1 Introduction

Online news service platform has become a popular electronic news reading channel in recent years. However, a large amount of news is released all the time, which makes it difficult for users to quickly find the news to their interests. Therefore, by learning users’ characteristics and preferences, personalized news recommender can select different news for each user separately and improve the quality of news service effectively.

Figure 1 shows the framework of personalized news recommendation system, which includes three main components, news data collection, news recommendation prediction modeling and personalized news display. Firstly, the recommender constructs the user profile by processing and analyzing raw news data and user behavior data. Secondly, it combines different news recommendation methods to model news and users, fully extracting the content features of news and mining user preferences to generate news and user representation. Then, it ranks the candidate news according to the prediction model to generate a personalized recommendation list. In addition, the user interface will display different kinds of news for each user and collect user feedback to update the recommendation results, which achieves personalized recommendation effect. Although the personalized news recommendation system has made great progress, it still needs to further improve the recommendation performance, including more fully mining news semantics, more fine-grained extraction of user preferences and more efficient news recommendation model.

Fig. 1
figure 1

The framework of personalized news recommendation system

Full size image

Personalized news recommendation technologies are crucial for news platforms to help reduce news overload and enhance user experience [1]. The core lies in how to accurately match users’ interests and candidate news. As is shown in Fig. 2, firstly, the news encoder and user encoder extract the features of news articles and user interests, and get the embedded vector representation of news and users by deep learning or traditional machine learning methods. And then, the model ranks the candidate news according to the matching degree between the semantic information of news and user’s interest preference information. The higher the score, the higher the probability of user clicking on the candidate news. The prediction model of PNR can integrate multiple sources of information such as user browsing records, reading time, location, and social information, to provide users with personalized, accurate and comprehensive news recommendation services.

Fig. 2
figure 2

The prediction model of personalized news recommendation system

Full size image

There have been many related review papers introducing PNR technologies. For example, Li et al. [2] introduced traditional news recommendation technologies based on handcraft features, including collaborative filtering, content-based and hybrid recommendation methods. Huang et al. [3] and Yu et al. [4] mentioned news recommendation methods on deep learning and reinforcement learning separately. Tian et al. [5] mainly analyzed the news recommendation algorithm on deep learning. Wang et al. [6] and Meng et al. [7] summarized news recommendation approaches from the perspectives of personalized recommendation framework and mobile recommendation, respectively. Different from the above methods, this paper comprehensively considers various characteristics and influencing factors on PNR, further discusses the research progress of PNR and provides a more comprehensive and systematic survey of PNR:

  • We provide a three-part framework and discuss the research of each part separately in order to understand the detailed process of personalized news recommendation more clearly.

  • We make a systematic classification on PNR technologies, emphatically summarize the news recommendation methods based on different types of graph structure learning, and comprehensively analyze the characteristics and advantages of deep learning technology from the core objects of news recommendation (i.e., users and news).

  • Lastly, the challenges and future directions for PNR researches are discussed.

The architecture of this paper is organized as follows. Section 2 gives a classification on PNR and analyses the differences between personalized and traditional news recommendation. Section 3 introduces news datasets source and data processing methods. Section 4 focuses on analyzing and discussing the key technologies on PNR. Section 5 summarizes the ranking methods, interface display and evaluation metrics. Section 6 puts forward the challenging problems and future research directions on PNR. Section 7 makes a conclusion.

2 An Overview on Personalized News Recommendation Methods

Personalized news recommendation is highly time-sensitive and easily influenced by contextual and social factors. Unlike general item recommendation such as movies or music where user preferences may remain stable over time, PNR models need to capture user preferences in time, update and adjust the recommended news content dynamically by tracking user’s recent interactions or incorporating temporal information, such as publication date, recency or popularity, into their recommendation algorithms. In addition, due to the explosion of news data, there is a higher demand for the scalability of the system. By summarizing, the comparison results between personalized news and other recommendations are shown in Table 1.

Table 1 Comparison between news and other recommendations
Full size table

2.1 A Taxonomy on Personalized News Recommendation

2.1.1 Time-Based News Recommendation

Time-based news recommendation includes timeliness and real time. Timeliness is the essential feature of news recommendation, which is different from other recommendations. News published for a long time often loses its value as news. According to the current behavior of users (such as pulling and sliding), the news recommendation system updates the recommendation results in real time, quickly responds to the changes of users’ interests, and enhances the instant experience of users. Liu et al. [8] propose a dual-task deep neural network model, which uses an extended time module to refine news embedding and learns user vector representation by predicting the active time of users on each news. Experiments show that this method improves the timeliness of news recommendation and promotes the dissemination of the latest news, but weakens the dominant position of users’ interest to a certain extent. Therefore, it is necessary to consider the balance between news timeliness and user preferences in the process of news recommendation.

2.1.2 Location-Based News Recommendation

The location of mobile users is not fixed, so considering the location of users reading news can get users’ current preferences more accurately, and it is also more in line with the needs of users. Therefore, Yuan et al. [9] introduce the location of news events into news recommendation models and propose a news recommendation algorithm considering geographical position. Firstly, an algorithm is designed to extract the location of news events and a vector space model is used to represent news feature vectors. Then user interest models are constructed for news sets with and without geographical locations, respectively. Chen et al. [10] propose a location-aware personalized news recommendation method with explicit semantic analysis called LP-ESA, which uses users’ personal interests and geographical context to recommend news. However, the topic space based on Wikipedia in LP-ESA has some problems, such as high dimension, sparsity and redundancy. Therefore, LP-DSA method is proposed, which uses deep neural network to extract dense, abstract, low-dimensional and effective feature representations for users, news and locations. Xu et al. [11] propose a news recommendation framework called MobiFeed specially designed for mobile users, which introduces path prediction into location-based mobile news recommendation and recommends location-related news to users in real time according to their moving track. Location-based news recommendation is helpful for users to find news in their neighborhood and capture what is happening in the current surrounding environment. However, the research in this area tends to improve the accuracy of location matching and often ignores the location-aware user preferences.

2.1.3 Social Networks-Based News Recommendation

Social network is a graph composed of individual nodes and edges reflecting specific relationships between nodes, which provides users with a platform to make friends and share information, and plays a role in information dissemination and circulation to a certain extent. Social information usually contains the latest information of users and their friends’ activities, which reflects the dynamics and diversity of users’ interests. Saravanapriya et al. [12] propose a multi-label convolution neural network that predicts users’ multi-label interests by mining social media and recommends the most popular news articles according to the tags that users are interested in. Ashraf et al. [13] model the relationship between users’ social media preferences and news categories, and rank news by users’ interests obtained from social media. Yang et al. [14] integrate knowledge graph (KG) and social network into news recommendation, quantify the social network structure with improved sampling mechanism, and use random walk sampling strategy to obtain neighbors in social networks. Experiments show that news recommendation with social factors can dynamically capture the changes of users’ interests, and then improve the news recommendation effect.

2.1.4 Session-Based News recommendation

Session-based news recommendation aims to model sequential information based on user preferences in a short time and provide personalized reading suggestions for users according to their short-term sessions. Moreira et al. [15] propose a session-based news recommendation deep learning meta-architecture called Chameleon, which uses recurrent neural network (RNN) to model users’ sequence interest. Meng et al. [16] combine environment, breaking news and news content, and propose a session-based context-aware interest drift network (CaIDN), which uses bidirectional attention recurrent network to effectively capture the drift of users’ reading interest from various aspects and improve the dynamics and diversity of users’ interest. However, the existing session-based news recommendation methods focus on extracting features from news articles and interaction sequence information, ignoring the semantic structure relationship between news articles. Sheu et al. [17] propose a session-based context-aware graph embedding framework for news recommendation, which enriches the entity semantics in articles by KG and further refines the article embedding by graph convolution network (GCN).

2.1.5 Multi-modal news Recommendation

The existing news modeling methods usually only learn the news representation from the news text, while ignoring the visual information in the news (such as pictures and animations). In fact, users click on news not only because they are interested in news headlines, but also they may be attracted by multi-modal features, such as images, video or audio. Therefore, learning multi-modal representation by integrating visual and text information is particularly important for news modeling and click prediction. Wu et al. [18] use pre-trained visual language models to encode the news text and the region of interest images extracted from the news images, and propose a multi-modal news recommendation method. Experiments show that news information with multi-modal features can effectively improve the recommendation performance. Xun et al. [19] use visual semantic modeling to capture the visual impression information perceived by users when browsing news, so as to have a deeper understanding of the process of users reading news. Experiences show that news recommendation information with multi-modal features can describe news content more comprehensively and improve the effect and accuracy of news recommendation.

2.2 Difference Between Personalized and Traditional News Recommendation

Different from the recommendation method based solely on non-personalized factors such as news popularity and news freshness, personalized news recommendation needs to deeply consider each user’s interest preference, and provide personalized service according to the matching degree between news content, location, category and user preference. Compared with the traditional news recommendation system, PNR mainly has the following characteristics.

  • Diversity. Diversification of recommendation results has a long-term impact on user experience and participation, and is an important factor in providing high-quality personalized news recommendation. However, most of the existing news recommendation methods only seek to improve the accuracy of recommendation and often ignore the diversity. Therefore, Wu et al. [20] propose a diversity-aware news recommendation method, which generates a news recommendation list with diversity perception in an end-to-end manner and uses diversity perception regularization method to encourage the model to make controllable diversity perception recommendation, thus achieving a good balance between the accuracy and diversity of news recommendation. Because news usually has a certain type of emotional tendency, Wu et al. [21] propose an emotional diversity-aware news recommendation method, which integrates emotional information into news modeling through emotional-aware news encoders and models user interests based on the emotional orientation of candidate news, effectively recommending news with different emotions to users to improve the diversity of news recommendation.

  • Timeliness. News timeliness is life of the news. Transmitting news to readers at the fastest speed is the core of news release and dissemination. Liu et al. [8] design a timeliness module to refine news representation in order to emphasize the impact of news freshness (i.e., the elapsed time between the ‘publish-time’ of the news article and the ‘click-time’ when the user clicked to open it for reading). Moreover, it proposes a multi-task learning framework (to conduct news recommendation and active-time prediction simultaneously) to reinforce news recommendation. Considering the correlation between user interest and time change, Qin et al. [22] use forgetting curve to construct a time-based function and integrate it into user interest modeling for time-weighted update, so as to realize real-time update of user interest modeling and improve the timeliness of news recommendation.

  • Popularity. Current methods usually use news headlines, summaries, entities and other information or add some auxiliary tasks to the multi-task learning framework to predict the click-through rate (CTR). However, there are few methods to integrate the popularity of predicted news and the attention of users to popular news into the prediction results. Wang et al. [23] propose a popularity-enhanced news recommendation (PENR) method, which adds the score of predicting news popularity to the final CTR prediction, and uses news popularity to simulate the tendency of users to pay attention to hot news. Because news popularity is affected by many different factors (such as news content and freshness), Qi et al. [24] propose a method combining news content, news freshness and real-time click-through rate to predict the popularity of candidate news, and predict the popularity of news recommendation in a more comprehensive time perception way. In addition, the popularity-aware user encoder generates user interest embedding according to the content and popularity of click news, eliminates the popularity deviation in user behaviors, and learns more accurate user interest representation to capture the personalized preferences of different users in popular news.

3 Data Collection and Processing

3.1 News Datasets

The statistical information of datasets commonly used is shown in Table 2, mainly including Adressa [25], Digg [26], Plista [27], Globo [28] and MIND [29] datasets, which are, respectively, introduced as follows.

  1. 1.

    Adressa. It consists of news logs collected by Adresseavisen website in three months, including full version and small datasets version. Among them, the full version contains 3,083,438 users, 48,486 articles and 27,223,576 clicks within 10 weeks. The other includes 561,733 users, 11,207 articles and 2,286,835 hits in one week.

  2. 2.

    Digg. It contains 3553 news items collected by the Institute of Information Science of the University of Southern California on the homepage of Digg website in June 2009, including digg_votes table and digg_friends table. Among them, the former has 139,409 users and 3,018,197 votes and the latter has 1,731,658 link relationships between 71,367 users.

  3. 3.

    Plista. It is published in RecSys2013 and constructed from 1,095,323 articles, 14,897,978 users and 84,210,795 reading records collected by 13 news websites in Germany in June.

  4. 4.

    Globo. It is retrieved from Brazilian news portal website called Globo, including 314,000 users, 46,000 articles and 3 million news clicks. It only provides pre-trained news embedding and there is no original news text information.

  5. 5.

    MIND. It is composed of the real news logs of 1 million users collected by Microsoft News website in 6 weeks, including MIND and MIND-small versions. MIND contains 161,031 news articles, 1,000,000 users and 24,155,470 behavior logs. MIND-small has 93,698 news articles, 50,000 users and 230,117 behavior logs.

Table 2 The statistics of common datasets of news recommendation
Full size table

3.2 Data Processing

The above news datasets provide us with the raw data for personalized news recommendation, that is, news and user information. The news information includes news ID, title, body, category, topic and so on. The user information includes user ID, click history, reading time, query history and so on. However, these raw data may have problems such as data exception, missing and redundancy. Therefore, data pre-processing is an indispensable process before prediction modeling. By processing these data, it can be suitable for subsequent news and user modeling and ranking tasks. Pre-processing operations mainly include data cleaning and filtering, extraction and quantification, filling and storage.

  • Data cleaning and filtering. It is a process of checking and verifying data, aiming at eliminating duplicates and noise in the raw data to ensure the quality and accuracy of data.

  • Data extraction and quantification. It mainly includes word segmentation of news texts, eliminating stop words and extracting news entities of topics, places, people and others for accurate news modeling. Quantification refers to transforming unstructured data into structured data to ensure the consistency of data, which is convenient for subsequent processing and calculation. For example, the user’s click behavior can be quantified by matrix representation [30]. Personalized heterogeneous graph [31] constructed from user behavior information can be learned by heterogeneous graph pooling method to understand the differences between different types of nodes and aggregate node characteristics and graph topology information in heterogeneous graph.

  • Data filling. Because there are few explicit data such as comments and scores in news recommendation, we usually use implicit data such as user click behavior [32,33,34,35,36], residence time [37], query and browsing history [38] to fill the data, which effectively alleviates the data sparsity and cold start problems of news recommendation.

  • Data storage. Due to the rich diversity of topics covered in news articles, effective news classification and clustering play a crucial role in news data storage, which can significantly enhance the efficiency and accuracy of news recommendation to a certain extent. Some studies use topic modeling techniques, such as latent Dirichlet allocation (LDA) [39], to identify topics from news texts and classify or cluster them to generate topic vector representation. Some machine learning algorithms, such as k-means clustering [40], Naive Bayesian classifier [41] and support vector machine(SVM) [42], are also used to extract features from news texts and classify or cluster them. Yuan et al. [43] use vector space model and TF-IDF algorithm to construct news feature vectors, and use K-means clustering algorithm to cluster these vectors. On this basis, the classification and clustering of news is convenient to determine the candidate recommended news set better. For a large number of news of the same topics, it is necessary to select several candidate recommended news, so that multiple kinds of news can appear on limited display webpages to ensure the diversity of recommendation results. In addition, users’ reading interest is often affected by news trends, so some hot news [23] and breaking news [16]should also be put into the candidate recommended news set in order to tap users’ attention to mass news.

4 The Technologies of Personalized News Recommendation

Personalized news recommendation is a vital technology that aims to provide individuals with news articles that cater to their specific reading interests. Based on different techniques and characteristics, we present the categorization of PNR technologies in Fig. 3.

Fig. 3
figure 3

The categorization of PNR technologies

Full size image

4.1 News Recommendation Based on Traditional Methods

Traditional news recommendation methods are mainly divided into three categories, which are collaborative filtering-based, content-based and hybrid recommendation methods. The collaborative filtering recommendation algorithm aims at recommending news browsed by similar users or news similar to the clicked news. That is, it recommends news by measuring the similarity among users or news. This method usually only uses descriptive features (such as ID and other attributes) to generate user and news embedding, which has data sparsity and cold start problems. Unlike general item recommendation, due to the fast update speed of news, the latest news is rarely read or commented on by users, which makes the system has insufficient users’ ratings, metadata, and textual content and makes it difficult to gather sufficient user interaction data to create reasonable recommendations. It is also one of the reasons for the cold start problems. Dong et al. [44] use collaborative filtering algorithm to predict user ratings and add hot news parameters to improve the correlation coefficient formula, alleviating the sparsity of user rating matrix data. The content-based recommendation method aims at recommending news similar to the content of the clicked news. The core lies in mining semantic features of the news texts. Liu et al. [45] use Bayesian model to learn user’s interest representation according to the click distribution of news with different topics. However, the over-reliance on users themselves makes it difficult to mine user’s potential interests. Hybrid recommendation approaches refer to the combination of above two or more recommendation algorithms. Bansal et al. [46] integrate topic model, Bayesian model and collaborative filtering method into a unified framework to recommend articles that users may comment on. However, it still has some problems such as data heterogeneity and sparsity.

4.2 News Recommendation Based on Deep Learning

Deep learning (DL) has become a new upsurge in the era of artificial intelligence and has been widely used in news recommendation systems [47]. It can automatically learn effective features from complex content and describe users and news representations by nonlinear network, which solves the problem that traditional recommendation algorithms rely too much on handcraft feature extraction. The models based on deep learning mainly include automatic encoder (AE) [48], convolutional neural network (CNN) [15, 34], recurrent neural network (RNN) [35, 36, 49], attention mechanism (AM) [50,51,52] and so on. These models have shown superior performance in different news and user modeling.

4.2.1 News Modeling

News modeling could capture the features of news articles and understand their rich text content, which is a key step in personalized news recommendation method. The news modeling method based on deep learning aims to automatically learn news representation from the original input. For example, Okura et al. [48] use a denoising automated encoder (AE) to learn news representations from news texts. Moreira et al. [15] utilize CNN network to convolute news content to generate news embedding representation. Zhu et al. [34] use two maximally pooled parallel CNN networks to learn the hidden feature representation of news from news headlines. CNN is widely used in news modeling, but it is difficult to capture long-distance text interaction and is not suitable for long-sequence news recommendation tasks. Therefore, some studies use attention mechanism to extend neural network, and construct news representation by selecting important information, so as to improve the accuracy of text feature extraction. Wu et al. [50,51,52] propose a news recommendation method based on multi-head self-attention mechanism, which can enhance the representation ability of news features by capturing the interaction between distant words. In addition, they also propose to use personalized attention network to learn the semantic representation of news headlines, use self-attention mechanism to learn the semantic representation of words in news headlines and text, and use interactive attention network to model the relationship between headlines and text. In recent years, BERT [53], Transformer [54] and other large-scale pre-training language models have been widely used in news recommendation [55,56,57,58,59]. For example, Zhang et al. [57] connect news texts in series into BERT model, and capture word-level and news-level multi-granularity user–news matching signals to enhance text expression. Huang et al. [58] propose an adaptive transformer model to learn the deep interaction between users and candidate news, and effectively integrate historical clicked news and candidate news to capture their inherent relevance. However, when large pre-training models input multi-domain information, there may be a problem that shallow feature coding for compressing category and entity information is incompatible with deep BERT coding. Therefore, Bi et al. [59] propose a multi-task learning framework, which integrates multi-domain information into BERT to improve the ability of news coding.

Therefore, state-of-the-art news modeling methods can effectively capture the multi-granularity semantics of news articles. They utilize attention mechanisms to focus on important information, weigh the importance of different textual components and emphasize the most relevant information. They utilize neural network architectures, such as CNN or RNN, to capture complex relationships and sequential dependencies. They benefit from pre-trained models for enhancing semantic understanding. By combining these NLP techniques, news recommendation models can capture comprehensive semantics of news content and provide more accurate and effective recommendations. Table 3 summarizes the news information representation and news modeling technologies based on deep learning methods in recent years. Although the above methods based on deep learning can automatically learn news representation, they do not make full use of related entities and their relationship information. Therefore, some studies try to construct graph data structure to mine the potential knowledge-level connections of news, which will be described in detail in Sect. 4.3.

Table 3 Comparison of different news modeling methods based on deep learning
Full size table

4.2.2 User Modeling

User modeling refers to inferring users’ personal interest, which is a key step in personalized news recommendation system. User modeling usually infers users’ interests and preferences from their historical click behavior. For example, Wu et al. [32] use news-level attention networks to learn user representations from the representations of clicked news. Zhang et al. [33] use attention network to aggregate different information of clicked news and candidate news to model users. However, the above method does not fully consider the influence of the sequence information of users’ historical reading, and it can better reflect the change and diversity of users’ interests over a period of time.

In order to further consider the user’s click order, some studies use RNN to model the dependencies of different click orders to better simulate user interests [34,35,36, 49]. Okura et al. [48] use GRU network to learn user representations from the clicked news representations. Zhu et al. [34] use RNN network based on attention mechanism to capture richer hidden sequence features of users’ clicks. However, although the above method enhances the dynamic representation of users’ interests, it is somewhat weak in capturing users’ global interests. Therefore, An et al. [35] propose a hybrid news recommendation method called Neural News Recommendation with Long-and Short-term User Representation, which learns short-term interest embedding of users through GRU network and models long-term interest of users through user ID embedding.

The above methods mainly rely on the information of users’ click behavior to model users. However, users’ click behavior is very noisy, so it is difficult to infer users’ interests comprehensively and accurately only from click feedback. Therefore, some studies incorporate other types of user information to strengthen user interest modeling [15, 38, 60, 61]. One is to add contextual content to enhance user modeling. For example, Moreira et al. [15] introduce context information such as time, device and location, and use UGRNN network to learn user representation. Another approach is to consider multiple types of user behavior. For example, Wu et al. [38] consider a variety of user behaviors such as news click, search query and web browsing, and learn user embedding from each behavior as different interest characteristics. Wu et al. [61] consider user clicks and reading behavior, simulate user clicks preferences from the headlines of clicked news and imitate user reading satisfaction from the subjects of clicked news. In addition, some studies combine a variety of explicit and implicit feedback to infer positive and negative user interests, enhancing the ability of user interest modeling [62,63,64,65]. Wu et al. [62] use strong feedback representation to extract positive and negative user interests from implicit weak feedback, which realizes accurate user interest modeling. Wu et al. [63] propose an implicit negative feedback news recommendation method, which distinguishes positive and negative news clicks according to the reading residence time of user clicks and learns user representations by additional attention networks. However, existing methods usually encode the clicked news independently and then aggregate it into the user embedding, which ignores the word-level interaction between different clicked news from the same user. Qi et al. [66] propose a fine-grained fast user modeling framework, which models user interests from fine-grained behavior interactions, and infers user interests by detailed clues of interaction behaviors. Table 4 summarizes the user information representation and user modeling technologies based on deep learning methods in recent years. Although the above method based on deep learning can automatically learn news representation, it does not fully consider the high-order relationship between users and news interaction. Therefore, some researches attempt to construct graph data structures to mine deeper user interest characteristics, which will be described in detail in Sect. 4.3.

Table 4 Comparison of different user modeling methods based on deep learning
Full size table

4.3 News Recommendation Based on Graph Structure Learning

Graph structure is a nonlinear and complex data structure. In news recommendations, graph structure is usually used to establish the interactive relationship between multiple users and news, in which the high-order connectivity between users and news contains abundant feature information. In recent years, the learning ability of graph neural network (GNN) in graph structure has gradually become prominent, and it has attracted wide attention because of its powerful feature expression ability based on node features and graph structure data [37, 67,68,69]. In news recommendation, graph neural network has strong representation ability in modeling high-order connectivity between users and news. This section mainly introduces personalized news recommendation models based on different types of graph structure data, including user–news interaction diagram, knowledge graph and social relationship diagram.

4.3.1 News Recommendation Based on User–News Interaction Diagram

Bipartite graph is a graph structure that describes the interaction between different users and items. In news recommendation, the user–news interaction graph uses the graph structure information composed of user nodes and news nodes to recommend news that may be of interest to target users through adjacent users, as is shown in Fig. 4. Ge et al. [67] model the interaction between users and news as a graph structure and design a two-hop graph learning module, which aggregates news and users’ neighbors by graph attention network (GAT) to enhance the expression ability of corresponding features. On the basis, Hu et al. [68] disengage the potential preference factors of users by neighborhood routing mechanism, which improves the expressiveness and interpretability of representation. In fact, graph structure can integrate multi-source heterogeneous data in news recommendation system. Therefore, some studies describe users and news information as heterogeneous graph (HG), and use advanced graph learning methods to further enrich user and news graph representation [37, 69]. Hu et al. [69] construct a user–news–topic heterogeneous graph to explicitly model the interaction among users, news and potential topics. The potential topic information can effectively alleviate the sparsity of data and enrich the semantic representation of news. Similarly, Ji et al. [37] incorporate the active time of users on the page into the news representation, and propose a temporal sensitive heterogeneous graph neural network (TSHGNN), which includes two sub-networks. One uses CNN and LSTM variant to learn the reading residence time of users and takes the click sequence feature as the time dimension feature. The other uses GNN to encode high-order structural information by taking the structural features of user–news–topic heterogeneous graphs as spatial dimension features. It makes full use of the time characteristics of user–news interaction, deeply models the dynamic user interests and increases the accuracy and timeliness of recommendation.

Fig. 4
figure 4

The user–news interaction diagram

Full size image

In these above approaches, each user is typically represented by only one node in the global user–news graph. To model user interests more richly, Wu et al. [70] propose a user modeling method called User-as-Graph, which models each user as a personalized heterogeneous graph constructed by user behavior information, and use heterogeneous graph pooling method (HG-Pooling) to learn user interest embedding, which fully models the correlation between user behaviors and provides finer-grained information for inferring user interests. HG-Pooling method not only summarizes the node characteristics and graph topology information in heterogeneous graphs, but also understands the differences between different types of nodes, which learns the user interest representation in heterogeneous graphs in a more efficient, flexible and fine-grained way. Remarkably, the above research focuses on how to extract fine-grained information from user graphs, but does not fully consider the necessary feature interaction between candidate news and users. Therefore, Mao et al. [71] propose a dual-interactive graphical attention network (DIGAT) composed of news graphs and user graphs. In news graphs, Semantic-augmented Graph (SAG) is used to fuse relevant semantic information to enrich the representation of a single candidate news. In user diagrams, the news–topic graph is used to model user history information to represent multi-level user interests. Besides, it designs a two-graph interaction process to perform effective feature interaction between the news graph and user graph, and learn news-user matching representation more accurately. However, many researches often only consider users’ click behavior to model user representation. In order to enrich users’ interest characteristics, Ma et al. [31] construct a multi-behavior user–news interaction graph by considering six different types of user behavior information (no click, click, praise, focus, comment and sharing), and propose a graph-based behavior-aware interactive news recommendation method (GBAN). By constructing a weighted multi-behavior interactive heterogeneous graph, it makes full use of the diversified relationship between users and news, and introduces core features into the behavior graph to measure the concentration of user interests, thus reasonably balancing the accuracy and diversity of personalized news recommendation systems.

Therefore, news recommendation based on user–news interaction diagram can make full use of the node and side information in the graph structure by GNN, and learn the relationship between nodes and the rule of information transmission to enhance the representation ability of the model. Different from general recommendation models, graph-based news recommendation models tailor specifically to news articles and related entities, which captures similarities between news articles, user–news interactions and other related relationships. In addition, news articles often have temporary or sequential dependencies. So graph-based news recommendation models can incorporate time information by introducing time-aware edges or considering the order of user interactions, which enables the models to capture the evolving user interest and dynamic preferences over time.

4.3.2 News Recommendation Based on Knowledge Graph

Knowledge graph (KG) contains powerful relational ability and rich semantic features, as is shown in Fig. 5. By leveraging external knowledge through KG, we can enrich the semantic understanding and fully explore the underlying knowledge connections within news articles, resulting in more fine-grained information representation. Therefore, incorporating KG into personalized news recommendation systems can further improve the accuracy, diversity, and interpretability of recommendations. Wang et al. [72] utilize knowledge-aware convolutional neural network (KCNN) to learn news representation from news headlines and entities, which combines semantic and knowledge layer representation of news. It regards words and entities as multiple channels and keeps their alignment during the convolution process. Multi-channel alignment mechanism eliminates the heterogeneity of words, entities and entity context embedding space, captures the potential knowledge-level relationship among news more comprehensively and obtains richer news content. Similarly, Ren et al. [73] use the above KCNN component to extract news features, and combine them with knowledge graph to construct a dual attention network, which comprehensively considers word-level and news-level attention mechanism integrating words, entities and entity contexts. Besides, it uses multi-head attention mechanism to fuse the two, which better models the diversity of user interests.

Fig. 5
figure 5

The structure of knowledge graph

Full size image

To fully consider the importance of high-order neighbor information, Sheu et al. [17] propose a context-aware graph embedding (CAGE) for session-based news recommendation, which uses one-hop neighbors of entities to construct sub-graphs to generate news semantic embedding, and further refines news article-level embedding combined with neighborhood structure information between articles. However, these embedding vectors mainly condense the low-level interaction between entities, which is difficult to identify whether two entities appear in the same news. Lee et al. [74] introduce topic relationship and design a knowledge level news encoder, which constructs a topic-rich sub-graph from the news title by adding two hop neighbors between entities and learns the representation by using graph neural network. By adding two-hop neighbors, it fully mines the topic relationship between entities and enriches the modeling of entity correlation. In addition to news headlines, it is also applicable to any type of text information, such as text content or news summaries.

However, these models only use single data such as news headlines, which do not make full use of news bodies, summaries, categories and other contents that can provide contextual information for headline entities. Therefore, the news semantic space is not rich enough. Then, some researches try to enrich the semantic features of news in knowledge graphs by constructing a multi-feature learning framework or combining news representations from multiple perspectives. Sun et al. [75] combine various news features (title, abstract, category, subcategory) with linked external entities to learn news representation, and select important words and features by using word-level and feature-level attention networks, which enriches the expression ability of news features and improves the accuracy of news recommendation. Xu et al. [76] establish a multi-view news framework by introducing a variety of news information, including news headlines, abstracts, categories and knowledge graph features, using the Knowledge Graph Interaction Network (KGIN) and multi-head attention mechanism to learn news representation and capture the relationship between entities and their neighbors.

Most methods usually learn user representations from the clicked news to reflect their existing interests, ignoring the potential interests and paying little attention to news that users may be interested in in the future. Therefore, Qiu et al. [77] propose a graph neural news recommendation method with user existing and potential interest modeling (GREP). It first digs the headlines of historical clicks to encode the existing interests of users, and explores the potential interests of users by finding out the entities related to the entities in historical clicks in the knowledge graph, thus enriching the expression of user interests. By using knowledge graphs to learn news articles and user representations, to a certain extent, it improves the expression ability of news semantics and user interest features. However, different users will have different interests in the same news article. If we can identify the entities related to user interests directly, the efficiency and interpretability of news recommendation will be further improved. Therefore, Tian et al. [78] propose a news recommendation method based on knowledge pruning recurrent graph convolution network. It does not model news article representation, but directly uses relevant entities in news to model user’s interest representation. In addition, not all knowledge auxiliary information provided in the knowledge graph is related to user interests. So it directly models user interests by pruning a large number of irrelevant knowledge graph information, so that user interests can be matched more accurately.

Knowledge graph, as a richer structured representation, contains different types of entities and relationships. Different from general recommendation models, KG-based news recommendation models focus on creating representations of news-related entities in KG, such as news topics, events and attributes. By mining the semantic relationships between domain-specific entities, KG enables the system to make more accurate and context-relevant news recommendations. Secondly, KG helps to identify the links between different topics, categories or fields. By looking for entities related to the entities in the user’s historical click news in KG, the system can infer implicit relationships, explore the potential interests of users, and recommend articles that may not be related to the user’s direct interests, so as to facilitate the prediction of the user’s future interests. Finally, with the emergence of new articles, topics or entities, they can be integrated into KG, so that the recommendation system can continuously update the latest information and provide users with personalized news recommendation services with more dimensions.

4.3.3 Social Information-Based News Recommendation

Social information usually contains up-to-date information about users and their friends, which helps to reflect the dynamics and diversity of user interests, as is shown in Fig. 6. Yang et al. [79] combine social network with knowledge graph, and adopt random walk sampling strategy to obtain neighbor information of target user in social networks to enrich user interest modeling. Zhu et al. [80] design a social information encoder, which first extracts the hidden features of user interests and friend relationship, then constructs a social relationship graph and inputs it into the graph convolution network to learn the embedding of user node information and friend relationship side information, and finally generates user interest representation. In addition, it also considers the common news clicked by public users, which effectively alleviates the cold start problem faced by traditional models. Experiments show that the news recommendation incorporating social information can obtain more abundant user information and further model more dynamic and diverse interests.

Fig. 6
figure 6

The structure of social networks

Full size image

This section mainly introduces the research on news recommendation based on graph structure learning. Table 5 summarizes the key technologies of different models on graph structure learning.

Table 5 Comparison of different modeling methods based on graph structure learning
Full size table

5 Personalized Ranking, User Interface and Evaluation Metrics

Due to the limitation of the amount of news that recommendation results show and the influence of different positions on the recommendation effect, personalized news ranking and recommendation results display also play a vital role in improving the performance of news recommendation systems.

5.1 News Ranking

News ranking aims to rank candidate news according to users’ specific preferences. Common news ranking methods mainly include two categories, relevance-based and reinforcement learning-based.

Relevance-based news ranking methods lie in how to accurately measure the similarity between candidate news and user interests. For example, Qin et al. [22] use the Jaccard similarity to measure the relevance between the candidate news and user preferences. Wu et al [50] compute the inner product between news and user vectors to predict the click probability. However, using only dot-product to compute the relatedness between candidate news and user features makes recommender tend to recommend news that is similar to those previously clicked news and ignores the importance of candidate news itself. For example, some breaking news and hot news might be popular with users who have different interests. Xin et al. [81] design a gated aggregator to predict the click probability, which combines candidate news and user features adaptively. Sun et al. [82] design a multi-level prediction module and predict click probabilities at three different semantic levels, including the text level, category level, and subcategory level, which fully explores the fine-grained matching between candidate news and user preferences. Wu et al. [83] calculate the click score by weighting the positive correlation between candidate news and global interests embedding and the negative correlation between candidate news and immediate interest embedding, so as to fully consider the time diversity of the click sequence behavior to make better updates. Zhang et al. [33] design an intra-domain and cross-domain matching mechanism, which combines the complementary information of different fields (such as title, abstract and text) to capture the multi-domain matching representation between each clicked news and candidate news, obtaining fine-grained semantic matching information. However, these methods usually only consider the correlation between candidate news and user interests, which makes it difficult to provide diversified news recommendation to users. Wang et al. [23] and Qi et al. [24] both introduce news popularity score to predict users’ attention to hot news and combine the personalized matching score, which ultimately affects the click probability of the candidate news. Wu et al. [61] integrate click prediction based on news headlines and satisfaction prediction based on news content to recommend news that users may click on and be satisfied with, which provides users with more diversified suggestions.

Compared with the relevance-based method, reinforcement learning (RL)-based ranking method is more suitable for improving long-term user experience. It aims to find the best ranking strategy to maximize long-term rewards. Li et al. [84] adopt contextual bandits method and adjust the news selection strategy based on context information and user feedback to maximize the total number of clicks. On the basis, Shen et al. [85] add a deep neural memory enhancement mechanism and model and track the historical state of each user according to their previous interaction which leads to obtain user preference only through a small amount of interaction information. Zheng et al. [86] use DQN (Deep Q-Learning Network) network to capture users’ dynamic interest, take user activity as a supplement to reward to get more user feedback information, and combine effective exploration strategy to avoid recommending too much old news. Similarly, Song et al. [87] design a deep Q-learning framework called DEN-DQL based on double exploration networks to deal with both current and future reward, and combine a more effective exploration strategy to avoid recommending uninterested news. Therefore, RL-based news ranking method is more suitable for exploring more diversified user interests and improving long-term user experience and participation.

5.2 User Interface

A friendly user interface with a reasonable layout is also an important component of the news recommendation system, which can not only enhance the user experience but also make it easier to get user feedback timely. There are two main types of user interfaces, namely website side and mobile side.

The website side is suitable for users in relatively quiet places such as offices and coffee shops. It has many advantages, including large screen size, rich content, better visual effects and embedded advertisements. The famous domestic news websites include Tencent, NetEase, Sina, Sohu and so on. The well-known news Websites abroad are Yahoo! News, Guardian, NBC, Google News, etc. Different news website layouts have different characteristics. For example, Sohu embeds a user feedback module in the lower right corner of its homepage, so that users can feedback their reading experience at any time and make suggestions on products. After users click the button on More like this or Less like this, Yahoo! website will pop up keywords for users to choose and feedback. At the same time, website news contents are very rich, including entertainment, military, technology, education and other different types of news. Besides, they may also have today’s breaking news, hot review news and other columns. Users can flexibly choose news that matches their interests or the latest and hot review news, fully mining users’ potential intentions and reading interests to enhance the diversity of news recommendations.

The mobile side is suitable for users to browse news at fragmented time and place. With the limitation of screen size and display information, different mobile news devices need to adjust the layout according to the screen sizes of themselves, make news recommendations adaptively and ensure a better user experience. In addition, the location relevance and device portability of the mobile side enable users to browse news anytime and anywhere, which is helpful for obtaining user feedback accurately and timely. For example, TikTok designs two modules called personalized recommendation and intra-city recommendation at the top. It lists most of news that users are interested in, so users can quickly update the recommended contents by modifying location. In addition, the mobile side usually combines social networks and social media to provide users with a platform on which they can make friends and share information, which plays an important role in news promotion and dissemination. By summarizing, the comparison results between web side and mobile side are shown in Table 6.

Table 6 Comparison between website side and mobile side
Full size table

5.3 Evaluation Metrics

The performance of personalized news recommendations is mainly reflected in the accuracy, diversity and response time of recommendation results. Classical evaluation indexes include precision, recall, F1-score, ROC (receiver operating character), AUC and so on. At present, the specific calculation formulas of F1 and AUC indexes which are widely used are as follows:

$$\begin{aligned} \textit{ F1-Score }=\frac{2 \times \textit{ Precision } \times \textit{ Recall }}{\textit{ Precision }+ \textit{ Recall }} \end{aligned}$$
(1)

where Precision indicates the click probability of the user on the recommended result. Recall shows the probability that news that users are interested in will be recommended.

$$\begin{aligned} A U C=\frac{\left. \mid \left\{ (i, j) \mid {\text {Rank}}\left( p_i\right) <{\text {Rank}}\left( n_j\right) \right) \right\} \mid }{N_p N_n} \end{aligned}$$
(2)

where \(N_p\) and \(N_n\) are the numbers of positive and negative samples. \(p_i\) is the predicted score of the i-th positive sample and \(n_j\) is the score of the j-th negative sample.

Because of the huge amount of news, users usually pay more attention to the news at the top of the recommendation list, some research methods weigh the recommendation results according to the ranking list. Common evaluation metrics based on ranking measurement include MAP (mean average precision), MRR (mean reciprocal rank) and NDCG (normalized discounted cumulative gain), etc. The specific calculation formulas are as follows:

$$\begin{aligned} AP_u@K=\frac{1}{\left| pos_u\right| } \sum _{i\in \textit{pred}_u @ K} \frac{p_i^{\left( \textit{pred}_u @ k \cap \textit{oos}_u\right) }}{p_i^{\textit{pred}_u}} \end{aligned}$$
(3)

where \(p_i^{\text {pred}_u}\) indicates the position of item i in the forecast list. \(p_i^{\left( \text {pred}_u @ k \cap pos\right) }\)expresses the position of item i in Ture Positive (TP) example set. If item i is not a TP example, return 0.

$$\begin{aligned} M A P @ K=\frac{\sum \limits _{u \in U} A P_u @ K}{|U|} \end{aligned}$$
(4)

where U represents all users. It means that the AP scores of all users are averaged. The larger the MAP value, the more related news in the recommendation list and the higher the ranking of related news.

$$\begin{aligned} M R R=\frac{1}{|U|} \sum _{u\in U} \frac{1}{\textit{rank}_u} \end{aligned}$$
(5)

where \(u\in U\) shows traversing all users. \({rank}_u\) refers to the location of the first TP example in the user recommendation list.

$$\begin{aligned} \text {NDCG} @ K=\frac{\sum \limits _{i=1}^k\left( 2^{r_i}-1\right) / \log _2(1+i)}{\sum \limits _{u \in U} 1 / \log _2(1+i)} \end{aligned}$$
(6)

where \(r_i\) is the relevance score of the i-th news. If user clicks on the i-th news, the value of \(r_i\) is 1.

To obtain higher user satisfaction, we should also consider other news recommendation indicators, such as topic diversity, novelty, fairness and popularity. Gabrilovich et al. [88] evaluate the recommendation results through novelty, which refers to the user’s ability to recommend non-hot news. The most direct way to measure novelty is according to the similarity between news, that is, the smaller the similarity between the news in the recommendation list and clicked news, the higher the novelty. Zheng et al. [86] use an in-list similarity (ILS) function to measure the diversity of results. Wu et al. [21] use a set of emotional diversity metrics to measure the emotional differences between clicked news and candidate news. Wu et al. [89] employ the accuracy of sensitive attributes (such as gender) prediction as a fairness indicator to weigh whether a news recommendation system is fair to different user groups or news publishers. In addition, some research methods usually use two or more evaluation metrics to comprehensively evaluate the performance of news recommendation systems and further improve the user’s experience, such as combining AUC, MRR and NDCG [71, 75].

6 Challenges and Future Directions

With a comprehensive review of the existing methods, we can see that personalized news recommendation technology has made substantial progress in recent years. However, there are still many challenges and unresolved problems. In this section, we will discuss the challenges and future research directions of PNR.

6.1 Challenges

6.1.1 Privacy Protection

Existing news recommendation methods usually rely on centralized storage of user behavior data for model training. Because of the privacy sensitivity of user behavior, centralized storage of user data may increase privacy issues and risks. Federated learning is a privacy protection framework, which allows multiple clients to cooperate in training models without sharing their private data. Qi et al. [90] propose a unified privacy-protected news recommendation framework, which utilizes user data stored locally in user clients to train the model and provide services to users in a privacy-protected way. However, the computational cost of directly learning the existing news recommendation model in a federated manner is high for the user client. Yi et al. [91] propose an efficient privacy protection news recommendation method based on federated learning framework. Instead of training the whole model, the news recommendation model was decomposed into a large news model maintained in the server and a lightweight user model shared on the server and the client.

6.1.2 Fake News

Although social media platforms have made great efforts in fake news detection, false information still spreads widely on media platforms such as Twitter. Therefore, it is necessary to adopt a fake news mitigation strategy. Most of the existing fake news processing work focuses on the mitigation of the whole social network, while ignoring the specific mitigation strategies to prevent users from sharing fake news. Wang et al. [92] propose a authenticity-aware and event-driven news recommendation model, which recommends corrective real news to users to effectively expose fake news. In addition, online social groups usually disseminate false news materials in the form of sharing or forwarding. Stitini et al. [93] provide an overview of traditional techniques for detecting fake news and modern methods for multi-classification using unlabeled data, aiming at detecting fake news sent on social networks to improve the trust quality and transparency of social network-based news recommendation systems.

6.1.3 Debiasing

Debiasing is an important responsibility to improve news recommendations. The deviation of news clicking behavior may bring noise to users’ interest modeling and model training, and then damage the performance of news recommendation model. Besides, the deviation of user behavior data coding is propagated to the recommendation model, which will be further amplified in the recommendation cycle. Therefore, Yi et al. [94] propose a personalized news correction method called DebiasRec with deviation perception ability, which is used to process deviation information to achieve more accurate user interest reasoning and model training. Different users may be affected by the same biased information in different ways, so considering users’ personalized preference for biased information is also helpful in eliminating the influence of biased information. Wu et al. [95] use a bias-aware click model to capture the impact of news position bias on web pages, so as to infer users’ unbiased interest in candidate news more accurately. Wu et al. [89] propose a fair-perceived news recommendation method with decomposition antagonistic learning and orthogonal regularization to alleviate the homogeneous news recommendation caused by the bias of sensitive user attributes (such as gender and identity). In addition, many different types of biases (such as exposure biases and selection biases) are rarely studied in the field of news recommendation. Therefore, it is of great significance for future research to understand how different deviations affect user behavior and learn how to eliminate their influence on model training and evaluation.

6.2 Future Directions

6.2.1 News Recommendation with More Accurate User Modeling

How to achieve more accurate user modeling has always been a challenging research focus in PNR. First, the existing news recommendation methods usually only learn single user embedding from users’ historical behaviors to express their reading interests. However, users’ interests are usually diverse and fine-grained and single user embedding is not enough to fully model users’ interests. Li et al. [96] propose a poly attention scheme to extract multiple interest vectors for each user and introduce agreement regulation to make the interest vectors more diversified. Wang et al. [97] design a novel parallel interest network to extract potential multiple interest embedding to prepare for subsequent news sequence modeling. Therefore, how to refine user embedding for multi-granularity user interest modeling is very worthy of study. Secondly, users’ interests are dynamic. Existing methods usually learn users’ long-term and short-term interests, respectively, and recommend candidate news related to recently clicked news. However, users may also read news that does not fit their usual interests, such as breaking news and hot news. Nguyen et al. [98] propose a model to combine user-specific and global news features for personalized and global recommendations. Zhu et al. [80] take into account the common news clicked by others to enrich user interest modeling. However, users’ interests often change with time. Wang et al. [99] propose a novel intention-aware model to capture users’ reading intentions and interest drifts for accurately predicting the next piece of news the users may be interested in. In fact, fully exploring user’s interest drift process will greatly improve the performance of news recommendation. So further fully considering the phenomenon of user interest drifts and accurately predicting their future interests is still a direction worth exploring. Third, there is usually noise in the clicking behavior of users. On the one hand, users will click on news by mistake that they are not interested in, so there will be some interactive noise. Wang et al. [100] use frequency-aware contrastive learning to seek the essential features of users to improve the robustness of noisy data. On the other hand, the recommendation system may recommend a variety of news that users are interested in, but only click on one of them, so that noise is generated with negative implicit feedback. Hu et al. [65] design denoising aggregators to refine the positive and negative implicit feedback sequences to further mitigate the impact of noise which is common in implicit feedback. Noise always exists in users’ click behaviors, so we need to further consider comprehensively various behaviors and feedback of users in order to explore users’ potential interests reasonably.

6.2.2 News Recommendation with More Multi-modal Information

In most news recommendations, only text content is often used, while visual information is ignored. In fact, users choose news not only due to the attention in headlines but also the attraction of multi-modal features (such as images, audio and video). Xun et al. [19] design a visual semantic modeling module to capture the visual impression perceived by users when browsing news, fully understand the process of users reading news, and provide an extended MIND data set by adding snapshot expression images to promote the future research of multi-modal news recommendation. At present, there are still few researches on news recommendation based on multi-modal. Therefore, introducing multi-modal features and modeling the multi-modal relationship between candidate news and click behaviors is a new exploration direction.

6.2.3 News Recommendation with Information Interaction

Enhancing the information interaction in personalized news recommendation is mainly manifested in the following three aspects. First, the existing news modeling usually compresses different views of news, such as title, abstract and category, into a single news vector, which makes it difficult for different views in different news to interact with each other and is not enough to fully model news. Zhang et al. [101] propose a multi-operator attention to fully consider global interaction among different views and use SC-CNN to capture the interaction between temporal information and local view. Second, most matching works usually learn the news representation of each candidate news independently, but ignore the interaction of multiple candidate news displayed to users together, which makes it hard to distinguish them better. Sun et al. [82] design a candidate interaction module to obtain competition information among different candidate news, extract the unique features of each candidate news and further model the interactions of candidate news explicitly, which could help distinguish the clicked candidate news from unclicked ones. Third, current researches usually encode the news clicked by users independently and aggregate them to generate user embedding. However, these methods ignore the interaction among different news clicks from the same user, which contains a wealth of detailed clues. Qi et al. [66] put forward a fine-grained and fast user modeling framework to capture behavior interactions within and between news, which further excavates the fine-grained interests of users. In the future research, we should not only study how to achieve more accurate user and news modeling, but also pay more attention to different types of interaction information between users and news, which is also essential to improve recommendation performance.

6.2.4 News Recommendation with Richer Context Information

Personalized news recommendation technology is also influenced by different types of contextual information to some degree, such as reading time interval, location, social network, news life cycle and emotions, which is of great significance for mining user interests and modeling user preferences. Yun et al. [102] combine the extracted emotional features with news content features, and discuss the influence of emotional information in user modeling. Meng et al. [60] put forward a deep common attention network called DCAN, which combines the attention of user preferences and news life cycle to simulate the dual influence on users clicking news. In addition, social network and media also play an important role in the sharing and spread of news, which helps to fully mine users’ potential and diverse interests. Therefore, constructing a heterogeneous social network to fully integrate the information and interaction of different users is a promising research direction in the future.

6.2.5 News Recommendation with Diversity

While diversity is considered in general recommendation models, diversity-based news recommendation models place a stronger focus on providing diverse recommendations due to the wide range of news topics and sources, catering to the users’ varied interests [20, 21].There are three main research directions to improve the diversity of news recommendation. The first is news recommendation with temporal and spatial diversity, which recommends news different from historical clicked news to make the recommendation results better meet the user’s diversified preference. Wu et al. [83] propose a temporary diversity-aware sequential news recommendation method that helps to recommend candidate news that is different from recently clicked news to increase the diversity of users’ interests. The second is fine-grained diversity, which can not only diversify the content and topics of news, but also integrate various factors such as publishers, places, opinions and emotions to provide higher quality and diversity-aware news recommendation results. The third is to fully balance the accuracy and diversity of personalized recommendation system. Ma et al. [31] propose a graph-based behavior-aware network, and introduce the core features and relevance features of behavior graphs to measure the concentration of users’ interests, so as to meet users’ needs for news diversity. In addition, some models could try to use specific multi-objective optimization methods to design diversity loss function, such as weighted multi-objective optimization, multi-objective adaptive balance strategy and evolutionary algorithm, to balance the accuracy and diversity of news recommendation. In conclusion, improving the diversity of news recommendations is still a hot topic in the future.

6.2.6 News Recommendation with High Efficiency

Most researches ignore the efficiency of the proposed algorithms or technologies, so we should consider the efficiency and accuracy of the expected results at the same time, obtain acceptable response time and evaluate each technology for offline and online real-time mode, so as to clarify the technical efficiency of using a single or multiple different datasets. Besides, we can also try the method of knowledge distillation to compress large-scale models to improve the efficiency of news recommendations.

7 Conclusion

We review the latest research progress of personalized news recommendation. Firstly, the characteristics and workflow of personalized news recommendation system are introduced. Secondly, the framework of personalized news recommendation system is described, and the knowledge and technologies involved in each part are expounded. Among them, the news recommendation technology based on graph structure learning is the focus of our study. Then, the evaluation metrics of personalized news recommendation system are introduced. Finally, the challenges and prospects faced by the current researches are pointed out.

This paper has two main contributions. Firstly, the researches on personalized news recommendation are analyzed, and the related technologies based on graph structure learning are discussed in detail. Secondly, Sect. 6 puts forward the challenges faced by PNR and the future directions. At present, there are still some difficult problems in PNR, such as recommendation performance evaluation, model fusion and optimization, privacy protection, layout design, system commercial value evaluation and so on. The research on personalized news recommendation needs further development, from static features to dynamic interest drifts, from single modal to multi-modal and from personalization to combining personalization and diversity. More technologies such as GNN and knowledge graphs, more context information like click time, location, and social relationship can be explored, and experience from related other recommendation fields(such as session recommendation, sequential recommendation)can be borrowed. In addition, we also need large-scale diversified datasets, such as the datasets containing location information to expand the researches on location-aware news recommendation to adapt to the variability of the environment, which will be a new breakthrough in the future researches.