Recommender systems in social platforms attract attention in part because of their potential impact over political phenomena, such as polarization or fragmentation of online communities. These research topics are also important because of the need for understanding systemic effects in view of upcoming risk-oriented AI regulation in the EU and the US. A common approach leverages outcomes of recommendations to audit recommender systems. A different approach is that of explainability, seeking to render recommendation mechanisms intelligible to humans, potentially enabling both auditing and actionable design tools. This second approach is particularly challenging in the context of online systems of political opinions because of the intrinsic unobservability of opinions. In this article we leverage multi-dimensional political opinion estimation of large online populations (along a left-right dimension but also along other political dimensions) to investigate latent spaces in representation learning computed by recommender systems. We train a recommender based on ubiquitous collaborative filtering principles using data on content sharing on Twitter by a large population, evaluating accuracy and extracting a latent space representation leveraged by the recommender. On the other hand, we leverage multi-dimensional political opinion inference to position users in political spaces representing their opinions. We then show for the first time the relation between latent representations leveraged by a recommender system and the spatial representation of users. We show that some dimensions learned by the recommender capture ideological positions of users, bridging politics and algorithmics in our social and algorithmic system, opening a path towards political explainability of AI.

To understand why internet users spread fake news online, many studies have focused on individual drivers, such as cognitive skills, media literacy, or demographics. Recent findings have also shown the role of complex socio-political dynamics, highlighting that political polarization and ideologies are closely linked to a propensity to participate in the dissemination of fake news. Most of the existing empirical studies have focused on the US example by exploiting the self-reported or solicited positioning of users on a dichotomous scale opposing liberals with conservatives. Yet, left-right polarization alone is insufficient to study socio-political dynamics when considering non binary and multi-dimensional party systems, in which relevant ideological stances must be characterized in additional dimensions, relating for example to opposition to elites, government, political parties or mainstream media. In this article we leverage ideological embeddings of Twitter networks in France in multi-dimensional opinions spaces, where dimensions stand for attitudes towards different issues, and we trace the positions of users who shared articles that were rated as misinformation by fact-checkers. In multi-dimensional settings, and in contrast with the US, opinion dimensions capturing attitudes towards elites are more predictive of whether a user shares misinformation. Most users sharing misinformation hold salient anti-elite sentiments and, among them, more so those with radical left- and right-leaning stances. Our results reinforce the importance of enriching one-dimensional left-right analyses, showing that other ideological dimensions, such as anti-elite sentiment, are critical when characterizing users who spread fake news. This lends support to emerging accounts of social drivers of misinformation through political polarization, but also stresses the role of the entanglement between fake news, anti-elite polarization, and the role of scientific authorities in public debate.

Ideological scaling is an ubiquitous tool for inferring political opinions of users in social networks, allowing to position a large number of users in left-right or liberal-conservative scales. More recent methods address the need, highlighted by social science research, to infer positions in additional social dimensions. These dimensions allow for the analysis of emerging divisions such as anti-elite sentiment, or attitudes towards globalization, among others. These methods propose to embed social networks in multi-dimensional attitudinal spaces, where dimensions stand as indicators of positive or negative attitudes towards several and separate issues of public debate. So far, these methods have been validated in the context of individual national settings. In this article we propose a method to embed a large number of social media users in multi-dimensional attitudinal spaces that are common to several countries, allowing for large-scale comparative studies. Additionally, we propose novel statistical benchmark validations that show the accuracy of the estimated positions. We illustrate our method on Twitter friendship networks in France, Germany, Italy, and Spain.

Traditionally, public opinion on different issues of public debate has been studied through polls and surveys. Recent advancements in network ideological scaling methods, however, have shown that digital behavioral traces in social media platforms can be used to mine opinions at a massive scale. This has yet to be shown to work beyond one-dimensional opinion scales, which are best suited for two-party systems and binary social divides such as those observed in the US. In this article, we use multidimensional ideological scaling for coupled with referential attitudinal data for some nodes. We show that opinions can be mined in a multitude of issues: from social networks, embedding them in ideological spaces where dimensions stand for indicators of positive and negative opinions, towards issues of public debate. This method does not require text analysis and is thus language independent. We illustrate this approach on the Twitter follower network of French users leveraging political survey data.

The prevalence of algorithmic recommendations has raised public concern about undesired societal effects. A central threat is the risk of polarization, which is difficult to conceptualize and to measure, making it difficult to assess the role of Recommender Systems in this phenomenon. These difficulties have yielded two types of analyses: 1) purely topological approaches that study how recommenders isolate or connect types of nodes in a graph, and 2) spatial opinion approaches that study how recommenders change the distribution of users on a given opinion scale. The former analyses prove inadequate in settings where users are not classified into categorical types (e.g., in two-party systems with binary social divides), while the latter rely on synthetic data due to the unobservability of opinions. To overcome both difficulties we present the first analysis of friend recommendations acting on real-world sub-graphs of the Twitter network where users are embedded in multidimensional ideological spaces and in which dimensions are indicators of attitudes towards issues in the public debate. We present a polarization metric adapted to these dual topological and spatial states of social network, and use it to track both the evolution of polarization on Twitter networks where the graph evolves following well-known Recommender Systems, and opinions co-evolve following a DeGroot opinion model. We show that different recommendation principles can sometimes drive or mitigate polarization appearing in real social networks.

A growing literature on ideology estimation through scaling methods in social networks restricts the scaling procedure to nodes that provide interpretability of the resulting feature space. On Twitter, for example, it is common to consider the subnetwork of parliamentarians and their followers. While effective in inferring meaningful ideological features, this restriction limits interesting applications such as country-wide measurement of polarization and its evolution. We propose two methods to propagate ideological features beyond these sub-networks: based on homophily (linked users have similar ideology), and based on structural similarity (nodes with similar neighborhoods have similar ideologies). In our methods, we leverage the concept of ideological coherence of a neighborhood as a parameter for propagation. Using Twitter data, we produce an ideological scaling for 370K users, and analyze the two propagation methods on a population of 6.5M users. We find that, when coherence is considered, the ideology of a user is better estimated from those with similar neighborhoods, than from their immediate neighbors.

The Heterogeneous Information Network (HIN) formalism is very flexible and enables complex recommendations models. We evaluate the effect of different parts of a HIN on the accuracy and the diversity of recommendations, then investigate if these effects are only due to the semantic content encoded in the network. We use recently-proposed diversity measures which are based on the network structure and better suited to the HIN formalism. Finally, we randomly shuffle the edges of some parts of the HIN, to empty the network from its semantic content, while leaving its structure relatively unaffected. We show that the semantic content encoded in the network data has a limited importance for the performance of a recommender system and that structure is crucial.