A Look Into News Avoidance Through AWRS: An Avoidance-Aware Recommender System

Neural content-based models have become the leading method for personalized news recommendation, surpassing traditional systems that relied on manual feature engineering (Wu et al., 2023). These Neural News Recommender (NNR) models typically include a news encoder that transforms input features into embeddings and a user encoder that creates user-level representations from clicked news (Wu et al., 2023, 2019d, 2019c; Okura et al., 2017; An et al., 2019; Wu et al., 2022b). The recommendation score is computed by comparing the candidate news embedding against the user embedding (Wang et al., 2018; Wu et al., 2019b). These models are trained using point-wise classification objectives with negative sampling (Huang et al., 2013; Wu et al., 2019d).

Personalized NNR systems focus on matching news content to user preferences, often limiting exposure to diverse viewpoints and creating ”filter bubbles” (Iana et al., 2024; Heitz et al., 2022; Pariser, 2011; Li and Wang, 2019). This leads to homogeneous news consumption, reinforcing users’ initial stances. To address these issues, researchers have explored various methods to diversify recommendations. One approach involves re-ranking personalized recommendation sets (Gharahighehi and Vens, 2021) to increase diversity. Another approach involves multi-task training of NNRs (Moreira et al., 2019; Qi et al., 2022; Wu et al., 2022a; Choi et al., 2022), where the primary personalization objective is coupled with auxiliary objectives that promote aspect-based diversification.

Recent advancements in news recommenders aim to enhance content-based personalization by considering additional aspects: (1) Sentiment - SentiRec (Wu et al., 2020b) recommends news with diverse sentiments. (2) Popularity - PP-Rec (Qi et al., 2021) incorporates news popularity to address cold-start and diversity issues. (3) Global Representations - GLORY (Yang et al., 2023) uses a global news graph and gated graph neural networks to enrich news representations. (4) Recency - LANCER (Bae et al., 2023) introduces the concept of lifetime to improve news recommendation models.

As previously mentioned, news avoidance has become a growing concern for news platforms, particularly those utilizing recommender systems due to the creation of ”filter bubbles” (Pariser, 2011). We decided to investigate how avoidance affects the behavior of news article clicks. Subsequent sections will illustrate the intrinsic relationship between avoidance, exposure, and the number of clicks a news article receives. This aspect, overlooked in prior studies to the best of our knowledge, can enhance news recommendations by providing insights into the interaction between exposure and avoidance, especially regarding popular items.

In the news domain, the item with the most clicks is often highly exposed and less avoided. We argue that news article avoidance conveys significant information about user preferences. For instance, during an election, election-related articles may be popular but often avoided by those fatigued by political news. If a user reads a political article that is largely avoided by others, it suggests a strong interest in that topic. Conversely, clicking on a popular football article, one that is little avoided, might indicate less about the user’s preference and more about its current popularity. In other words, largely avoided news articles provide more information about user tastes than less avoided ones. This understanding allows us to incorporate such contextual information into recommender systems, improving the understanding of user preferences. The following sections will formally define each of the concepts discussed above.

Example: At time $t_{1}$, we know that we have 100 impression lists, and the news article $n_{174}$ has been displayed in 50 of them. Thus, we have $n_{E}(n_{174},t_{1})=50$ and $n_{I}(t_{1})=100$. Hence, the exposure per impression of this news article $n_{174}$ is $EPI(n_{174},t_{1})=\frac{50}{100}=0.5$.

Avoidance reflects the deliberate non-engagement with specific news content by readers. This phenomenon may arise from various factors, such as a lack of interest in the topic, discomfort with the content, or distrust towards the source. Understanding the patterns of avoidance is crucial for analyzing user behavior and optimizing content delivery to enhance reader engagement.

Example: Given that the news article $n_{174}$ at time $t_{1}$ received 20 clicks out of 50 exposures, its avoidance is calculated as $Av(n_{174},t_{1})=1-\frac{n_{clk}(n_{174},t_{1})}{n_{E}(n_{174},t_{1})}=1-\frac{20}{50}=0.6$.

Relevance indicates how well a news article aligns with readers’ interests, but measuring it is complex and requires specific data. Key factors in measuring relevance include: (1) time since publication, (2) news text embedding, (3) exposure per impression $EPI(n,t)$, (4) avoidance $Av(n,t)$, and (5) number of clicks $clk(n,t)$. Each factor is weighted within the model to evaluate relevance, offering a robust framework for assessing the alignment of a news article with user interests. More details on the calculation of relevance can be found in Section 5.3.

Given a user $u$ and a candidate news article $n_{c}$, our goal is to compute an interest score $Int_{s}$ that quantifies user $u$’s potential engagement with $n_{c}$. We evaluate a collection of candidate news articles $\mathbf{N}_{c}=[n_{c1},n_{c2},\ldots,n_{cL}]$ and recommend the highest-ranking articles to user $u$. User $u$ has a history of clicked news articles $\mathbf{H}_{u}=[h_{1},h_{2},\ldots,h_{M}]$. Each news article is characterized by its title ($\mathcal{T}$), abstract ($\mathcal{A}$), category ($\mathcal{C}_{cat}$), and associated entities ($\mathbf{E}_{i}=[e_{1},e_{2},\ldots,e_{k}]$).

Our news encoder, inspired by (Qi et al., 2021, 2022; Wu et al., 2019b), extracts detailed information from text and entities within a news article using a multi-head attention architecture. It incorporates embeddings from the title, category, and entities, generating embeddings with Pretrained Language Models (PLMs) following (Iana et al., 2024). We used RoBERTa Base (Liu et al., 2019) for the MIND-small dataset, NB-BERT Base (Kummervold et al., 2021) for the Adressa one-week dataset, and a finetuned version of the DeBERTa model (He et al., 2021) called Japanese DeBERTa V2¹¹1https://huggingface.co/ku-nlp/deberta-v2-base-japanese. For all models, only the last four PLM layers were fine-tuned. The final news embedding n combines $\textbf{n}_{t}$ (title), $\textbf{n}_{c}$ (category), and $\textbf{n}_{e}$ (entity). For the Nikkei and Adressa one-week datasets, only title and category information were used.

As discussed in Section 3.3, measuring relevance at a given time $t$ involves considering several factors: the content of the article, the time elapsed since publication, avoidance and exposure per impression values, and the number of clicks. Therefore, the Avoidance-aware Relevance Predictor, illustrated in Figure 5 (a), evaluates the relevance of each previously clicked news article $n$. This assessment incorporates the time elapsed, encoded using the Time2Vec (Kazemi et al., 2019) model, resulting in the time elapsed embedding $\bm{t}_{el}$. The news embedding $\bm{n}$, obtained from the news encoder described in Section 5.2, along with the number of clicks at time $t$ denoted as $n_{clk}(n,t)$, and the impact of avoidance and exposure per impression are integrated into the user engagement embedding $\bm{ue}$, as detailed in Section 4. Consequently, the avoidance-aware relevance scores ($\textbf{r}_{aw}$) for each news article are computed as follows:

The core idea of AWRS is to integrate avoidance-awareness. Leveraging the CAUM model’s architecture (Qi et al., 2022), which facilitates the extraction of how closely candidate news articles relate to previously clicked ones, AWRS enhances the candidate-aware attention network with user engagement embeddings. This enhancement refines the comprehension of user preferences by examining the categories of news articles users interact with, as well as their levels of exposure and avoidance. This approach yields a nuanced depiction of user interests.

Inspired by (Wu et al., 2019d), we use negative sampling during model training. For each clicked news item (positive sample), we randomly sample $K$ non-clicked items (negative samples) from the same impression and shuffle their order. The click probability score of the positive item is $\hat{y}^{+}$, and the scores for the $K$ negative items are $[\hat{y}^{-}_{1},\hat{y}^{-}_{2},\ldots,\hat{y}^{-}_{K}]$. Then, these scores are normalized using the softmax function given by

We frame the click prediction task as a pseudo $(K+1)$-way classification problem. The loss function is the negative log-likelihood of all positive samples $S$: $\mathcal{L}=-\sum_{i\in S}\log(p_{i})$. This method trains the model to distinguish between clicked and non-clicked news items.

Table 2 presents a performance comparison between the baselines and our proposed AWRS model. As shown in the table, our model demonstrates improvements in nearly all key metrics, attributed to its enhanced awareness of avoidance and exposure. For instance, if a user opts to read an article about chess, a typically avoided topic, it suggests a strong interest in this subject. Conversely, an article about the Oscars, which is generally less avoided, might be clicked due to its current popularity rather than genuine interest. It is crucial to note that such patterns may change over time. For example, during a specific period involving controversial news about chess, articles on chess might become less avoided, whereas Oscars articles might be more avoided if they are far from the premiere date. Therefore, we argue that avoidance and time are closely related. This understanding allows us to delve deeper into our discussion for each dataset.

We aim to address research questions RQ2, RQ3, and RQ4 by exposing an ablation study, as demonstrated in the following sections.

From our exploration of current methodologies in news article recommendation systems, we identify two notable approaches. Firstly, there are methods like PP-REC (Qi et al., 2021) that leverage contextual information such as time to enhance recommendations. Secondly, methods such as those discussed in GLORY (Yang et al., 2023) utilize global information derived from static global graphs. Our proposed model aligns with the former category, leveraging time-varying avoidance data of news articles to enhance user recommendations. However, like all models, our approach has its drawbacks. For instance, one limitation of our model is its increased memory requirements due to the need for extensive per-unit-time information such as avoidance ($Av$) and Exposure Per Impression ($EPI$). Despite this increased memory demand, we did not observe a corresponding increase in computational training time for our models.