Our analysis is focused on the Liar dataset, a dataset with over 10000 statements with labels indicating their truthfulness.

We have explored the dataset in Exploration.ipynb, doing an analysis of the statements based on their feature, trying to get some insight into possible correlations with the truth classification and to better understand the dataset.

While performing this analysis, we have realized that there is not enough data to achieve the level of generality we hoped for in the original proposal. Therefore, our focus has shifted towards a feature analysis of the lies, as well as the design of classifiers for statements from politifacts.

Note that the feature analysis we performed might not generalize, since the dataset is relatively small. However, the original publishers of the dataset have obtained promising results by training various machine learning models, which is also our goal.

We have collected more data about the statements using the politifacts API. As mentioned above, we plan to use this API to do generalize our analysis from the Liar dataset to other political statements. To this end, we have followed our inital plan to perform feature augmentation based on sentiment analysis. We have experimented with two approaches: the python nltk library and google cloud. From our analysis, the former does not seem to produce very reliable metrics as is, but has a Bayes classifier that we plan to train on this dataset. On the other hand, the latter has consistently provided accurate metrics in our experiments and we plan to use it to augment the dataset.

These experiments can be found in SentimentAnalysis.ipynb.

In the following weeks, we will use the insight we gained in the dataset to train machine learning models in order to analyse the statements from politifacts. We plan to use their API to analyse political statements and do an analysis on them. This is related to our intial goal, of finding the trustworthiness of various media sources, but limited in scope to political statements and their source.

• Until this point, most of the features we have analyzed have been within the categorical features of the dataset. We will expand this by accounting for the date of statements as well as the best classification we can make for their sentiments. We have already implemented some sentiment classifers, we now wish to see how they can influence our model.

• We will also look for other relevant features to query from the json files hosted on Politifacts.

• We want to improve our predictive model hopefully by accounting for more underlying features. We also want to see if this model is at all generalizable, so we want to take some reference some other fake news datasets and see if we can make predictions on those datasets with our model (will require a lot of cleaning to fit our model format).

-If we find that our model is generalizable in this way, then we will analyze which features are important in our model prediction to create our data story.

-If these results are not generalizable, we would like to isolate specific speakers, topics, or pairs of speakers to see if we can find some underlying patterns to make a convincing data story that may not be in the form of a general model. In doing this, we want to externalize from our dataset and see if our findings are applicable for other statements of this type.

• Based upon our week 2 findings, we want to construct some good visuals and a compelling data story. We do not yet know if our data story will be about general findings or if they will require a case-study approach relating to specific speakers and/or subjects.