1 Introduction
With the advent of the digital age and the widespread use of smart mobile devices, the amount of information available to people has increased dramatically, leading to a growing problem of information overload. To address the drawbacks of information overload and recommend truly useful information to users in a personalized way, recommendation systems have become a powerful tool for solving this problem, with their applications becoming increasingly widespread and attracting considerable attention and research. Faced with massive amounts of information, recommendation systems can personalize recommendations based on users' information needs and interests, suggesting information and products that users may be interested in. Compared to traditional search engines, recommendation systems can automatically discover users' interests by collecting their preferences and learning, eliminating the need for users to manually input their interests. This makes recommendation systems simpler and easier to use while guiding users to discover their information needs.
Despite long-term development and evolution, recommender system algorithms still have some shortcomings. For example, limited data collection capabilities cannot cope with the current massive amounts of information, and poor continuity in the data collection process leads to severe data sparsity. This inevitably makes it difficult for recommender system algorithms to analyze user behavior and mine user interests. At the same time, most existing recommender system algorithms focus on analyzing user and recommended product attributes to assist in recommendations, often neglecting to consider the user's context, such as the type of application the user previously viewed, the number of times the user has launched a particular application, and the dwell time, resulting in poor recommendation performance. Furthermore, recommender system algorithms also face the problem of recommendation homogenization. Most recommender systems are better at recommending resources similar to the user's historical interests; the recommended content lacks novelty, and repeated recommendations of similar resources fail to guide users to discover more new potential interests, thus limiting the recommender system's capabilities and compromising its accuracy.
To address the limitations of traditional recommendation systems and their lack of consideration for the user's context during the recommendation process, this paper aims to further optimize traditional recommendation system algorithms by proposing a multi-dimensional application recommendation system based on user feature clustering and combined scenario features. This system will provide personalized application recommendations based on different user groups, application software characteristics, and user context information. Specifically, firstly, singular value decomposition (SVDecomposition) is performed on the feature matrix of the user group to obtain a dimensionality-reduced and denoised user vector representation. SVDecomposition of user group features removes less important components from the data, thus simplifying the representation of user features. This simplification also makes the proposed recommendation system more efficient. Secondly, hierarchical clustering analysis is performed on the feature vectors of the user group to cluster users with similar user features, and the mean vector of each user group is used as its feature representation. Based on this, the algorithm combines user context factors, taking the user as the starting point and the context as a supplement, and predicts the best application recommendation list based on the Bayesian model, thereby achieving more optimized application software recommendations for users and improving the accuracy and efficiency of users' use of massive amounts of information.
Regarding the performance and scalability of the recommender system, to meet the demands of high-concurrency processing of massive amounts of data, this paper uses a distributed cluster server to run the proposed system, which greatly improves the performance in data storage and model computation. Furthermore, the computing cluster gives the recommender system good scalability and plays a crucial role in its stability during operation.
To verify the effectiveness of the proposed algorithm, this paper combines the proposed user feature hierarchical clustering based on singular value decomposition with contextual feature information, and compares the recommendation quality and model quantification performance of the Bayesian model recommendation system algorithm with traditional recommendation system algorithms, thereby verifying the effectiveness and advantages of the system.
Section 2 introduces the relevant theoretical research of the algorithm involved; Section 3 describes the details of the proposed algorithm and its innovation based on existing research; Section 4 introduces the architecture and specific implementation of the recommender system; Section 5 presents the experimental verification results of the recommender system in this paper; and the final section summarizes the paper.
2. Algorithm Introduction and Related Theoretical Research
2.1 Application of Clustering Algorithms in Recommendation Systems
Clustering algorithms are exploratory analytical methods that do not require pre-defined clustering criteria. Starting from sample data, clustering algorithms automatically discover and cluster the attributes of spatial entities, grouping samples that initially lacked categories into different classes (clusters). The goal is to ensure that samples belonging to the same class are similar to each other, while samples belonging to different classes are sufficiently distant in terms of features. Furthermore, clustering algorithms have good scalability for large datasets, making them extremely suitable for recommendation systems with high data demands. In addition, the interpretability of clustering results and their high efficiency provide a solid foundation for their application in recommendation systems. Shinde and Kulkarni proposed a personalized recommendation system algorithm based on user rating clustering. This algorithm first categorizes users according to their product ratings, and then provides personalized product recommendations to user groups with similar ratings, interests, and preferences, thus making the recommendation system more accurate. Verification has shown that clustering algorithms effectively improve the quality of product recommendations in recommendation systems. Hsu uses a clustering algorithm in its recommendation system for recommending English learning courses to students. It first clusters students with different learning habits into different student groups, and then optimizes the recommendation system algorithm to recommend suitable English courses to student groups with different learning habits, thus making the recommendation system more human-centered.
However, the clustering algorithms mentioned above have two drawbacks when applied to recommender systems. First, none of these methods use matrix factorization to reduce the dimensionality and denoise the user feature information before clustering. This results in redundant and weakly representative features with high dimensionality, thus impairing the efficiency of the recommender system. Second, these methods do not fully utilize the contextual information of the user, limiting the recommender system's ability to make recommendations.
2.2 Application of Contextual Information in Recommender Systems
Contextual information can assist recommendation systems in analyzing personalized information such as a user's background and interests. In recommendation system models, context can be additional information reflecting a user's decision-making, such as the number of times a user opens a certain type of application, the length of time a user stays in a certain type of application, and the types of applications the user previously used. For example, the applications a user previously used can help predict the applications the user might use next. By acquiring contextual information about the user in real time, and building upon existing user and application data, the system can analyze the user's information needs in a more humanized and intelligent way, thereby improving the effectiveness of application recommendations.
However, current context-based recommender algorithms all rely on user context information to analyze the features of recommended products, rather than analyzing the user's own preferences. Hu et al., in their recommender system for television programs, utilized user context information such as browsing history, purchase history, search history, and even mouse movements to analyze the features of television programs and better recommend them to users. Abbas et al., on the other hand, used user ratings and reviews of YouTube videos to obtain better feature analysis and positioning of videos, thereby assisting the recommender system in recommending more accurate and suitable YouTube videos to users.
However, the aforementioned algorithms all utilize user contextual information to analyze product features, while existing recommendation system algorithms rarely use user contextual information to analyze user preferences. Furthermore, user ratings and reviews not only require additional user participation, but different users also have different rating and review criteria. This makes the collection and utilization of such user contextual information difficult, thus impacting the recommendation system's effectiveness.
3. Algorithm Design
To provide more accurate and efficient application recommendations for users, this paper proposes a multi-dimensional recommendation system based on user feature clustering and contextual features. First, singular value decomposition is performed on user features. Then, hierarchical clustering is applied to the decomposed user features, and the mean vector of each user class is used as the feature information representing that class of users. Finally, combining the user's contextual information, a Bayesian model is used to predict applications that the user may be interested in and sort them in descending order of probability to form an application recommendation list.
Figure 1. Schematic diagram of user-feature matrix M
3.1 Singular Value Decomposition of User Features
First, the user features required by the recommendation system will be initialized in matrix form. Each user is represented by a T-dimensional vector containing T features (A1 - AT), such as age, gender, interests, and location. Therefore, N users (U1 - UN) will form an N-row, T-column user feature matrix, as shown in Figure 1.
However, for the T feature dimensions representing user characteristics, not every feature dimension can effectively represent user information. Some features contain noise or a large amount of redundant information, thus lacking good representational ability. Furthermore, different features have different levels of importance; not all feature dimensions have the same importance when representing users. In addition, the initial feature dimension T may be large, resulting in a high-dimensional user feature matrix. Directly inputting a large user feature matrix into the recommendation system algorithm will lead to low algorithm efficiency. To solve these problems, this paper's algorithm uses singular value decomposition (SVD) to decompose the user feature matrix, thereby achieving dimensionality reduction and noise reduction.
Singular Value Decomposition (SVD) can decompose the user-feature matrix M into a matrix product of the user-reduced feature space matrix U, the singular value diagonal matrix Σ, and the reduced feature space-feature matrix V, as shown in Equation (1). Through SVD, the features with the strongest representational power are retained, while less influential or weaker features are removed. By retaining the main influencing factors, the features can be made to have a stronger representational power for the user. Furthermore, the decrease in feature dimension after SVD results in a lower-dimensional user-reduced feature space matrix U, making subsequent recommendation operations more efficient.
(1)
3.2 Hierarchical Clustering of User Features
After performing singular value decomposition on the user feature matrix, the algorithm uses hierarchical clustering to cluster users based on their features, grouping them into user groups with similar characteristics and interests. The algorithm flow of the bottom-up hierarchical clustering algorithm is shown in Figure 2. First, the algorithm initializes all user data as a leaf node. In each iteration, the algorithm calculates the similarity (e.g., Euclidean distance) between leaf nodes; subsequently, this similarity information determines whether they will be merged into a new cluster. The hierarchical clustering algorithm continuously iterates and merges until the iteration termination condition is met, such as the algorithm generating a specified number of clusters, or the distance between clusters meeting certain conditions.
The algorithm in this paper inputs the user feature information after singular value decomposition into a hierarchical clustering algorithm for iteration. Clustering iteration stops when the distance between the mean vectors of each cluster is greater than or equal to a distance threshold λ. After each iteration, the Euclidean distance between the mean vectors of each cluster is calculated. In Figure 3, the λ value on the left represents the minimum distance between the mean vectors of each class in the current clustering state. If the pre-set clustering iteration stopping threshold λ is 1.5, the clustering algorithm will stop iterating when four clusters are generated, because at this point, the Euclidean distance between the clusters is greater than the pre-set threshold λ. At this time, users G1, G2, and G8 will be grouped into one class, users G3, G4, and G9 into another class, users G5 and G7 into another class, and user G6 will be in a separate class. The satisfaction of the iteration stopping condition means that each user cluster has a certain degree of difference, thus avoiding users being ambiguously and incorrectly assigned to user groups they should not belong to.
Figure 2. Hierarchical clustering training flowchart
Figure 3. Genealogical diagram of the clustering process.
Table 1. Contextual features and their representation examples
Using hierarchical clustering, users with similar characteristics are grouped into the same class. Then, the mean vector of each class is used to represent the characteristics of users in that class for subsequent recommendation calculations, as shown in formula (2). The user feature representation of the c-th class is Uc, which is the mean of the user feature vectors of all I users assigned to that class.
(2)
In this way, although users within the same category share largely similar characteristics, there are still some differences and diversity among users within the same category, meaning they are not completely identical. Taking the mean vector of features for each user category can effectively fuse the features of users within the same category. This appropriately introduces the diversity brought by users within the same category without destroying the user's original interests. As a result, the algorithm can use the average features of users within the same category who have roughly the same interests to help that user discover other points of interest shared by users within the same category. This guides the user to discover the potential interests that users within the same category may possess, making the recommendation algorithm more effective.
3.3 Combination of User Characteristics and Contextual Factors and Use of Bayesian Prediction Models
Unlike traditional recommender system algorithms, the algorithm in this paper is not limited to user characteristics (such as age, gender, and hobbies) when analyzing and recommending applications using data features. Building upon traditional user and application features, it combines real-time user context information to create an application recommender system that utilizes multi-dimensional information. By combining user context factors, the recommender system can integrate real-time user context with existing user features into a multi-dimensional feature space, enabling a more human-centered and intelligent analysis of user needs and optimizing application recommendation performance. Some of the user context features used in this algorithm and their representation examples are shown in Table 1.
After obtaining user features, contextual features, and application features of user groups obtained through singular value decomposition and hierarchical clustering, this paper improves the original collaborative filtering recommendation algorithm by using a multidimensional feature composed of user features, contextual features, and application features, and employing a Bayesian model for application recommendation. In the algorithm, the user attribute set is represented as U={U1, U2, U3…, UN}, the context set is represented as S={S1, S2, S3…, SN}, and the recommended application list is represented by A.
The algorithm aims to recommend appropriate applications to users with specific user and contextual characteristics by calculating the probability of recommending different applications under different user attributes and contexts, as shown in formula (3):
(3)
According to formula (3), the algorithm can calculate the probability of different applications being used for a given user feature and its contextual features. In formula (3), based on the probability of U given that A is used, the probability of S given that A is used, and the probability of U and S occurring simultaneously, the probability values of different applications A being used given user feature U and contextual feature S can be calculated. The higher the probability obtained, the more worthy the application is to be recommended. Therefore, the top K applications with the highest probabilities are the K applications most worthy of recommendation. Thus, the algorithm obtains an application recommendation list sorted in descending order of probability.
Meanwhile, the proposed algorithm can accept and process the characteristic that the user's situation information is easily changed over time. The algorithm can update the probability information in formula (3) through real-time user behavior information such as direct clicks, favorites, dwell time, and number of clicks, so as to achieve real-time updates of the algorithm and make the application recommendation have better real-time performance.
3.4 Overall Algorithm Flow
The overall flow of the algorithm in this paper is shown in the curly braces in Figure 4. First, the algorithm performs singular value decomposition on the user data, and then performs hierarchical clustering iteration on the dimensionality-reduced and denoised user features obtained from the decomposition until the termination condition is met, resulting in user clusters that are far apart. The mean vector of each cluster is used as the user feature in that cluster. Then, using a Bayesian model, the algorithm combines user features and contextual features to predict and rank the probability of application usage, ultimately forming an application recommendation list.
Figure 4 System Architecture
4 System Architecture and Implementation
4.1 System Overall Architecture
The overall architecture of this system is divided into five parts, as shown in Figure 4. The bottom-level basic data layer mainly contains basic user attribute data, such as age, gender, location, and mobile device attributes; basic application information, such as the category of an application, its rating, etc.; and user application access data, i.e., the contextual information required by the algorithm. To obtain the necessary data, data collection is required. Its main goal is to synchronize the data from the business system to the data storage system for later processing and use.
The second layer of the recommendation system in this paper is an offline computing layer, whose responsibility is to perform data processing and model training. First, the collected data is cleaned and organized; second, the model is trained according to the algorithm flow described in Section 3 of this paper; finally, a model that can recommend applications to users is formed. The algorithm flow is shown in parentheses in Figure 4.
The recommendation system's third to fifth layers are the storage and interface encapsulation layer, the business processing layer, and the application layer, respectively. First, the system encapsulates the data interface based on the recommended ranking content; then, if necessary, the business processing layer will make necessary fine-tuning adjustments to the recommended content information according to market changes and other information; finally, the business-intervention-based recommendation data is transmitted to the front-end user. The entire system must meet the following requirements:
(1) Establish timely and comprehensive internet information processing capabilities, and conduct basic data collection on various types of commonly used application software by exploring network resources;
(2) Based on the cloud computing framework, establish the technical capabilities for horizontal expansion of distributed storage and large-scale data parallel computing, realize the horizontal expansion of the system, and facilitate the streaming processing of large amounts of data;
(3) Establish a highly cohesive and loosely coupled architecture, reuse basic capabilities, improve the interface design between systems, and reduce the strong correlation between crawling, analysis, visualization and application;
(4) While enhancing the accuracy of recommendations, continuously improve the user experience and use the proposed algorithm to improve the accuracy and recall of recommendations.
This paper's system optimizes traditional recommendation systems by clustering user features obtained from singular value decomposition and combining them with contextual information. A Bayesian model is then used to analyze and predict the applications needed by users, thereby enhancing application recommendation effectiveness. Simultaneously, the system fully considers operational performance and scalability, as well as its real-time processing capabilities to handle real-time data feedback and quickly adapt to changes in user attribute characteristics. By building the application recommendation system on a distributed architecture, it achieves parallel processing of massive amounts of data, effectively improving the system's throughput performance.
4.2 System Construction
The following technologies were used in constructing the multi-dimensional application recommendation system proposed in this paper: The system uses MySQL to store user attribute data and HBase to store user profile data. HDFS distributed file system is used for data storage. The system adopts a Spark-based underlying platform, utilizing Spark Streaming for real-time computation, feature vector calculation, and generation of sorted lists. Kafka is used for real-time data cluster forwarding. Because the system requires extensive and complex data retrieval from the database, the ElasticSearch distributed system architecture is a good solution. Furthermore, the visualization platform uses the popular J2EE application architecture. The platform layer uses the Java standard library and a J2EE server. The J2EE project runs on Tomcat and is integrated with the Spring MVC framework for web development. By utilizing and coordinating these technologies, the system can efficiently meet the required requirements. The system can support real-time data acquisition and processing, and can handle at least millions of data points during data collection and processing. When making application recommendations, this system can support at least hundreds of concurrent operations while maintaining system availability. During real-time data queries, latency will not exceed 4 seconds. Meanwhile, to meet system scalability requirements and increase data processing capacity, this system has an additional backup server to prevent overload due to excessive data processing volume.
Figure 5. Comparison of clustering effects with and without singular value decomposition.
5. Analysis of Experimental Results
This paper conducts thorough and systematic testing on the implemented application recommendation system. First, it verifies that the adopted singular value decomposition and hierarchical clustering methods have good performance. Then, it verifies the overall effectiveness of the proposed algorithm from two aspects: recommendation results and metrics. Furthermore, this paper also tests the system performance, thereby verifying the system's efficiency and stability.
5.1 Comparison of Clustering Effectiveness Before and After Singular Value Decomposition
To verify the effectiveness of performing singular value decomposition (SVD) before hierarchical clustering, this paper performs hierarchical clustering on user features with and without SVD, and randomly selects the user cluster containing user #1056, calculating the average attribute difference among users in the same group. The comparison shows that the average difference of data without SVD before hierarchical clustering can be as high as 0.6, while the maximum difference range after SVD is 0.4. This indicates that hierarchical clustering of user feature data after SVD to eliminate redundant information retains the main information of the original data, resulting in more refined data and better clustering performance. Figure 5 further visualizes the clustering effect of the algorithm: the number of misclassifications in the hierarchical clusters after SVD is significantly reduced, demonstrating the effectiveness of clustering after SVD.
Table 2 Comparison of the effects of using different application algorithms to recommend applications to users
Figure 7. Comparison of recall rates of different clustering algorithms
5.2 Hierarchical Clustering Based on Singular Value Decomposition and Traditional Clustering Algorithms
To further verify the effectiveness of the hierarchical clustering algorithm based on singular value decomposition (SVD), the proposed method is compared with traditional machine learning clustering algorithms, and the recall rate of user clustering results is used as an indicator to measure the effectiveness of different algorithms. This paper performs user clustering on Mahout, a platform within the Hadoop ecosystem, as shown in Figure 6. The comparison methods used in the experiment are decision tree model, K-means algorithm, and traditional hierarchical clustering method without SVD. By inputting the user dataset M′, which has undergone preprocessing steps such as removing erroneous data, into different clustering algorithms, four different clustering results (R1, R2, R3, and R4) are obtained. Clustering result R4 is obtained by first decomposing the data M′ using the SVD algorithm and then inputting it into the hierarchical clustering algorithm.
Figure 6 Comparison Method
In the validation process, this paper uses 10-part cross-validation to verify the three traditional clustering algorithms and the proposed algorithm, and calculates the recall rate. As shown in Figures 6 and 7, the average recall rate of the traditional K-means clustering method is approximately 57%, while the average recall rate of the decision tree algorithm reaches approximately 60%. The average recall rate of the method without singular value decomposition before hierarchical clustering is approximately 68%, while the average recall rate of the method after singular value decomposition and hierarchical clustering reaches approximately 73%, which is about 5% higher than the former. This fully demonstrates the effectiveness of the proposed method compared to traditional clustering algorithms.
5.3 Comparison and Analysis of Recommendation Result Quality of Recommendation Algorithms
As shown in Figure 4, the algorithm flowchart illustrates the process. After user feature data undergoes singular value decomposition and hierarchical clustering, the algorithm uses the mean vector of each user cluster to represent the user features within that cluster. Subsequently, the algorithm combines contextual features and outputs a final list of recommended applications for the user through a Bayesian model.
To compare the recommendation results of our proposed algorithm with those of traditional algorithms, we use collaborative filtering-based and association rule-based recommendation algorithms as comparison methods. In collaborative filtering-based algorithms, the algorithm analyzes the preferences of other users with similar characteristics to recommend applications that the user might like. Association rule-based algorithms, on the other hand, utilize discovered association rules between applications to recommend applications that the user might be interested in based on information about the applications the user has used. Neither of the compared methods utilizes singular value decomposition to process feature representation, hierarchical clustering, or the user's contextual information. Table 2 shows the recommendation performance of different algorithms for the same user.
Taking user #0013 randomly selected from the collected dataset as an example, Table 2 shows the top 5 recommendation results with the highest probability values among the application prediction results of different algorithms running on the distributed framework. Among the three methods, only the method presented in this paper recommends applications that match the user's actual usage. Furthermore, compared to other algorithms, the other applications recommended by the algorithm presented in this paper also show a high correlation with the user's actual usage, demonstrating the superiority of the proposed method in recommendation performance.
5.4 Comparison and Analysis of Quality Quantification Indicators of Recommendation Results from Recommendation Algorithms
To further quantify and compare the recommendation result quality of our proposed algorithm with existing traditional recommendation algorithms, we calculated the Mean Absolute Error (MAE) based on the relevance of the recommended applications, i.e., the degree of matching between the actual used application and the recommended application. The MAE value is lower when the recommended application and the actual used application belong to the same category than when they belong to different categories; therefore, a smaller MAE value indicates higher search result quality. Furthermore, the Root Mean Squared Error (RMSE) generated by the Bayesian model and the recall value of the recommendation results were also used as quantitative comparison indicators of the algorithm's recommendation result quality. A smaller RMSE value and a larger recall value both indicate that the model has a more ideal performance. The experimental comparison results are shown in Figure 8.
Figure 8 Comparison of quantitative metrics for recommendation algorithm results
In the validation process, this paper employs the commonly used 10-part cross-validation method and averages the results from the 10 sets. As clearly shown in Figure 8, compared to traditional recommendation algorithms, the proposed algorithm has the lowest MAE and RMSE values and the highest Recall value. The MAE and RMSE values of the proposed method are both less than 0.6, while the RMSE value of the association rule-based algorithm is approximately 0.7, and the RMSE value of the collaborative filtering-based algorithm is close to 0.8. Simultaneously, the Recall value generated by the proposed algorithm is significantly improved compared to the two comparative methods. Therefore, this fully demonstrates the effectiveness and accuracy of the recommendation results from a quantitative perspective.
5.5 Performance Analysis of the Recommendation System
In the actual testing process, this paper also conducted thorough testing and analysis on the performance of the constructed system to detect its system load capacity in a real operating environment.
During performance testing, 2000 virtual users were added to test the system's average response time. Figure 9 shows the average response time for a single-page user click. The test results showed that the system's average response time was approximately 1.4 seconds, the longest response time was approximately 3.8 seconds, and the shortest response time was approximately 0.1 seconds, meeting the system's concurrency requirements.
Figure 9 Average Transaction Response Time
Meanwhile, this paper also conducted performance tests on the constructed system, and the results are summarized in Table 3. The test results fully demonstrate the efficiency and reliability of the constructed system. In summary, the recommendation system presented in this paper shows a significant improvement in application recommendation compared to the compared methods, with higher recommendation effectiveness and high efficiency and reliability in system performance. Through experimental verification and result analysis, compared to the compared methods, the algorithm optimization in this system makes application recommendations more accurate, thus better recommending suitable applications to users. Furthermore, the system can accept real-time updated data, improving its intelligence and reducing operational complexity. Finally, the system architecture used in this paper ensures the system's high efficiency, reliability, and high scalability, enabling the system to be more widely applied.
Table 3 System Performance Records
5.6 Discussion and Analysis
This paper improves the accuracy of application recommendation systems by using singular value decomposition (SVD) to reduce the dimensionality and denoise user feature information, and hierarchical clustering to cluster user information, while also considering the contextual information of users. Compared to methods in the literature that do not use matrix factorization to reduce the dimensionality and denoise user feature information before clustering, the proposed method uses SVD to remove redundant components in user feature information, enhance the representation ability of feature information, and improve the algorithm's running efficiency.
Furthermore, hierarchical clustering of user feature information and the use of cluster mean vectors as representations of user classes effectively enhance feature diversity without compromising the overall characteristics of user groups, thereby improving the recommendation effectiveness of the recommendation system. Compared to methods in the literature that utilize user contextual information to analyze product features, our method uses user contextual information to analyze user preference features, thus better assisting the recommendation system in recommending more suitable applications to users. In conclusion, our proposed recommendation system algorithm has better application recommendation capabilities compared to the methods described above.
6. Summary
This paper addresses the problem of traditional recommender systems failing to consider the user's context when recommending applications. It proposes a novel multi-dimensional recommender system based on user feature clustering combined with contextual features. The key technology of the proposed algorithm lies in firstly using singular value decomposition (SVD) to reduce the dimensionality and noise of user features, resulting in less redundant user data with stronger representational power. The lower dimensionality also makes the proposed recommender system more efficient. Subsequently, hierarchical clustering of users is performed based on the decomposed user features, and the mean vector of each cluster is used as the feature information representing the users within that cluster. This approach, without destroying the similarity of users within a cluster, fuses the features of users within the cluster by calculating the cluster mean, appropriately introducing diversity into user features and avoiding overly monotonous recommendations that could lead to user fatigue. Finally, the algorithm uses the obtained user features combined with contextual features to form a multi-dimensional feature set and employs a Bayesian model to predict the probability of recommended applications, thus obtaining a list of recommended applications ordered in descending order of recommendation probability, achieving higher-quality application recommendations to users. At the same time, the algorithm can adapt to user contextual features that change over time, thus enabling real-time recommendations.
This paper implements the proposed recommendation system based on a distributed architecture, and verifies it with sufficient test data, showing that the system not only has accurate application recommendation capabilities, but also high efficiency and stability, laying a solid foundation for the application of the proposed recommendation system.
Authors: Wu Jiashu1,2 Rao Huaxiao1,3 Fan Xiaopeng1 Wang Yang1*
1. Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
2. School of Computing and Information Systems, University of Melbourne
3. University of Chinese Academy of Sciences
This article is reprinted from "Integration Technology".
