The CSN used may be improved to account for factors other than similarity. Further, other factors may influence a consumer’s trust in a specific source, and consumers may only be able to be nudged to a certain point due to the influence of their peers. Moreover, distrust of the recommendation system may also develop over time. Other concerns include the fact that in this model, users would continue to be exposed to low quality news until their reading profile is further developed.

"Trustworthy misinformation mitigation with soft information nudging": Future Work

User reading profiles changed over time to sources of higher quality and of more moderate political leaning, as demonstrated in the graphs below.

University of Michigan - Ann Arbor

1,814,682 articles were selected from the NELA-GT-2018 dataset (a dataset from a variety of sources with labels pertaining to quality and ideology). A CSN (described above) was used as a relationship graph - a TF-IDF matrix was generated for all articles in the dataset, and the cosine similarity between each article vector pair was computed. Pairs with a cosine similarity greater than 0.85 were extracted to create a directed graph with edges whose thickness indicates the similarity between two groups. Quality scores and political leaning scores were then generated for each article using the labels in the dataset. User profiles were generated with a variety of trusted sources. The simulation was ran over a set time period and the changes in news consumption habits were measured.

Learn Before

Related

Learn After