A knowledge graph partitioning method for edge cache based on frequent pattern mining

Abstract

The present invention provides an edge cache knowledge graph partitioning method based on frequent pattern mining, which belongs to the field of data processing technology. The present invention uses the frequent query patterns mined from the query graph to color the initial knowledge graph, and divides the knowledge graph into different color areas by improving the exchange algorithm, so that the divided knowledge graph meets the minimum cutting edge while meeting the edge cache size. When the user visits next time, the information of the surrounding required service scenarios can be quickly responded to achieve near-optimal cache performance. The present invention reasonably deploys the knowledge graph to the edge server, so that surrounding users can quickly obtain the required information, greatly reducing the backhaul pressure of obtaining data from the cloud, and providing users with a better interactive experience.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology and relates to an edge cache knowledge graph partitioning method based on frequent pattern mining. When a user visits next time, the information of the surrounding required service scenarios can be quickly responded to achieve near-optimal cache performance. Background Art

[0002] With the rapid development of big data and artificial intelligence technologies, knowledge graphs, with their powerful semantic processing and open interconnection capabilities, break the data isolation in different scenarios and provide users with a better interactive experience. However, when users request information about surrounding services (medical, education, government, shopping malls, etc.), large-scale knowledge graphs will put tremendous pressure on the backhaul network, thereby reducing the quality of user experience. In addition, user queries are very frequent in urban scenarios, and it is very challenging to quickly respond to frequent query requests in nearby edge servers. Mobile edge caching allows application providers to deploy their applications and related data on distributed edge servers to provide services to nearby users. Popular content can be cached at the edge of the network very close to mobile users. Users do not need to obtain the requested content from the cloud center after multi-hop transmission through the backhaul and core network, but obtain the requested content through single-hop wireless transmission, which greatly reduces end-to-end latency. Therefore, it is essential to deploy knowledge graphs on distributed edge servers by using mobile edge caching.

[0003] Existing knowledge graph partitioning algorithms usually use the minimum cut edges or points and consider whether the partition is balanced as the objective function. They do not pay attention to whether the content of the same partition is related, and do not take into account the differences in edge server cache capacity. In actual application scenarios, it is impossible to reasonably distribute the knowledge graph to the edge server while minimizing the number of cross-partition edges and effectively improving the knowledge graph query processing performance.

[0004] Based on this, the present invention utilizes the hot spot information obtained from frequent query patterns to divide the initial knowledge graph into different areas and meet the current server cache capacity. When the user visits next time, it can quickly respond to the information of the surrounding required service scenarios and achieve near-optimal cache performance. Summary of the invention

[0005] In order to solve the problem of how to reasonably deploy large-scale knowledge graphs in edge servers, the present invention provides an edge cache knowledge graph partitioning method based on frequent pattern mining. This method combines the hot spot information obtained by the frequent pattern mining pattern to color the initial knowledge graph, and divides the knowledge graph into different color areas through an improved exchange algorithm. Finally, the hot content of the knowledge graph is deployed in the adjacent edge server while satisfying the minimum cutting edge and being suitable for the edge server cache size.

[0006] The technical solution of the present invention:

[0007] A method for partitioning edge cache knowledge graph based on frequent pattern mining includes the following steps:

[0008] Step 1: The edge server generates a query graph G based on historical queries Q = {V Q ,E Q ,LE Q ,LV Q}, where V Q Representative point set; E Q represents edge set; LE Q is the edge label, indicating the frequency; LV Q is a point label, indicating the category attribute.

[0009] Step 2: Based on the query graph G Q Perform frequent query pattern mining to mine frequent query patterns. The query pattern is defined as P = {V P ,E P ,LE P ,LV P}, where V P represents the query mode point set; E P represents the query pattern edge set; LE P is the query mode edge label, where the edge labels are all empty; LV P It is a query pattern point label, which represents the category attribute. There are no repeated point labels in the query pattern, and it pays more attention to the connection between entities of different categories.

[0010] Step 3: Initially color the knowledge graph based on frequent query patterns, fix the points on the frequent query patterns to the colors of the edge servers to which they belong, randomly initialize the colors of other nodes according to the threshold of the edge cache capacity, and then apply heuristic local search for multi-way partitioning, and push the knowledge graph to a low-energy state (minimum cut edge).

[0011] Step 4: Divide the knowledge graph into its corresponding edge servers according to the partitions of different colors. If the knowledge graph expanded by the frequent query pattern in the current server is a non-connected graph, a directed edge will be established between the two nodes in the knowledge graph according to the connection of the nodes in the frequent query pattern, and the directed edge will be named with the current server name.

[0012] In step 2, the specific steps of frequent query pattern mining are as follows:

[0013] Step 2.1: Initialize the result set and candidate set to be empty, and query graph G Q All frequent edges in are placed in the frequent edge set.

[0014] Step 2.2: Calculate the path graphs of all frequent edges and merge the path graphs generated by all frequent edges into the candidate set.

[0015] Step 2.3: The query pattern generated by the frequent edge e is p. Put the query pattern p into the empty stack S and pop it out of the stack in a loop. If the query pattern at the top of the stack cannot be expanded or has been completely expanded, the expanded query pattern is incorporated into the result set and the query pattern at the top of the stack is popped out of the stack. If the query pattern at the top of the stack can be further expanded, the query pattern p is expanded and recorded as p′, and the path graph expansion of p′ is recorded as pg G (p′).

[0016] Step 2.4: Initial Path Map pg G (p′) is set to empty, and all edges in query pattern p′ are added to the query pattern edge set. The path graph of frequent edge e in the query pattern edge set is merged into the query pattern path graph pg in a loop. G In (p′), the path graph expansion is completed.

[0017] Step 2.5: If the path graph pg G If the minimum independent individual support of (p′) is greater than the specified threshold, the expanded query pattern p′ is put into the stack and the expansion continues. When the stack is empty, the loop ends and the frequent edge e is removed from the frequent edge set.

[0018] Step 2.6: Repeat steps 2.3 and 2.4 until the frequent edge set is empty, and finally output the frequent query patterns in the result set that meet the frequent threshold.

[0019] In step 3, the specific steps of multi-path knowledge graph division are as follows:

[0020] Step 3.1: Find the optimal exchange node partner from the neighbor set of node u (excluding the frequent query mode nodes). If the optimal exchange node cannot be found in the neighbor set, find the optimal exchange node in the sample set;

[0021] Step 3.2: Find the optimal switching node as follows:

[0022] For node u in the neighbor set or sample set, a circular calculation is performed to see whether the energy of node u after exchanging with node v is greater than that before the exchange.

[0023] d u (π v ) α +d v (π u ) α >d u (π u ) α +d v (πv ) α

[0024] Among them, π v Indicates the color of node v, d u (π v ) represents the number of nodes in the neighborhood of node u that have the same color as node v; α is the parameter of the energy function. By adjusting the value of α, the graph can obtain the best result in actual partitioning.

[0025] Step 3.3: In order to avoid falling into the local optimum, simulated annealing technique is used. A temperature parameter T is introduced, where the temperature parameter decreases as time increases.

[0026] If the new energy new=d u (π v ) α +d v (π u ) α With temperature T r The product is greater than the original energy old = d u (π u ) α +d v (π v ) α , and it is the maximum value, then the current node v is the optimal exchange node.

[0027] Step 3.4: Repeat steps 3.2 and 3.3 until all nodes in the neighbor set or sample set are completely traversed, and the temperature T is gradually reduced. r , until T r When it reaches the lower limit 1, the temperature will not decrease any more.

[0028] Step 3.5: If the optimal exchange node is not empty, exchange the colors of node v and node u.

[0029] Beneficial effects of the present invention: Compared with other knowledge graph partitioning algorithms, the present invention mines frequent query patterns based on the items in the query graph to extract hot information that users are interested in in the current server; and the mining process does not consider similar nodes, but pays more attention to the connection between different categories, which is more in line with realistic application scenarios. At the same time, taking into account the different cache sizes of edge servers, the number of nodes of the same color is set according to the cache size of the edge server during the initial coloring, and a multi-directional knowledge graph partitioning algorithm is used to make the divided subgraph have the minimum cut edge to match the number of regional nodes with the cache size of the edge server. The present invention reasonably deploys the knowledge graph to the edge server, so that surrounding users can quickly obtain the required information, greatly reducing the backhaul pressure of obtaining data from the cloud, and providing users with a better interactive experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Process scenario diagram of the edge cache knowledge graph partitioning method based on frequent pattern mining.

[0031] Figure 2 The logical flow chart of the edge cache knowledge graph partitioning method based on frequent pattern mining.

[0032] Figure 3 An example diagram of the query graph pattern provided by the present invention and the path graph corresponding to the pattern.

[0033] Figure 4 An example diagram of the pattern expansion process and its corresponding path graph expansion provided by the present invention.

[0034] Figure 5 This is an example diagram of the exchange of two color nodes in the knowledge graph partitioning algorithm provided by the present invention. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the specific implementation modes of the present invention will be described in detail below with reference to the accompanying drawings.

[0036] Figure 1 This is a process scenario diagram of the edge cache knowledge graph partitioning method based on frequent pattern mining of the present invention. The edge server generates frequent query patterns based on historical query requests of nearby users. After receiving the edge frequent query patterns, the cloud server performs a series of processing to partition the initial knowledge graph. Afterwards, the cloud server deploys the partitioned part of the knowledge subgraph to the corresponding edge cache. When the user sends a query request again, the edge server quickly obtains the information accessed by the user from the cache and responds quickly.

[0037] The edge cache knowledge graph partitioning method based on frequent pattern mining provided by the present invention is as follows: Figure 2 As shown, the following steps are included:

[0038] Step 1: Generate a query graph based on historical query requests. The edge server generates a query graph for the current server based on the historical query requests of surrounding users. The query graph is defined as G Q = {V Q ,E Q ,LE Q ,LV Q}, where V Q Representative point set; E Q represents the edge set, (v i ,v j )∈E represents node v i and v j The undirected edge between them, and v i ,v j ∈V;LEQ (v i ,v j ) is the edge (v i ,v j ) tag, in the present invention, without loss of generality, LE is defined Q (v i ,v j ) contains only one label indicating the frequency; LV Q (v i ) is point v i In the present invention, without loss of generality, LV is defined as Q (v i ) contains only one label to represent the category attribute. Based on the query graph, frequent query patterns can be effectively mined and the connection between hot information in the knowledge graph can be found.

[0039] Step 2: Frequent query pattern mining. Given a query graph G Q = {V Q ,E Q ,LE Q ,LV Q} and frequent threshold τ, the frequent query pattern P is mined in the query graph, where the minimum independent individual support σ of the frequent query pattern P is G (P) ≥ τ. The minimum independent individual support is the minimum value of the number of query graph nodes corresponding to all nodes in the query graph when the query pattern nodes are mapped to the query graph. Frequent query patterns pay more attention to the connection between entities of different categories, considering the significance of pattern mining in real application scenarios. If and only if V P ,E P All are query graphs G Q The points and edges in G, and the point label attributes are consistent with the query graph G Q Same; when the edge label attributes in the frequent query pattern P are all empty and there are no points with the same category attributes, define P = {V P ,E P ,LE P ,LV P} is the query graph G Q = {V Q ,E Q ,LE Q ,LV Q}'s frequent query patterns.

[0040] The frequent query pattern mining described above has the following specific steps:

[0041] Step 2.1: Initialize the result set and candidate set to be empty, and query graph G Q All frequent edges in are placed in the frequent edge set fEdges.

[0042] Step 2.2: Loop through the path graph pg of all frequent edges e in the frequent edge set fEdges G (e) And merge the path graphs generated by all frequent edges into the candidate set candidate.

[0043] Step 2.3: The query pattern generated by the frequent edge e is p. Put the query pattern p into the empty stack S and pop it out of the stack in a loop. If the query pattern at the top of the stack cannot be expanded or has been completely expanded, the expanded query pattern is incorporated into the result set result and the query pattern at the top of the stack is popped out of the stack. If the query pattern at the top of the stack can be further expanded, the query pattern p is expanded and recorded as p′, and the path graph expansion of p′ is recorded as pg G (p′).

[0044] Step 2.4: Initial Path Map pg G (p′) is set to empty, and all edges in the query pattern p′ are added to the query pattern edge set pEdges, and the path graph of the frequent edge e in the query pattern edge set is merged into the query pattern path graph pg in a loop. G In (p′), the path graph expansion is completed.

[0045] Step 2.5: If the path graph pg G If the minimum independent individual support of (p′) is greater than the specified threshold τ, the expanded query pattern p′ is put into the stack and the expansion continues. When the stack is empty, the loop ends and the frequent edge e is removed from the frequent edge set.

[0046] Step 2.6: Repeat steps 2.3 and 2.4 until the frequent edge set is empty, and finally output the frequent query patterns in the result set that meet the frequent threshold.

[0047] Step 3: Initial coloring of the knowledge graph. For the edge server to be deployed, set the initial coloring number C according to its cache size. n ={c1,c2,…,c i ,…,c n}, i = {1, 2, ..., n}. For the frequent query patterns belonging to the current edge server, the points are placed in color color i , and randomly place the remaining uncolored c i Set the node to color color i The colors of the frequent query pattern points will not be exchanged in the subsequent partitioning algorithm, and the colors of the remaining nodes will change with the query graph G. Q The energy of the system is defined as the number of edges between nodes with different colors, while the energy of a node is the number of neighbors with different colors. To keep the size of the partitions fit within the edge cache, nodes cannot change their colors individually, they only exchange colors with each other.

[0048] Step 4: Knowledge graph partitioning. After the initial assignment of node colors, heuristic local search is applied to push the partition configuration to a low-energy state, i.e., the minimum cut edge. Local search is performed in parallel by all nodes in the graph: each node tries to change its color to the most dominant color among its neighbors. However, the node colors in the frequent query pattern are locked and cannot be changed, and in order to maintain the size of the partition, the nodes cannot change their colors individually, they just exchange colors with each other. Each node iteratively selects another node from its neighbors or a random sample, and studies the gain in graph energy after the color exchange. If the color exchange reduces the energy, the two nodes exchange their colors. Otherwise, they retain their own colors. When applying local search, the key issue is to ensure that the algorithm does not fall into a local optimum. To this end, the simulated annealing technique is used.

[0049] Given a knowledge graph H×{V,E}, a k-way knowledge graph partitions the point set V into k subsets. A partition function π:V→{1,…,k} is used to assign a color to each node. Therefore, π u is the color of node u. Nodes with the same color form a partition. We use N u Represents the neighbor set of node u, and N u (c) is defined as the set of neighbors of u with color c:

[0050] N u (c) = {v∈N u :π v =c}

[0051] The number of neighbors of node u is d u Indicates that d u (c)=|N u (c)| is the number of neighbors of u with color c. The energy of the system is defined as the number of edges between nodes with different colors (equivalent to edge cuts). Therefore, the energy of a node is the number of its neighbors with different colors, and the energy of the graph is the sum of the node energies:

[0052]

[0053] Divide the sum by 2, because the sum counts each edge twice. Now we can describe the problem of partitioning the multi-way knowledge graph as: Find the optimal partition π * Make

[0054]

[0055] st|V(c i )|<λ i ,

[0056] where V(c) is the set of nodes with color c, λi is the threshold of edge server i’s capacity.

[0057] The specific steps for multi-path knowledge graph division in step 4 are as follows:

[0058] Step 4.1: Find the optimal exchange node partner from the neighbor set of node u (excluding the frequent query mode nodes). If the optimal exchange node cannot be found in the neighbor set, find the optimal exchange node in the sample set;

[0059] Step 4.2: Find the optimal switching node as follows:

[0060] For node u in the neighbor set or sample set, a circular calculation is performed to see whether the energy of node u after exchanging with node v is greater than that before the exchange.

[0061] d u (π v ) α +d v (π u ) α >d u (π u ) α +d v (π v ) α

[0062] Among them, π v Indicates the color of node v, d u (π v ) represents the number of nodes in the neighborhood of node u that have the same color as node v; α is the parameter of the energy function. By adjusting the value of α, the graph can obtain the best result in actual partitioning.

[0063] Step 4.3: In order to avoid falling into the local optimum, simulated annealing technique is used. A temperature parameter T is introduced, where the temperature parameter decreases as time increases.

[0064] If the new energy new=d u (π v ) α +d v (π u ) α With temperature T r The product is greater than the original energy old = d u (π u ) α +d v (π v ) α , and it is the maximum value, then the current node v is the optimal exchange node.

[0065] Step 4.4: Repeat steps 4.2 and 4.3 until all nodes in the neighbor set or sample set are completely traversed, and the temperature T is gradually reduced. r , until T r When it reaches the lower limit 1, the temperature will not decrease any more.

[0066] Step 4.5: If the optimal exchange node is not empty, exchange the colors of node v and node u.

[0067] Step 5: Determine whether the knowledge graph subgraph divided into the current partition is a connected graph. If not, establish a directed edge named after the current server name between the unconnected frequent query pattern nodes. In subsequent queries, the current server name can be used to quickly retrieve strongly associated knowledge, effectively improving retrieval efficiency. Finally, the divided knowledge graph subgraph is deployed to the corresponding server.

[0068] Figure 3 is an example query graph and one of the query patterns and the path graph corresponding to the query pattern. The path graph can be defined as: given a query pattern P = {V P ,E P ,LE P ,LV P} and query graph G Q , the path graph of P can be expressed as pg G (P), each belonging to pg G A node v of (P) is connected to its reachable neighbors (i.e., the distance is less than a threshold δ), and for any domain adjacent to v (i.e., all nodes containing adjacent labels), there is at least one node connected to v. Figure 3 The path graph of query pattern AB is shown, where 0-4 represents the query graph G Q The path in , means that node u0 can reach node u4 (i.e. u4 is a neighbor of u0). In the example, δ = 10 and τ = 2. The frequency level considers each node in the query pattern P separately, which means the minimum value of the number of nodes in the path graph from each type of node mapping F(v), i.e.

[0069] Figure 4 The query pattern expansion process of the example query graph and its corresponding path graph expansion. Specifically, in order to obtain pg G (ABC), we first get pg G (AB) and pg G (BC) and merge them. Since there are no different nodes (i.e. nodes with label B) in their common domain, there is no need to eliminate invalid nodes and calculate pg G (ABC) = 2. Similarly, when pg G (ABC) and pg G(AC) When merging, because the common domain (i.e., the domain with label A) has no invalid nodes, we continue to process another common domain (i.e., the domain with label C) and find that nodes u3, u6, and u7 are invalid, so they and their corresponding associated edges can be deleted. When node u6 is removed, node u0 does not satisfy the properties in the path graph definition, so node u0 is also removed. Finally, the frequency degree is obtained Path diagram The setting τ=2 is not satisfied.

[0070] Figure 5 For the example of two color swaps, the color swap of nodes p and q in the left graph is acceptable because the nodes change from the state of having 1 and 0 neighbors of the same color to 1 and 3 neighbors of the same color, respectively. If α=1, nodes u and v in the right graph (each node is in the state of having two neighbors of similar colors) will not swap their colors because 2+2>1+3. However, if α>1, nodes u and v will swap their colors. Although this swap does not directly reduce the total edge cuts of the graph, it increases the probability of future color swaps of the two nodes.

[0071] The technical means disclosed in the present invention are not limited to the technical means disclosed in the above-mentioned embodiments, but also include technical solutions composed of any combination of the above-mentioned technical features. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.

Claims

1. A method for partitioning edge cache knowledge graph based on frequent pattern mining, characterized in that: The method comprises the following steps: Step 1: The edge server generates a query graph G based on historical queries Q = {V Q , E Q , L.E. Q , LV Q }, where V Q Representative point set; E Q represents edge set; LE Q is the edge label, indicating the frequency; LV Q is a point label, indicating the category attribute; Step 2: Based on the query graph G Q Perform frequent query pattern mining to mine frequent query patterns. The query pattern is defined as P = {V P , E P , L.E. P , LV P }, where V P represents the query mode point set; E P represents the query pattern edge set; LE P is the query mode edge label, where the edge labels are all empty; LV P It is a query pattern point label, which indicates the category attribute. There is no repeated point label in the query pattern, which pays more attention to the connection between entities of different categories. Step 3: Initially color the knowledge graph based on frequent query patterns, fix the points on the frequent query patterns to the color of the edge server to which they belong, randomly initialize the colors of other nodes according to the threshold of the edge cache capacity, then apply heuristic local search for multi-way partitioning, and push the knowledge graph to a low-energy state; Step 4: Divide the knowledge graph into its corresponding edge servers according to the partitions of different colors. If the knowledge graph expanded by the frequent query pattern in the current server is a non-connected graph, a directed edge will be established between the two nodes in the knowledge graph according to the connection of the nodes in the frequent query pattern, and the directed edge will be named with the current server name.

2. According to the method for partitioning edge cache knowledge graph based on frequent pattern mining according to claim 1, it is characterized in that: In step 2, the specific steps of frequent query pattern mining are as follows: Step 2.1: Initialize the result set and candidate set to be empty, and query graph G Q All frequent edges in are placed in the frequent edge set; Step 2.2: Calculate the path graphs of all frequent edges and merge the path graphs generated by all frequent edges into the candidate set; Step 2.3: The query pattern generated by the frequent edge e is p. Put the query pattern p into the empty stack S and pop it out of the stack in a loop. If the query pattern at the top of the stack cannot be expanded or has been completely expanded, merge the expanded query pattern into the result set and pop the query pattern at the top of the stack. If the query pattern at the top of the stack can be further expanded, the query pattern p is expanded and recorded as p′, and the path graph expansion of p′ is recorded as pg G (p′); Step 2.4: Initial Path Map pg G (p′) is set to empty, and all edges in query pattern p′ are added to the query pattern edge set. The path graph of frequent edge e in the query pattern edge set is merged into the query pattern path graph pg in a loop. G In (p′), the path graph expansion is completed; Step 2.5: If the path graph pg G If the minimum independent individual support of (p′) is greater than the specified threshold, the expanded query pattern p′ is put into the stack and the expansion continues; when the stack is empty, the loop ends and the frequent edge e is removed from the frequent edge set; Step 2.6: Repeat steps 2.3 and 2.4 until the frequent edge set is empty, and finally output the frequent query patterns in the result set that meet the frequent threshold.

3. The method for partitioning edge cache knowledge graph based on frequent pattern mining according to claim 1 or 2, characterized in that: In step 3, the specific steps of multi-path knowledge graph division are as follows: Step 3.1: Find the optimal exchange node partner from the neighbor set of node u. If the optimal exchange node cannot be found in the neighbor set, find the optimal exchange node in the sample set. Step 3.2: Find the optimal switching node as follows: For node u in the neighbor set or sample set, a circular calculation is performed to see whether the energy of node u after exchanging with node v is greater than that before the exchange. d u (p v ) α +d v (p u ) α >d u (p u ) α +d v (p v ) α Among them, π v Indicates the color of node v, d u (π v ) represents the number of nodes in the neighborhood of node u that have the same color as node v; α is the parameter of the energy function. By adjusting the value of α, the graph can obtain the best result in actual partitioning; Step 3.3: In order to avoid falling into the local optimum, simulated annealing technology is used; a temperature T parameter is introduced, where the temperature parameter decreases as time increases; If the new energy new=d u (π v ) α +d v (π u ) α With temperature T r The product is greater than the original energy old = d u (π u ) α +d v (π v ) α , and it is the maximum value, then the current node v is the optimal exchange node; Step 3.4: Repeat steps 3.2 and 3.3 until all nodes in the neighbor set or sample set are completely traversed, and the temperature T is gradually reduced. r , until T r When it is the lower limit 1, the temperature will no longer decrease; Step 3.5: If the optimal exchange node is not empty, exchange the colors of node v and node u.

Images