For crawling tweets we use the GetOldTweets Library1 to bypass the limitations of the o cial Twitter API. This allows us to to download historical Twitter data within a speci c date range for a particular query. The queries we used for crawling Twitter data are built on topics that were published by two Indonesian factchecking websites2. The tweets are in the Indonesian language. We gathered data from the rst day of the 2019 Indonesian presidential campaign (September 23, 2018) to a week after the election result was publicized (May 28, 2019).

We selected 281 topics related to the presidential elections from the above referenced fact-checking websites with their corresponding supporting URLs. For each topic we created a query. For example, for the supporting URL that examines whether the 23 European Union ambassadors support Prabowo-Sandi or 1https://github.com/Jefferson-Henrique/ GetOldTweets-python

2https://cekfakta.tempo.co/ (Cek Fakta Tempo from Tempo) and https://turnbackhoax.id/ (Turn Back Hoax from Ma ndo) not3, we used the topic \European Ambassadors Support Prabowo" as the query to extract the relevant tweets.

To ensure alignment between the extracted tweets and the supporting URL, tweets from the rst time the news aired in social media until its seventh day are selected. After removal of duplicate tweets, this resulted in a set of 8,784 tweets for the 281 topics. For annotation, tweets that one retweet, one like, and one reply, or less are removed resulting in a set of 2,360 that we use for annotation. 3.2

We recruited 10 native Indonesian speakers to annotate the data. They do not have political job, political a liation, or belong to a political party to facilitate the impartiality. Having 2,360 tweets as original data set, and two annotators per tweet, each annotator had to label 472 tweets.

The information provided to the annotators was the topic, the supporting URL, and the tweet text. One topic is linked to one supporting URL and to multiple tweets. We wrote an extensive annotation guideline for Indonesian false news and validated it in several short iterations before starting the actual annotation process. 4 Annotators are asked to assign one of four classes to each tweet:

True: Tweets that relate to the topic and are true or accurate according to the supporting URLs; False: Tweets that relate to the topic and are false or inaccurate according the to supporting URLs; Misleading: Tweets that relate to the topic and have accurate information according to supporting URLs but lead to wrong conclusions; Other: Tweets that do not relate to the topic or are not discussed within supporting URLs.

While misleading news is sometimes considered a subset of false news, we decided to distinguish it separately for text classi cation. According to [ 21 ], misleading news tends to use correct facts and data, but how the news is delivered or how conclusions are drawn is false and therefore leads to the wrong interpretation. This is consistent with other de nitions that misleading news conceives false facts by topic changes, irrelevant information, and equivocations to mislead the audience [ 22 ].

3https://cekfakta.tempo.co/fakta/111/fakta-atau-hoaxbenarkah-23-dubes-uni-eropa-dukung- prabowo-sandi, determined to be false news

4The annotation guideline can be found here: https: //github.com/rayansuryadikara/false_news_detection_and_ dissemination_analysis

896 648 189 627 2,360

The annotation process was conducted in two stages. In the rst stage, two annotators annotated the data. In the second stage, a third annotator (the rst author of this paper) acted as a nal judge for any tweet where two previous annotators disagreed. We analyzed the inter-rater reliability of the annotated data using Cohen's . Out of 10 annotator pairs, there are ve pairs with moderate agreement ( = 0:41 0:60), four pairs with fair agreement ( = 0:21 0:40), and one pair with slight agreement ( = 0:01 0:20). The highest score is 0.52 and the lowest is 0.07. As a whole, we obtain fair agreement with a mean of 0.33. The statistics of the annotated data are outlined in Table 1. 3.3

We extract two di erent networks from our Twitter collection of 8,748 tweets. The rst is the mention network. In literature, it is suggested that mentioning other usernames in a tweet represents a more direct form of communication than what is obtained from a network based on follower connections [ 18 ]. The second network that we create is the hashtag co-occurrence network The frequency of use for a hashtag indicates its popularity. In the 2019 Indonesian presidential election, there are certain hashtags created to support or oppose certain gures, such as #jokowiamin to support Joko Widodo, the incumbent, and #2019gantipresiden (\2019 change the president") to support Prabowo, the challenger.

The mention network is a weighted directed network where posting usernames are de ned as the source and mentioned usernames are the target of a directed link. Link weight is determined by how many times the source username mentions the target username. The hashtag co-occurrence network is a weighted undirected network in which two hashtags are connected if they occur together in a tweet. Link weight is determined by counting how many times the tags co-occur.

In our experiments, we visualize the two networks to analyse how true news and false news spread in presidential election settings. For the network data, misleading news will be merged under false news to keep it straightforward and to simplify the contrasting visualization between true news and false news. In doing so, we actually model both networks as a multigraph in which two nodes can be connected based on how often they communicate or co-occur in both true and fake news. 4

Features. For the content-based classi cation, we compare three types of features: orthography features, sentiment lexicon features, and word n-grams.

Social media such as Twitter is a common example wherein there the conventions of orthographies are sometimes lacking [ 6 ]. Therefore, orthography patterns are commonly used for social media analysis [ 8, 17 ]. We de ne ve orthography features: counts of exclamation marks (E), question marks (Q), uppercase letters (U), lowercase letters (L), and emojis (M).

For sentiment features, we use the Indonesian Sentiment Lexicon (InSet) [ 9 ] which comprises 3,609 positive words and 6,609 negative words5. The sentiment scores range from -5 to 5, where negative scores indicate negative words and positive scores indicate positive words. Words with score 0 are disregarded since the lexicon excludes neutral category. Along with InSet, we use an Indonesian abusive lexicon [ 8 ],which comprises 126 words that are considered abusive.6 Thus, we have three sentiment lexicon features: the positive word count (P), the negative word count (N), and the abusive word count (A). Before applying the sentiment lexicons, we apply stop words removal and text normalization7. The stop words dictionary is adopted from [ 20 ].8 The text normalization dictionary comprises of 11,034 terms which are mapped to a normalized form. The dictionary is a continuous, collective work from researches [ 1, 8, 16 ] on the Indonesian language. In addition to lemmatization, the dictionary also facilitates Indonesian abbreviations, slangs, misspelled words, and even political gures' names. Therefore, the normalized form often consists of more than one word.

For the word n-gram features the text was lowercased, and URLs and punctuation were removed. For mentioned usernames and hashtags, we removed the @ and # symbols while the usernames and the hashtag words themselves were kept because both are instrumental parts of tweets to be identi ed and distinguished [ 13, 15 ]. Some of the usernames and hashtags are also included in the text normalization dictionary and therefore are normalized as well. We used six subsets of word n-grams to create vocabularies: Unigram, bigram, trigram, uni-bigram, bi-trigram, and uni-bitrigram. In all n-gram feature sets we use tf-idf as term weight.

si ers as prior work on Indonesian text classi cation [ 1, 8 ]: Multinomial Naive Bayes (MNB), Support Vector Machines (SVM) with SGD optimization [ 25 ], and Random Forest (RF), all implemented in Scikit-learn. We used the default hyperparameter settings for each classi er. For SVM, this means that C = 1. For RF, the number of estimators is 100 with no maximum depth for the trees. The nal precision and recall scores of each set of text feature are the average scores of these three classi ers. Meanwhile, F1 scores are calculated according to average precision and recall scores.

Voting ensembles. We assembled the results of from each experiment with di erent text features. The nal precision, recall, and F1 scores of each ensemble follow the same approach with the text feature sets after the voting ensemble is performed. We use majority voting: the numbers for each label are compared and the most voted label is selected. If there is not one label with the most votes, the class will be determined according to a text feature that has the best performance. We construct two di erent ensembles: Ensemble I is arranged from all combinations of each feature, Ensemble II is arranged from the best combination of each feature. 4.2

We aim to analyse how true news and false news spread between actors in the two networks described in Section 3.3. For visualization, we use Gephi [ 3 ], an opensource tool for social network analysis. While we do not directly model the precise di usion of the news 5https://github.com/fajri91/InSet as the network evolves, we do believe that these two 6https://github.com/okkyibrohim/ methods provide crucial insights in the reach of di erid-multi-label-hate-speech-and-abusive-language-detection ent types of news and the network e ects involved in 7https://github.com/okkyibrohim/ the process. id-multi-label-hate-speech-and-abusive-language-detection/ blob/master/new_kamusalay.csv Community detection is a method capable of 8https://github.com/stopwords-iso/ partitioning the network into communities (more tightly connected groups with fewer connections to other communities). Here, we use the well-known Louvain modularity maximization algorithm to perceive the potential of lter bubble e ects in a community [ 7 ].

Filter bubbles are a phenomenon in which a person is exposed to ideas, people, facts, or news that adhere to or are consistent with a particular political or social ideology, leaving alternative ideas unconsidered and in some cases outrightly rejected [ 12 ]. We propose to systematically identify every community to see the type of news circulating in that community.

Centrality measures assign a ranking to nodes in a network based on their topological position in the network. Here, we choose to use betweenness centrality to identify the most in uential nodes. Betweenness centrality measures for a particular node how many other nodes are connected via a shortest path that runs through that node. Therefore for the mention network, the node or username acts as an important hub in receiving and spreading information to other nodes [ 14 ]. On an individual node level, betweenness centrality captures information from neighboring users who both consume and generate false news. For the hashtag co-occurrence network, the hashtag is also an important hub where it frequently co-occurs lot with other hashtags. 5

Experimental settings. We evaluate our classi ers in two di erent types of experimental settings. The rst setting is the data set with three classes, namely True News, False News, and Misleading News. The second setting is the data set with four classes: True, False, Misleading, Other, and Unclear (where the three annotators all assigned a di erent label) While the 3class setting is easier for the classi er to learn, the 5-class setting is more realistic because it includes the tweets that are irrelevant but will occur in a real Twitter stream as well. We used a xed random train{test split of the data for evaluation of the models, with 20% of the data for testing.

Comparison of feature sets. We nd that in the 3-class classi cation, the best n-gram feature set is the combination of unigrams and bigrams; in the 5-class classi cation the best n-gram feature set is the use of bigrams alone. The best orthography feature set for the 3-class classi cation is the feature set with counts of exclamation marks, question marks, lowercase letters, and emojis; for the 5-class classi cation having the uppercase letter count instead of the question mark count is the most e ective set. Of the sentiment lexicons, using a combination of positive and negative sentiment words gives the best results for both settings. The assemble of the best feature combinations performed the best in the 3-class, while the assemble from all feature combinations performed the best in the 5-class. We compare the best feature combination for each feature type in Table 2.

The table shows that the n-gram features outperform orthographies and sentiment lexicons in each setting and each class. The ensemble methods are also not able to improve over the n-gram features alone. Nevertheless, the ensembling method allows orthography and sentiment lexicons to be included as features in text classi cation with better performance than independently, especially from social media sphere.

Final quality of text classi cation With the best text features in the 5-class setting (which is more di cult, but also more realistic than the 3-class setting), we obtain precision scores of 55% for true news, 71% of false news, and 68% for misleading news. Recall is 85% for true news, 62% for false news, and 26% for misleading news. The low recall for misleading news is caused by the small number of items in this category.

We analyzed the full collection of 8,784 tweets where the unannotated data set (6,424 tweets) is labelled by the SVM classi er with SGD optimization in the 5-class setting with the best-performing feature set (word bigrams). We then do the social network analysis on the automatically labelled dataset, which we discuss in the next section. 5.2

Table 3 shows the counts of nodes and edges (full network, and for true and false news) in the labelled Twitter networks. The last line of the table shows the number of communities. For the 10 largest communities, the distribution of true and false news by community as well as the top 10 in uential actors are shown in Figure 1 and 2 for the mention network and in Figure 5 and 6 for the hashtag co-occurrence network.

The distributions are stacked column of true news and false news, listing the number of nodes and edges in each discovered community or actor (usernames for the mention network work and hashtags for the hashtag co-occurrence network). True news is de ned by blue color while false news is de ned by orange color.

In the visualization, communities are represented by colours and betweenness centrality determined node size, as shown in Figure 3 and 4 for the mentioned net

analysis of the hashtag co-occurrence network for the 2019 Indonesian presidential elections on Twitter, the interesting ndings are:

True news is more strongly associated with top in uential hashtags.

False news is more strongly associated with sentiment-induced hashtags than with hashtags about events or occurrences. Examples are 2019gantipresiden (2019 change the president, shown in Figure 7 and 8), indonesianeedsprabowo and 01jokowilagi (01 Jokowi again), which show support for both candidates. These results con rm the nding of previous work [ 5 ] that emotions are important in detecting false information.

There is a community formed (Community 3) where only false news circulate in it. This community is lled with many slandering hashtags towards the incumbent Jokowi, such as jaekingoflies (Jae is one of derogatory title to Jokowi), jaengibuldimanalagi (Where does Jae lie again) and uninstalljaenow. However, none of them is a hashtag with enough in uence. The inclined \true news" in uential hashtags are very general terms and not directly about the presidential election, such as hoax and Indonesia. Hashtag hoax is especially noteworthy because any tweet which includes this hashtag mostly warns that the topic is a hoax, therefore ghting back hoax and is categorized as true news. The particular case of this hashtag was also outlined in the annotation guideline.

The mentioned-based network shows that the in uential users are not only receive more false news, but also spread them as well. These usernames consists of unveri ed and veri ed ones, with the top two in uential usernames are veri ed and \false news" inclined. This indicates that accounts with veri cation mark are not always clean from hoaxes.

Meanwhile, the hashtag-based network shows that supportive or sentiment-induced hashtags tend to relate more with false news, rather than more general events or terms. This indicates that these hashtags are more prone to information bias. Especially the supportive hashtags for each candidate, where users show fanatic support and attack the opposite candidate as well, often with false information.

As a reminder, these results illustrate the circumstances of the 2019 Indonesian presidential election event on Twitter. Furthermore, the news are selected based on fact-checking websites, which con rming circulating, trending topics on social media whether it is true or false. 6