• Information systems → Collaborative filtering ; Web searching and information discovery; Data mining; Document representation; 1

Due to the explosive growth of information, online news services have become increasingly important for people to get information and understand the outside world. Online news platforms, such as Google News1 and MSN News2, contain rich content and contextual information, pertaining to groups of society, politics, entertainment and so on. Although the enormous amount of news streams can be widely applicable to diferent people preferences, it can also cause information overwhelming to users. Due to the time sensitiveness of news articles, users’ interactions with news articles are highly sparse, which results in the data sparsity problem of recommendation systems. To address this challenge, some previous studies such as [ 12, 16 ] utilize rich content features in news to model users’ preference. Recently, external Knowledge Graph (KG) information, which contains rich knowledge facts and well-structured relations, is also incorporated to alleviate the data sparsity issue and improve the explainability of recommender systems [ 26 ]. By using the rich information from an extra KG, the data sparsity problem can be alleviated to some extent.

However, some news specific information is missing in a generic KG, including the collaborative relations of entities encoded in news articles and browsing behaviors of users. Such collaborative relations reveal the context similarity of entities in news and have been rarely explored in news recommendation. For instance, entities that frequently co-occur in articles or clicked by same users are usually strongly related in the news domain. A news article like “ Rihanna shows support for LeBron James in Game 7 vs. Celtics", may indicate a new relation between Rihanna and James that does not exist in a generic KG. Also, all of the previous studies overlook the semantic topics of the article itself. We find topics are also 1https://news.google.com/?hl=en-US&gl=US&ceid=US/ 2https://www.msn.com/en-us/news important factors to attract users’ attention and can benefit to construct a more powerful graph if well utilized. Moreover, as for the information of KG used in recommendation systems, previous graph based studies indiscriminately utilize the KG entities [ 17, 26 ], while ignore the fact that some entities and relations in a generic KG may contain irrelevant information for news recommendations. e.g., The birthday of Donald Trump is an uninformative relation for news recommendations and including it may even result in the ineficiency problem.

Based on above considerations, in this paper, we propose to utilize the collaborative information from news content and user behaviors to construct a more powerful knowledge graph, named News Graph (NG). We first remove the news-irrelevant relations in the original KG, then add two new types of information into NG, i.e., article topic entities and the collaborative edges among KG nodes. For topic entities, we consider both the explicit and implicit topics, i.e., categories of news articles and LDA [ 15 ] topics. As for the collaborative edges, we consider the associations among entities from content information of articles and users’ reading behaviors. In particular, the collaborative edges are extracted in three ways: (1) co-occurrence in the same news; (2) clicked by the same user; and (3) clicked by the same user in the same browsing session. Due to the selection of news-relevant relations and enhancement of collaborative information, the resulting news graphs is expected to possess a stronger capacity of representing news articles and users’ reading behaviors, thus it is news domain-oriented. To verify it, we have conduct experiments on four diferent news recommendation tasks, including personalized item recommendation, news category classification, news popularity prediction and local news detection. Results consistently demonstrate that leveraging our news graph is much more efective than leveraging a generic knowledge graph.

Our contributions are summarized as follows: • To the best of our knowledge, it is the first time that a domain specific graph, i.e., news graph, is constructed for serving news recommendations. Compared with a generic knowledge graph, we remove the news-irrelevant relations, while add some news topic entities and collaborative relations to make the graph more suitable for news recommendations. • To construct the news graph, we propose a News Relation Selection (NRS) algorithm to select the news-relevant relations. Meanwhile, we incorporate more news content and user behaviors into the news graph. Specifically, we construct three new types of collaborative relations for entities, i.e., co-occurring in the same news, clicked by the same user and clicked by the same user in the same browsing session. • Extensive experiments are conducted on a real news dataset.

The results demonstrate the efectiveness of our news graph for multiple news recommendation tasks, including item recommendations, article category classification, article popularity prediction and local news detection.

In this section, we introduce how to construct the NG in detail, including the construction of the news-relevant KG, collaborative relations and topic entities. We present an illustrative overview of NG in Fig. 1.

Politics

News Graph

sports Entity.Belongto.Topic Entity.Belongto.Topic

Entity.Belongto.Topic Entity.Belongto.Topic LJeaBmroesn We use a news corpus from MSN News2 ranging from Nov. 2018 to Apr. 2019, which contains 621,268 news articles and 594,529 distinctive news entities. To incorporate the extra knowledge information, we adopt Microsoft Satori3, which is a large scale commercial knowledge graph. For eficiency consideration, we search the one hop neighbors of all occurred entities in our news corpus in Microsoft Satori KG and extract all triples in which the confidence of relations linked among entities are greater than 0.8. The basic statistics of the extracted knowledge graph in our news corpus are shown in Table 1.

However, we observe that many relations in KG, e.g., the Birthday of Donald Trump, may not be very relevant to news recommendation tasks, but lead to the increase of millions of irrelevant triples in NG ( comparing the number of Triples in KG and News-Relevant KG in Table 1). Including enormous irrelevant relations not only makes the knowledge graph less efective to provide news-related information, but also makes some explainable recommender models such as [ 28 ] harder to search good knowledge paths for reasoning. we propose a News Relation Selecting (NRS) algorithm to filter out the news irrelevant relations. The details of NRS algorithm are shown in Algorithm 1 on the facing page. The basic idea of this algorithm is that we search at most 2-hop neighbors in the abovementioned KG for entities which appear in news articles (we called it news entities henceforth). During the search, if it reaches another news entity, we increase the weight for the relation that links these

News topics are important factors to attract the attention of users. Not every news article contains knowledge entities. Sometimes users click on the article simply because they like the topics. To ifll in the gaps where articles do not contain knowledge entities or contain non-informative entities, we propose to leverage news topics to make a supplement of the information of entities. We consider two types of topics information of news articles, i.e., the explicit and implicit topics of articles. As classified by editors, the category labels of articles are the best explicit topic information of articles. However, sometimes the simple category information may not be comprehensive enough to represent the topics of articles, especially when the articles do not have category labels. Hence we also utilize the Latent Dirichlet Allocation (LDA) model [ 15 ] to get 4https://en.wikipedia.org/wiki/Mayday_(Taiwanese_band) 5https://en.wikipedia.org/wiki/Jay_Chou

FC layers

Share Parameters

Weight 1 1 (softmax) 128 (ReLU)

Sum

Weight 2 1 (softmax) 128 (ReLU)

...

Input 1

Input 2 the implicit topics of each article. We add the two types of news topic entities, i.e., category and LDA entities, as the special entity nodes in NG. The linkage relations between an entity and its article topic entities are established only when the number of linkage time is over five times. Detailed statistics of topic entities are shown in Topic Entities in NG part in Table 1. 3

The goal of this paper is to propose a domain-specific knowledge graph for better news recommender systems. To demonstrate the efectiveness of the news graph, we design a series of simple experiments to compare the consumption of a general knowledge graph and our news graph. we conduct experiments on four typical news recommendation tasks, i.e. personalized article recommendations, news category classification, news popularity prediction and local news detection tasks. We use a real-world news reading dataset from MSN News2 for experiments. We collect the user-item interaction logs from Jan.1, 2019 to Jan.28, 2019, which contain 24, 542 news articles, 665, 034 users, and a total number of 6,776,611 impressions. 3.1

For both general knowledge and our news graph, in order to consume knowledge entities, we adopt an attentive pooling component as depicted in Fig. 2 to merge all entities included in one news article into one embedding vector. The input of the attentive pooling component are entity embeddings learned from TransE [ 3 ], while the output is a merged knowledge-aware vector. Then the original document representation is enhanced by this merged knowledge-aware vector. In this paper, we mainly focus on verifying the efectiveness of NG, hence we adopt a simple but eficiency graph embedding method, i.e. TransE, to get the entity embedding vector in KG and NG for a fair comparison. Definitely we can explore other advanced methods to study the influence of graph embedding method in feature work.

Formally, suppose a news article n contains m entities {e1, e2, ..., em }. The original document vector (DV) of n is vd (which can be generated by any models, such as DSSM [ 9 ] or BERT [ 5 ]). The proposed Weight n 1 (softmax) 128 (ReLU)

Input n (1) (2) (3) (4) model framework is designed to generate another document representation vn that contains the entity information, which can supplement the original document representation vd . Given a graph (KG or NG), we first adopt the widely used graph embedding method TransE [ 3 ] to obtain the entity embedding ve ∈ RDe . To obtain a ifxed-length vector representation v+ for entities, we aggregate the embedding of entities via the attentive pooling component. The normalized attention weight αej of an entity ej is defined as: aej αej = =

W2Tσ (W1Tvej + b1) + b2,

exp(aej ) Ím k=1 exp(aek ) where aej is the attention weight before normalization, which is computed by a two-layer attention network. W1 and W2 are weight matrices of vej , b1 and b2 are the bias value in the two attention layers respectively. σ is the activation function, and we use ReLU in our model. Then the merged entity representation vector v+ is represented as: v+ = m Õ

αek · vek .

k=1 The new document vector vn is computed by applying a non-linear transformation of the concatenated vector of vd and v+: vn = σ W3T(vd ⊕ v+) + b3 where ⊕ denotes the concatenation of two vectors. In the following sections, a series experiments are designed to verify the superiority of the NG over the KG in learning useful vn . The overall architecture used for diferent tasks are illustrated in Fig. 3, more details will be introduced in the next section. 3.2

To verify the usefulness of NG for news recommendations, we conduct experiments on four news-related tasks: • Personalized item recommendation task: given a user and a candidate article, it predicts the probability that the user will click on the article. • Category classification task: to predict the category label 6 that a news article belongs to. We have 15 top-level categories in the news corpus, including US News, Entertainment, Sports, Lifestyle, Money, Celebrities, Royals News, World News, Travel, Autos, Politics, Health, Video, Weather and Food&Drink. • Popularity prediction task: we split the articles into 4 balanced groups according to its click-through ratio (which indicates its popularity level). The task is to predict the popularity level given a news article. • Local news detection task: it predicts whether a news article reports an event that happens in a local context that would not be an interest of another locality.

For the item recommendation task, we compute the click through ratio (CTR) based on a concatenation of the user vector (UV) and the document vector (DV). UV is computed by a simple time-decayed averaging of DV of the user’s clicked articles. The DV vn is derived by the method described in Section 3.1. A two-layer feed-forward 6In this task, the topic nodes, i.e., category entities are not enabled in NG construction.

CTR 1 (Sigmoid) 128 (ReLU) UV 128 (ReLU)

DV (a) Item recommendation without entity

CTR 1 (Sigmoid) 128 (ReLU) UV

128 (ReLU) DV

Attentive Pooling

EV

UV: user vector DV: document vector EV: entity vector m: number of class

Trainable modules Fixed features

Vector concatenation

Score m (Softmax) 128 (ReLU)

DV

Score m (Softmax) 128 (ReLU) DV

Attentive Pooling

EV (b) Item recommendation with entity (c) Document classification without entity (d) Document classification with entity User

News Candidate

User

News Candidate

News Candidate

News Candidate neural network is used to get the CTR prediction score. For optimization we use a ranking loss, i.e., for each positive user-item pair, we randomly sample five negative items and to maximize the softmax likelihood of the positive pair. The running model architectures are shown in Fig.3(a,b). For the rest of tasks, we treat them as classification problems (binary classification for local news detection, multi-class classification for news category classification and news popularity prediction), and only take the document vector as input. The loss function is cross entropy 7. The running model architectures are shown in Fig.3(c,d). 3.3

For news recommendation task, Area Under Curve (AUC) [ 4 ] and Normalized Discounted Cumulative Gain at rank k (NDCG@k) [ 10 ] is utilized to evaluate the model performance across seven days. For the multi-classification problem, including category classification and popularity prediction tasks, we adopt Accuracy (ACC) and F1-Score (micro) as the evaluation metrics. As for the binary local news detection task, we use three metrics AUC, F1-Score and ACC. For news recommendation task, we use the first two weeks’ data to construct the user click history, the third week for training and the evenly split the data in last week for validation and test. For the other three tasks, we randomly split the news articles into: 8 : 1 : 1 for training, validation and testing respectively. 3.4

For each method, grid search is applied to find the optimal settings. We report the result of methods with its optimal hyperparameter settings. The dimensions of article embedding learned by topic model LDA [ 1 ] and learned by the DSSM model [ 9 ] are set to 90, and the embedding dimensions of entity and article are set to 90 for a fair comparison. The learning rate is set to 0.005 and batch size is 500. 7https://ml-cheatsheet.readthedocs.io/en/latest/loss_functions.html#cross-entropy We already release the source code at https://github.com/danyangliu/NewsGraphRec. 3.5

To verify the usefulness of our proposed NG, we compare the performance of the following models.

• KG: entity embeddings are learned from KG. • NG: entity embeddings are learned on our constructed NG. • DV: only using the original article vector vd (which is learned from LDA model [ 15 ] and DSSM [ 9 ], we concatenate the vectors from these two models as the original document vector). • DV+KG: we concatenate DV and KG vectors as the final document vector, i.e., vn based on KG. • DV+NG: we concatenate DV and NG vectors as the final document vector, i.e., vn based on NG.

The results of the news recommendation task are shown in Table 2. The results of category classification, popularity prediction and local news detection tasks are presented in Table 3. With comparisons of all baseline methods, we have the following observations: • In personalized article recommendation task, we demonstrate the performance of all methods on seven consecutive days. We can see that DV and DV+KG models have comparable performance. Overall DV+KG is better than DV model on both AUC and NDCG@10, but not consistently across the seven days. The model DV+NG performs best than all baseline models consistently on everyday. This indicates that the information learned in our proposed NG is really helpful for the prediction of users’ news preference. Meanwhile, the results verify our assumption that a general knowledge graph may contain uninformative relations or miss some domain specific relations/entities, thus is sub-optimal for news recommendation.

Traditional recommender systems mostly sufer from several inherent issues such as data sparsity and cold start problems. To address the above problems, researchers usually incorporate extra information to increase the capacity of data and make a supplement to the current model. Our work is highly related with knowledgeenhanced recommendation, which had been intensively studied in recent years in alleviating the data sparsity and providing the explainability virtue for recommender systems.

The extra information in knowledge base, such as Freebase [ 2 ], Google Knowledge Graph [ 21 ] and Bing Satori [ 18 ], Knowledge Valut [ 6 ], YAGO [ 8 ], Probase [ 29 ], had been proved to be critical for many real-world tasks, such as question answering, document representation [ 20 ] and graph-based recommendation [ 19 ]. These knowledge graph are constructed from the large volume of noisy text data. For example, both Google Knowledge Graph and Bing Satori developed entity databases for hundreds of millions of entities [ 23 ]. They extracted entities from real world objects and concepts including people, places, books, movies, events and so on. An entity may have some properties and relationships to other entities. In addition, external knowledge graph contains much more fruitful facts and connections about items [ 2 ]. For example, CKE [ 31 ] proposes a general framework to jointly learn from the auxiliary knowledge graph, textual and visual information. As for news recommendation, DKN [ 26 ] is proposed to incorporate knowledge embedding and text embedding from news content. RippleNet [ 25 ] proposes to simulate how users’ preferences propagate over the knowledge graph and uses a memory neural network to capture users’ high-order preferences based on knowlege entities. It motivates several new projects which so far achieve state-of-the-art performance in knowledge graph-based recommendations [ 27 ].

Another related task is knowledge graph embedding, which has been extensively investigated in recent years. There are also a large body of relational approaches for modeling the relational patterns on knowledge graphs [ 3, 13, 22, 30 ]. As for aggregating the node representation in graph based models, previous studies learn the low-dimensional representations of graph vertices with preserving graph topology structure, node content, and other information. For example, GCN [ 11 ] utilizes a localized graph convolutions for a classification task. GAT [ 24 ] uses self-attention network for information propagation, which utilizes a multi-head attention mechanism to increase model capacity. GCN and GAT are popular architectures of the general graph networks.

The major diference between prior work and ours is that they directly use the generic knowledge graph, while in our work, we design a domain specific knowledge graph, called news graph. We utilize collaborative relations of entities (e.g.,entities frequently appear in same news articles or clicked by same users), and the topic context of news article to construct a more powerful graph for news recommendations. 5

In this paper, we demonstrate the necessity of using a domain specific graph for news recommendation, which yet has not been explored in the literature. We propose to construct a news graph by considering the collaborative relations and topic entities, as well as ifltering out the irrelevant relations of news reading applications for eficiency consideration. Currently, we adopt a simple attentionbased model to demonstrate the efectiveness of our proposed news graph. In the future, we will design more flexible and efective models to fuse the knowledge entities in a news article for precise document understanding. Meanwhile, we will explore how the news graph can benefit more tasks related to recommender systems, such as the model explainability [ 7 ] and item candidates retrieval [ 14 ].

The authors would like to thank Microsoft News for providing technical support and data in the experiments, and Jiun-Hung Chen (Microsoft News) and Ying Qiao (Microsoft News) for their support and discussions.