WO2022190384A1 - 特徴量選択プログラム、装置、及び方法 - Google Patents
特徴量選択プログラム、装置、及び方法 Download PDFInfo
- Publication number
- WO2022190384A1 WO2022190384A1 PCT/JP2021/010196 JP2021010196W WO2022190384A1 WO 2022190384 A1 WO2022190384 A1 WO 2022190384A1 JP 2021010196 W JP2021010196 W JP 2021010196W WO 2022190384 A1 WO2022190384 A1 WO 2022190384A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- feature amount
- feature
- concept
- hypotheses
- selection
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims description 27
- 238000003860 storage Methods 0.000 claims description 17
- 238000010187 selection method Methods 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 5
- 238000010586 diagram Methods 0.000 description 16
- 238000005516 engineering process Methods 0.000 description 10
- 238000010276 construction Methods 0.000 description 9
- 238000013473 artificial intelligence Methods 0.000 description 8
- 238000005304 joining Methods 0.000 description 7
- 239000000284 extract Substances 0.000 description 6
- 238000004140 cleaning Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000004891 communication Methods 0.000 description 2
- 238000005520 cutting process Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000000692 Student's t-test Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 238000012353 t test Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/48—Program initiating; Program switching, e.g. by interrupt
- G06F9/4806—Task transfer initiation or dispatching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
- G06N5/022—Knowledge engineering; Knowledge acquisition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
- G06N5/045—Explanation of inference; Explainable artificial intelligence [XAI]; Interpretable artificial intelligence
Definitions
- the disclosed technology relates to a feature quantity selection program, a feature quantity selection device, and a feature quantity selection method.
- explainable AI Artificial Intelligence
- feature quantities explanatory variables
- a technology has been proposed that selects feature values to be used in the model from among many feature values. .
- the selected feature amount is not necessarily a feature amount that improves the interpretability of the model output. There is a problem.
- the disclosed technology aims to select feature quantities that improve the interpretability of model output.
- the technology disclosed identifies feature amounts of a higher concept, which is a lower concept of the feature amounts included in the feature amount set.
- the disclosed technique is a hypothesis expressed by a combination of at least one or more feature amounts including the feature amount of the subordinate concept, and a plurality of hypotheses each having different feature amounts of the subordinate concept, for the objective variable It is determined whether success or failure satisfies a predetermined condition.
- the technology disclosed herein selects the feature amount of the broader concept as the feature amount to be added to the feature amount set when a predetermined condition is satisfied.
- it has the effect of being able to select feature quantities that improve the interpretability of the model's output.
- FIG. 10 is a diagram for explaining a range for cutting out a feature amount from a knowledge graph;
- FIG. FIG. 4 is a diagram showing a set of triples included in the knowledge graph; It is a figure which shows an example of learning data.
- It is a functional block diagram of a feature selection device. It is an example of learning data to which a feature amount of a higher concept is added. It is a figure which shows an example of upper-lower-order correspondence TB.
- FIG. 10 is a diagram for explaining selection of a feature amount of a higher concept; It is a figure which shows an example of a rule set.
- 2 is a block diagram showing a schematic configuration of a computer functioning as a feature selection device;
- FIG. 6 is a flowchart showing an example of feature amount selection processing
- FIG. 11 is a diagram for explaining another example of conditions for selecting a feature amount of a broader concept
- FIG. 11 is a diagram for explaining another example of conditions for selecting a feature amount of a broader concept
- FIG. 10 is a diagram showing an example of a knowledge graph for explaining another example of construction of learning data
- FIG. 10 is a diagram showing another example of learning data
- graph format data such as a knowledge graph composed of nodes corresponding to feature values and edges connecting the nodes with attributes indicating the relationship between the features (hereafter simply referred to as "
- a feature amount is selected from a graph.
- FIG. 1 shows an example of a graph representing part of the data related to the above question "whether or not a certain professional baseball player wins the title".
- elliptical circles represent nodes
- values (character strings) within nodes represent feature values
- arrows connecting nodes represent edges
- values (character strings) written together with edges represent attributes.
- a graph is a set of triples represented by three elements: an edge and a node on the starting point side and a node on the ending point side connected by the edge.
- FIG. 1 shows an example of triples represented by three elements: an edge and a node on the starting point side and a node on the ending point side connected by the edge.
- the first column is the value of the feature amount corresponding to the node on the starting point side of the edge (first node)
- the second column is the attribute of the edge
- the third column is the node on the ending point side of the edge (second node). node).
- the feature amount of the first node is represented by the attribute of the edge and the value of the feature amount of the second node.
- the graph can extend in many layers in the depth direction and in many columns in the width direction, so it is possible that the graph contains a huge amount of features. Therefore, it is not realistic to select all feature values included in the graph. Therefore, it is necessary to cut the graph in an arbitrary range as the range for selecting the feature amount.
- a simple method for cutting a graph in an arbitrary range as shown in the dashed line in Fig. 1, a node corresponding to a specific feature value and a feature value range corresponding to a node directly connected by an edge can be cut off. That is, a set of triples whose elements are nodes corresponding to specific feature values is specified.
- the value of the specific feature amount is the value of the feature amount of player names such as "professional baseball player A” and "professional baseball player B".
- learning data as shown in FIG. 3 is constructed from the clipped range of the graph.
- "joined team” and “hometown” are explanatory variables
- "title” is an objective variable.
- an explanation such as “If you are from Hiroshima, Okayama, Tottori, Shimane, or Yamaguchi, and you join team X, it is easier to win the title” is obtained as the output of the model. Such explanations are redundant and the interpretability of the model output is not good. Note that the data for the home prefectures of Tottori, Shimane, and Yamaguchi are not shown in FIG.
- Attributes associated with edges included in the graph also include attributes indicating higher-level and lower-level conceptual relationships between features. Utilizing this, as indicated by the dashed-dotted line in FIG. 1, the feature quantity of the higher concept of the previously selected feature quantity is specified.
- the attributes including "part of” included in FIG. 1 are an example of attributes indicating the conceptual relationship between higher and lower levels. For example, the triple node “Hiroshima Prefecture” – edge “part of” – node “Chugoku region” is “Hiroshima prefecture is a part of Chugoku region”, that is, Hiroshima prefecture is a lower concept and Chugoku region is a higher concept.
- AIC is an index represented by the sum of a logarithmic likelihood term indicating the likelihood of a model generated from the selected feature amount and a term indicating the number of selected feature amounts.
- the log-likelihood term of AIC may be smaller when each feature value of the subordinate concept is selected. have a nature.
- the AIC itself may also be smaller than in the case of selecting the feature amount of the higher concept. In such a case, the decision to select the feature amount of the higher concept is not made. However, even in the latter case, we would like to leave the possibility of selecting the feature quantity of the higher concept.
- the feature quantity selection device 10 functionally includes a learning data construction unit 12, a specification unit 14, a selection unit 16, and a generation unit 18. Also, in a predetermined storage area of the feature selection device 10, a knowledge graph 20 and a TB (table) 22 corresponding to upper and lower levels are stored.
- the knowledge graph 20 is, as shown in FIG. 4 is a graph showing data to be inferred by.
- the learning data constructing unit 12 acquires feature amounts included in a specific range cut from the knowledge graph 20 as a feature amount set.
- the learning data constructing unit 12 constructs learning data using the feature amounts included in the feature amount set.
- the learning data constructing unit 12 includes, in the knowledge graph 20, nodes corresponding to specific feature value values and edges directly connected to the nodes, as indicated by broken lines in FIG. Cut a range that includes and contains nodes.
- the value of the specific feature amount is the value of the feature amount "player name" such as "professional baseball player A" and "professional baseball player B".
- the learning data constructing unit 12 collects a set of triples included in the cut range graph (for example, FIG. 2) for each triple containing a specific feature value as an element, and generates learning data as shown in FIG. to build.
- the learning data constructing unit 12 extracts triples containing “professional baseball player A” as an element for professional baseball player A, and extracts attributes associated with edges included in the extracted triples as feature values. be the item name of Also, the learning data constructing unit 12 sets the value of the feature amount corresponding to the other node included in the extracted triple as the value corresponding to the item name of the feature amount. Note that a combination of the item name of the feature amount and the value of the feature amount is an example of the feature amount of the technology disclosed herein.
- the learning data construction unit 12 stores the item and value of the feature amount of the added superordinate concept in the learning data.
- Add FIG. 5 shows an example of the learning data shown in FIG. 3 to which the feature amount of the higher concept is added. In FIG. 5, the portion indicated by the dashed line is the feature amount of the higher concept added.
- the identifying unit 14 identifies the feature amount of the superordinate concept, which is the subordinate concept of the feature amount included in the feature amount set acquired by the learning data constructing unit 12 . Specifically, for each feature amount included in the feature amount set, the specifying unit 14 connects a node corresponding to the value of the feature amount with an edge associated with an attribute indicating a higher-level and lower-level conceptual relationship. Determine whether the node exists. If the corresponding node exists, the identifying unit 14 identifies the feature quantity corresponding to that node as the feature quantity of the higher concept.
- the attribute containing "part of” is an example of an attribute that indicates the conceptual relationship of higher and lower levels. Therefore, from the relationship between the nodes connected by the edges associated with the attribute "part of", the identifying unit 14 determines that the feature value "hometown-Hiroshima prefecture" is the lower-level concept, the feature of the higher-level concept Specify the quantity "Region - Chugoku Region”. Similarly, the specifying unit 14 specifies the feature quantity “region-Chugoku region” of the superordinate concept, which is the subordinate concept of the feature quantity “hometown-Okayama prefecture”. The identifying unit 14 stores the identified feature amount of the higher concept in association with the feature amount of the lower concept in, for example, a higher/lower level correspondence TB 22 as shown in FIG.
- the selection unit 16 selects a plurality of hypotheses expressed by a combination of at least one or more feature amounts including the feature amounts of the lower-level concepts, each having different feature amounts of the lower-level concepts, with a predetermined success or failure with respect to the objective variable. Determine whether the conditions are met. When the success or failure of the hypothesis satisfies a predetermined condition, the selection unit 16 selects the feature amount of the higher concept as the feature amount to be added to the feature amount set.
- the selection unit 16 selects the feature quantity of the superordinate concept. determine whether or not For example, the selection unit 16 extracts the feature quantity of the lower concept associated with the feature quantity of the higher concept for each feature quantity of the higher concept stored in the higher-lower correspondence TB 22 .
- the feature amount of the superordinate concept is x super
- the feature amount of the subordinate concept is x sub
- the feature amount other than the subordinate concept included in the feature amount set is x nonsub .
- the value of the feature quantity x * is v, it is expressed as x * -v.
- the feature amount of the subordinate concept of x super -i is x sub -j 1 , x sub -j 2 , . number).
- the selection unit 16 determines that the hypothesis holds that the condition x super ⁇ i and x nonsub ⁇ a affects the objective variable y, and selects x super .
- x super is “region”, i is “Chugoku region”, x sub is “home prefecture”, j 1 is “Hiroshima prefecture”, j n is “Okayama prefecture”, x nonsub is the "joining team” and a is the "team X”.
- the hypotheses that include the feature values of the subordinate concepts are hypotheses that a professional baseball player whose home prefecture is Hiroshima Prefecture and whose joining team is Team X is likely to win a title, . . .
- it is hypothesized that a professional baseball player who joins team X will easily win a title.
- the selection unit 16 determines that the hypothesis holds that a professional baseball player from the Chugoku region who joins team X is likely to win a title. . Then, the selection unit 16 selects the feature amount “region-Chugoku region” of the superordinate concept as the feature amount to be added to the feature amount set.
- x super is “region”, i is “Tohoku region”, x sub is “home prefecture”, j 1 is “Aomori prefecture”, ..., j n is “Fukushima prefecture”, x " nonsub " is “joining team”, and "a” is "team Y”.
- a professional baseball player whose home prefecture is Aomori Prefecture and whose joining team is Team Y, which is a hypothesis including the feature value of the subordinate concept, is likely to win the title.
- the selection unit 16 determines that the hypothesis that a professional baseball player who is from the Tohoku region and joins team Y is likely to win a title is not valid, and the superordinate concept feature value "region - Tohoku region". are not selected as features to be added to the feature set.
- the selection unit 16 calculates the effect of each hypothesis on the objective variable in order to test each of the above hypotheses. For example, if the target variable is a binary classification problem, the ratio of the number of learning data that are positive examples to the target variable (hereinafter referred to as the "number of positive examples") to the number of learning data, and the number of all positive examples
- the impact may be calculated, such as by a t-test based on the proportion of positive cases for each hypothesis. Influence may also be calculated using explainable AI techniques such as, for example, WideLearning (see references 1 and 2).
- Reference 1 JP-A-2020-46888
- Reference 2 Hiroaki Iwashita, Takuya Takagi, Hirofumi Suzuki, Keisuke Goto, Kotaro Ohori, Hiroki Arimura, "Efficient Constrained Pattern Mining Using Dynamic Item Ordering for Explainable Classification", arXiv:2004.08015 , https://arxiv.org/abs/2004.08015
- the importance is a value that increases as the number of positive examples increases.
- the selection unit 16 determines that the hypothesis that the condition affects the objective variable holds true.
- the generation unit 18 associates a condition expressed by a combination of at least one feature quantity included in the feature quantity set to which the selected superordinate concept feature quantity is added, and an objective variable that holds under the condition. Generate the attached rule. For example, the generation unit 18 can use WideLearning described in the selection unit 16 to generate rules. Specifically, as described above, the generation unit 18 calculates the degree of importance for each condition represented by an exhaustive combination of feature amounts, Generate a set of rules using each of the conditions.
- the generation unit 18 assigns to each rule included in the rule set an index according to the number of positive examples of the learning data corresponding to the conditions included in the rule, and outputs the rule set.
- FIG. 8 shows an example of an output rule set.
- the example of FIG. 8 shows an example in which the number of positive examples is given as an index for each condition under which a certain objective variable is established.
- the index is not limited to the number of positive cases that meet the conditions, but may be the ratio of the number of positive cases that meet the conditions to the total number of positive cases.
- the selection unit 16 uses WideLearning to generate and test hypotheses
- the generation unit 18 uses the hypotheses generated by the selection unit 16 and the calculated importance of each condition. may be used to generate a rule set and an index for each rule.
- the rule set is used in explainable AI, and the correctness of the inference target data with respect to the objective variable is output as an inference result according to the degree of compatibility between the inference target data and the rule set.
- the rule to which the data to be inferred matches is the explanation that indicates the basis of the inference result.
- the feature quantity of the higher concept is added without replacing the feature quantity of the lower concept included in the original set of feature quantities. For this reason, an increase in the amount of information may result in redundant explanations and a decrease in the interpretability of the output of the model. Therefore, as described above, the generation unit assigns an index corresponding to the number of positive examples to each rule, so that rules with higher importance can be checked with priority by sorting in order of index.
- a rule including a feature amount of a higher concept includes a rule including a feature amount of a lower concept for the feature amount of a higher concept, so the number of positive examples is larger than that of a rule including a feature amount of a lower concept. Therefore, by sorting in order of index, it is possible to preferentially confirm rules that include feature amounts of higher concepts.
- the feature quantity selection device 10 can be realized, for example, by a computer 40 shown in FIG.
- the computer 40 includes a CPU (Central Processing Unit) 41 , a memory 42 as a temporary storage area, and a non-volatile storage section 43 .
- the computer 40 also includes an input/output device 44 such as an input unit and a display unit, and an R/W (Read/Write) unit 45 that controls reading and writing of data to and from a storage medium 49 .
- the computer 40 also has a communication I/F (Interface) 46 connected to a network such as the Internet.
- the CPU 41 , memory 42 , storage unit 43 , input/output device 44 , R/W unit 45 and communication I/F 46 are connected to each other via bus 47 .
- the storage unit 43 can be implemented by a HDD (Hard Disk Drive), SSD (Solid State Drive), flash memory, or the like.
- a feature amount selection program 50 for causing the computer 40 to function as the feature amount selection device 10 is stored in the storage unit 43 as a storage medium.
- the feature quantity selection program 50 has a learning data construction process 52 , a specification process 54 , a selection process 56 and a generation process 58 .
- the storage unit 43 also has an information storage area 60 in which information forming each of the knowledge graph 20 and the upper/lower correspondence TB 22 is stored.
- the CPU 41 reads the feature amount selection program 50 from the storage unit 43, develops it in the memory 42, and sequentially executes the processes of the feature amount selection program 50.
- the CPU 41 operates as the learning data construction unit 12 shown in FIG. 4 by executing the learning data construction process 52 . Further, the CPU 41 operates as the specifying unit 14 shown in FIG. 4 by executing the specifying process 54 . Further, the CPU 41 operates as the selection unit 16 shown in FIG. 4 by executing the selection process 56 . Further, the CPU 41 operates as the generation unit 18 shown in FIG. 4 by executing the generation process 58 . Further, the CPU 41 reads information from the information storage area 60 and develops each of the knowledge graph 20 and the upper/lower correspondence TB 22 in the memory 42 . As a result, the computer 40 executing the feature quantity selection program 50 functions as the feature quantity selection device 10 . Note that the CPU 41 that executes the program is hardware.
- the function realized by the feature quantity selection program 50 can also be realized by, for example, a semiconductor integrated circuit, more specifically an ASIC (Application Specific Integrated Circuit) or the like.
- the feature quantity selection device 10 executes the feature quantity selection process shown in FIG. Note that the feature amount selection process is an example of the feature amount selection method of the technology disclosed herein.
- step S12 the learning data constructing unit 12 cuts out from the knowledge graph 20 a range that includes the node corresponding to the value of the specific feature amount and the nodes that are directly connected to the node by an edge. Then, the learning data constructing unit 12 acquires a set of feature amounts included in the cut range, and builds learning data from the acquired set of feature amounts.
- step S14 for each feature amount included in the feature amount set acquired in step S12, the specifying unit 14 determines that the node corresponding to the value of the feature amount has an attribute indicating a higher-order and lower-order conceptual relationship. Determine whether there is a node connected by the associated edge. If the corresponding node exists, the identifying unit 14 identifies the feature quantity corresponding to the node as the feature quantity of the higher concept. Then, the specifying unit 14 associates the feature quantity of the specified superordinate concept with the feature quantity of the subordinate concept, and stores them in the superordinate/lower order correspondence TB 22 .
- step S16 the selection unit 16 extracts the feature amount of the lower concept associated with the feature amount of the higher concept for each feature amount of the higher concept stored in the upper/lower correspondence TB 22 . Then, if the hypothesis that the condition including the feature amount of the lower-level concept affects the objective variable holds for all the conditions including the feature amount of the lower-level concept, the selection unit 16 selects the feature amount of the lower-level concept. Select the feature value of the superordinate concept and add it to the feature value set. In addition, the learning data constructing unit 12 adds the items and values of the feature amount of the added broader concept to the learning data constructed in step S12.
- step S18 the generation unit 18 generates a condition represented by a combination of at least one feature amount included in the feature amount set to which the selected broader concept feature amount is added, and under the condition, A rule is generated in association with objective variables that are established.
- step S20 the generation unit 18 assigns to each rule included in the rule set an index according to the number of positive examples of learning data corresponding to the conditions included in the rule, and outputs the rule set. , the feature quantity selection process ends.
- the feature amount selection device identifies feature amounts of a higher-level concept with the feature amounts included in the feature amount set as lower-level concepts. Then, the feature amount selection device selects a plurality of hypotheses expressed by a combination of at least one feature amount including the feature amount of the subordinate concept, wherein each hypothesis has different feature amounts of the subordinate concept. satisfies a predetermined condition. When a predetermined condition is satisfied, the feature amount selection device selects the feature amount of the higher concept as the feature amount to be added to the feature amount set. Thereby, the feature quantity selection device can select a feature quantity that improves the interpretability of the output of the model.
- the feature values of the higher concept corresponding to the feature values of the lower concept are selected.
- a predetermined ratio for example, 0.8
- the corresponding feature of the higher concept is selected. You may do so.
- FIG. 11 since four hypotheses out of five hypotheses including the feature amount of the lower concept are established, it is determined that the hypothesis in which the feature amount of the lower concept is replaced with the feature amount of the higher concept also holds.
- the feature amount of the higher concept thereof may be selected. This takes into consideration the bias in the number of learning data that correspond to each hypothesis. For example, it is determined that the hypothesis holds when the percentage of positive cases under each condition is equal to or greater than a predetermined value (for example, 0.8). As shown in FIG.
- the number of cases in parentheses written together with each hypothesis is "the number of positive examples of the condition/the number of learning data corresponding to the condition".
- the present invention is not limited to this. Presence or absence of a specific attribute and the number of specific attributes may be used as the feature quantity. Also, data cleaning processing or the like may be performed on learning data constructed from these feature amounts.
- FIG. 13 shows a portion of the knowledge graph relating to professional baseball player C.
- the learning data construction unit sets a value (for example, 1) indicating TRUE as a feature indicating the presence or absence of the specific attribute. Extract as quantity.
- the learning data construction unit sets a value indicating FALSE (for example, 0) to indicate the presence or absence of the specific attribute.
- the learning data constructing unit extracts the number of triples having a specific attribute as an element, which is included in the set of triples forming the knowledge graph, as a feature amount indicating the number of specific attributes.
- the upper diagram of FIG. 14 shows an example of learning data constructed from the knowledge graph shown in FIG. In the example of FIG. 14, the specific attribute is inside the " " of the item name of the feature amount.
- the lower diagram of FIG. 14 shows learning data after data cleaning processing, deletion of feature amounts not used for hypotheses, and addition of superordinate concept feature amounts.
- the presence or absence of "prefecture of origin”, the number of "prefecture of origin”, the presence or absence of "height”, the number of "height”, and the presence or absence of "biography” are deleted by cleaning the data, and "height"
- the value of indicates an example deleted as a feature quantity that is not used in the hypothesis.
- the lower diagram of FIG. 14 shows an example in which "local area” is added as a feature amount of the superordinate concept of "home prefecture".
- the present invention is not limited to this.
- the program according to the technology disclosed herein can also be provided in a form stored in a storage medium such as a CD-ROM, DVD-ROM, USB memory, or the like.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Medical Informatics (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
Description
参考文献1:特開2020-46888号公報
参考文献2:Hiroaki Iwashita, Takuya Takagi, Hirofumi Suzuki, Keisuke Goto, Kotaro Ohori, Hiroki Arimura, "Efficient Constrained Pattern Mining Using Dynamic Item Ordering for Explainable Classification", arXiv:2004.08015, https://arxiv.org/abs/2004.08015
12 学習データ構築部
14 特定部
16 選択部
18 生成部
20 ナレッジグラフ
22 上位下位対応TB
40 コンピュータ
41 CPU
42 メモリ
43 記憶部
49 記憶媒体
50 特徴量選択プログラム
Claims (20)
- 特徴量集合に含まれる特徴量を下位概念とする上位概念の特徴量を特定し、
前記下位概念の特徴量を含む、少なくとも1以上の特徴量の組み合わせで表現される仮説であって、前記下位概念の特徴量が各々異なる複数の仮説の、目的変数に対する成否が所定の条件を満たす場合に、前記上位概念の特徴量を前記特徴量集合に追加する特徴量として選択する
ことを含む処理をコンピュータに実行させるための特徴量選択プログラム。 - 前記所定の条件は、前記複数の仮説のうち、所定割合以上の仮説が成立する場合である請求項1に記載の特徴量選択プログラム。
- 前記所定の条件は、前記複数の仮説のうち、所定割合以上の仮説が成立すると共に、前記下位概念の特徴量を前記上位概念の特徴量に置き換えた仮説が成立する場合である請求項1に記載の特徴量選択プログラム。
- 前記上位概念の特徴量を特定する処理は、特徴量の値に対応するノードと、上位下位の関係を含むノード間の関係を示す属性が対応付けられたエッジとを含むグラフにおいて、前記特徴量集合に含まれる特徴量の値に対応するノードに、前記上位下位の関係を示す属性が対応付けられたエッジで接続されるノードに対応する特徴量を特定することを含む請求項1~請求項3のいずれか1項に記載の特徴量選択プログラム。
- 前記特徴量集合は、前記グラフにおいて、特定の特徴量の値に対応するノードと前記エッジで直接接続されるノードに対応する特徴量である請求項4に記載の特徴量選択プログラム。
- 選択された前記上位概念の特徴量が追加された前記特徴量集合に含まれる少なくとも1以上の特徴量の組み合わせで表現される条件と、前記条件の下で成立する目的変数とを対応付けたルールの集合を生成することをさらに含む処理を前記コンピュータに実行させるための請求項1~請求項5のいずれか1項に記載の特徴量選択プログラム。
- 前記ルールの集合に含まれるルールの各々に、前記ルールに含まれる条件に該当するデータであって、前記目的変数に対する正例となるデータの数に応じた指標を付与して出力することをさらに含む処理を前記コンピュータに実行させるための請求項6に記載の特徴量選択プログラム。
- 特徴量集合に含まれる特徴量を下位概念とする上位概念の特徴量を特定する特定部と、
前記下位概念の特徴量を含む、少なくとも1以上の特徴量の組み合わせで表現される仮説であって、前記下位概念の特徴量が各々異なる複数の仮説の、目的変数に対する成否が所定の条件を満たす場合に、前記上位概念の特徴量を前記特徴量集合に追加する特徴量として選択する選択部と、
を含む特徴量選択装置。 - 前記所定の条件は、前記複数の仮説のうち、所定割合以上の仮説が成立する場合である請求項8に記載の特徴量選択装置。
- 前記所定の条件は、前記複数の仮説のうち、所定割合以上の仮説が成立すると共に、前記下位概念の特徴量を前記上位概念の特徴量に置き換えた仮説が成立する場合である請求項8に記載の特徴量選択装置。
- 前記特定部は、特徴量の値に対応するノードと、上位下位の関係を含むノード間の関係を示す属性が対応付けられたエッジとを含むグラフにおいて、前記特徴量集合に含まれる特徴量の値に対応するノードに、前記上位下位の関係を示す属性が対応付けられたエッジで接続されるノードに対応する特徴量を特定することを含む請求項8~請求項10のいずれか1項に記載の特徴量選択装置。
- 前記特徴量集合は、前記グラフにおいて、特定の特徴量の値に対応するノードと前記エッジで直接接続されるノードに対応する特徴量である請求項11に記載の特徴量選択装置。
- 選択された前記上位概念の特徴量が追加された前記特徴量集合に含まれる少なくとも1以上の特徴量の組み合わせで表現される条件と、前記条件の下で成立する目的変数とを対応付けたルールの集合を生成する生成部をさらに含む請求項8~請求項12のいずれか1項に記載の特徴量選択装置。
- 前記生成部は、前記ルールの集合に含まれるルールの各々に、前記ルールに含まれる条件に該当するデータであって、前記目的変数に対する正例となるデータの数に応じた指標を付与して出力する請求項13に記載の特徴量選択装置。
- 特徴量集合に含まれる特徴量を下位概念とする上位概念の特徴量を特定し、
前記下位概念の特徴量を含む、少なくとも1以上の特徴量の組み合わせで表現される仮説であって、前記下位概念の特徴量が各々異なる複数の仮説の、目的変数に対する成否が所定の条件を満たす場合に、前記上位概念の特徴量を前記特徴量集合に追加する特徴量として選択する
ことを含む処理をコンピュータが実行するための特徴量選択方法。 - 前記所定の条件は、前記複数の仮説のうち、所定割合以上の仮説が成立する場合である請求項15に記載の特徴量選択方法。
- 前記所定の条件は、前記複数の仮説のうち、所定割合以上の仮説が成立すると共に、前記下位概念の特徴量を前記上位概念の特徴量に置き換えた仮説が成立する場合である請求項15に記載の特徴量選択方法。
- 前記上位概念の特徴量を特定する処理は、特徴量の値に対応するノードと、上位下位の関係を含むノード間の関係を示す属性が対応付けられたエッジとを含むグラフにおいて、前記特徴量集合に含まれる特徴量の値に対応するノードに、前記上位下位の関係を示す属性が対応付けられたエッジで接続されるノードに対応する特徴量を特定することを含む請求項15~請求項17のいずれか1項に記載の特徴量選択方法。
- 前記特徴量集合は、前記グラフにおいて、特定の特徴量の値に対応するノードと前記エッジで直接接続されるノードに対応する特徴量である請求項18に記載の特徴量選択方法。
- 特徴量集合に含まれる特徴量を下位概念とする上位概念の特徴量を特定し、
前記下位概念の特徴量を含む、少なくとも1以上の特徴量の組み合わせで表現される仮説であって、前記下位概念の特徴量が各々異なる複数の仮説の、目的変数に対する成否が所定の条件を満たす場合に、前記上位概念の特徴量を前記特徴量集合に追加する特徴量として選択する
ことを含む処理をコンピュータに実行させるための特徴量選択プログラムを記憶した記憶媒体。
Priority Applications (5)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP21930234.6A EP4307184A4 (en) | 2021-03-12 | 2021-03-12 | PROGRAM, DEVICE AND METHOD FOR FEATURE SELECTION |
PCT/JP2021/010196 WO2022190384A1 (ja) | 2021-03-12 | 2021-03-12 | 特徴量選択プログラム、装置、及び方法 |
JP2023505066A JPWO2022190384A1 (ja) | 2021-03-12 | 2021-03-12 | |
CN202180095326.5A CN117321611A (zh) | 2021-03-12 | 2021-03-12 | 特征量选择程序、装置以及方法 |
US18/461,265 US20230409377A1 (en) | 2021-03-12 | 2023-09-05 | Feature selection program, feature selection device, and feature selection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/010196 WO2022190384A1 (ja) | 2021-03-12 | 2021-03-12 | 特徴量選択プログラム、装置、及び方法 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US18/461,265 Continuation US20230409377A1 (en) | 2021-03-12 | 2023-09-05 | Feature selection program, feature selection device, and feature selection method |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022190384A1 true WO2022190384A1 (ja) | 2022-09-15 |
Family
ID=83227672
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2021/010196 WO2022190384A1 (ja) | 2021-03-12 | 2021-03-12 | 特徴量選択プログラム、装置、及び方法 |
Country Status (5)
Country | Link |
---|---|
US (1) | US20230409377A1 (ja) |
EP (1) | EP4307184A4 (ja) |
JP (1) | JPWO2022190384A1 (ja) |
CN (1) | CN117321611A (ja) |
WO (1) | WO2022190384A1 (ja) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170105867A1 (en) * | 2015-10-16 | 2017-04-20 | Hung-Nan Hsieh | Remotely Operated Snore-Stopping Pillow |
JP2017146869A (ja) * | 2016-02-19 | 2017-08-24 | 富士ゼロックス株式会社 | 情報検索プログラム及び情報検索装置 |
JP2017174022A (ja) * | 2016-03-22 | 2017-09-28 | 株式会社日立製作所 | 目的変数に対応する説明変数群を決定するシステム及び方法 |
US20190138806A1 (en) * | 2017-11-03 | 2019-05-09 | Tata Consultancy Services Limited | Signal analysis systems and methods for features extraction and interpretation thereof |
WO2020053934A1 (ja) * | 2018-09-10 | 2020-03-19 | 三菱電機株式会社 | モデルパラメタ推定装置、状態推定システムおよびモデルパラメタ推定方法 |
JP2020046888A (ja) | 2018-09-18 | 2020-03-26 | 富士通株式会社 | 学習プログラム、予測プログラム、学習方法、予測方法、学習装置および予測装置 |
-
2021
- 2021-03-12 EP EP21930234.6A patent/EP4307184A4/en not_active Withdrawn
- 2021-03-12 JP JP2023505066A patent/JPWO2022190384A1/ja active Pending
- 2021-03-12 CN CN202180095326.5A patent/CN117321611A/zh active Pending
- 2021-03-12 WO PCT/JP2021/010196 patent/WO2022190384A1/ja active Application Filing
-
2023
- 2023-09-05 US US18/461,265 patent/US20230409377A1/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170105867A1 (en) * | 2015-10-16 | 2017-04-20 | Hung-Nan Hsieh | Remotely Operated Snore-Stopping Pillow |
JP2017146869A (ja) * | 2016-02-19 | 2017-08-24 | 富士ゼロックス株式会社 | 情報検索プログラム及び情報検索装置 |
JP2017174022A (ja) * | 2016-03-22 | 2017-09-28 | 株式会社日立製作所 | 目的変数に対応する説明変数群を決定するシステム及び方法 |
US20190138806A1 (en) * | 2017-11-03 | 2019-05-09 | Tata Consultancy Services Limited | Signal analysis systems and methods for features extraction and interpretation thereof |
WO2020053934A1 (ja) * | 2018-09-10 | 2020-03-19 | 三菱電機株式会社 | モデルパラメタ推定装置、状態推定システムおよびモデルパラメタ推定方法 |
JP2020046888A (ja) | 2018-09-18 | 2020-03-26 | 富士通株式会社 | 学習プログラム、予測プログラム、学習方法、予測方法、学習装置および予測装置 |
Non-Patent Citations (5)
Title |
---|
H. AKAIKE: "Information theory and an extension of the maximum likelihood principle", 2ND INTERNATIONAL SYMPOSIUM ON INFORMATION THEORY, 1973, pages 267 - 281 |
HIROAKI IWASHITATAKUYA TAKAGIHIROFUMI SUZUKIKEISUKE GOTOKOTARO OHORIHIROKI ARIMURA: "Efficient Constrained Pattern Mining Using Dynamic Item Ordering for Explainable Classification", ARXIV:2004.08015, Retrieved from the Internet <URL:https://arxiv.org/abs/2004.08015> |
R. MIYASHIROY. TAKANO: "Mixed Integer Second-Order Cone Programming Formulations for Variable Selection in Linear Regression", EUROPEAN JOURNAL OF OPERATIONAL RESEARCH, vol. 247, 2015, pages 721 - 731 |
See also references of EP4307184A4 |
SUZUMURA, SHINYA ET AL.: "Modified Bonferroni method for Selective Inference", IEICE TECHNICAL REPORT, vol. 116 (PRMU2016-59), no. 209 (IBISML2016-14), 13 October 2016 (2016-10-13), pages 39 - 45, XP009549313, ISSN: 0913-5685 * |
Also Published As
Publication number | Publication date |
---|---|
CN117321611A (zh) | 2023-12-29 |
JPWO2022190384A1 (ja) | 2022-09-15 |
EP4307184A4 (en) | 2024-05-01 |
EP4307184A1 (en) | 2024-01-17 |
US20230409377A1 (en) | 2023-12-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP4863778B2 (ja) | 情報処理装置、および情報処理方法、並びにコンピュータ・プログラム | |
JP5159713B2 (ja) | システムの構成要素の設計アーキテクチャを自動設計する自動設計装置、自動設計方法及び自動設計プログラム | |
US8943084B2 (en) | Method, program, and system for converting part of graph data to data structure as an image of homomorphism | |
JP2022502744A (ja) | コード分類のためのハイブリッド機械学習モデル | |
JP4479908B2 (ja) | データソート処理プログラム、データソート処理方法およびデータソート処理装置 | |
JPWO2017090114A1 (ja) | データ処理システム及びデータ処理方法 | |
JP2010128583A (ja) | テスト容易化設計装置ならびにその方法およびプログラム | |
WO2022190384A1 (ja) | 特徴量選択プログラム、装置、及び方法 | |
JP7199203B2 (ja) | ゲームプログラムを検査するためのシステム、方法、プログラム、機械学習支援装置、及びデータ構造 | |
Capra et al. | Towards evolving petri nets: a symmetric nets-based framework | |
CN116663491B (zh) | 基于bdd求解功能覆盖组条件约束语句的方法、设备和介质 | |
Sayed Ahmed | On notions of representability for cylindric‐polyadic algebras, and a solution to the finitizability problem for quantifier logics with equality | |
Čermák et al. | Combining incremental strategy generation and branch and bound search for computing maxmin strategies in imperfect recall games | |
Bartlett et al. | New progressive variable ordering for binary decision diagram analysis of fault trees | |
JP6511793B2 (ja) | テストケース生成プログラム、テストケース生成方法及びテストケース生成装置 | |
JP2022188894A (ja) | 相関ルール生成プログラム、装置、及び方法 | |
Chevrette | G-LINK: a probabilistic record linkage system | |
KR102007126B1 (ko) | 결손된 운영 데이터의 복원 방법 및 복원 장치 | |
JP4073303B2 (ja) | プログラム分割方法とそれを実施するプログラム | |
Richerby et al. | Graph searching in a crime wave | |
EP2856396A2 (en) | Buildable part pairs in an unconfigured product structure | |
CN106294127A (zh) | 基于聚类的面向对象软件测试用例生成方法 | |
Locodi et al. | Introducing a graph topology for robust cooperation | |
Fatemi et al. | Rating and generating Sudoku puzzles based on constraint satisfaction problems | |
CN110782006A (zh) | 一种基于交叉生成复杂神经网络的方法及系统 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21930234 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2023505066 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 202180095326.5 Country of ref document: CN |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2021930234 Country of ref document: EP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 2021930234 Country of ref document: EP Effective date: 20231012 |