Recently many literatures have studied the problem of rumor detection on social media and proposed various automatic detection algorithms. In this ongoing work report we exploit the power of the crowd and formulate the reviewer selection problem, which aim to nd reliable reviews for a possible rumor. Our reviewer selection scheme can be considered complementary to existing methods. We give theoretical analysis and provide a greedy algorithm with approximation guarantee. We conduct experiments on a Twitter dataset about rumors, which validates the e ectiveness and e ciency of our algorithm.
Nowadays people increasingly rely on the Internet to learn what is happening around the world. Among tons of stories and pages available online, news recommendation systems provide users with personalized news articles. However, social media and news platforms, seeking to please users, can shunt information that they guess their users will like hearing, but inadvertently isolate what they know into their own lter bubbles. [Par11] Rumors and fake news often propagate within lter bubbles and some argue that they have a ected the outcome of the 2016 U.S. presidential election.
Although many articles have investigated the lter bubble problem, little attention has been paid to recommendation systems themselves. In this paper we examine the role of recommendation algorithms on the formation of lter bubbles. In the following two sections we will analyze if and how content-based and collaborative ltering news recommendation algorithms cause the lter bubble problem. Most current recommendation systems use a hybrid model which takes both news content and user feedback into consideration, but it is more reasonable to study them separately to get an insight of the problem. 2
content-based Methods Most content-based news recommendation algorithms map users and news into the same feature space and calculate their similarity with certain distance metric. The most common approach is to apply an LDA based topic model to generate news representations.
We train the basic Latent Dirichlet allocation (LDA) model [BNJ03] on a public news dataset1, which includes 142568 articles from 15 media outlets. The topic number is set to 100. We use the Python topic modeling library gensim2 for implementation. After training we obtain the topic distribution of each news document.
Fig. 1 shows the topic distribution of news related to President Trump from liberal and conservative websites. To visualize the 100-D data we apply Principal Component Analysis (PCA) for dimensionality reduction. It proves that as for topic distributions, there is minor di erence between news articles with leftleaning and right-leaning political stance.
We further extract news on two unrelated issues, climate change and border wall, and analyze the topic distributions of news from di erent ideological perspectives. The result is shown in Fig. 2.
In Fig. 2, articles with opposing views on "climate change" issue (green and yellow dots) occupy the same region in the feature space, and articles with di erent opinions on "border wall" have similar properties, which indicates that LDA model cannot distinguish 1https://www.kaggle.com/snapcrack/all-the-news 2https://radimrehurek.com/gensim/ di erent opinions on the same issue. A user who has read an article about excessive regulation on business is very likely to get recommendation about climate change. A direct corollary will be pure LDA contentbased recommendation systems do not lead to the lter bubble problem. 3
Collaborative ltering methods proved to be successful in many domains, from movie recommendation to shopping recommendation. Typical collaborative ltering algorithms are built on a user-itemrating matrix, and use two separate feature spaces for users and items. User and item representation can be traditional matrix factorization based vectors [KNK13, Kor08] or neural network based embeddings [WDZE16, HLZ+17].
In this paper, we use the classic LFM (Latent Factor Model) as a representative for collaborative ltering algorithms. [KBV09, MS08] LFM models the preference y^ui as the dot product of latent factor vectors pu and qi, representing the user and the item, respectively.
y^ui = pu>qi
We collect data from Twitter using Twitter standard API.3 We select 11 popular U.S. news media with di erent political leanings listed in Table 1. You can refer to Wikipedia pages4 to learn about the liberalconservative divide in U.S. politics. For each of them we retrieve the most recent 200 tweets and query for their retweeters. After removing users with less than 3 3https://developer.twitter.com/en/docs/tweets/search/apireference/get-search-tweets.html
4https://en.wikipedia.org/wiki/Social liberalism, https://en.wikipedia.org/wiki/Social conservatism retweets and news with no retweets, we get a dataset of 7119 users, 2065 news and 55746 retweets (viewed as positive ratings). Since no negative feedback can be retrieved on Twitter, we randomly choose user-item pairs as negative samples. Finally we use Matrix Factorization to generate latent factor vectors for users and news. Here we set number of factors to 5. We also try other latent dimension numbers and they show a similar pattern. LFM is implemented in Python using Surprise5.
To further investigate the situation, we choose two unrelated categories, international politics and California wild re, and inspect the news representations from liberal and conservative media. The result is shown in Fig. 4. manual inspection, we nd that for California wild re, nearly all news websites tell exactly the same story and there is no ideological perspective on this event, but news from outlets with di erent political leaning are still mapped to di erent regions in the feature space. Articles on California wild re from liberal media are mapped close to articles on international politics from liberal media. So is conservative media.
Up to now we have found an explanation for lter bubbles. Users tend to read (or retweet for Twitter) news with similar political leaning with their own, and when an article is published it will rst be read by like-minded people, which nally leads to a dead lock. News articles sharing common audience will be mapped to the same region in the feature space, even if the articles are about di erent topics, and then users will be recommended with what like-minded people tends to read, and the lter bubble will be reinforced. Finally we have strong lter bubbles and leave users isolated from di erent ideological perspectives. 4
In this paper we analyze the role of recommendation systems in lter bubbles. We analyze topic distributions of news under LDA on di erent issues and from di erent sources, and discover that typical contentbased news recommendation algorithms cannot distinguish di erent opinions on the same topic. We analyze news representations under Latent Factor Model and indicate that collaborative ltering algorithms tend to map news from the same outlets or with similar political leaning into contiguous regions, thus leaving users in lter bubbles. 5
For content-based methods, we only test on original LDA. In our future work we will try topic modeling combined with sentiment analysis and opinion mining.
After guring out where the lter bubbles comes from, we should consider how to overcome this problem. We need to strike a balance between breaking the lter bubbles and still making users enjoy what they see. A reinforcement learning framework, which learn users' open-mindedness or tolerance on di erent topics and adjust recommendation policy accordingly, could be a reasonable consideration. [BNJ03]