JP2020119487A - Deep-learning forecast data reflection system using olap result with pivot table structure - Google Patents

Abstract

To provide a system for supporting accurate decision making by adding future prediction data to an OLAP result capable of multidimensional analysis, and providing the segmented future prediction as well as multi-dimensional analysis of the historical data.SOLUTION: A prediction reflection server 30 includes: a data set receiving unit for receiving a data set of the pivot table structure; a dimension extraction unit for generating a row dimension tree and a column dimension tree having a data dimension in a hierarchical structure in a data set of a pivot table structure; a frame converting unit which combines the row dimension tree and column dimension tree to generate a merge tree, and forms a data set of a record structure from the merge tree; a data predictor for predicting the data set of the record structure using deep learning; and a data reconstruction unit reconstructing the merge tree using the record data set to which prediction data has been added and reconstructing the pivot table using the reconstructed merge tree.SELECTED DRAWING: Figure 2

Description

INDUSTRIAL APPLICABILITY The present invention detects a dataset having a time-series characteristic from a dataset having a pivot table structure generated based on online analytical processing (OLAP) and performs deep learning (deep learning). ) Method and reflect the prediction data set in the pivot table structure, the present invention relates to a deep learning prediction data reflection system using the OLAP result of the pivot table structure.

Generally, an online analytical processing system, that is, an OLAP system is a typical example of a decision support system, and multidimensional analysis capable of analyzing data by using a plurality of criteria (dimensions). To enable.

Prior art OLAP systems provide pre-built data from various aspects, in a stylized form. However, while the OLAP system helps analyze historical data by multiple criteria to understand the phenomenon, it cannot provide future predictions. Should the OLAP system be able to subdivide the data and also show future expectations for the data, it would be useful for decision making.

However, OLAP systems typically display datasets in a pivot table structure to present the data in various aspects (or multidimensional). In the pivot table, the horizontal axis (row) and the vertical axis (column) are configured by respective dimension names (item names, etc.), and the cells where the horizontal axis (row) and the vertical axis (column) are in contact are configured as a data area. It is in table format. That is, the pivot table indicates data values by combining a plurality of dimensions (multidimensional).

That is, the structure of the pivot table (Tabular) that is the OLAP result changes in depth and form according to the multidimensional structure. In the pivot table structure format, it is difficult to predict the data, and even if the data is predicted, it is difficult to reflect the prediction result in the pivot table structure format.

Korean Registered Patent No. 1034428 (2011.05.12. Announcement) Japanese Published Patent No. 2002-007435 (2002.01.11. Announcement) Korean Registered Patent No. 09696656 (2010.07.14. Announcement) Korean Registered Patent No. 0424144 (2004.03.24. Announcement)

The present invention is to solve the above-mentioned problem, and an object thereof is to provide a data set having a time series characteristic with respect to a data set of a pivot table structure generated based on OLAP (Online Analytical Processing). It is to provide a deep learning prediction data reflecting system using the OLAP result of the pivot table structure, which detects and predicts by the deep learning method and reflects the prediction data set in the pivot table structure.

In particular, an object of the present invention is to extract a dimension of a hierarchical structure from a pivot table structure and use it to reconstruct a pivot table structure data set as a record structure (or data frame structure) data set and reconstruct it. It is to provide a deep learning prediction data reflection system using the OLAP result of the pivot table structure, which performs the deep learning prediction using the created data set.

In order to achieve the above object, the present invention relates to a deep learning prediction data reflection system using OLAP results of a pivot table structure,
A data set receiving unit for receiving a data set having a pivot table structure; a dimension extraction unit for generating a row dimension tree and a column dimension tree having data dimensions in a hierarchical structure from the data set having a pivot table structure; A frame conversion unit that combines the column dimension tree to generate an integrated tree and forms a record-structured data set from the integrated tree; a data prediction unit that predicts a record-structured data set using deep learning; And a data reconfiguring unit that reconfigures the integrated tree by using the record data set to which the prediction data is added and reconfigures the pivot table by using the reconfigured integrated tree.

Further, in the present invention, in the deep learning prediction data reflection system using the OLAP result of the pivot table structure, the dimension extraction unit may include a row dimension tree and a row dimension tree based on a hierarchical structure of the row dimension and the column dimension of the pivot table. A column dimension tree is configured, the root node of each dimension tree is set as a virtual node, each dimension name of the pivot table is associated with one node to generate each node, and the relationship between the upper dimension and the lower dimension is established. Set the upper and lower relationships between the nodes based on the relationship, set each node in the dimension tree to have a dimension name, and set each node as an upper node or lower node based on the upper or lower relationship of the dimension. It is characterized by being configured hierarchically.

Further, in the present invention, in the deep learning prediction data reflecting system using the OLAP result of the pivot table structure, the dimension extracting unit is configured such that each node of the upper category in the dimension tree is likewise all nodes of the lower category. It is characterized in that it is configured to have.

Further, the present invention is a deep learning prediction data reflecting system using an OLAP result of a pivot table structure, wherein the dimension extracting unit connects a data node having a data value to a major node to a dimension tree having a measure. The dimensional tree is generated by the number corresponding to the number of data sets of the measure in the data area of the pivot table.

Further, in the present invention, in the deep learning prediction data reflection system using the OLAP result of the pivot table structure, the frame conversion unit sets the dimension tree having a measure as a lower tree and the remaining trees as upper trees, It is characterized in that the two-dimensional trees are integrated so that the leaf node of the upper tree becomes the root node of the lower tree.

Also, in the present invention, in the deep learning prediction data reflection system using the OLAP result of the pivot table structure, the frame conversion unit sets the dimension name of the measure as a field for each measure, and the data value of the measure. Is set as the field value of the relevant field, the category of the upper node of the major node is set as the field, and the dimension name of the relevant category is set as the field value of the relevant field. It is characterized in that one record is constructed using and.

Further, the present invention is a deep learning prediction data reflecting system using an OLAP result of a pivot table structure, wherein the data prediction unit targets a field having time series data in a record structure data set and measures the next cycle. The data value of the field is predicted, but the prediction data is obtained for each combination of the remaining dimensions excluding the time series dimension. For the data of the time series dimension, the predicted data for each combination in the next cycle is recorded as a record. It is characterized by forming.

Further, in the present invention, in the deep learning prediction data reflecting system using the OLAP result of the pivot table structure, the data reconstructing unit reflects the dataset of the record structure of the prediction data in the integration tree, and the integration tree is the pivot table. It is characterized by being reconstructed as a structure data set.

As described above, according to the deep learning prediction data reflection system using the OLAP result of the pivot table structure according to the present invention, it is possible to add and display future prediction data to the OLAP result that enables multidimensional analysis. In addition to providing multidimensional analysis on past data, it also provides subdivided forecasts for the future, which has the effect of supporting accurate decision making.

Further, according to the deep learning prediction data reflecting system using the OLAP result of the pivot table structure according to the present invention, the pivot table structure is reconfigured as a record structure to perform the deep learning prediction, and accordingly, the multi-dimensional structure can be obtained. It is possible to overcome problems that are difficult to predict due to changes in depth and form.

It is a block diagram of the composition of the whole system for carrying out the present invention. FIG. 6 is a block diagram of a configuration of a deep learning prediction data reflecting system using the OLAP result of the pivot table structure according to the embodiment of the present invention. It is an illustration figure of the pivot table which concerns on one Embodiment of this invention. FIG. 6 is a view showing a row dimension tree according to an exemplary embodiment of the present invention. FIG. 6 is a view showing an example of a column dimension tree according to an exemplary embodiment of the present invention. FIG. 5 is a view showing an example of a lower-dimensional tree instance according to an embodiment of the present invention. FIG. 7 is an exemplary diagram of a row dimension tree including a sum node according to an exemplary embodiment of the present invention. FIG. 9 is a view showing an example of a tree in which a total tree according to an exemplary embodiment of the present invention is separately configured. FIG. 6 is a view showing an integrated tree according to an exemplary embodiment of the present invention. FIG. 7 is a view showing an integrated tree including a total node according to an exemplary embodiment of the present invention. FIG. 4 is a view showing an example of a data set having a record structure according to an embodiment of the present invention. It is an illustration figure of the dataset of the record structure in which the prediction data was reflected which concerns on one Embodiment of this invention. FIG. 7 is an exemplary diagram of an integrated tree in which prediction data according to an exemplary embodiment of the present invention is reflected. FIG. 7 is an exemplary view of an integrated tree that reflects prediction data according to an embodiment of the present invention, and that includes a total node.

Hereinafter, specific contents for carrying out the present invention will be described with reference to the accompanying drawings.

Further, in describing the present invention, the same parts are designated by the same reference numerals, and the repeated description thereof will be omitted.

First, an example of an entire system for carrying out the present invention will be described with reference to FIG.

As shown in FIG. 1, an overall system for implementing the present invention provides a client 20 provided in a user terminal 10, a database 80 for storing data, analyzes data, and provides an analysis result in a pivot table format. The analysis server 50 and the prediction reflection server 30 configured to reflect the deep learning prediction data in the data of the pivot table structure and reconfigure the data. In addition, the user terminal 10 connects to the analysis server 50 or the prediction reflection server 30 via a network (not shown).

First, the client 20 is a program system for a client provided in the user terminal 10 and has a user interface via a Web browser. That is, the user browses and analyzes data online via a web browser or a screen interface such as a web browser. At this time, the user terminal 10 receives the user's command or the like, executes the command, and displays the processing result on the screen or the Web browser.

On the other hand, the user terminal 10 is a computer terminal having a computing function, such as a personal computer (PC: personal computer), a notebook, a tablet PC, a phablet, a PDA, or a smartphone. The user terminal 10 is connected to the analysis server 50 or the prediction reflection server 30 via the network, and the client 20 on the user terminal 10 can perform online data processing work.

Further, the client 20 requests the analysis server 50 for online processing work such as data request and analysis, acquires the analysis result from the analysis server 50, and displays it on the Web browser. In particular, the client 20 displays the data set acquired from the database 80 in the pivot table format.

Further, the client 20 requests the prediction reflection server 30 to reflect the prediction on the data of the pivot table structure, acquires the reflection result from the prediction reflection server 30, and displays it on the Web browser. At this time, it is displayed in a pivot table structure in which the prediction result is reflected.

On the other hand, the functions of the client 20 include, for example, an analysis request for a data set, a display of an analysis result in a pivot table format, a prediction reflection for the data of the pivot table, a reconstruction request, and a display of the result. It is realized by the script function. That is, the function of the client 20 can be realized in the program system by a script based on the Web standard such as HTML5.0.

Next, the database 80 is an ordinary database (DB) for storing data, and includes a database management system (DBMS: DataBase Management System) for managing data, and stores, deletes, retrieves data, etc. Is performed using a query (or query statement). In particular, the database 80 is a commercialized database and provides a data query service using a general query function for processing a data set.

In particular, the database 80 is a database that stores big data. Also, preferably, the database 80 is a relational database (RDB).

Next, the analysis server 50 acquires a data set from the database 80, generates analysis data (data set) having a pivot table structure, and transmits these analysis result data to the client 20 for display. Preferably, the analysis is performed at the request of the client 20 and the result is transmitted.

That is, the analysis server 50 is a normal OLAP server that performs online analysis processing (OLAP: Online Analytical Processing). Preferably, the results of the analysis are organized in a pivot table structure. The pivot table has a data area formed by rows and columns to display a data set, and a dimension area formed at one side (preferably left side) of a row or at one end (preferably upper end) of a column to display a dimension name. It is a table format composed of,. In the cells of the data area, data values and statistical values (for example, number, total, average, etc.) based on the combination of the row dimension and the column dimension are input and displayed.

Next, the prediction reflection server 30 reflects the prediction data on the analysis result analyzed by the analysis server 50, and transmits the reflection result data to the client 20 for display. Preferably, the analysis work is performed at the request of the client 20, and the result is transmitted.

That is, the analysis server 50 extracts the hierarchical structure of rows and columns from the pivot table structure, reconfigures the data of the pivot table structure as the data of the record structure, and performs deep learning prediction using the reconstructed data. Then, the data of the record structure that reflects the prediction result is reconstructed as the pivot table structure again.

On the other hand, the functions of the client 20 and the prediction reflection server 30 described above are examples, and may be realized in various ways according to the implementation technology of the server and the client. That is, the client 20 and the prediction reflection server 30 may be one prediction reflection system, and the functions thereof may be distributed according to the performance. For example, the client 20 may simply have the Web browser function and the interface function, and all the functions may be constructed by the prediction reflection server 30. That is, the prediction reflection server 30 can have not only the interface function of the pivot table and the reconstruction function of the record structure but also all the deep learning prediction functions. According to another example, the prediction reflection server 30 has only a deep learning prediction function, and the client 20 can have a pivot table interface function and a record structure reconfiguration function. That is, the functions can be distributed in various forms depending on the server-client implementation method.

Next, with reference to FIG. 2, a configuration of the deep learning prediction data reflecting system 30 using the OLAP result of the pivot table structure according to the embodiment of the present invention will be described.

The deep learning prediction data reflecting system 30 according to the present invention can be realized by a server-client system.

The conventional deep learning method is a method in which data is collected/refined/preprocessed and then directly coded using a deep learning library or the like. The predictive data reflection system 30 according to the present invention directly uses the analysis result generated by the analysis server 50 (for example, i-META, i-STREAM, etc.), and sets the structured data (or the data set of the pivot table structure). ) Is converted to a frame format (record format) dataset, time series elements and sequential elements are automatically recognized in the dimension, and the deep learning prediction result is automatically converted to structured data using the dimension arrangement information. Reflect on.

As shown in FIG. 2, a deep learning prediction data reflection system 30 according to an exemplary embodiment of the present invention includes a data set receiving unit 31 that receives a data set having a pivot table structure, and a data dimension from a data set having a pivot table structure ( A dimension extraction unit 32 for extracting a data dimension in a hierarchical structure (or a tree structure), a frame conversion unit 33 for converting a pivot table structure data set into a record structure data set, and a deep structure for a record structure data set. The data prediction unit 34 that performs prediction using learning and the data reconstruction unit 35 that reconstructs the record data set to which the prediction data is added as a pivot table data set.

First, the data set receiving unit 31 receives a data set having a pivot table structure (or a pivot table data set).

As shown in FIG. 3, the pivot table structure includes a data area formed by rows and columns to display a data set, and one side (preferably left side) of a row or one end (preferably upper end) of a column. And a dimension area for displaying the dimension name.

That is, the dimension area is divided into a column dimension area displayed at the upper end of the column of the data area and a row dimension area displayed on the left side of the row of the data area. In the example of FIG. 3, <electronic material division>, <home entertainment division>, <sales quantity>, <won unit price>, etc. are the dimension names of the columns in the column dimension area, and the dates <20150101>, <20150102>. >,. ． ． <Domestic demand>, <Export>, etc. are the dimension names of rows in the row dimension area.

On the other hand, each dimension such as the column dimension and the row dimension has a hierarchical structure. The lowest dimension (or the lowest dimension name) has a one-to-one correspondence with each row or column of each data area. The lowest dimension is distinguished by the dimension of its parent or ancestor. Therefore, the row or column of each data area corresponding to each lowest dimension is divided (defined) by the lowest dimension and its parent or ancestor dimension. In the example of FIG. 3, the data set in the first column of the data area is divided by the dimension of <sales quantity> of <electronic material division>. The data set in the second column is divided according to the dimension of <Unit price of won> in <Electronic Materials Division>. Further, the data set in the first row is divided by the dimension of <domestic demand> of <20180101>.

In addition, in the cells of the data area, data values and statistical values (for example, the number, total, average, etc.) of combinations of row dimensions and column dimensions are input and displayed.

The lowest dimension of either the row dimension or the column dimension of the pivot table is set to <measure>. The measure dimension indicates the type of data value in the data area. In the example of FIG. 3, the lowest dimension of the column dimension is set to measure (or measure dimension). The data set in the first column of the data area indicates <sales volume>, and the data set in the second column indicates <won unit price>.

Next, the dimension extraction unit 32 generates a dimension tree of rows and columns showing a hierarchical structure of dimensions from the data set having the pivot table structure.

That is, the dimension extracting unit 32 generates a dimension tree indicating a hierarchical structure of row and column dimensions from a data set having a pivot table structure.

In the example of FIG. 3, the row dimension has a hierarchical structure of <sales date> and <sales category name>. At this time, <sales date> is called an upper dimension, and <sales category name> is called a lower dimension. The upper and lower parts of the hierarchical structure are relative concepts.

At this time, <sales date> and <sales category name> are referred to as dimension categories, and the value of each dimension is referred to as a dimension name (or dimension value). The line category of <sales date> has dimension names such as <20180101>, <20180102>, and <20180104>, and the line category of <sales category name> has two dimension names such as <domestic demand> and <export>. To have.

In the example of FIG. 3, the column dimension has a hierarchical structure of <business division name> and <measure>. At this time, <business category name> and <major> are dimensional categories, respectively. The <Category Division Name> column category has dimension names such as <Electronic Material Business Division> and <Home Entertainment Business Division>, and the <Major> category has two dimension names such as “sales volume” and “won unit price”. It has a dimension name, in particular the dimension name of a measure is called the measure name.

The dimension extraction unit 32 constructs a row dimension tree and a column dimension tree, respectively, based on the hierarchical structure of the row dimension and the column dimension of the pivot table. That is, a virtual node is arbitrarily set as the root node of each dimension tree. Each dimension name in the pivot table is associated with one node to generate each node, and the upper and lower relationships between the nodes are set based on the relationship between the upper dimension and the lower dimension. That is, each node of the dimension tree has a dimension name, and each node is hierarchically configured as an upper node or a lower node based on the upper or lower relation of the dimension.

On the other hand, in the pivot table structure, each of the dimension names of the upper category is configured to have all the dimension names of the lower category. Therefore, in the dimensional tree, each node of the upper category similarly has all the nodes of the lower category.

FIG. 4 is a graph showing a row dimension tree generated from the row dimension structure of the pivot table of FIG. In the example of FIG. 4, each of the nodes <20180101>, <20180102>, and <20180104> of the upper category <sales date> similarly has all the lower nodes <domestic demand> and <domestic demand> of the lower category <sales category name>. Export>

FIG. 5 is a graph showing a column dimension tree generated from the column dimension structure of the pivot table of FIG. In the example of FIG. 5, each of the nodes <electronic material division>, <home entertainment division>, etc. of the upper category <division name> is similarly all of the lower nodes <sales quantity of the lower category <major>. >, <Unit price of won>.

Further, as shown in FIG. 5, in a dimensional tree having measures, leaf nodes are major nodes, but data value nodes (or data nodes) may be connected to the major nodes. At this time, the data set of the data node corresponds to the data set of the measure in the data area of the pivot table. There are as many measure data sets as there are row dimensions (the lowest dimension of rows).

As shown in FIG. 6, as many dimension trees as the number of data sets of the measure in the data area can be created. The data set (measure data set) in the first row of the data area in FIG. 3 is {25,1000,20,900}, and by using the data set of the measure as the value of the data node, the first data in FIG. Can be generated like the second dimension tree.

Further, preferably, the dimension extracting unit 32 additionally generates a row dimension that is not included in the row dimension tree structure, that is, a dimension that is parallel in the dimensional structure of the pivot table, as a total node. In addition, a dimension tree corresponding to the total of the data sets of the measure in the data area is also generated. In particular, the sum node is added (or creates another tree structure) to the tree structure to have the same level as the tree level of the lowest dimension corresponding to the sum dimension in the dimension structure of the pivot table. That is, the total dimension and the lowest dimension corresponding to the total dimension are configured to have a sibling relationship.

As shown in the example of FIG. 7a, <20180101 Total>, <20180102 Total>,. ． ． The row dimension such as is a structure arranged in a tree structure, and does not correspond to the tree structure of the previously obtained dimension. Add these nodes to the tree structure as sum nodes or generate another sum tree. FIG. 7a shows the dimension tree obtained earlier with the addition of summation nodes, and FIG. 7b shows the summation tree separately constructed.

At this time, the dimension tree is also generated for the data set of the measure in the data area corresponding to the total. That is, a sum tree for the measure data set (eg, {40, 1500, 50, 1300}) corresponding to the sum is also generated.

Next, the frame conversion unit 33 creates an integrated dimension tree by combining the previously created row and column dimension trees, and creates a record-structured data set (hereinafter referred to as a record data set) from the integrated dimension tree.

First, the frame conversion unit 33 sets the two-dimensional trees so that the leaf node of the upper tree becomes the root node of the lower tree by setting the dimension tree having the measure as the lower tree and the remaining trees as the upper tree. Integrate.

FIG. 8 shows an example in which the row dimension tree of FIG. 4 and the column dimension tree of FIG. 5 are integrated to generate an integrated dimension tree. Since the measure dimension exists in the column dimension tree, the column dimension tree becomes the lower tree and the row dimension tree becomes the upper tree.

On the other hand, as many subtrees as the number of major datasets are generated. At this time, the major data set of each data node corresponds to each row dimension. Therefore, each subtree integrates the row-dimensional leaf node corresponding to the data set of its own data node as the root node.

For example, as shown in FIG. 8, the leaf node <domestic demand> or <export> of the upper tree is combined as the root node of the lower tree (column dimension tree). The data set of the first lower tree is {25,1000,20,900}, and the row dimension corresponding to the data set is {<20180101>, <domestic demand>}. Therefore, the first lower tree joins the leaf node <domestic demand> of the upper tree, which is divided into {<20180101>, <domestic demand>}, as the root node.

On the other hand, preferably, the frame conversion unit 33 generates an integrated dimensional tree including the total nodes. That is, a major data set (eg, {40, 1500, 50, 1300} etc.) corresponding to the total is also generated in the integrated tree. FIG. 9 shows an integration tree that includes a totals node.

Further, when the integrated dimension tree (or integrated tree) is created, the frame conversion unit 33 creates a data set having a record structure (data frame format) from the data set of the integrated tree. At this time, the record structure is composed of a set of records composed of many fields. Therefore, a field or field name is configured, and a field value corresponding to each field is set to generate a data frame.

That is, for each measure, the dimension name of the measure is set as a field, and the data value of the measure is set as the field value of the field. Further, the category (category name) of the upper node of the major node is set as a field, and the dimension name of the category is set as the field value of the field. At this time, the root node is excluded. Therefore, one record is constructed using the dimension name of the upper node of the major node and the data value of the major.

At this time, the tree structure including the total node is excluded. That is, only when the tree structure does not include the total node, the record structure data set is generated from the integrated dimension tree.

In the example of FIG. 6, the dimension names <sales quantity> and <won unit price> of the top major category are set as fields (field names), and the data values <25> and <1000> are set as the field values of the fields. .. Further, the categories <sales date>, <sales category name>, and <business division category name> of the upper node of the major category are defined as respective fields (field names). Further, the dimension names <20180101>, <domestic demand>, and <electronic material division> of the upper nodes of the major node are set as the field values of the field.

Therefore, the record is composed of {<20180101>, <domestic demand>, <electronic material division>, <25>, <1000>}. At this time, the field name is composed of {<sales date>, <sales category name>, <business division category name>, <sales quantity>, <won unit price>}.

When records are generated for the major nodes of all the major categories of the integration tree, a record structured data set can be generated. FIG. 10 shows a data structure of a record structure generated from the integrated tree of FIG.

According to another embodiment, the frame conversion unit 33 may go through the upper tree and the lower tree to generate a dataset having a record structure without generating all the lower trees corresponding to all the major datasets. At this time, only one lower tree is generated without connecting the data nodes to the lower tree. That is, the nodes of the lower tree are composed only of nodes for dimensions.

The method of forming the field is similar to that of the above-described embodiment.

Then, it goes around as follows to make each record.

_{row level2 → ... → row leveln →} column level1 → ... → column leveln
That is, the value (dimension name) of each node is input to the record while starting from the root node in the upper tree and sequentially traveling to the leaf nodes. When it reaches the leaf node of the upper tree, it starts from the root node of the lower tree and travels to the lower node. At this time, touching the major node inputs a data value in the major field. At this time, in addition to the node in which the rows and columns of the data area are circulated, data of the same dimension is input as the data value of the measure field.

Next, the data prediction unit 34 performs prediction on the record data set by using deep learning (LSTM: Long Short Term Memory). Preferably, in the data structure of the record structure, the data value of the major field in the next cycle is predicted for the field having the time series data.

Preferably, the time series fields are fields created by dimension names.

In the example of FIG. 10, the <sales date> field is a field created by the row dimension, and since the dates have a time-series order, the next analysis cycle for the <sales date> next cycle <20180105>. Can measure the major data of.

At this time, the prediction data is obtained for each combination of the remaining dimensions excluding the time-series dimension, and the data of the time-series dimension forms the data predicted for each combination in the next cycle as a record. At this time, the field value of the dimension is composed of the next cycle of the time series dimension, the combination of the remaining dimensions, and the predicted data.

Next, the data reconstruction unit 35 reconstructs the record data set to which the prediction data is added as a data set having a pivot table structure. That is, the data set having the record structure of the prediction data is reflected in the integrated tree, and the integrated tree is reconfigured as the data set having the pivot table structure. At this time, the reverse process of the process of forming the data structure of the record structure described above is performed.

The prediction data is reflected in the integrated tree from the data set having the record structure in which the prediction data in FIG. 11 is reflected. FIG. 12 shows an integrated tree in which the prediction data is reflected.

As shown in FIG. 12, a node of a new hierarchy having a dimension next to the cycle of the time series dimension is generated and inserted into the integrated tree. At this time, all nodes included in the category of the time series dimension are included.

That is, a node corresponding to the new record generated in FIG. 11 is generated, and the generated node is added to the original integrated tree to reconfigure.

On the other hand, preferably, as shown in FIG. 13, a part for the total node (a partial tree of the total node, or a total node and its subordinate tree) is also added to the new node (new partial tree). At this time, the lower-level structure (lower-level tree) of the total node is generated by adding all the lower-level tree structures of other sibling nodes (particularly, the sibling nodes generated by the new record).

The summation of the lower trees is performed by the following method. Similarly, the structure of the subordinate tree is generated. Then, only the data values that are leaf nodes differ, but by summing up all the data values of leaf nodes at the same position (leaf nodes of other sibling nodes), the leaf node at that position (the subtree structure of the total node) Data value of leaf node) is generated.

The data reconfiguring unit 35 also generates the structure of the pivot table from the reconfigured integrated tree. The pivot table is reconstructed by performing the reverse process of the process of forming the integrated tree from the structure of the pivot table.

Although the invention made by the present inventor has been specifically described based on the embodiments, the present invention is not limited to the above embodiments, and various modifications can be made without departing from the scope of the invention. Needless to say.

This patent application is a basic research project carried out with the support of the Korea Research Foundation as a financial resource of the Korean government in 2017 (No. 2017M3C4A7083282).

10 User Terminal 20 Client 30 Prediction Reflecting Server 31 Dataset Receiving Section 32 Dimension Extracting Section 33 Frame Converting Section 34 Data Predicting Section 35 Data Reconstructing Section 50 Analysis Server 80 Database

Claims (6)

In the deep learning prediction data reflection system using the online analysis processing (OLAP) result of the pivot table structure,
A data set receiving unit for receiving a data set having a pivot table structure;
A dimension extraction unit for generating a row dimension tree and a column dimension tree having data dimensions in a hierarchical structure from a data set having a pivot table structure;
A frame conversion unit that combines the row-dimensional tree and the column-dimensional tree to generate an integrated tree, and forms a record-structured data set from the integrated tree;
A data predictor for predicting a record structured data set using deep learning;
And a data reconfiguring unit for reconfiguring the integration tree using the record data set to which prediction data is added and reconfiguring a pivot table using the reconfigured integration tree. A deep learning prediction data reflection system using the OLAP result of a table structure. The dimension extraction unit configures a row dimension tree and a column dimension tree based on a hierarchical structure of row dimensions and column dimensions of the pivot table, sets a root node of each dimension tree as a virtual node, and the pivot table Each dimension name of is associated with one node to generate each node, and the upper and lower relations between the nodes are set based on the relation between the upper dimension and the lower dimension. 2. The pivot result of the pivot table structure according to claim 1, wherein each node is hierarchically configured as an upper node or a lower node based on a higher or lower dimension relationship. Deep learning prediction data reflection system using. The pivot table structure OLAP according to claim 2, wherein the dimension extraction unit is configured such that each node of the upper category similarly has all the nodes of the lower category in the dimension tree. Deep learning prediction data reflection system using results. The dimension extraction unit connects a data node having a data value to a measure node with respect to a dimension tree having measures, and generates as many dimension trees as the number of data sets of measures in the data area of the pivot table. The deep learning prediction data reflecting system using the OLAP result of the pivot table structure according to claim 2, characterized in that. The frame conversion unit integrates the two dimensional trees so that the leaf tree of the upper tree becomes the root node of the lower tree by setting the dimensional tree having the measure as the lower tree and the remaining trees as the upper tree. The deep learning prediction data reflecting system using the OLAP result of the pivot table structure according to claim 4, characterized in that. The frame conversion unit sets, for each measure, the dimension name of the measure as a field, sets the data value of the measure as the field value of the field, and sets the category of the upper node of the major node as the field and sets the category. The pivot according to claim 1, wherein one record is configured by using the dimension name of the upper node of the major node and the data value of the measure by using the dimension name of the field as the field value of the field. A deep learning prediction data reflection system using the OLAP result of a table structure.

Images