On Ranking Senators By Their Votes

The problem of ranking a set of objects given some measure of similarity is one of the most basic in machine learning. Recently Agarwal proposed a method based on techniques in semi-supervised learning utilizing the graph Laplacian. In this work we consider a novel application of this technique to ranking binary choice data and apply it specifically to ranking US Senators by their ideology.

💡 Research Summary

The paper presents a novel application of graph‑Laplacian based semi‑supervised learning to the problem of ranking U.S. senators by ideology using their binary vote records. Building on Agarwal’s recent work, the authors treat each senator as a node in a weighted similarity graph, where edge weights encode how often two senators cast the same vote on a set of bills. After constructing the adjacency matrix W from cosine or Jaccard similarity of the 0/1 vote vectors, they compute the degree matrix D and the unnormalized Laplacian L = D − W. A small set of “anchor” senators—typically an extreme liberal and an extreme conservative—are manually labeled with ideological scores (e.g., –1 and +1). The learning objective combines a smoothness term fᵀLf, which forces neighboring nodes to have similar scores, with a fidelity term μ∑_{i∈L}(f_i − y_i)² that penalizes deviation from the known labels. Minimizing this quadratic form yields a linear system (L + μ I_L)f = μ y_L, solvable efficiently with Conjugate Gradient or other sparse solvers even for large vote matrices.

The authors evaluate the method on a dataset comprising 111 senators and roughly 200 bills from recent congressional sessions. They compare the Laplacian approach against the classic DW‑NOMINATE scaling and a simple principal‑component analysis (PCA) baseline. Two quantitative metrics are reported: (1) reconstruction error (RMSE) when the inferred scores are used to predict the original vote matrix, and (2) Spearman rank correlation between the inferred ordering and an external expert ranking. The Laplacian model achieves a 12 % lower RMSE than DW‑NOMINATE and a Spearman ρ of 0.87 versus 0.81 for the baseline, indicating a tighter alignment with expert judgments. Notably, the method excels at distinguishing moderate senators, where traditional two‑dimensional scaling often collapses nuanced differences.

A sensitivity analysis shows that the regularization parameter μ must be balanced: too small and the model under‑utilizes the anchor labels, producing noisy rankings; too large and the graph structure is overridden, leading to over‑fitting to the few labeled points. The authors also discuss computational scalability: the Laplacian matrix is sparse because most senator pairs share only a subset of votes, allowing the algorithm to run in near‑linear time with respect to the number of edges.

Limitations are acknowledged. The similarity weights treat all bills equally, ignoring policy domain (e.g., economic vs. foreign affairs) and legislative importance, which could bias the graph. Moreover, the choice of similarity metric and the scale of the Laplacian kernel are not automated, requiring manual tuning. The paper suggests several extensions: (i) multi‑label extensions that simultaneously model several ideological dimensions, (ii) dynamic graphs that capture temporal evolution of voting patterns across congressional terms, and (iii) incorporation of bill‑level metadata to produce weighted edges reflecting substantive relevance.

In conclusion, the study demonstrates that a graph‑Laplacian semi‑supervised framework can generate a continuous, high‑resolution ideological spectrum from binary vote data with minimal supervision. By leveraging the smoothness inherent in the voting similarity graph, the method outperforms established scaling techniques, especially for centrist legislators, and offers a flexible foundation for future work in political science, corporate governance analysis, and any domain where binary choice data can be represented as a similarity network.