Intelligent hierarchical cache management method and system fusing multi-modal interaction features

Abstract

The application belongs to the technical field of storage management, and particularly relates to an intelligent hierarchical cache management method and system fusing multi-modal interactive features, to solve the technical problems that a traditional cache strategy is difficult to adapt to complex user behaviors and lacks context and multi-modal fusion. The management method comprises the following steps: S1, constructing a user interactive trajectory graph; S2, obtaining an initial context process vector; generating a fused feature vector; generating an initial cache level affinity vector of a data object to each level of cache; S3, obtaining a revised context process vector; S4, calculating an asymmetric risk cost of the data object when a miss occurs in different cache levels; and solving a cache placement or eviction strategy under a stochastic programming framework. The application can improve the overall performance of the cache system, reduce the access delay of a key task, and improve the quality of service perception of users.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of storage management, and particularly relates to an intelligent hierarchical cache management method and system fusing multi-modal interaction features. BACKGROUND

[0002] Hierarchical cache is applied to content distribution network, edge computing and various distributed storage systems. Traditional cache management strategies, such as least recently used, least frequently used and variants, are simple to implement, but their prediction accuracy is limited when facing today's increasingly complex user behavior patterns. These methods are difficult to capture the context dependency of user behavior and cannot distinguish the deep intention behind different interaction behaviors, thus often leading to unsatisfactory cache hit rate, especially in scenarios where user interest changes rapidly, the overall performance and user experience will be affected.

[0003] In recent years, intelligent cache decision methods based on machine learning and deep learning have emerged, which can use richer data features (such as content attributes of data objects and user historical behavior) to predict future access patterns. However, existing methods usually ignore the continuity and process of user behavior, and fail to model multiple user interactions as a coherent "trajectory" or "task flow", thus losing valuable context process information. In addition, there is insufficient fusion of multi-modal interaction information between users and data objects, making it difficult to explore the synergistic effect of different modalities in predicting user intent. More importantly, traditional methods treat all cache miss events as equal cost, ignoring the importance difference of different data objects in the user task flow, i.e. a miss of a critical data may have more serious negative impact than multiple misses of non-critical data, lacking asymmetric evaluation of cache decision risk.

[0004] Therefore, there is an urgent need for a new intelligent hierarchical cache management method that can deeply understand user interaction processes, fuse multi-modal features and perform risk perception. SUMMARY

[0005] The application provides an intelligent hierarchical cache management method and system fusing multi-modal interaction features, to solve the technical problems that traditional cache strategies are difficult to adapt to complex user behavior and lack context and multi-modal fusion.

[0006] In a first aspect, the application provides an intelligent hierarchical cache management method fusing multi-modal interaction features, comprising:

[0007] S1, obtaining historical multi-modal interaction data of a user and a data object, and constructing a user interaction trajectory graph according to an event sequence in the historical multi-modal interaction data;

[0008] S2, performing preliminary embedding on the user interaction trajectory graph by using a graph neural network to obtain an initial context process vector; and based on the initial context process vector, fusing multi-modal features through an attention mechanism to generate a fused feature vector; concatenating the fused feature vector with a vector representing the state of each level of cache and inputting the result into a prediction network to generate an initial cache level affinity vector of the data object for each level of cache;

[0009] S3, adjusting the node features of the user interaction trajectory graph based on the initial cache level affinity vector, and performing re-embedding calculation by using the graph neural network to obtain a revised context process vector;

[0010] S4, evaluating the topological importance of the data object in the user interaction trajectory graph according to the revised context process vector, and calculating the asymmetric risk cost of the data object when a miss occurs at different cache levels; when receiving a cache request, combining the initial cache level affinity vector, the asymmetric risk cost and the storage size of the data object, and solving the cache placement or eviction strategy in a stochastic programming framework to maximize the risk-adjusted expected hit benefit of the system.

[0011] Further, in step S1, historical multi-modal interaction data of users and data objects is obtained, and a user interaction trajectory graph is constructed according to the event sequence therein, including:

[0012] defining user nodes and data object nodes;

[0013] extracting interaction events between user nodes and data object nodes, the interaction events including clicking, browsing for more than 30 seconds, collecting, and commenting;

[0014] if there is an interaction event between a user node and a data object node, a edge is established between them, and the type and occurrence time of the interaction event are taken as the attributes of the edge;

[0015] assigning an initial weight to the edge according to the type of the interaction event.

[0016] Further, the corresponding initial weight of clicking is 0.2, the corresponding initial weight of browsing for more than 30 seconds is 0.5, the corresponding initial weight of collecting is 0.8, and the corresponding initial weight of commenting is 1.0.

[0017] Further, in step S2, the user interaction trajectory graph is preliminarily embedded by using a graph neural network to obtain an initial context process vector, including:

[0018] a graph convolution network (GCN) is used as the graph neural network, and two layers of convolution are used to aggregate neighbor node information to generate a 128-dimensional initial context process vector for each node.

[0019] Further, in step S2, based on the initial context process vector, a fusion feature vector is generated by attention mechanism fusion of multi-modal features, including:

[0020] The text modal feature and the image modal feature of the data object are extracted, and the corresponding modal features are converted into a 768-dimensional text feature vector and a 2048-dimensional image feature vector through a BERT model and a ResNet-50 model respectively;

[0021] The initial context process vector is taken as a query vector, and key vectors and value vectors of the text feature vector and the image feature vector are generated; the query, key, and value vectors are projected to a unified dimensional space through a learnable linear transformation; the similarity between the query vector and the key vector is calculated through the attention mechanism to obtain the attention weight of each modal feature, and the value vector is weighted and summed using the attention weight to obtain the fusion feature vector.

[0022] Further, in step S2, the fusion feature vector is concatenated with a vector representing the state of each level of cache and input into a prediction network to generate an initial cache level affinity vector of the data object for each level of cache, including:

[0023] For the L1, L2, and L3 three levels of cache, a state vector is constructed respectively, and the state vector contains three normalized indexes of current capacity occupancy, hit rate in the past one hour, and average access delay;

[0024] The fusion feature vector and the state vector of the three levels of cache are concatenated in sequence;

[0025] The concatenated vector is input into a prediction network, and the prediction network is a multi-layer perceptron containing three fully connected layers and a Softmax output layer. The prediction network outputs a 3-dimensional probability distribution vector as the initial cache level affinity vector.

[0026] Further, in step S3, based on the initial cache level affinity vector, the node features of the user interaction trajectory graph are adjusted, including:

[0027] The initial cache level affinity vector is taken as a new feature;

[0028] For each data object node in the user interaction trajectory graph, the fusion feature vector and the initial cache level affinity vector are concatenated to generate a new vector with expanded dimensions, and the original fusion feature vector of the data object node is replaced, completing the adjustment of the fusion feature vector for subsequent re-embedding calculation.

[0029] Further, in step S4, the topological importance of the data object in the user interaction trajectory graph is evaluated according to the revised context process vector, and the asymmetric risk cost of the data object when a miss occurs in different cache levels is calculated accordingly, including:

[0030] Based on the revised context process vector, the edge weight in the user interaction trajectory graph is updated by calculating the vector similarity between connected nodes in the user interaction trajectory graph; and the topological importance score of each data object node is iteratively calculated using the PageRank algorithm on the user interaction trajectory graph after the edge weight is updated;

[0031] Set the basic miss cost value for L1, L2, and L3 cache levels;

[0032] Multiply the topological importance score of each data object node by the basic miss cost value of each cache level to obtain the asymmetric risk cost of the data object when a miss occurs in the corresponding cache level.

[0033] Further, in step S4, the cache placement or eviction strategy is solved under the stochastic programming framework to maximize the risk-adjusted expected hit revenue of the system, including: defining decision variables When the data object is placed in the cache level , , otherwise 0; set the hit basic revenue for each cache level; the objective function is the maximization of , where is the expected revenue of the data object placed in the cache level , and the calculation method is: , where is the initial cache level affinity of the data object to the cache level , is the asymmetric risk cost of the data object when a miss occurs in the cache level ; the constraint conditions include: the sum of the sizes of all placed data objects in each cache level does not exceed the total capacity of the corresponding cache level; each data object is placed in at most one cache level;

[0034] An integer linear programming solver is used to solve the programming problem to obtain the cache placement decision.

[0035] Secondly, the present invention provides an intelligent hierarchical cache management system that integrates multimodal interaction features, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent hierarchical cache management method integrating multimodal interaction features is implemented.

[0036] The beneficial effects are as follows: This invention acquires historical multimodal interaction data of users to construct a user interaction trajectory map, capturing the continuity and contextual dependencies of user behavior. By leveraging graph neural networks and attention mechanisms, it deeply integrates multimodal interaction information such as clicks and browsing duration, thereby more accurately discovering user intent and predicting data access trends. Utilizing an asymmetric risk cost assessment mechanism, it analyzes the topological importance of data objects in the user's task flow, evaluating the differentiated impact of different data cache misses on user experience, thus prioritizing the caching of critical data. Through a comprehensive stochastic programming framework, it jointly optimizes multiple dimensions such as initial cache level affinity, asymmetric risk cost, and data object storage size, formulating a caching strategy that maximizes the expected hit benefit after risk adjustment, improving the overall performance of the caching system, reducing access latency for critical tasks, and thus improving the user's perceived service quality. Attached Figure Description

[0037] Figure 1 A flowchart for an intelligent hierarchical cache management method that integrates multimodal interaction features;

[0038] Figure 2 This is a schematic diagram illustrating feature embedding and attention fusion.

[0039] Figure 3 A diagram illustrating the optimization of cache placement decisions;

[0040] Figure 4 A schematic diagram of an intelligent hierarchical cache management system that integrates multimodal interaction features. Detailed Implementation

[0041] An embodiment of the intelligent hierarchical cache management method integrating multimodal interaction features provided by this invention:

[0042] like Figure 1 As shown, the intelligent hierarchical cache management method integrating multimodal interaction features includes:

[0043] S1: Obtain historical multimodal interaction data between the user and the data object, and construct a user interaction trajectory map based on the event sequence.

[0044] Specifically, raw behavior data of users, such as page browsing, clicking, swiping, dwell time, collection, purchase, etc., is collected from server logs or application buries, and constitutes historical multi-modal interaction data. A single session of a user is regarded as an event sequence, and each event in the event sequence corresponds to a data object, such as a product page or a news article. In a user interaction trajectory graph, each data object is abstracted as a node, and the node features include content attributes of the data object, such as product category, price, article theme, etc. If a user continuously accesses data object B from data object A in a session, a directed edge is established between data object A and data object B, and the edge features can encode interaction type, dwell time, and other process information. By aggregating massive user session data, a large-scale directed heterogeneous graph is formed.

[0045] In an optional embodiment, in step S1, historical multi-modal interaction data of users and data objects is obtained, and a user interaction trajectory graph is constructed according to the event sequences therein, including:

[0046] Defining user nodes and data object nodes;

[0047] Extracting interaction events between user nodes and data object nodes, the interaction events including clicking, browsing time longer than 30s, collection, and comment;

[0048] If there is an interaction event between a user node and a data object node, a directed edge is established between them, and the type and occurrence time of the interaction event are taken as attributes of the edge;

[0049] According to the type of the interaction event, an initial weight is assigned to the edge.

[0050] For example, the initial weight corresponding to clicking is 0.2, the initial weight corresponding to browsing time longer than 30s is 0.5, the initial weight corresponding to collection is 0.8, and the initial weight corresponding to comment is 1.0.

[0051] Specifically, a corresponding graph node is created for each user and each data object (such as an article or a video) in the system. By analyzing historical interaction records, interaction events between users and data objects are identified, for example: if user A has once clicked data object B, then browsed for 45s and collected, these three behaviors are identified as an interaction event between user A and data object B.

[0052] Based on the identified interaction events, an edge is established between the user A node and the data object B node, which records the details of the interaction, such as the event type is click, the browsing time exceeds 30s and collection, and their respective timestamps. According to the preset rules, weights are assigned to the interaction behaviors, for example, the weight of click is set to 0.2, the weight of browsing time exceeding 30s is set to 0.5, and the weight of collection is set to 0.8. If the user also makes a comment, the weight of this behavior is the highest 1.0, in this way the importance of different interaction behaviors is represented, forming a complete user interaction trajectory graph.

[0053] S2, using a graph neural network to preliminarily embed the user interaction trajectory graph to obtain an initial context process vector; and based on the initial context process vector, fusing multi-modal features through an attention mechanism to generate a fused feature vector; concatenating the fused feature vector with a vector representing the state of each level of cache and inputting the result into a prediction network to generate an initial cache level affinity vector of the data object for each level of cache.

[0054] Specifically, a graph neural network is used to learn the constructed user interaction trajectory graph. The graph neural network can be a graph convolution network (GCN) or a graph attention network (GAT). The graph neural network aggregates the information of neighbor nodes and edges through multiple iterations to update the representation of any data object node in the user interaction trajectory graph, so that the node representation vector can detect the context information in the interaction trajectory, such as what the user usually looks at before accessing it and where it goes afterwards. The node embedding obtained after the graph neural network processing is the initial context process vector. The initial context process vector is concatenated with the original multi-modal feature vector (such as content features) of the data object itself, and input into an attention network. The attention network learns the weights of different feature dimensions for predicting the user's future intention, and outputs a fused feature vector that contains both process information and content information by weighting and summing each feature according to the weights.

[0055] Suppose there are three levels of cache L1, L2 and L3, the real-time state features of each level of cache are obtained, such as current occupancy rate, recent hit rate, average access delay, etc., and these features form the state vectors of each level of cache. The fused feature vector of the data object is concatenated with the state vectors of all levels of cache to form a longer combined feature vector. The combined feature vector is input into a feedforward neural network, such as a multilayer perceptron (MLP) containing several hidden layers. The output layer of the feedforward neural network has as many nodes as the number of cache levels, and uses a Sigmoid or Softmax activation function, thereby generating a probability vector. Each dimension of the probability vector represents the possibility of placing the data object in the corresponding cache level to produce an access hit, and the probability vector is the initial cache level affinity vector.

[0056] In an optional embodiment, in step S2, the user interaction trajectory graph is preliminarily embedded by using a graph neural network to obtain an initial context process vector, including:

[0057] A graph convolution network GCN is used as the graph neural network, and neighbor node information is aggregated by two layers of convolution to generate a 128-dimensional initial context process vector for each node.

[0058] Specifically, the constructed user interaction trajectory graph is input into a two-layer graph convolution network. For any data object node in the graph, the network aggregates the feature information of all directly connected user nodes through two rounds of information transmission to generate a 128-dimensional vector for the data object node. The vector detects the context environment of the data object node in the interaction network, i.e., the initial context process vector.

[0059] In an optional embodiment, in step S2, based on the initial context process vector, multi-modal features are fused by an attention mechanism to generate a fused feature vector, including:

[0060] Text modal features and image modal features of the data object are extracted, and the corresponding modal features are converted into a 768-dimensional text feature vector and a 2048-dimensional image feature vector by a BERT model and a ResNet-50 model respectively;

[0061] The initial context process vector is taken as a query vector to generate a key vector and a value vector for each of the text feature vector and the image feature vector. The query, key, and value vectors are projected to a unified dimensional space by a learnable linear transformation. The similarity between the query vector and the key vector is calculated by an attention mechanism to obtain an attention weight of each modal feature, and the value vector is weighted and summed by using the attention weight to obtain a fused feature vector.

[0062] Specifically, the text content of the data object and the contained picture are extracted. The text content is processed into a 768-dimensional text feature vector by a BERT model, and the picture is processed into a 2048-dimensional image feature vector by a ResNet-50 model.

[0063] To fuse features from different sources, an attention mechanism is adopted: the 128-dimensional initial context process vector is taken as the query vector, and the 768-dimensional text feature vector and the 2048-dimensional image feature vector are respectively transformed by linear transformation to generate the corresponding key vector and value vector. After linear transformation, the three vectors are mapped to the same dimensional space, and the similarity between the query vector and the two key vectors is calculated to obtain the attention weight of the text feature and the image feature respectively. For example, if the context shows that the user pays more attention to the text content, the text feature will obtain a higher attention weight. According to the attention weight, the value vectors are weighted and summed to generate a fusion feature vector that comprehensively represents the data object, as shown in Figure 2

[0064] In an optional embodiment, in step S2, the fusion feature vector is spliced with the vector representing the state of each level of cache to input the prediction network to generate the initial cache level affinity vector of the data object for each level of cache, including:

[0065] For L1, L2 and L3 caches, state vectors are constructed respectively, and the state vectors contain three normalized indexes of current capacity occupancy, hit rate in the past one hour and average access delay;

[0066] The fusion feature vector is spliced with the state vectors of the three levels of cache in sequence;

[0067] The sequentially spliced vector is input into the prediction network, and the prediction network is a multi-layer perceptron containing three fully connected layers and a Softmax output layer. The prediction network outputs a 3-dimensional probability distribution vector as the initial cache level affinity vector.

[0068] Specifically, the running state of L1, L2 and L3 caches is monitored in real time: for example, at a certain time, the capacity occupancy of L1 cache is 90%, the hit rate in the past one hour is 95%, and the average access delay is 15ns. These raw data are normalized to form a state vector of L1 cache, such as 0.9, 0.95 and 0.2. State vectors of L2 and L3 caches are also generated.

[0069] ​For a specific data object, the generated fusion feature vector is taken out, and the fusion feature vector is spliced with the obtained L1, L2, and L3 state vectors in turn to form a longer combined vector. The combined vector contains not only the characteristics of the data object itself, but also the global state of the current cache system. The combined vector is input into a pre-trained multi-layer perceptron prediction network, and the multi-layer perceptron prediction network outputs a three-dimensional probability vector, such as 0.7, 0.2, and 0.1, which represents the initial suitability or affinity of the data object placed in the L1, L2, and L3 caches.

[0070] S3, based on the initial cache level affinity vector, adjusting the node features of the user interaction trajectory graph, and using the graph neural network to re-embed the calculation to obtain a corrected context process vector.

[0071] Specifically, the generated initial cache level affinity vector is attached to the existing fusion feature vector of the corresponding data object node in the user interaction trajectory graph as a new feature. After completing the feature update, the graph neural network is run again. Since each node is supplemented with its own cache value information, the graph neural network will not only consider the original interaction context when aggregating neighbor information, but also include the cache affinity information of the neighbor nodes in the calculation. The information is propagated and iterated on the graph again, so that each node's embedding representation (i.e., the corrected context process vector) can indirectly perceive the cache potential of other nodes in the neighbor and even the entire interaction graph, thereby obtaining a more global and optimized process representation.

[0072] In an optional embodiment, in step S3, based on the initial cache level affinity vector, adjusting the node features of the user interaction trajectory graph, includes:

[0073] Taking the initial cache level affinity vector as a new feature;

[0074] For each data object node in the user interaction trajectory graph, the fusion feature vector and the initial cache level affinity vector are spliced to generate a new vector with expanded dimensions, and the original fusion feature vector of the data object node is replaced, completing the adjustment of the fusion feature vector for subsequent re-embedding calculation.

[0075] Specifically, after calculating the initial cache level affinity vector of each data object for the three-level cache, the initial cache level affinity vector is taken as a new feature representing the cache characteristics of the data object. Taking a certain data object as an example, assuming that its initial cache level affinity vector is 0.7, 0.2, and 0.1, this three-dimensional vector will be used to enhance the representation of the data object in the user interaction trajectory graph.

[0076] For each data object node, find its original fusion feature vector in the user interaction trajectory graph, and concatenate the newly obtained initial cache level affinity vector to the end of the original fusion feature vector. For example, if the original fusion feature vector is 200-dimensional, after concatenation it becomes a new 203-dimensional vector. The new vector not only contains the original information of the data object, but also incorporates the adaptability evaluation of different cache levels. Replace the original fusion feature vector of the corresponding node in the user interaction trajectory graph with the new vector, complete the update of the fusion feature vector of the nodes in the entire user interaction trajectory graph, and the updated user interaction trajectory graph will be used for more accurate embedding calculation in the next round.

[0077] S4, according to the revised context process vector, evaluate the topological importance of the data object in the user interaction trajectory graph, and calculate the asymmetric risk cost of the data object when the cache request is received. Joint the initial cache level affinity vector, the asymmetric risk cost and the storage size of the data object, solve the cache placement or eviction strategy under the random programming framework to maximize the risk-adjusted expected hit benefit of the system.

[0078] Specifically, based on the revised context process vector, the network centrality index of each node in the user interaction trajectory graph is calculated, such as PageRank centrality or betweenness centrality. The higher the network centrality index score of a node, the more core and critical position it occupies in the user interaction trajectory, and the higher its topological importance. Define a risk cost function that maps the topological importance of the node to a basic risk value. The higher the topological importance of the node, the greater the basic risk value. Set different penalty factors for different cache levels, for example, the penalty factor of the cache level closest to the source data is the smallest, and the penalty factor of the farthest cache level is the largest. The asymmetric risk cost of a certain data object when a cache miss occurs in a certain cache level is equal to the basic risk value of the data object multiplied by the penalty factor of the cache level.

[0079] For a cache request, the risk-adjusted expected hit reward is defined as the objective function, which is equal to the sum of the expected hit rewards of the data objects in each level of cache minus the asymmetric risk cost of a miss. The expected hit reward is obtained by multiplying the initial cache level affinity vector with the base reward (e.g. saved access time) brought by a hit in each level. The cache decision problem is modeled as a multi-level knapsack problem: where each level of cache is a knapsack with limited capacity, the data objects are the items to be put in, the size of an item is the storage size of a data object, and the value of an item is the risk-adjusted expected hit reward. When a new object needs to be placed and the cache is full, the object with the lowest value per unit size (i.e. risk-adjusted expected hit reward divided by size) is evicted to make room for the new object. This decision process ensures that in the limited cache space, the objects with the highest overall value are always kept and accepted.

[0080] In an optional embodiment, in step S4, the topological importance of the data object in the user interaction trajectory graph is evaluated according to the revised context process vector, and the asymmetric risk cost of a miss of the data object in different cache levels is calculated accordingly, including:

[0081] Based on the revised context process vector, the edge weights in the user interaction trajectory graph are updated by calculating the vector similarity between the connected nodes in the user interaction trajectory graph; and the topological importance score of each data object node is iteratively calculated using the PageRank algorithm on the user interaction trajectory graph after the edge weights are updated;

[0082] The base miss cost value is set for the L1, L2, and L3 three-level caches;

[0083] The topological importance score of each data object node is multiplied by the base miss cost value of each level of cache to obtain the asymmetric risk cost of a miss of the data object in the corresponding cache level.

[0084] Specifically, the edge weights of the user interaction trajectory graph are updated using the revised context process vector. For any edge connecting a user and a data object in the user interaction trajectory graph, the cosine similarity of the revised context process vectors of the two nodes is calculated, and the cosine similarity value is taken as the new weight of the edge, which is more accurate than the fixed weight based on the interaction type. After the weights of all edges are updated, the PageRank algorithm is run on the new user interaction trajectory graph, and the topological importance score of each data object node is calculated according to the link structure and weight of the user interaction trajectory graph. The higher the score, the more central and popular the data object is in the network.

[0085] To represent the risk of cache misses, a base cost value is assigned to different cache levels. Since L1 cache is the fastest and has the greatest impact from a miss, its base cost value is set to the highest at 10.0, followed by L2 at 5.0, and L3 at the lowest at 2.0. This method uses the topological importance score of each data object node as the base cost value. The asymmetric risk cost of a data object missing at a certain cache level is equal to the topological importance score of each data object node multiplied by the asymmetric risk cost of a miss at that cache level. For example, a data object with an importance score of 0.05 would have asymmetric risk costs of 0.5, 0.25, and 0.1 for misses in L1, L2, and L3, respectively. This demonstrates that the cost of a cache miss for important data is much higher than for ordinary data.

[0086] In an optional embodiment, in step S4, solving for the cache placement or eviction policy within a stochastic programming framework to maximize the risk-adjusted expected hit payoff of the system includes: defining decision variables. When data object Placed in the cache level hour, Otherwise, the base hit benefit is 0; set the base hit benefit for each level of cache. The objective function is Maximize, where, For data objects Placed in the cache level The expected return is calculated as follows: In the formula, For data objects For cache levels Initial cache level affinity, For data objects At the cache level The asymmetric risk cost of a cache miss; constraints include: the sum of the sizes of all data objects placed in each cache level does not exceed the total capacity of the corresponding cache level; each data object can be placed in at most one cache level;

[0087] An integer linear programming solver is used to solve the planning problem and obtain the cache placement decision.

[0088] Specifically, this method mathematizes the caching decision problem for each data object in the system. Should it be placed in the cache level? Establish a binary decision variable Set base reward values ​​for hits in the L1, L2, and L3 caches, for example, 20, 10, and 5 respectively. The goal is to maximize the system's total reward, which is the sum of the risk-adjusted expected rewards of all data objects, i.e., maximizing... For a single data object Placed in the cache level Expected returns After determining the maximization objective, the following physical constraints must also be satisfied: the total size of all data objects placed in the L1 cache cannot exceed the capacity of L1, and the same applies to L2 and L3; no single data object can be placed in multiple cache levels simultaneously. The complete mathematical model, including the objective function and constraints, is then input into an integer linear programming solver, such as... Figure 3 As shown. The solver will use an optimization algorithm to find a set of optimal solutions. Values, i.e., which ones The set of solutions, which should be 1 and which should be 0, directly corresponds to the final cache placement or eviction policy, such as deciding to put data object A into L1 and evict data object B.

[0089] An embodiment of the intelligent hierarchical cache management system integrating multimodal interaction features provided by this invention:

[0090] like Figure 4 As shown, the intelligent hierarchical cache management system integrating multimodal interaction features includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent hierarchical cache management method integrating multimodal interaction features is implemented.

[0091] The intelligent hierarchical cache management system that integrates multimodal interaction features also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0092] In addition, in the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

Claims

1. An intelligent hierarchical cache management method integrating multimodal interaction features, characterized in that, include: S1, acquire historical multimodal interaction data between users and data objects, and construct a user interaction trajectory map based on the event sequence therein; S2, the user interaction trajectory map is initially embedded using a graph neural network to obtain an initial context process vector; and based on the initial context process vector, multimodal features are fused through an attention mechanism to generate a fused feature vector; the fused feature vector is concatenated with a vector representing the state of each level of cache and then input into a prediction network to generate the initial cache level affinity vector of the data object to each level of cache; S3, Based on the initial cache level affinity vector, adjust the node features of the user interaction trajectory graph, and re-embed the graph neural network to obtain the corrected context process vector; S4. Based on the corrected context process vector, evaluate the topological importance of the data object in the user interaction trajectory graph, and calculate the asymmetric risk cost of the data object when it misses at different cache levels. When a cache request is received, the initial cache level affinity vector, the asymmetric risk cost, and the storage size of the data object are combined to solve for a cache placement or eviction policy within a stochastic programming framework, in order to maximize the system's risk-adjusted expected hit benefit. This includes defining decision variables. When data object Placed in the cache level hour, Otherwise, the base hit benefit is 0; set the base hit benefit for each level of cache. The objective function is Maximize, where, For data objects Placed in the cache level The expected return is calculated as follows: , For data objects For cache levels Initial cache level affinity, For data objects At the cache level The asymmetric risk cost of a miss; The constraints include: the sum of the sizes of all data objects placed in each cache level does not exceed the total capacity of the corresponding cache level; each data object can be placed in at most one cache level; An integer linear programming solver is used to solve the planning problem and obtain the cache placement decision.

2. The intelligent hierarchical cache management method integrating multimodal interaction features according to claim 1, characterized in that, In step S1, historical multimodal interaction data between the user and the data object is acquired, and a user interaction trajectory map is constructed based on the event sequence therein, including: Define user nodes and data object nodes; Extract interaction events between user nodes and data object nodes, including clicks, browsing time exceeding 30 seconds, favorites, and comments; If there are interaction events between user nodes and data object nodes, then an edge is established between them, and the type and occurrence time of the interaction event are used as attributes of the edge; Assign initial weights to edges based on the type of interaction event.

3. The intelligent hierarchical cache management method integrating multimodal interaction features according to claim 2, characterized in that, The initial weight for a click is 0.2, the initial weight for a browsing time of more than 30 seconds is 0.5, the initial weight for a favorite is 0.8, and the initial weight for a comment is 1.

0.

4. The intelligent hierarchical cache management method integrating multimodal interaction features according to claim 1, characterized in that, In step S2, the user interaction trajectory graph is initially embedded using a graph neural network to obtain an initial context process vector, including: A graph convolutional network (GCN) is used as the graph neural network. The neighbor node information is aggregated through two convolutional layers to generate a 128-dimensional initial context process vector for each node.

5. The intelligent hierarchical cache management method integrating multimodal interaction features according to claim 4, characterized in that, In step S2, based on the initial context process vector, multimodal features are fused through an attention mechanism to generate a fused feature vector, including: Text modal features and image modal features of the data objects are extracted, and the corresponding modal features are converted into 768-dimensional text feature vectors and 2048-dimensional image feature vectors using the BERT model and ResNet-50 model, respectively. The initial context process vector is used as the query vector to generate key and value vectors for text and image feature vectors, respectively. The query, key, and value vectors are projected into a unified dimensional space through a learnable linear transformation. The similarity between the query vector and the key vector is calculated through an attention mechanism to obtain the attention weights of each modality feature. The value vectors are then weighted and summed using the attention weights to obtain the fused feature vector.

6. The intelligent hierarchical cache management method integrating multimodal interaction features according to claim 5, characterized in that, In step S2, the fused feature vector is concatenated with the vector representing the state of each cache level and then input into the prediction network to generate the initial cache level affinity vector of the data object to each cache level, including: For the L1, L2, and L3 caches, state vectors are constructed respectively. The state vectors contain three normalized metrics: current capacity utilization, hit rate in the past hour, and average access latency. The fused feature vector is then concatenated sequentially with the state vector of the three-level cache. The concatenated vectors are input into the prediction network, which is a multilayer perceptron containing three fully connected layers and one Softmax output layer. The prediction network outputs a 3-dimensional probability distribution vector, which serves as the initial cache layer affinity vector.

7. The intelligent hierarchical cache management method integrating multimodal interaction features according to claim 1, characterized in that, In step S3, based on the initial cache level affinity vector, the node features of the user interaction trajectory map are adjusted, including: The initial cache level affinity vector is used as a new feature; For each data object node in the user interaction trajectory map, the fused feature vector is concatenated with the initial cache level affinity vector to generate a new vector with expanded dimensions, and the original fused feature vector of the data object node is replaced to complete the adjustment of the fused feature vector for subsequent re-embedding calculation.

8. The intelligent hierarchical cache management method integrating multimodal interaction features according to any one of claims 1-7, characterized in that, In step S4, based on the corrected context process vector, the topological importance of the data object in the user interaction trajectory graph is evaluated, and the asymmetric risk cost of the data object when a miss occurs at different cache levels is calculated accordingly, including: Based on the corrected context process vector, the edge weights in the user interaction trajectory graph are updated by calculating the vector similarity between connected nodes in the user interaction trajectory graph; on the user interaction trajectory graph with updated edge weights, the PageRank algorithm is used to iteratively calculate the topological importance score of each data object node. Set a base miss generation value for the L1, L2, and L3 caches; The asymmetric risk cost of a data object is obtained by multiplying the topological importance score of each data object node by the base miss cost of each cache level.

9. An intelligent hierarchical cache management system integrating multimodal interaction features, characterized in that: It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent hierarchical cache management method integrating multimodal interaction features as described in any one of claims 1-8 is implemented.

Images