🔬 physics

Recurrent patterns of user behavior in different electoral campaigns: A Twitter analysis of the Spanish general elections of 2015 and 2016

By analyzing millions of tweets from Spain's 2015 and 2016 general elections, the study reveals that collective user behaviors, including activity correlations, power-law distributions, and network topologies, exhibit robust and recurrent patterns across different electoral campaigns.

Original authors: Samuel Martin-Gutierrez, Juan C. Losada, Rosa M. Benito

Published 2026-02-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Samuel Martin-Gutierrez, Juan C. Losada, Rosa M. Benito

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 two massive, chaotic concerts happening in the same city, just six months apart. The first concert is the Spanish General Election in December 2015, and the second is the "encore" in June 2016. The bands (political parties) changed their setlists slightly, and the crowd's energy shifted, but the venue (Twitter) and the general vibe remained surprisingly similar.

This paper is like a scientific study of the crowd's behavior at these two concerts. The researchers didn't just listen to the music; they analyzed millions of text messages (tweets) to see how people interacted, shouted, and shared information.

Here is the breakdown of their findings, translated into everyday language:

1. The "Crowd Pulse" is Predictable

The researchers looked at the daily activity of Twitter users. They found that the "heartbeat" of the conversation was almost identical in both elections.

The Analogy: Think of the election campaign like a rollercoaster ride. The activity starts low, slowly climbs as the campaign heats up, hits a massive, screaming peak on Election Day, and then drops off.
The Finding: Even though the specific politicians were different, the shape of the rollercoaster was the same. If you plotted the activity of the 2015 crowd against the 2016 crowd, they lined up perfectly. It suggests that human behavior during elections follows a very strict, repeatable script.

2. The "Super-Connector" Rule (Power Laws)

The study looked at how the number of messages relates to the number of people talking.

The Analogy: Imagine a party. If you invite 10 people, you get 10 conversations. If you invite 100 people, do you get 100 conversations? No, you get way more because people start talking to each other, not just the host.
The Finding: The number of tweets, retweets (sharing), and mentions (tagging) follows a "Power Law." This means that as more people join the conversation, the activity explodes disproportionately. In 2015, adding more people caused a massive explosion of activity. In 2016, it was a bit more linear, but the rule still held: the more people involved, the louder the room gets, and it gets louder faster than you'd expect.

3. The "Echo Chamber" Architecture

The researchers mapped out who was talking to whom, creating a giant map of connections.

The Analogy: Imagine a room full of people wearing different colored shirts (representing different political parties).
The Finding: The map showed that people mostly talked to others wearing the same color shirt.
- Mentions: People would shout across the room to their own team.
- Retweets: This was even stronger. People almost never shared a message from the "other team." It was like a series of isolated echo chambers.
- The Twist: When two parties (Podemos and IU) decided to form a coalition (a team-up) for the 2016 election, the "wall" between their colored shirts disappeared. They started talking and sharing with each other much more. This proved that the network structure is flexible enough to change when political alliances change, but the "us vs. them" pattern remains the default.

4. The "Megaphone" Effect

The study measured "efficiency"—how many people a single person could reach with one message.

The Analogy: Imagine a regular person shouting in a crowd versus a politician with a megaphone.
The Finding: Politicians are naturally more "efficient." Because they have huge followings, one tweet from them ripples through the network much further than one from a regular user. However, the pattern of how this efficiency works is the same for everyone. Whether you are a regular user or a politician, the distribution of influence follows the same mathematical curve.

5. The "Silence" Before the Storm

In Spain, there is a law called "Electoral Silence" (the day before the election where campaigning is banned).

The Analogy: It's like the moment the DJ stops the music right before the final song.
The Finding: On this day, the volume of conversation dropped significantly. Politicians stopped shouting, and because the "leaders" went quiet, the regular crowd quieted down too. This shows how much the crowd relies on the politicians to drive the conversation.

The Big Takeaway

The main conclusion of the paper is that human behavior in elections is surprisingly robotic.

Even though the political landscape changed (new parties, coalitions, different leaders), the underlying "software" of how people interact on Twitter didn't change. The patterns of shouting, sharing, and ignoring the opposition were so consistent that the researchers could predict how the 2016 crowd would behave just by looking at the 2015 data.

It suggests that while politics is messy and emotional, the way we communicate about it online follows strict, repeatable laws of physics.

1. Problem Statement and Objective

The paper addresses the challenge of understanding collective user behavior in social media during political processes. While previous studies have focused on predicting election outcomes or analyzing specific user interactions, this research aims to identify recurrent behavioral patterns across different electoral campaigns.

The authors leverage the unique opportunity provided by the Spanish general elections of December 2015 and June 2016. These two events occurred in rapid succession with a similar socio-political context but involved different political actors (notably, the formation of the "Unidos Podemos" coalition in 2016). The objective is to determine if the underlying dynamics of Twitter interactions are robust to changes in the social and political landscape or if they are highly sensitive to specific campaign contexts.

2. Methodology

Data Collection

Source: Twitter Streaming API.
Keywords: Neutral keywords were used to avoid bias (e.g., 20D, 20D2015, #EleccionesGenerales2015 for 2015; 26J, 26J2016, #EleccionesGenerales2016 for 2016).
Scope: Data was collected for over two months surrounding each election, with a core analysis period of 18 days (15 days of the official campaign + reflection day + election day + day after).
Volume: Approximately 1.79 million tweets (2015) and 1.75 million tweets (2016).
Political Accounts: Specific lists of accounts for the main parties (PP, PSOE, Podemos, Cs, and IU) were compiled to distinguish between regular users and political elites.

Network Construction

Two types of directed, weighted networks were constructed:

Mention Networks: A directed link from user $j$ to $i$ exists if $j$ mentions $i$ (includes @username).
Retweet Networks: A directed link from $j$ $j$ to $i$ $i$ exists if $j$ $j$ retweets an original post by $i$ $i$ .
- Note: The methodology links retweeters directly to the original poster, ignoring intermediate retweeters, based on API metadata limitations at the time.

Analytical Techniques

Time Series Analysis: Daily activity counts (tweets, retweets, mentions) were correlated between the two elections.
Power Law Fitting: The relationship between the number of interactions (tweets, retweets, mentions) and the number of unique users was fitted to a power law ( $E(t) \propto N(t)^\alpha$ ).
Distribution Analysis: User activity distributions were fitted to discrete power laws using Maximum Likelihood Estimation (MLE) to determine the exponent $\gamma$ .
Network Topology:
- Community Detection: Louvain algorithm to identify political clusters.
- Degree Distributions: Analysis of in-degree ( $k_{in}$ ) and out-degree ( $k_{out}$ ) distributions and their exponents.
- Assortativity: Calculation of degree assortativity and assortative mixing ( $r_m$ ) to measure homophily (preference for connecting with similar nodes).
Efficiency Metric: Defined as $\eta = \text{RTs} / \text{Posts}$ , measuring the collective response to an individual's output.

3. Key Contributions

Cross-Election Correlation: The study provides empirical evidence of a significant linear correlation ( $r^2 = 0.99$ ) between the daily activity time series of two distinct but temporally close elections, suggesting a high degree of predictability in collective behavior regardless of specific political actors.
Power Law Dynamics: It establishes that daily interaction volumes (tweets, retweets, mentions) follow a power law with respect to the number of unique users. The exponent $\alpha$ shifted from slightly super-linear ( $\approx 1.2$ ) in 2015 to approximately linear ( $\approx 1.0$ ) in 2016, indicating a change in how user growth translates to activity volume.
Robustness of Network Topology: The study demonstrates that the topological properties of mention and retweet networks (degree distributions, assortativity patterns) remain consistent across elections, even when the political landscape changes (e.g., coalition formation).
Coalition Detection: The methodology successfully detected the formation of the "Unidos Podemos" coalition through network analysis. The assortative mixing metric showed a distinct increase in interaction between Podemos and IU in 2016 compared to 2015, validating the approach for detecting political alliances.
Elite vs. Regular User Efficiency: The paper quantifies that political accounts are significantly more efficient at propagating messages (higher $\eta$ ) than regular users, a pattern consistent across both elections and all parties.

4. Key Results

Temporal Evolution: Both elections showed a similar temporal profile: low activity before the campaign, a rise during the campaign, a peak on election day (accounting for ~30-36% of total messages), and a drop on "reflection day" (electoral silence).
Power Law Exponents:
- Activity vs. Users: In 2015, activity grew super-linearly with users; in 2016, it was linear.
- User Activity Distribution: The exponent $\gamma$ for daily user activity fluctuated within a narrow range ( $2.0 < \gamma < 2.5$ ) for both elections, indicating stable collective behavior.
Network Structure:
- Communities: Networks naturally clustered by political party.
- Lack of Debate: There was a near-total lack of cross-party communication (retweets) and very low cross-party mentions. The diagonal of the adjacency matrix (self-links) was orders of magnitude heavier than off-diagonal links.
- Assortativity: Networks were highly assortative (users connected to their own party). The "coalition partition" in 2016 showed higher assortativity than the "party partition," confirming the tight integration of the Podemos-IU coalition.
Efficiency: Political accounts consistently showed a higher probability of high efficiency ( $\eta > 1$ ) compared to the general user base.
Disruptive Events: Fluctuations in degree assortativity were observed to correlate with specific events (e.g., debates, viral videos), showing that while the structure is robust, daily dynamics are sensitive to exogenous shocks.

5. Significance and Implications

Universality of Behavioral Patterns: The findings suggest that the "physics" of political communication on Twitter is governed by robust, recurrent patterns that persist despite changes in political actors or specific campaign narratives. This implies that social media dynamics in elections may be predictable based on structural properties rather than just content.
Methodological Validation: The study validates the use of network metrics (specifically assortative mixing and efficiency distributions) as tools for detecting political phenomena, such as the formation of coalitions, purely from interaction data.
Political Insight: The results confirm a "lack of debate" in the Spanish political sphere on Twitter, where parties operate in echo chambers. However, the formation of a coalition was the only significant structural change detected, highlighting the sensitivity of these metrics to formal political alliances.
Future Applications: The authors propose that these recurrent patterns can serve as a baseline to detect anomalous behaviors in future elections, potentially identifying atypical uses of social media (e.g., bot networks or coordinated inauthentic behavior) that deviate from the established "normal" dynamics.