A Deep Learning-Based Active Content Caching Method for End-User Devices

Abstract

This invention provides a deep learning-based method for proactively caching content on end-user devices, applied in the field of wireless communication, to address the problem of proactively caching personalized content on end-user devices in wireless edge networks. Based on deep learning technology, this invention comprehensively considers two types of end-users in real-world scenarios—cold-start users and non-cold-start users—and mines the content preference characteristics of these two types of users, pre-caching content matching their preferences to the user devices. This invention reduces content transmission overhead in network communication through deep learning technology, lowering the average latency and total energy consumption for end-users accessing content. By comprehensively analyzing, mining, and predicting the content preferences of both types of end-users and proactively caching relevant content to user devices, this not only reduces data traffic load on the network but also effectively improves network service quality and user experience.

Description

Technical Field

[0001] This invention relates to active content caching technology, and more particularly to an active content caching technology for end-user devices. Background Technology

[0002] Artificial intelligence (AI) technology is driving the rapid development of wireless communication. It is envisioned that 6G wireless communication will complete the evolution from "Internet of Everything" to "Intelligent Internet of Everything." While AI technologies based on deep learning and big data analytics require enormous computing and communication resources, embedding AI model training and inference capabilities into the edge of wireless communication networks allows for seamless integration of sensing, communication, computing, and intelligence, thereby improving the effectiveness and security of wireless communication networks. Traditional cloud computing suffers from high latency and other problems, failing to meet user needs. Cloud computing includes powerful and centralized data centers with high computing and caching capabilities; however, the number of centers is limited, and they are located far from end users, unable to store and process all data and requests, leading to network congestion and data transmission delays. When cloud computing fails to meet the established expectations of wireless communication networks, increasing research is shifting its focus to wireless edge networks.

[0003] In wireless edge networks, resources are distributed closer to end-user devices. Due to the limited resource capacity on edge devices, content caching technology has received significant attention for improving edge intelligence. Content caching is a method for resource management in edge network devices, enabling rapid response to user requests and alleviating network traffic and bandwidth congestion. Proactive content caching, in particular, analyzes and predicts content of interest to end-users based on its popularity or their browsing behavior, and caches it in advance during off-peak hours, thereby improving access speed and reducing transmission costs and time.

[0004] Content caching technology faces many challenges. For example, end-user requests are highly diverse, and the similarity between users may affect the access demand for certain content. Furthermore, the mobility of users' geographical locations and the dynamic nature of changing interests can also impact the performance of content caching. How to meet personalized user content preferences and further improve the content distribution efficiency and user experience of wireless edge networks is a pressing issue. Existing technologies include numerous studies on content caching for user devices. For instance, Wang F, Wang F, and Liu J, "Intelligent video caching at network edge: A multi-agent deep reinforcement learning approach," in IEEE INFOCOM2020-IEEE Conference on Computer Communications, IEEE, 2020, doi:10.1109 / INFOCOM41043.2020.9155373, proposes using multiple edge servers to adaptively learn user content needs to improve content service quality; Choi M, Molisch AF, and Han DJ,

[0005] "Probabilistic caching and dynamic delivery policies for categorized content and consecutive user demands," in IEEE Transactions on Wireless Communications, 2020, doi:10.1109 / TWC.2020.3044076, proposes predicting users' content needs for the next stage based on content access history and proactively caching the corresponding content. Huang Z, Hu B, and Pan J, "Caching by user preference with delayed feedback for heterogeneous cellular networks," in IEEE Transactions on Wireless Communications, 2020, doi:10.1109 / TWC.2020.3035377, simultaneously consider whether users can access content and the timeliness of feedback on content access experience, and formulate a latency-tolerant content caching scheme. However, the above research still has the following problems:

[0006] (1) The main research focuses on mining users’ long-term interests and preferences, but ignores the dynamics of changes in users’ short-term and current preferences. The modeling of users’ content interests and preferences is not comprehensive enough and cannot fully describe the content needs that users may be interested in in the future.

[0007] (2) Cold-start users with limited historical access content are not considered. In wireless edge network scenarios, end users include both non-cold-start users with rich access history and cold-start users with almost no interaction information. The content preference modeling methods proposed in current research are not applicable to cold-start users.

[0008] (3) It does not consider that similar content popularity trends may exist among geographically proximate user groups. When actively caching content, the popularity of content varies with time and region. Existing work has not enriched the modeling of content preference features for certain users by incorporating the content preference features of geographically proximate users. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a corresponding active caching method for device content for terminal users in wireless edge networks based on different terminal user types, while ensuring network performance and user experience.

[0010] The technical solution adopted by this invention to solve the above-mentioned technical problems is a method for actively caching content on end-user devices based on deep learning, comprising the following steps:

[0011] 1) For each end user, the edge server first obtains the corresponding historical access content sequence of the user. If the number of contents in the historical access content sequence is below the set value, the user is considered a cold start user and proceeds to step 3). Otherwise, the user is considered a non-cold start user and proceeds to step 2.

[0012] 2) For non-cold start users, the historical access content sequence is represented by a feature vector to obtain the historical access content feature sequence; then the content sequence accessed by the user during the most recent content request service and the latest content accessed by the user are obtained; the latest content accessed by the user is the last content in the historical access content sequence, and the content sequence accessed by the user during the most recent content request service is the content sequence within the last lifecycle of the historical access content sequence.

[0013] The historical access content feature sequence is input into the long-term preference model of non-cold start users. The long-term preference model of non-cold start users outputs the long-term content preference features of users to the fusion module based on the self-attention mechanism.

[0014] The content sequence accessed by the user during the most recent content request service is input into the short-term preference model of non-cold start users. The short-term preference model of non-cold start users outputs the user's short-term content preference feature y2 based on the temporal convolutional network to the fusion module.

[0015] The user's most recently accessed content is directly output to the fusion module as the user's current content preference feature y3.

[0016] The fusion module weights and fuses users' long-term content preference characteristics, short-term content preference characteristics, and current content preference characteristics to obtain the dynamic preference characteristics of non-cold start users;

[0017] 3) For cold start users, non-cold start users whose geographical location is less than a set threshold are added to the set of users close to the cold start user. Then, the dynamic preference features of each non-cold start user in the set of users close to the cold start user are weighted and fused to obtain the dynamic preference features of the cold start user.

[0018] 4) The edge device assigns a value to the user content preference feature vector h based on the user's type and the corresponding dynamic preference features. Then, it uses the similarity between the terminal user content preference feature vector h and the cached content vector C in the edge device to obtain the probability p that the cached content in the edge device is accessed by the user.

[0019] p = softmax(hC) T )

[0020] Where T represents the transpose operation and softmax is the normalization exponential function;

[0021] 5) Based on the recommended content quantity value v in the cache settings, the top v most frequently accessed content from the edge device will be cached in advance to the end user device.

[0022] This invention comprehensively considers two types of end users in real-world scenarios. First, it determines whether a user is a cold-start user or a non-cold-start user based on their historical access content. For non-cold-start users, based on their historical access content, it mines their long-term, short-term, and current interests, comprehensively considering these three dimensions of interest characteristics to obtain the final user content preference features. For non-cold-start users, since their own historical access content is almost nonexistent, it is impossible to extract useful information from it. This invention considers that similar content trends exist in geographically proximate user clusters. Therefore, this invention models the content preferences of non-cold-start users by fusing the content preference features of users in geographically proximate user clusters. After obtaining the end-user content preference features, it calculates the similarity between cached content on edge devices and user content preferences to select content that the user may be interested in in the future and caches it on the user's device in advance.

[0023] The beneficial effects of this invention are:

[0024] (1) Effectiveness. To address the congestion and load issues caused by rapidly increasing traffic to communication networks, this invention proposes an effective proactive content caching method. By mining end-users' content interests and preferences, content that users are interested in is cached on user devices in advance during off-peak hours, thereby providing better and faster services, effectively improving network latency, and enhancing user experience quality.

[0025] (2) Comprehensiveness. The content proactive caching method proposed in this invention comprehensively considers two types of end users in real-world scenarios: cold-start users and non-cold-start users. Based on the different characteristics of the historical content accessed by these two types of users, methods for mining their content interest and preference features are considered separately.

[0026] (3) Personalization. This invention recognizes the dynamic nature of user interests over time and their diversity across individuals. Therefore, the methods used in this paper to mine user content preference characteristics are all based on the user's own historical browsing content, which can more accurately grasp each user's personalized content preferences. In particular, for non-cold start users, this invention improves the accuracy of personalized feature modeling from three dimensions: long-term interests, short-term interests, and current interests. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of an embodiment scenario.

[0028] Figure 2 This is a flowchart of the method for proactively caching content on end-user devices. Detailed Implementation

[0029] The wireless communication scenarios applied in the embodiments are as follows: Figure 1 As shown, wireless edge networks contain numerous edge servers and devices, as well as end-user devices such as mobile phones, tablets, and computers. All of these require sending and receiving data, and various services and applications also need to collect and process data in real time. The transmission of this information over the communication network can lead to network congestion. By deploying proactive content caching technology in wireless edge networks, user access requests can be quickly sent locally, and content that users may access in the future can be cached on user devices in advance during off-peak hours. This significantly improves network performance, reduces content transmission overhead, and enhances the quality of user experience.

[0030] This invention takes into account the following factors that affect the quality of user experience:

[0031] First, there is the network latency problem caused by the massive information transmission and processing in communication networks. Today's communication networks are facing a surge in the number of wireless devices, including mobile phones, wearable devices, connected cars and various sensors. All of these devices need to send and receive data. The massive information transmission and exchange leads to network congestion and traffic congestion, making it impossible to meet users' access needs in real time during peak data access periods.

[0032] Secondly, there is the issue of the dynamism and diversity of user content preferences. Because user interests are dynamic and vary over time, encompassing multiple dimensions such as long-term, short-term, and current interests, statically considering user content preferences alone is insufficient to accurately grasp these changes. Traditional content caching methods based on popularity also fail to meet the personalized content needs of each user.

[0033] Finally, there's the issue of cold-start users in wireless edge networks. In real-world applications, there are newly joined users who have almost no content access history—cold-start users. Traditional methods for mining user content preferences based on historical access data are not suitable for these users. However, cold-start users cannot be ignored in wireless communication scenarios. How to model the content preferences of these users and proactively cache content is also a factor affecting the quality of end-user experience.

[0034] Based on the above points, this invention designs a method for proactively caching content on end-user devices using deep learning technology. This method comprehensively considers two types of end-users in real-world scenarios: cold-start users and non-cold-start users, and models and mines their content preference features separately. For non-cold-start users, it considers changes in content preferences across three dimensions: long-term interests, short-term interests, and current interests, mining the user's dynamic content preference features. For cold-start users, considering that user clusters in similar geographical areas will have similar content popularity trends, it fuses the content preferences of non-cold-start users in the cluster to model their content preferences. By pre-caching content that users may access in the future to their devices, the user experience quality is effectively improved.

[0035] The following is combined Figure 2 This document provides a detailed description of an active content caching method for end-user devices based on deep learning technology, specifically including the following steps:

[0036] 1) For each end user, the edge server first obtains the corresponding historical access content sequence C = {c1, c2, ..., c...} i ,…,c m},like If 1≤m≤2, the user is considered a cold start user; otherwise, the user is considered a non-cold start user. m is the total number of historical accessed content.

[0037] 2) For users who are not cold-started, first process each piece of content c in the historical access content sequence C. i Feature modeling is performed, and the multimodal features of the content are fused to obtain its vector representation e. i ∈R 1×d Where d is the dimension of the feature vector, R is the real number field, and the content feature sequence corresponding to the historical access content set C is E =

[0038] {e1,e2,…,e i ,…,e m}, E∈R m×d ;

[0039] The lifecycle of each content service request for a non-cold start user is divided according to the session control. Each session represents one content service request, thus defining the sequence of content accessed by the user within the lifecycle t of each request. The historical content access sequence C represents the user's long-term content preferences; the content sequence {c} accessed by the user during their most recent content service request... m-t ,…,c m} represents the user's short-term preferences; while the content the user has recently visited, c m This represents the user's current preferences.

[0040] Specifically, the method for modeling long-term preferences of non-cold-start users is as follows:

[0041] This model mines long-term content preferences of non-cold-start users based on the self-attention mechanism. The input to the long-term preference model for non-cold-start users is the sequence of user's historical access content features, E = {e1, e2, ..., e...}. m}, E∈R m×d It learns the positional features of each content in the sequence through a position encoder, and its similarity is calculated using Attention, which is a scalable dot product, i.e.:

[0042]

[0043] Here, Q represents the query vector of the input content vector, and K and V represent the keys and values ​​of the input content vector, respectively. These parameters are used to adjust for excessively large click results, thereby making gradient changes more stable. Therefore, the output feature y1 of the long-term preference model for non-cold-start users is defined as:

[0044] y1 = Attention(EW) Q EW K EW V )

[0045] Among them, W Q W K and W V The parameters for the long-term preference model of non-cold-start users are represented by the weight matrices of the query vector, key value, and item value, respectively, which are determined through continuous optimization during model training. To facilitate the subsequent fusion of multi-dimensional content preference features, the output feature y1 represents the user's long-term content preference features, where y1∈R. 1×d .

[0046] Specifically, the method for modeling short-term preferences of non-cold-start users is as follows:

[0047] This paper utilizes Temporal Convolutional Networks (TCNs) to mine short-term content preference features of non-cold-start users. The input to the short-term preference model for non-cold-start users is the sequence of content features accessed by the user during their most recent content request service request {c}. m-t ,…,c m For a convolution kernel q of size n, where the dilation factor of dilated convolution is defined as 0, the user's short-term content preference features can be expressed as:

[0048]

[0049] Where k represents the content index in the content feature sequence accessed during the most recent content request service process, j represents the convolution range of the current convolutional layer, and q j This represents a convolution process with a convolution range of j, where l represents the number of layers in the network, and c... k～o×l This represents the k to o×l content feature sequences participating in the convolution. Similarly, to facilitate the subsequent fusion of multi-dimensional content preference features, this module also adds a pooling layer; the output feature y2 is the user's short-term content preference feature, y2∈R. 1×d .

[0050] Specifically, the method for modeling the current preferences of non-cold-start users is as follows:

[0051] The content recently visited by users m This represents their current preference. Therefore, c m The corresponding content feature vector represents the user's current content preference features, i.e., y3 = e m The output feature y3 of the current preference model for non-cold-start users is the user's current content preference feature, where y3∈R. 1×d .

[0052] Specifically, the fusion model integrates the content preference features of non-cold start users across three dimensions to obtain a comprehensive and accurate representation of user content preference features.

[0053] The fusion model weights and fuses long-term content preference features y1, short-term content preference features y2, and current content preference features y3, resulting in the final dynamic preference features h of non-cold-start users. n It can be represented as:

[0054] h n = f(θ1y1+θ2y2+θ3y3)

[0055] Where θ1, θ2, and θ3 are the weights of long-term content preference feature y1, short-term content preference feature y2, and current content preference feature y3, respectively, and are parameters learned by the fusion model through training. n ∈R 1×d .

[0056] 3) For cold-start users, since they have almost no historical browsing content, there is very little information available for mining content preference features. Therefore, the modeling method for their content preference features should be considered separately. This invention believes that similar content popularity trends exist in user clusters with similar geographical locations. Therefore, the modeling method for the content preference features of cold-start users in this invention is to integrate the content preference features of non-cold-start users in user groups with similar geographical locations. The location acquisition function is defined as l(·), and the cold-start user set is defined as U. c The non-cold start user set is defined as U n For the i-th cold start user and the j-th non-cold start user Calculate the geographical distance between them:

[0057]

[0058] || is the absolute value operation;

[0059] Non-cold-start users whose geographical location is below a set threshold will be added to the set of nearby users of the cold-start user. Where b represents the number of similar users in the i-th set of similar users. (Get) Each user Corresponding content feature preferences This invention uses an attention mechanism to fuse the content preference features of these similar users. The final output of the cold start user content preference feature model is the dynamic preference feature h of the cold start user. c :

[0060]

[0061] Among them, a k This represents the learned attention weights.

[0062] 4) In wireless edge networks, users can quickly retrieve desired content from edge devices via D2D and other communication methods. The cached content in the edge device is vectorized into a representation C∈R. m×d Where m is the amount of cached content. Then, based on the user type, the user content preference feature vector h is assigned a value using the user's corresponding dynamic preference features, h = h0. c or h = h n The similarity between the end-user content preference feature vector h and the cached content vector C is calculated, representing the probability p that the cached content in the edge device is accessed by the user.

[0063] p = softmax(hC) T )

[0064] Where T represents the transpose operation.

[0065] 5) Based on the number of contents v set in the cache, the top v contents with the highest access probability will be cached in advance to the end user device.

[0066] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for proactive content caching on end-user devices based on deep learning, characterized in that, Includes the following steps: 1) For each end user, the edge server first obtains the user's historical access content sequence. If the number of contents in the historical access content sequence is below a set value, the user is considered a cold start user and proceeds to step 3); otherwise, the user is considered a non-cold start user and proceeds to step 2). 2) For non-cold start users, the historical access content sequence is represented by a feature vector to obtain the historical access content feature sequence; then the content sequence accessed by the user during the user's most recent content request service and the latest content accessed by the user are obtained; The most recently accessed content by a user is the last content in the historical access content sequence, and the content sequence accessed by a user during their most recent content request service is the content sequence within the last lifecycle of the historical access content sequence; The historical access content feature sequence is input into the long-term preference model of non-cold start users. The long-term preference model of non-cold start users outputs the long-term content preference features of users to the fusion module based on the self-attention mechanism. The sequence of content accessed by the user during their most recent content request is input into the short-term preference model for non-cold-start users. The short-term preference model for non-cold-start users outputs the user's short-term content preference features based on a temporal convolutional network. To the fusion module; The user's most recently viewed content is directly used as the user's current content preference feature. Output to the fusion module; The fusion module weights and fuses users' long-term content preference characteristics, short-term content preference characteristics, and current content preference characteristics to obtain the dynamic preference characteristics of non-cold start users; 3) For cold start users, non-cold start users whose geographical location is less than a set threshold are added to the set of similar users of the cold start user. Then, the dynamic preference features of each non-cold start user in the set of similar users are weighted and fused to obtain the dynamic preference features of the cold start user. 4) Edge devices, based on user type, utilize the dynamic preference features corresponding to that user to analyze the user's content preference feature vector. Assign values, and then utilize the end-user content preference feature vector. With the cached content vector in edge devices The similarity between them gives the probability that cached content in the edge device will be accessed by a user. : ； in, This indicates the transpose operation. It is a normalized exponential function; 5) Based on the recommended content quantity value v in the cache settings, the top v most frequently accessed content from the edge device will be cached in advance to the end user device.

2. The method as described in claim 1, characterized in that, The long-term preference model for non-cold-start users is as follows: ； in, This is a sequence of features of the user's historical accessed content. and These are all parameters from a long-term preference model for non-cold-start users. The output features of the long-term preference model for non-cold-start users are the users' long-term content preference features. The function is represented as: ； in, A query vector representing the input content vector. and These represent the key and item values ​​of the input content vector, respectively. These are the parameters for gradient adjustment.

3. The method as described in claim 1, characterized in that, The short-term preference model for non-cold-start users is as follows: ； in, For the pooling function, the input to the short-term preference model for non-cold-start users is the sequence of content features accessed by the user during their most recent content request service. , This represents the total number of historical accesses. This represents the content sequence number in the content feature sequence accessed during the most recent content service request, where t is the lifecycle of a content service request. This represents the user's short-term preferences and the content the user has recently visited. This represents the user's current preferences. Indicates the convolution range of the current convolutional layer. Indicates the range of convolution. The convolution process, Indicates the number of layers in the network. The kernel size is [size]. Indicates that participates in convolution. A sequence of content features, is the dilation factor of dilated convolution. The output features of the short-term preference model for non-cold-start users are the short-term content preference features of users.

Images