🤖 AI

Privacy-Aware Predictions in Participatory Budgeting

This paper proposes a privacy-preserving approach that utilizes project textual descriptions and anonymous historical voting records to predict funding outcomes for participatory budgeting proposals, thereby assisting organizers in managing large volumes of submissions without relying on personally identifiable information.

Original authors: Juan Zambrano, Clément Contet, Jairo Gudiño-Rosero, Felipe Garrido-Lucero, Umberto Grandi, César Hidalgo

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Juan Zambrano, Clément Contet, Jairo Gudiño-Rosero, Felipe Garrido-Lucero, Umberto Grandi, César Hidalgo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a city council has a big jar of money and wants to spend it on things the community actually wants, like new parks, better bus stops, or sports courts. This is called Participatory Budgeting. Citizens propose ideas, and then everyone votes on them.

However, there's a problem: In big cities, people might submit thousands of ideas. Imagine a city like Toulouse or Wrocław receiving over 1,000 project proposals. The city organizers can't possibly read every single one, understand the details, and figure out which ones are popular before the voting even starts. Usually, a small group of volunteers has to manually sort through this mountain of paperwork, which is slow, expensive, and sometimes a bit mysterious to the public.

This paper is about building a smart, privacy-friendly assistant to help city organizers sort through these proposals before the voting begins.

The Goal: A "Crystal Ball" for Project Popularity

The researchers wanted to create a tool that could look at a new project proposal and guess: "How many votes will this get?"

If the tool can predict which projects are likely to be popular, organizers can use that information to:

Filter the noise: Focus their attention on the most promising ideas.
Be transparent: Show the public exactly how they decided which projects to move forward with, rather than just saying "a committee decided."
Protect privacy: Crucially, they wanted to do this without looking at who voted for what or using personal data about the voters (like their age, income, or address). They only wanted to use the text of the project descriptions and past election results.

How They Tested It

The researchers acted like detectives, looking at two real cities: Toulouse, France and Wrocław, Poland.

They took data from an old election in each city (the "training" data).
They tried to predict the results of a newer election in the same city (the "test" data).
They used two different types of "brains" to make these predictions:
1. Classic Math Models: Traditional computer algorithms (like ElasticNet and XGBoost) that look for patterns in numbers and words.
2. Large Language Models (LLMs): Advanced AI chatbots (like GPT-4 and Llama) that are usually used for writing stories or answering questions.

The "Three Ways" to Ask the AI

The researchers tested three different ways to ask the AI to make a prediction, similar to how you might ask a friend for advice:

The "Zero-Shot" (The Guess): You ask the AI, "Here is a project description. How many votes will it get?" without giving it any history.
- Result: The AI was terrible at this. It was like asking someone to guess the weather in a city they've never visited without a forecast.
The "Full Context" (The List): You give the AI a list of all the projects from the previous year, but you strip away the long descriptions to save space. You just give it the names, costs, and categories.
- Result: This was okay, but not great. The AI missed the nuance of what the projects actually were.
The "RAG" (The Smart Researcher): This is the winner. You tell the AI: "Here is the new project. Now, look at the most similar projects from last year (based on their descriptions) and tell me how those did."
- Result: This worked surprisingly well. It's like asking a local expert, "I'm thinking of building a new park. Last year, we built a similar park in a similar neighborhood, and it got 5,000 votes. Based on that, how will this new one do?"

What They Found

Text Matters: The AI needed to read the actual descriptions of the projects to do a good job. If you just gave it the cost and the category (e.g., "Sports," "$10,000"), it couldn't predict the votes well. The words people use to describe their ideas hold the secret to their popularity.
Ranking is Easier than Counting: The AI wasn't perfect at guessing the exact number of votes (e.g., "It will get 4,321 votes"). However, it was very good at ranking. It could reliably say, "Project A is more popular than Project B."
AI vs. Humans: The AI's performance was comparable to the classic math models. It didn't magically beat them, but it could do the job just as well.
Privacy First: The system worked perfectly without ever needing to know who the voters were. It only needed the history of what won in the past.

The Bottom Line

The paper concludes that we can use AI to help city organizers manage the "avalanche" of project proposals in participatory budgeting.

Think of it like a pre-screening filter. The AI doesn't make the final decision or replace the voters. Instead, it acts as a helpful assistant that says, "Out of these 1,000 ideas, here are the top 300 that look like they will be the most popular based on history and description." This allows human organizers to focus their time on the most viable projects, making the whole democratic process faster, cheaper, and more transparent, all while keeping voter privacy intact.

Important Caveat from the Paper:
The authors warn that this tool shouldn't be used to replace the voting process or to make final decisions automatically. It's a support tool for the "shortlisting" phase. Also, they note that if people start writing their proposals specifically to "trick" the AI, the system might stop working as well. It's a tool to help humans, not to replace them.

Technical Summary: Privacy-Aware Predictions in Participatory Budgeting

Problem Statement
Participatory budgeting (PB) is a democratic process where citizens propose and vote on public investment projects. While computational social choice research has extensively focused on optimizing the voting phase (e.g., designing selection rules), the consolidation phase—where organizers must shortlist a large volume of proposals before voting—remains largely opaque and often relies on manual committee decisions without clear objectives. This phase poses significant scalability challenges due to the high volume of textual data (project titles and descriptions). Furthermore, existing AI approaches in this domain often rely on voter demographics or personally identifiable information (PII) to create "voter personas," raising privacy concerns and limiting applicability in contexts where such data is unavailable or legally restricted.

Methodology
The authors propose a privacy-preserving framework to predict project success (measured by vote counts) using only publicly available data: project textual descriptions and anonymous historical voting records. The study utilizes two multi-year datasets from European cities: Toulouse, France (2022 and 2024 campaigns) and Wroclaw, Poland (2016 and 2017 campaigns). These datasets were constructed by merging voting data from the Pabulib.org repository with project descriptions scraped from municipal websites.

The methodology compares two distinct approaches:

Classical Machine Learning (ML):
- Models: ElasticNet (regression) and XGBoost (random forests).
- Features: Tabular data (district, category, cost) combined with three textual preprocessing strategies: TF-IDF, Sentence-BERT (SBERT), and BGE-M3.
- Training/Evaluation: Models were trained on the earliest available campaign in each city and evaluated on the most recent one. Principal Component Analysis (PCA) was applied to reduce dimensionality.
- Baseline: A "No-Text" configuration trained solely on tabular data.
Large Language Models (LLMs):
- Models: GPT-4 Turbo and Llama 3.3 70B.
- Prompting Strategies: Three variants were tested to provide progressively richer context:
  - Zero-Shot: No historical data provided.
  - Full Context: List of previous projects (titles, categories, costs, districts, votes) without textual descriptions due to context window limits.
  - Retrieval-Augmented Generation (RAG): The model receives the target project plus the 20% most semantically similar projects from the historical dataset (based on cosine similarity), including their full descriptions and vote counts.
- Privacy Safeguards: The authors verified that the LLMs had no prior exposure to the specific voting results or full project descriptions of the test campaigns, ensuring predictions were not based on memorized training data.

Key Contributions

Datasets: The construction and release of two new multi-year PB datasets merging voting outcomes with rich textual descriptions for Toulouse and Wroclaw.
Privacy-Preserving Prediction: A demonstration that accurate project ranking can be achieved without using voter demographics or PII, relying solely on public project data and historical election outcomes.
Comparative Analysis: A systematic evaluation of classical ML models versus LLM prompting strategies (Zero-shot, Full Context, RAG) for predicting vote counts and ordinal rankings.
Shortlisting Utility: An analysis of how these predictors perform specifically for the "shortlisting" task (identifying top-k projects likely to pass a greedy selection rule), rather than just exact vote count prediction.

Results

Performance vs. Baseline: All proposed models significantly outperformed a random baseline (Impartial Culture).
Ordinal Ranking vs. Exact Counts: LLMs demonstrated higher accuracy in predicting the ordinal ranking of projects (Kendall's $\tau$ ) than in predicting exact vote counts (Mean Absolute Error).
LLM vs. ML: LLM-based approaches, particularly with RAG, achieved performance comparable to, and in some cases exceeding, the best classical ML models.
- In Toulouse, GPT-4 Turbo (RAG) achieved a Kendall's $\tau$ of 0.35, comparable to the best ElasticNet variant (0.36).
- In Wroclaw, Llama 3.3 (RAG) achieved a $\tau$ of 0.51, outperforming the best XGBoost variant (0.44).
Context Importance: Zero-shot prompts performed poorly, confirming that prior knowledge of the city is insufficient without specific historical data. RAG consistently yielded the most stable and accurate results among LLM strategies.
Shortlisting Efficacy: When evaluating the overlap of the top 30% of predicted projects with the actual funded projects (using the Jaccard index), the models showed substantial utility. For Wroclaw, LLM RAG variants reached a Jaccard index of nearly 0.5, indicating strong alignment with the projects that would pass a greedy selection rule.
Data Efficiency: The inclusion of textual descriptions was essential for outperforming random baselines and tabular-only models.

Significance and Claims
The paper claims that its methodology offers a practical, privacy-aware support tool for PB organizers to manage large volumes of proposals during the consolidation phase. By relying only on publicly available data, the approach builds trust and encourages participation.

The authors emphasize that their tool is designed to assist organizers, not replace the voting process. They explicitly state that LLMs should not be used as "out-of-the-box" shortlisting tools without technical setup (e.g., RAG implementation) and that the output should serve as an initial ranking for human review. The study highlights that historical election data from the same city can effectively replace demographic inputs for prediction, addressing a key limitation in previous "voter persona" research.

Finally, the authors acknowledge limitations, including the potential for strategic manipulation of project descriptions, the "black box" nature of LLMs, and the risk of cost bias (where higher-cost projects are favored). They conclude that while the approach shows promise for digital democracy, it requires further scrutiny and should be deployed as part of an AI-assisted, human-in-the-loop process.