow problem.
B. Stopping criterion and ρ
- For the stopping criterion, the authors use the one from [16].
- When the algorithm converges, Rk and Sk should be zero.
- This gives rise to the following stopping criterion: ‖Rk‖F ≤ pri, ‖Sk‖F ≤ dual.
- Lastly, the choice of ρ can be automated based on these residuals.
C. Overall algorithm
- Notice that when ρ = 0, the optimization (11) is a SDP relaxation of the OPF.
- This helps us not to get trapped, but the price the authors pay is a possible oscillation.
A. Two bus network
- It is shown in [3] that the feasible region becomes the two disjoint regions under some reactive power constraints on q1, q2, as shown by the black lines on the ellipse.
- Hence, a feasible power injection (rank 1 solution) is received in this simple network.
- Here the mesh network is a ring with 10 nodes and 10 links.
- Indeed, when their heuristic converges, it recovers the rank one solution although the SDP relaxation always generate a full rank solution.
V. CONCLUSION
- The authors propose a non-convex ADMM heuristic for the OPF.
- By introducing a redundant variable whose rank is one, the authors can split the minimization into two steps, where the first step is a convex optimization, and the second step is a rank constrained minimization.
- Then, the authors show that the second step, a non-convex optimization, can be carried out analytically.
- Moreover, the authors observe the convergence of their heuristic under the existence of hidden rank one solution in the SDP relaxation of the OPF.
- Inspired by this, the convergence proof under this 2Minimum eigenvalue of the solution from the SDP relaxation is greater than 0.01 in all cases.
Did you find this useful? Give us your feedback
Citations
5,390 citations
Cites methods from "Dimensionality reduction for visual..."
...Based upon preliminary releases of a soware implementation, UMAP has already found widespread use in the elds of bioinformatics [5, 12, 17, 46, 2, 45, 15], materials science [34, 23], and machine learning [14, 20, 21, 24, 19, 47] among others....
[...]
2,459 citations
1,809 citations
Cites methods from "Dimensionality reduction for visual..."
...The cell scatter plots were obtained using the UMAP method [13]....
[...]
1,180 citations
Cites background from "Dimensionality reduction for visual..."
...What sets UMAP apart in this comparison is its speed and ability to scale to large numbers of cells (Becht et al, 2018)....
[...]
830 citations
Cites methods from "Dimensionality reduction for visual..."
...Overall, the two-dimensional projection in Figure 2D provides a way to visualize the relationships among antibodies in the space of immune-escape mutations, similar to how dimensionality reduction techniques such as tSNE or UMAP help visualize high-dimensional single-cell transcriptomic data (Amir et al., 2013; Becht et al., 2018)....
[...]
References
30,124 citations
13,652 citations
7,741 citations
5,390 citations
4,141 citations
Related Papers (5)
Frequently Asked Questions (7)
Q2. What led to myeloid dendritic cells on one branch?
GMP then led to classical monocytes that further led to myeloid dendritic cells on one branch and to cells labeled as intermediate monocytes on another branch.
Q3. What is the common technique used in cytometry?
t-SNE is currently the most commonly-used technique and is efficient at highlighting local structure in the data, which for cytometry notably translates to the representation of cell populations as distinct clusters.
Q4. What was the density of events in the t-SNE space?
The density of events in the dimensionallyreduced space also appeared less uniform in t-SNE, with large clusters in the t-SNE space being less dense than the smaller ones.
Q5. What was the main characteristic of the datasets used?
For then Han et al. dataset the authors used the AUCell R package[15] , which computes the AUC of gene sets within each single cell, using gene sets from the Haemopedia[12] resource to annotate cell lineages.
Q6. What do they claim to be the advantages of t-SNE?
They claim that compared to t-SNE it preserves as much of the local and more of the global data structure, with a shorter runtime.
Q7. How long did UMAP take to intepret?
In addition, UMAP appeared more stable than t-SNE, being more consistent across distinct replicates and independent subsampling which should facilitate consistency in its intepretation (Figure S3).