JP2017211846A - Retrieval data management device, retrieval data management method, and retrieval data management program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a retrieval data management device, a retrieval data management method and a retrieval data management program, capable of reducing the number of edges in consideration of search performance.SOLUTION: A retrieval data management device includes: a storage unit configured to store data of a graph structure to which a multidimensional vector data of a retrieval object is connected by edges; a deletion processing unit configured to extract an extracted edge as an edge satisfying a predetermined condition from the data of the graph structure stored in the storage unit, connect pieces of multidimensional vector data identical to the extracted edge with a route different from the extracted edge, determine whether an alternate path having the number of edges equal to or less than a predetermined number is present or not, and when it is determined that the alternate path is present, delete the extracted edge from the data of the graph structure.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a search data management device, a search data management method, and a search data management program.

  In the field of data search such as image search, data having a graph structure in which search target data is virtually connected by connection elements is constructed, and data search is performed by tracing the connection elements. . The connection element is information indicating that two pieces of data are connected, and is referred to as an edge or a link (hereinafter referred to as an edge). The setting of the edge is performed according to a rule such that a feature amount of data to be searched is first obtained and data having similar feature amounts are connected. The feature amount is typically represented by multidimensional vector data, and that the feature amount is close is defined, for example, as a short distance between the multidimensional vector data. As the distance between the multidimensional vector data, for example, the Lp norm obtained for the difference between the vector elements is used (p = 1, 2,...).

  In addition, there are directed edges and undirected edges. A directed edge is an edge that can trace data in only one direction, and an undirected edge is an edge that can trace data in both directions. Since there are more search path patterns when the undirected edge is set, there are cases where the search can be performed at high speed and with high accuracy. However, for example, when a graph structure in which a directed edge is set is mechanically changed to an undirected edge, many edges may be set for specific data. When the number of edges becomes excessive, the number of man-hours for the search process increases, and the performance may be deteriorated.

  In view of these points, a technique is disclosed in which a predetermined number of data is left out of data connected to certain data of interest and the distance from the data of interest is short, and the connection with the remaining data is released (for example, , See Patent Document 1).

JP 2011-090352 A

However, while the conventional technique can effectively reduce the edges, the search performance tends to decrease because the search performance (search time) is not taken into consideration.
The present invention has been made in view of such circumstances, and an object of the present invention is to provide a search data management device, a search data management method, and a search data management program capable of reducing edges in consideration of search performance. One of them.

  One aspect of the present invention is a storage unit that stores graph structure data in which multidimensional vector data to be searched is connected by an edge, and an edge that satisfies a predetermined condition from the graph structure data stored in the storage unit. Whether a certain extracted edge is extracted, and the same multidimensional vector data as the extracted edge is connected by a route different from the extracted edge, and whether there is an alternative path having a predetermined number of edges or less And a deletion processing unit that deletes the extracted edge from the data of the graph structure when it is determined that the extracted edge exists.

  According to one embodiment of the present invention, edges can be reduced in consideration of search performance.

It is a block diagram centering on the search data management apparatus 1 of embodiment. 2 is a diagram illustrating an example of a hardware configuration of a control unit 20. FIG. It is a figure which shows an example of the content of the existing vector data. 4 is a diagram illustrating an example of the contents of a graph index 44. FIG. It is a flowchart which shows an example of the flow of a process by 26 A of new creation / addition process parts. It is a figure which shows a mode that a neighborhood vector is updated. It is a flowchart which shows another example of the flow of a process by 26 A of new creation / addition process parts. It is a flowchart which shows another example of the flow of a process by 26 A of new creation / addition process parts. FIG. 5 is a diagram conceptually illustrating a state where a vector having the shortest distance from the vector y is extracted from the vector set S and excluded from the vector set S. It is a figure which shows notionally vector N (G, s) of vector s. It is a figure which shows notionally a mode that a vector is added to the vector set S or the vector set R. It is a figure which shows notionally a mode that the vector with the longest distance with the vector y is deleted from the vector contained in the vector set R, and r is reset. It is a flowchart which shows an example of the flow of the process performed by the deletion process part 26B. It is a figure which shows notionally a mode that the alternative path | pass whose number of edges is below p is deleted. It is the figure which illustrated the method of selecting a search start vector.

  Hereinafter, embodiments of a search data management device, a search data management method, and a search data management program according to the present invention will be described with reference to the drawings. In the following description, “the distance between the vectors is short” and “the vectors are close”, and “the distance between the vectors is long” and “the vector is far” have the same meaning.

[overall structure]
FIG. 1 is a configuration diagram centering on a search data management apparatus 1 according to the embodiment. The search data management device 1 is connected to one or more client terminals CL via a network NW. The client terminal is a personal computer, a mobile phone such as a smartphone, a tablet terminal, or another terminal device. The network NW includes a wireless base station, a public line, a dedicated line, a provider terminal, the Internet, and the like. When the search data management device 1 receives the query data from the client terminal CL, the search data management device 1 searches for data similar to the query data and returns the search result to the client terminal CL. The data returned by the search data management device 1 may be the data itself (for example, image data generated by jpg or the like) or an identifier (URL or the like) for referring to the data. Further, the query data and the search target data may be any kind of data such as images, sounds, text data, and the like. In the following description, it is assumed that the search data management device 1 searches for images.

  The search data management device 1 includes, for example, a network interface 10, a control unit 20, an input / output device 30, and a data server 40. The network interface 10 is, for example, a NIC (Network Interface Card).

  The control unit 20 includes, for example, a vector data generation unit 22, an index search unit 24, and a graph index editing unit 26. The graph index editing unit 26 includes a new creation / addition processing unit 26A and a deletion processing unit 26B. These components of the control unit 20 are realized, for example, when a processor such as a CPU (Central Processing Unit) executes a program. Some or all of these functional units may be realized by hardware such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or software. And hardware may be implemented in cooperation. The contents of processing by these functional units will be described later.

  FIG. 2 is a diagram illustrating an example of a hardware configuration of the control unit 20. The control unit 20 has, for example, a configuration in which a CPU 20A, a RAM (Random Access Memory) 20B, and a program memory 20C are connected by an internal bus. Various interfaces such as a network interface 10, an input / output interface, and a memory interface are connected to the internal bus. The RAM 20B stores data to be processed by the CPU 20A, data of processing results, and the like. The program memory 20C includes a part or all of a storage device such as a ROM (Read Only Memory), a HDD (Hard Disk Drive), and a flash memory. Programs such as a vector data generation program 20Ca, an index search program 20Cb, and a graph index editing program 20Cc are stored in the program memory 20C. These programs do not need to be clearly separated, and some or all of them may be shared.

  Returning to FIG. 1, the input / output device 30 includes an input device such as a mouse, a keyboard, and a touch panel that accepts an input operation by an administrator of the search data management device 1, an LCD (Liquid Crystal Display), an organic EL (Electroluminescence) display device, Output devices such as speakers.

  The data server 40 is realized by a storage device such as an HDD or a flash memory. The data server 40 may be shared with the program memory 20C of the control unit 20, or may be realized by a storage device separate from the program memory 20C. The data server 40 may be realized by an external storage device such as a NAS (Network Attached Storage) device that can be accessed from the search data management device 1 via various networks. In this case, the “storage unit” in the claims may be interpreted to indicate the RAM 20B in which data acquired from the data server 40 is temporarily stored.

[Control unit]
Hereinafter, the contents of processing by each functional unit of the control unit 20 will be described. Note that the control unit 20 does not need to be realized by a single processor, and may perform distributed processing for each function. For example, the vector data generation unit 22, the index search unit 24, and the graph index editing unit 26 may be realized by separate processors.

  The vector data generation unit 22 generates multidimensional vector data, which is a feature amount of input data, based on input data received from the client terminal CL or the like. When the input data is an image, the multidimensional vector data is, for example, a local feature amount, BoF (Bag of Features), a color histogram, or a combination thereof. When the input data is text data, the multidimensional vector data is, for example, vector data of word appearance frequency, meaning vectors extracted using a neural network, and the like.

  The index search unit 24 searches the data server 40 using the multidimensional vector data (hereinafter referred to as new vector data) generated by the vector data generation unit 22 as a query, and extracts existing vector data 42 similar to the new vector data. To do. For example, existing vector data 42 and a graph index 44 are stored in the data server 40.

  The existing vector data 42 is multidimensional vector data generated from data to be searched. The existing vector data 42 is associated with data used to generate the existing vector data 42 or an identifier that refers to the data. The search data management device 1 sends to the client terminal CL the existing vector data 42 obtained by using the new vector data as a query, that is, data associated with the existing vector data 42 similar to the new vector data, or an identifier referring to the data. Send. FIG. 3 is a diagram illustrating an example of the contents of the existing vector data 42. As shown in the figure, the existing vector data 42 is the multidimensional vector data and the data used to generate the multidimensional vector data for the identification information (vector ID in the figure) of the multidimensional vector data, or This is data associated with an identifier that refers to data.

  The graph index 44 is information regarding edges connecting a plurality of existing vector data 42, and is data having a graph structure formed by a plurality of edges connecting any two of the existing vector data 42. The edges included in the graph index 44 in the present embodiment are bidirectional edges. FIG. 4 is a diagram illustrating an example of the contents of the graph index 44. As illustrated, the graph index 44 is data in which the vector ID of the existing vector data at both ends to which the edge is connected is associated with the edge ID that is identification information of each edge. The data structures shown in FIGS. 3 and 4 are merely examples. For example, data in which an edge ID of a connected edge is associated with a vector ID is stored in the data server 40 as a graph index 44. It may be stored. In other words, the format of the data structure is not an essential problem in the present embodiment, and data having any data structure may be stored in the data server 40 as long as the required properties are satisfied.

  The index search unit 24 resembles the existing vector data 42 within a predetermined distance with respect to the new vector data among the existing vector data 42 that can be traced by the edge defined by the graph index 44, similar to the new vector data. Extract as existing vector data. The distance between vectors is defined as, for example, the Lp norm obtained for the difference between vector elements (p = 1, 2,...). In addition, the index search unit 24 selects a predetermined number of existing vector data 42 in order from the shortest distance to the new vector data among the existing vector data 42 that can be traced by the edge defined by the graph index 44. It may be extracted as existing vector data similar to. Then, the index search unit 24 transmits the extracted existing vector data 42 to the client terminal CL via the network interface 10 and the network NW.

  Note that the search processing by the index search unit 24 is realized using, for example, the processing of the flowchart of FIG. In the flowchart of FIG. 8, a vector set R obtained from new vector data as a vector y is an example of a search result by the index search unit 24. In this case, the number of vectors included in the search result can be adjusted by adjusting the predetermined number ks to a desired value.

  With the functions of the vector data generation unit 22 and the index search unit 24, the client terminal CL can acquire data similar to the transmitted data from the search data management device 1. Note that the role of the search data management apparatus 1 may be up to extracting the existing vector data 42 similar to the new vector data. In that case, the data corresponding to the extracted existing vector data 42 or data is referred to. An apparatus different from the search data management apparatus 1 may have a function of transmitting the identifier to be transmitted to the client terminal CL.

  The new creation / addition processing unit 26A of the graph index editing unit 26 creates a new graph index 44 or performs addition processing. For this processing, for example, one of the two methods described below can be adopted. In the following description, a multidimensional vector may be simply referred to as a vector.

(New creation / addition method 1)
FIG. 5 is a flowchart showing an example of the flow of processing by the new creation / addition processing unit 26A. The process of this flowchart can be used when a new graph index 44 is created, and when new multidimensional vector data is added, the multidimensional vector data already associated with the graph index 44 and the new It can also be used for the process of adding new multidimensional vector data by executing it on multi-dimensional vector data. As a premise of the processing of this flowchart, it is assumed that vector IDs 1 to n are assigned to multidimensional vector data included in the existing vector data 42. Arguments i and j used in the flowchart are sequentially selected from 1 to n and executed. That is, the new creation / addition processing unit 26A first sets the argument i to 1, and repeatedly performs the processes of S100 to S106 until the argument j reaches n while incrementing the argument j one by one. Next, the new creation / addition processing unit 26A sets the argument i to 2, and repeats the processes of S100 to S106 until the argument j reaches n while incrementing the argument j one by one. This is repeated until the argument i reaches n.

  The new creation / addition processing unit 26A determines whether the vector i (the vector with the vector ID i; the same applies hereinafter) is an element of the neighborhood vector N (G, j) of the vector j or i = j. Is determined (S100). The neighborhood vector N (G, j) of the vector j is a set of vectors selected by a predetermined number kp in order from the shortest distance when viewed from the vector j. “G” is a code indicating data having a graph structure. Each vector that is an element of the neighborhood vector N (G, j) of the vector j and the vector j are connected by an edge. Accordingly, at least a vector that is an element of the neighborhood vector N (G, j) of the vector j is a vector that can be traced from the vector j through one edge. Note that the vector j may be an element of a neighborhood vector of another vector that is not an element of the neighborhood vector N (G, j) of the vector j. Hereinafter, this will be described on the assumption. When the vector i is an element of the neighborhood vector N (G, j) of the vector j, or when i = j, the processing of S102 to S106 is skipped.

  When the vector i is not an element of the neighborhood vector N (G, j) of the vector j and i = j, the new creation / addition processing unit 26A converts the vector i to the neighborhood vector N (G, j) of the vector j. It adds to an element (S102). Next, the new creation / addition processing unit 26A determines whether or not the number of vectors that are elements of the neighborhood vector N (G, j) of the vector j exceeds a predetermined number kp (S104). The predetermined number kp is an arbitrary natural number. For example, kp = 3. When the number of vectors that are elements of the neighborhood vector N (G, j) of the vector j exceeds the predetermined number kp, the new creation / addition processing unit 26A calculates the vector j from the neighborhood vector N (G, j) of the vector j. The vector having the longest distance from (the farthest from the vector j) is excluded (S106).

  FIG. 6 is a diagram illustrating how the neighborhood vector is updated. In the illustrated example, when vectors (1) to (3) were originally set in the neighborhood vector N (G, j) of the vector j, the vector i was added and the vector (3) was excluded. As a result, the neighborhood vector N (G, j) of the vector j is updated to N (G, j) *. By repeating such processing, the neighborhood vector N (G, x) is set for all the vectors x = 1 to n. Thereby, the graph index 44 is created.

(New creation / addition method 2)
7 and 8 are flowcharts showing another example of the processing flow by the new creation / addition processing unit 26A. As a premise of the processing of this flowchart, n-dimensional vector data having a vector ID from 1 to n assigned to the existing vector data 42 and having the graph index 44 already exist, and mn pieces of new multi-dimensional vector data already exist. It is assumed that multidimensional vector data is added (m> n). The argument y used in the flowchart is selected in order from (n + 1) to m, and the process is executed.

  The new creation / addition processing unit 26A executes a K neighborhood search process (S200). The K neighborhood search process will be described with reference to FIG. As shown in FIG. 8, the new creation / addition processing unit 26A sets the radius r of the hypersphere to ∞ (infinity) (S300), and extracts the vector set S from the existing vector set (S302). The hypersphere is a virtual sphere indicating a search range (a range that can be included in the elements of the neighborhood vector N (G, y) of the input vector y if it is within the range). Note that the vectors included in the vector set S extracted in step S302 may be included in the initial set of the vector set R at the same time.

  Next, the new creation / addition processing unit 26A extracts a vector having the shortest distance from the vector y among the vectors included in the vector set S and sets it as a vector s (S304). Next, the new creation / addition processing unit 26A excludes the vector s from the vector set S (S306). FIG. 9 is a diagram conceptually illustrating how a vector having the shortest distance from the vector y is extracted from the vector set S and excluded from the vector set S.

  Next, the new creation / addition processing unit 26A determines whether the distance d (s, y) between the vector s and the vector y exceeds r (1 + ε) (S308). Here, ε is an expansion element, and r (1 + ε) indicates whether or not the search range (a neighborhood vector of a vector in the range is included in the neighborhood vector N (G, y) of the input vector y). This is a value indicating the radius of the range to be determined. When the distance d (s, y) between the vector s and the vector y exceeds r (1 + ε), the new creation / addition processing unit 26A outputs the vector set R as a neighborhood vector N (G, y) of the vector y. Then, the process of the flowchart of FIG.

  When the distance d (s, y) between the vector s and the vector y does not exceed r (1 + ε), the new creation / addition processing unit 26A determines the vector that is an element of the neighborhood vector N (G, s) of the vector s. One vector that is not included in the vector set C is selected from them, and the selected vector u is stored in the vector set C (S312). FIG. 10 is a diagram conceptually showing the neighborhood vector N (G, s) of the vector s. The vector set C is provided for convenience in order to avoid duplicate search, and is reset to an empty set when the flowchart of FIG. 8 is started.

  Next, the new creation / addition processing unit 26A determines whether or not the distance d (u, y) between the vector u and the vector y is equal to or less than r (1 + ε) (S314). When the distance d (u, y) between the vector u and the vector y is equal to or less than r (1 + ε), the new creation / addition processing unit 26A adds the vector u to the vector set S (S316).

  Next, the new creation / addition processing unit 26A determines whether the distance d (u, y) between the vector u and the vector y is equal to or less than r (S318). If the distance d (u, y) between the vector u and the vector y exceeds r, the process proceeds to S330.

  When the distance d (u, y) between the vector u and the vector y is equal to or less than r, the new creation / addition processing unit 26A adds the vector u to the vector set R (S320). FIG. 11 is a diagram conceptually illustrating how a vector is added to the vector set S or the vector set R. In the illustrated example, vectors (4) and (5) whose distance from the vector y is equal to or less than r are added to the vector set R, and the vector (6) whose distance from the vector y exceeds r and is equal to or less than r (1 + ε). ) Is added to the vector set S.

  Then, the new creation / addition processing unit 26A determines whether or not the number of vectors included in the vector set R exceeds ks (S322). The predetermined number ks is a natural number that is arbitrarily determined. For example, ks = 3.

  When the number of vectors included in the vector set R exceeds ks, the new creation / addition processing unit 26A excludes the vector having the longest distance from the vector y among the vectors included in the vector set R from the vector set R. (S324).

  Next, the new creation / addition processing unit 26A determines whether or not the number of vectors included in the vector set R matches ks (S326). When the number of vectors included in the vector set R matches ks, the new creation / addition processing unit 26A determines the distance between the vector having the longest distance from the vector y among the vectors included in the vector set R and the vector y. Is set to a new r (S328).

  FIG. 12 is a diagram conceptually illustrating a state in which a vector having the longest distance from the vector y is excluded from vectors included in the vector set R, and r is reset. In the example of FIG. 12, in addition to the vector s, (4), and (5) added to R in FIG. 11, the vector (7) is included in R as a result of the previous loop processing. . In this case, the vector (5) having the longest distance from the vector y is excluded from R, and the distance between the vector (7) having the longest distance from the vector y and the vector y among the remaining vectors is set to r. Is done.

  Then, the new creation / addition processing unit 26A determines whether all vectors have been selected from the vectors that are elements of the neighborhood vector N (G, s) of the vector s and stored in the vector set C (S330). ). If all vectors have not been selected from the vectors that are elements of the neighborhood vector N (G, s) of the vector s and stored in the vector set C, the process returns to S312.

  When all the vectors are selected from the vectors that are elements of the neighborhood vector N (G, s) of the vector s and stored in the vector set C, the new creation / addition processing unit 26A indicates that the vector set S is an empty set. It is determined whether or not there is (S330). When the vector set S is not an empty set, the process is returned to S304. When the vector set S is an empty set, the new creation / addition processing unit 26A converts the vector set R into a neighborhood vector N (G, y) of the vector y. And the process of the flowchart of FIG. 8 ends (S332).

  By performing the processing shown in FIG. 8, it is possible to realize high-speed processing while suppressing consumption of the memory area as compared to the case of exhaustively processing all vectors. That is, the process shown in FIG. 8 proceeds mainly by following an edge defined by the graph index 44, and therefore, the processing man-hours can be reduced compared to exhaustively searching the existing vector data 42. In addition to the memory area necessary for the existing vector group, data relating to a vector set S serving as a seed, a vector set C for avoiding duplicate search, and a vector set R in which the maximum number of vectors is limited to ks Since it only has to be managed, consumption of the memory area can be suppressed. Furthermore, a vector sufficiently distant from the vector y as input data is unlikely to be added to the vector set S by processing, and even if it is included in the vector set S in S302, the search target is low, Since the possibility that it will not be increased increases, the number of processing steps can be further reduced.

  Returning to FIG. When the neighborhood vector N (G, y) of the vector y is obtained, the new creation / addition processing unit 26A selects one vector that is an element of the neighborhood vector N (G, y) of the vector y, (S202). Next, the new creation / addition processing unit 26A adds the vector y as an element of the neighborhood vector N (G, z) of the vector z (S204).

  Next, the new creation / addition processing unit 26A determines whether or not the number of vectors that are elements of the neighborhood vector N (G, z) of the vector z exceeds kp (S206). When the number of vectors that are elements of the neighborhood vector N (G, z) of the vector z exceeds a predetermined number kp, the new creation / addition processing unit 26A calculates the vector z from the neighborhood vector N (G, z) of the vector z. The vector having the longest distance from (the furthest from the vector j) is excluded (S208).

  Then, the new creation / addition processing unit 26A determines whether or not all vectors z that are elements of the neighborhood vector N (G, y) of the vector y have been selected (S210). When all the vectors z that are elements of the neighborhood vector N (G, y) of the vector y are not selected, the process returns to S202, and the vector z that is an element of the neighborhood vector N (G, y) of the vector y is obtained. If all of them are selected, the loop processing of S200 to S208 ends. Thereby, the graph index 44 is created.

  Here, the processes of (new creation / addition method 1) and (new creation / addition method 2) may be used in combination. That is, when the vector group is first given and the graph index 44 is created, the processing of (new creation / addition method 1) is performed, and then when the vector group is added (new creation / addition method 2) Processing may be performed. In that case, kp and ks may be the same or different. Alternatively, the processing of (New creation / addition method 1) may be performed each time a vector group is added.

(Deletion process)
The deletion processing unit 26B extracts an edge satisfying a predetermined condition from the graph index 44 created by the new creation / addition processing unit 26A, and connects the same vector as the extracted edge (extracted edge) through a different route. When there are alternative paths with the number of edges equal to or less than the predetermined number p, a process of deleting the extracted edges is performed. The predetermined condition is, for example, the determination condition of S402 in FIG.

  FIG. 13 is a flowchart illustrating an example of the flow of processing executed by the deletion processing unit 26B. As a premise of the processing of this flowchart, it is assumed that vector IDs 1 to n are assigned to the multidimensional vector data included in the existing vector data 42 (after the flowcharts of FIGS. 7 and 8 are executed). If so, vector IDs of 1 to m may be given). An argument e to be used in the flowchart is selected in order from 1 to n and the process is executed.

  First, the deletion processing unit 26B selects one vector from vectors that are elements of the neighborhood vector N (G, e) of the vector e and sets it as a vector f (S400). Next, the deletion processing unit 26B determines whether Rank {N (G, e), f} exceeds a predetermined number kr (S402). Rank {N (G, e), f} is an index value indicating how many times the vector f is closest to the vector e among the vectors that are elements of the neighborhood vector N (G, e) of the vector e. It is. The predetermined number kr is a value less than the predetermined number kp, for example. Accordingly, the process of “determining whether Rank {N (G, e), f} exceeds a predetermined number kr” is performed from the vector that is an element of the neighborhood vector N (G, e) of the vector e to the vector e. Is a process of extracting a vector having a long distance compared to other vectors. When Rank {N (G, e), f} does not exceed the predetermined number kr, the processing of S404 to S410 is skipped.

  When Rank {N (G, e), f} exceeds a predetermined number kr, the deletion processing unit 26B selects an alternative path search start vector (S404), and searches for a previous path from the search start vector (S406). ).

  Next, the deletion processing unit 26B determines whether there is an alternative path with the number of edges equal to or less than p (S408). When there is an alternative path having the number of edges equal to or less than p, the deletion processing unit 26B deletes the edge (e, f) connecting the vector e and the vector f from the graph index 44 (S410).

  FIG. 14 is a diagram conceptually illustrating a state where an alternative path having the number of edges of p or less is deleted. In the illustrated example, both the edge (e, f (1)) connecting the vectors e and f (1) and the edge (e, f (2)) connecting the vectors e and f (2) are both There is an alternate path. The number of edges of the alternative path of edge (e, f (1)) is 3, whereas the number of edges of the alternative path of edge (e, f (2)) is 4. Here, when p = 3, the edge (e, f (1)) is treated as erasable, and the edge (e, f (2)) is treated as irremovable.

  Next, the deletion processing unit 26B determines whether all vectors have been selected from the vectors that are elements of the neighborhood vector N (G, e) of the vector e (S412). If not all vectors are selected from the vectors that are elements of the neighborhood vector N (G, e) of the vector e, the process returns to S400. On the other hand, when all the vectors are selected from the vectors that are elements of the neighborhood vector N (G, e) of the vector e, the processing of the flowchart of FIG. 13 ends.

  With such processing, it is possible to reduce edges in consideration of search performance. In this type of technique, if a large number of edges are added to a vector, the search performance is rather degraded. Therefore, by preferentially deleting edges with a relatively long distance, the number of local searches can be reduced and the search process can be made more efficient. However, if an edge with a long distance is deleted indefinitely, the path to be detoured becomes long, and the search process may rather become inefficient. In addition, there is a case where there is no detour path and a vector that cannot be searched for is generated.

  In view of such circumstances, in the search data management device 1 of the present embodiment, after deleting the edge, the edge is deleted only when the same vector can be reached with a relatively small number of edges. . As a result, the number of edges can be reduced in consideration of search performance, and the number of edges can be maintained appropriately.

  Here, in the process of S <b> 404 in the flowchart of FIG. 13, as a technique for selecting a search start vector, the following technique may be mentioned.

(Search target node selection method 1)
The deletion processing unit 26B may comprehensively select all vectors connected to the vector e via edges as search start vectors. As a result, an alternative path having the number of edges of p or less can be found with certainty.

(Search target node selection method 2)
The deletion processing unit 26B may start a search for an alternative path using a vector closest to the vector f among vectors connected to the vector e via an edge as a search start vector. As a result, the load of the alternative path search process can be reduced most.

(Search target node selection method 3)
The deletion processing unit 26B substitutes, as a search start vector, a vector whose distance between the vector e and the vector f is shorter than the distance between the vector e and the vector f among the vectors connected to the vector e via an edge. Path searching may be started. As a result, the load of the alternative path search process can be reduced, and an alternative path having the number of edges equal to or less than p can be reliably detected as compared with (search target node selection method 2).

  In any of the methods, the graph search is performed using only edges from the shortest one to the rank kr among the edges assigned to the vector e (only edges from the shortest to the rank kr in order). As a search start vector). This is because even if an alternative path including an edge larger than the rank kr is found, the edge on the alternative path may be deleted by subsequent deletion processing.

  FIG. 15 is a diagram illustrating a method for selecting a search start vector. In this figure, when the (search target node selection method 2) is adopted, only the vector (8) is selected as the search start vector. On the other hand, when (search target node selection method 3) is adopted, vectors (8), (9), and (10) are selected as search start vectors.

  In any case, when the number of edges reaches p, the search for the route is terminated, thereby suppressing an increase in processing load. The deletion processing unit 26B exhaustively searches the vector after the search start vector as a starting point, and when the vector f cannot be reached before the number of edges reaches p, the number of edges starting from the search start vector Is determined to have failed to find an alternative path equal to or less than p, and the search from the search start vector is terminated.

  According to the search data management device 1 of the embodiment described above, the storage unit (40) that stores the data (42, 44) of the graph structure in which the multidimensional vector data to be searched is connected by the edge, and the storage unit Extracted edges satisfying a predetermined condition are extracted from the stored graph structure data (S402 in FIG. 13; edges (e, f)), and the same multidimensional vector data (e, f) as the extracted edges are extracted. It is determined whether or not there is an alternative path having a number of edges equal to or less than a predetermined number (p) and connected by a route different from the extracted edge. By providing the deletion processing unit (26B) to be deleted, it is possible to reduce edges in consideration of search performance.

  As mentioned above, although the form for implementing this invention was demonstrated using embodiment, this invention is not limited to such embodiment at all, In the range which does not deviate from the summary of this invention, various deformation | transformation and substitution Can be added.

DESCRIPTION OF SYMBOLS 1 ... Search data management apparatus, 10 ... Network interface, 20 ... Control part, 22 ... Vector data generation part, 24 ... Index search part, 26 ... Graph index edit part, 26A ... New creation / addition process part, 26B ... Deletion process Part, 30 ... I / O device, 40 ... data server, 42 ... existing vector data, 44 ... graph index

Claims (9)

A storage unit for storing data of a graph structure in which multidimensional vector data to be searched is connected by edges;
Extracting the extracted edge that is an edge satisfying a predetermined condition from the data of the graph structure stored in the storage unit,
Connecting the same multi-dimensional vector data as the extracted edge with a path different from the extracted edge, and determining whether there is an alternative path having a predetermined number of edges or less,
A deletion processing unit for deleting the extracted edge from the data of the graph structure when it is determined that it exists,
A search data management device comprising: The predetermined condition is that, among a plurality of edges connected to certain multidimensional vector data, a distance between connected multidimensional vector data is long in comparison with other edges.
The search data management device according to claim 1. The predetermined condition is that, when a plurality of edges connected to certain multidimensional vector data are ranked in ascending order of distance between connected multidimensional vector data, the predetermined condition is lower than a predetermined order.
The search data management device according to claim 1 or 2. The deletion processing unit, when searching for the alternative path, among the multidimensional vector data connected to one multidimensional vector data among the multidimensional vector data connected by the extracted edge, The multidimensional vector data closest to the other multidimensional vector data among the multidimensional vector data connected by the edge is set as a search start vector.
The search data management device according to any one of claims 1 to 3. The deletion processing unit, when searching for the alternative path, among the multidimensional vector data connected to one multidimensional vector data among the multidimensional vector data connected by the extracted edge, The multidimensional vector data closer to the other multidimensional vector data among the multidimensional vector data connected by the extracted edge than the multidimensional vector data is set as a search start vector.
The search data management device according to any one of claims 1 to 3. Creation of the graph structure data based on the distance between multiple multidimensional vectors, by limiting the number of other multidimensional vectors connected to the target multidimensional vector via edges to a predetermined number Further comprising
The search data management device according to any one of claims 1 to 5. A generating unit that generates multidimensional vector data as feature data of query data received from another device;
Based on the multidimensional vector data generated by the generation unit and the data of the graph structure, multidimensional vector data similar to the multidimensional vector data generated by the generation unit is extracted and returned to the other device Further comprising a search unit for
The search data management device according to any one of claims 1 to 6. Extract the extracted edge, which is the edge that satisfies the predetermined condition, from the data of the graph structure in which the multidimensional vector data to be searched is connected by the edge,
Connecting the same multi-dimensional vector data as the extracted edge with a path different from the extracted edge, and determining whether there is an alternative path having a predetermined number of edges or less,
If it is determined that it exists, the extracted edge is deleted from the data of the graph structure;
Search data management method. On the computer,
From the data of the graph structure in which the multidimensional vector data to be searched is connected by the edge, the extracted edge that is an edge satisfying a predetermined condition is extracted,
Connecting the same multi-dimensional vector data as the extracted edge with a route different from the extracted edge, and determining whether there is an alternative path having a predetermined number of edges or less,
When it is determined that it exists, the extracted edge is deleted from the data of the graph structure.
Search data management program.

Images