As internet technology advances, people′s news consumption habits are gradually shifting from traditional media outlets such as newspapers and television to the internet. Online news services are gaining popularity due to their convenience; however, they also pose the challenge of information overload. To address this, personalized news recommendation has become imperative. Nonetheless, effectively leveraging existing information to mitigate the cold-start problem and to accurately model both news representations and user interests remains a pressing issue for news recommendation systems.The integration of multi-information fusion and knowledge enhancement, drawing upon external knowledge bases, offers a promising solution to the user cold-start problem and enhances the efficacy of news recommendation systems. Traditional approaches in news recommendation tend to rely on textual information such as news headlines and content, often overlooking the potential of nontextual information such as user IDs, news features, and news popularity. Furthermore, the completeness of existing knowledge repositories remains a challenge. To address these concerns, this study introduces a novel news recommendation approach termed MIFKE (Multi Information Fusion and Knowledge Enhancement). This approach integrates both news textual and non-textual information.By incorporating external knowledge, it enhances the existing knowledge base, and finally models news content and user interests accurately to complete the recommendation process.The key contributions of this paper are as follows: 1) Comprehensively utilizing both the textual and non-textual information, and introducing a mixed-attention mechanism to facilitate more thorough information fusion. 2) Incorporating external knowledge bases and enhancing the user-entity association graphs with two additional types of relational connections for knowledge enhancement. 3)Validating the efficacy of the proposed MIF-KE approach in recommendation on public news datasets and demonstrating its ability to effectively mitigate the user cold-start problem in news recommendation. The first part of this paper introduces the research background and significance. It analyzes the use of textual information by existing news recommendation algorithms, underscores the importance of non-textual information, and emphasizes the necessity of combining textual and non-textual information. Finally, it explores the application of entity information in news recommendation and analyzes existing methods based on entity information. The second part details the proposed news recommendation framework, including problem definition, model framework, news encoder, user interest modeling, and the prediction and training of MIF-KE. The third part consists of the experimental analysis. This study performs data preprocessing on the mainstream news dataset MIND to obtain MIND_large and MIND_small, which are used to verify the effectiveness of MIF-KE and its ability to alleviate the cold-start problem, comparing it with eight other recommended methods. The experimental results indicate the following: 1) Deep learning-based methods outperform traditional machine learning methods because they can extract deeper user and news features and learn more complex feature representations. 2) Attention-based methods and method based on long-term and short-term interest both achieve promising results. Among them, NPA, based on news headlines and utilizing attention mechanisms to assign weights, outperforms general deep learning methods. NAML and LSTUR further integrate news content, categories, user IDs, and other information, outperforming NPA. This demonstrates the importance of multi-information fusion. 3) Graph structure can uncover potential complex relationships among user behaviors.4) MIF-KE utilizes multiple types of information (news headlines, content, categories, popularity, user IDs) and fully integrates them using mixed attention. Additionally, MIF-KE incorporates external knowledge bases and enhances knowledge through user entity graphs. While effectively making recommendations, this approach also mitigates the cold-start problem.This paper investigates the integration of word features, the introduction of non-textual information, the fusion of different types of information, and the impact of knowledge enhancement on the performance of MIF-KE. Experimental results show that: 1) Compared to single information extraction methods, word representations combining local and global contextual representations can improve recommendation effectiveness. Combining news headlines with content also facilitates accurate news modeling. 2) The fusion of multiple non-textual features can extract more valuable news representations, especially news categories and popularity, which significantly improve the recommendation effectiveness. 3) Knowledge enhancement, that is, the incorporation of two types of connection relationships, can improve the recommendation effect overall. 4) The mixed attention mechanism can extract more relevant information from text and learn news representations under the crossinfluence of various features. This paper also analyzes certain important parameters, including the number of layers in graph convolutional neural networks and the threshold of structural similarity scores. Experimental results demonstrate that: 1) Multi-layer GCN can extract higher-order information and effective node features, improving the recommendation effectiveness, but excessive layers may cause over-smoothing issues. 2) Excessive thresholds can reduce the number of effective connections, resulting in suboptimal model performance. Finally, through comparative experiments, this paper validates the ability of MIF-KE to alleviate the cold-start problem.