1 摘要

Despite the tremendous efforts by social media platforms and factcheck services for fake news detection, fake news and misinformation still spread wildly on social media platforms (e.g., Twitter). Consequently, fake news mitigation strategies are urgently needed. Most of the existing work on fake news mitigation focuses on the overall mitigation on a whole social network while ignoring developing concrete mitigation strategies to deter individual users from sharing fake news. In this paper, we propose a novel veracityaware and event-driven recommendation model to recommend personalised corrective true news to individual users for effectively debunking fake news. Our proposed model Rec4Mit (Recommendation for Mitigation) not only effectively captures a user’s current reading preference with a focus on which event, e.g., US election, from her/his recent reading history containing true and/or fake news, but also accurately predicts the veracity (true or fake) of candidate news. As a result, Rec4Mit can recommend the most suitable true news to best match the user’s preference as well as to mitigate fake news. In particular, for those users who have read fake news of a certain event, Rec4Mit is able to recommend the corresponding true news of the same event. Extensive experiments on real-world datasets show Rec4Mit significantly outperforms the state-of-theart news recommendation methods in terms of the capability to recommend personalized true news for fake news mitigation.

2 💡Note

2.1 论文试图解决什么问题？

Most of the existing work on fake news mitigation focuses on the overall mitigation on a whole social network while ignoring developing concrete mitigation strategies to deter individual users from sharing fake news.

2.2 这是否是一个新的问题？

No, fake news mitigation has been explored before.

2.3 这篇文章要验证一个什么科学假设？

By veracity-aware and event-driven personalized news recommendation, Rec4Mit can effectively and concretely mitigate fake news.

2.4 有哪些相关研究？如何归类？谁是这一课题在领域内值得关注的研究员？

2.4.1 Fake News Mitigation

Only focus on the mitigation on the whole network while ignoring the concrete mitigation on individual users.

1. Independent cascade (IC) and linear threshold (LT) model based approaches.
   • taking users as nodes and each edge between a pair of users indicates the influence between them.
2. Point process model based approaches.
3. Reinforcement learning (RL) based approaches.

2.4.2 News Recommendation

1. RNN, Attention, CNN, GNN, Knowledge distillation and RL: all of these methods are for conventional news recommendation and cannot be employed to fake news mitigation task which has unique data and goal.
2. Recommender systems for mitigating fake news: fact-check URL

2.5 🔴论文中提到的解决方案之关键是什么？

2.5.1 Problem Formulation

User-news interaction: each user's sequence of historical interactions (e.g., clicks or reading) with news in a certain time period.

1. News sequences interacted by all users:
   1. : pieces of news which are sequentially interacted by a given user .
   2. : the whole news set.
2. records the meta information.
3. Context : for each user , given the historical true and/or fake news pieces interacted by .
4. Goal:
   1. To learn the user's dynamic reading preference from and predict the label (i.e., true or fake) of each piece of candidate news based on its meta information.
      • To generate a recommendation list of predicted true news which can best satisfy the user's reading preference at the moment.

2.5.2 Methodology

The overall structure of Rec 4 Mit model.
Input: a context consisting of a sequence of interacted news pieces.
Output: predicted next news.

2.5.2.1 Event-veracity Disentangler

How to model the transition over latent events while avoiding the interference from news veracity related information (e.g., news content style)?

1. To avoid the interference between event information and label information of a piece of news.
2. An adversarial auto-encoder framework
   1. Encoder: BERT + 3-layer MLP with LeakyReLU
      1. Input: a piece of news .
      2. Output: news embedding .
         1. Embedding of news ID
         2. Embedding of news meta information : a concatenation of both news title embedding and news description embedding.
   2. Event decoder: 3-layer MLP with LeakyReLU and residual connection
      1. Input: news abstract representation
      2. Output: latent representation of events behind news
   3. Veracity decoder: 3-layer MLP with LeakyReLU and residual connection + news veracity classifier
      1. Input: latent representation of events behind news
      2. Output: the label representation
      3. Function:
         1. Forced to extract the veracity-specific information.
         2. Used to predict the label of each candidate news so that only true news would be recommended.
   4. Loss function module
      1. Reconstruction loss
      2. Label prediction loss:
         • binary cross entropy ,
      3. Adversarial loss:
         • To ensure the disentangled event representation is pure and does not contain label-related information.
      4. The overall loss: .

2.5.2.2 Event Detection and Transition Module

To identify which event (s) the user is focusing on and how her/his reading preference regarding event is changing along with her/his reading history.

1. Event detector: to detect the possible latent event (s) associated with each news.
   1. Input: the event representation of news
   2. Arch:
      1. A softmax layer to obtain an event distribution vector to indicate the possibility of belonging to each of the latent events: ,
      2. Split the event representation into event-specific representation vectors: .
   3. Output: event-specific representation vectors.
2. Event transition net: to model the complex transition relations within and between events over a sequence of context news read by a user.
   1. Event memory module
      • To store the detected event-specific representations for each piece of news in the context.
      1. Input: an event-specific representation tensor
      2. Arch: Attention-based aggregation
         1. Event representation
         2. Learnable position embedding
         3. the -th row of matrix :
      3. Output: event-aware matrix , where ..
   2. Event activator
      • To activate those events in the memory matrix which are closely related to the next target news.
      1. Input: the disentangled event representation of is used as the activation signal.
      2. Output: the user-aware and candidate news-specific context embedding.
         1. The candidate news-specific context embedding:
         2. The activation degrees of all the events behind :

2.5.2.3 Next-news Prediction Module

To predict the next news for user : #### 2.5.2.4 Model Optimization and Training

• The total loss:
  1. The loss generated during the next-item prediction task
     • Cross-entropy loss:
       • A contrastive pair from a news context
       • Taking the ground truth next news as the positive sample and randomly sampling n pieces of news from the news set as the negative sample set .
  2. The loss generated during the disentangling operation on context news
  3. The loss generated during the disentangling operation on candidate news

2.6 论文中的实验是如何设计的？

• Data preparation
  1. Extract the sequence of news (contain true and/or fake news) read by each user on Twitter.
     1. News content
     2. Social context
     3. Spatiotemporal information
  2. Order these pieces of news by the reading timestamp to form a news sequence.
  3. All the sequences together form the user-news interaction dataset.
  4. the meta information including the title and one-paragaph description of each news is extracted to form the news information table. ## 2.7 用于定量评估的数据集是什么？评估标准是什么？Baseline是什么？

1. 数据集: PolitiFact and GossipCop
2. 指标:
   1. The prediction accuracy to measure whether an recommendation approach can accurately recommend the right news to well match a user's reading preference.
      • Recall, mean reciprocal rank (MRR) and normalized discounted cumulative gain (NDCG)
   2. The ratio of true news included in the recommendation list
      • .
3. 基线: sequential/session-based news recommendation baselines

2.8 论文中的实验及结果有没有很好地支持需要验证的科学假设？

实验结果：Comparisons w.r.t. Prediction Accuracy.

实验结果：Comparisons w.r.t. the Ratio of True News.

2.9 这篇论文到底有什么贡献？

1. News recommendation to mitigate fake news.

2.10 下一步呢？有什么工作可以继续深入？

1. Extend to graph learning.