USQuery is a website designed to make congressional data more accessible. Here you can find information on how members of Congress have voted in present and past congresses. Detailed graphics provide easy to understand results of how many votes a bill had, and the distribution of votes by party and state. Each member of congress has their own page where you can see the list of votes they made for a specific Congress, as well as history of their past roles in Congress and general information about their office and contacts. Bills have their own page, showing the timeline of actions on said bill as well as information on what subjects the bill is related to, who sponsored/cosponsored the bill, and a summary of the contents of the bill.
Email: usquery.help@gmail.comData is collected from Congress API, and the official House and Senate websites. Some content is generated through the use of AI, these are marked with a yellow warning in order to be transparent of where content is AI generated. We track all voting members of US Congress, meaning that unfortunately, as of now any non voting members like representatives from US territories are not tracked. We plan to integrate these members in the near future.
Predictions on how many votes a bill can receive are made from a three step process; First, bill text is processed by collecting provided bill subjects from the Congress API, and then quantifying how liberal or conservative the bill is for each subject. Next, the quantified features are passed through a pre-trained neural network, outputting the probability for a senator or representative to vote "Yes" on the bill. Finally, using the probabilities, we run 1000 simulated votes and plot the distribution of total "Yes" votes for the House and Senate
Our model aims to classify whether a member of the House or Senate will vote "Yes" or "No" on a given bill. For the purpose of predicting if a bill will pass, both "Present" and "No Vote" votes are classified as a "No" vote. The scope of our training data is all house and senate votes from 2024 up until 2014. Due to limitations on token counts, some bills that had text files that were too long were omitted from the training data.
Using the raw bill text and provided bill subjects from Congress API, we quantify the feature values for each subject on a scale from very conservative (-20) to very liberal (20). This classification is done through a prompt given to Gemini 2.0, providing the raw bill text and subject list. There are over 500 subject columns in our training data. Each row of training data consists of both bill features and specific member features. Other columns include whether the member is part of the house or the senate, what state the member is from, and what party the member is from. Through rigorous testing of different model architectures, the best results we found were a testing accuracy of 87% and a loss of 0.3, a good loss value for a binary classification model. Further improvements will be made in the future, with plans to generate more quality data, and with a further scope in the data used.
Predictions give the user a distribution of 1000 sample votes, where each sample is a prediction of how many total votes the bill may get. Predictions are grouped by Senate/House membership, party, and state, and then a binomial random variable is produced using the probability of a yes vote given by the pre-trained mode. For each sample, all Senator and House member votes are predicted by the binomial random variables. The distribution of counts of total votes is then plotted, one plot for the Senate and one plot for the House.