Machine learning utilization support apparatus, method thereof, and program therefor

Abstract

To make it possible even for a non-expert in machine learning to select a proper machine learning task based on data of a processing target.SOLUTION: A machine learning utilization support apparatus according to an embodiment of the present invention has a reception unit, a conversion unit, a database, a retrieval unit, and a proposal unit. The reception unit receives a data set of a target of processing of machine learning tasks. The conversion unit convers each of samples of the data set to first data as a common data format. The database stores a candidate for a machine learning task and second data having the common data format and relating to explanation of the candidate in association with each other. The retrieval unit, based on the first data, retrieves the candidate for the machine learning task associated with the second data corresponding to the first data in the database. The proposal unit, based on the retrieval result of the candidate, generates proposal data including the candidate for the machine learning task and explanation of the candidate.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] An embodiment of the present invention relates to a machine learning utilization support device, a method, and a program. [Background technology]

[0002] In recent years, advances in machine learning (ML) technology have led to the use of machine learning to improve business efficiency in a variety of fields, including anomaly detection, visual inspection, and document understanding. In the field of machine learning, the mechanism for determining what output to return for input data is called a machine learning model, or simply a model, and the process of updating the model to match the input data is called learning. Machine learning is broadly categorized into supervised learning, unsupervised learning, and reinforcement learning depending on the characteristics of the problem, and when training a model, it is necessary to use an appropriate learning method and algorithm depending on the data in hand and the objectives. However, developing and adjusting models requires highly specialized knowledge and skilled engineers, and the process is time-consuming and costly.

[0003] In response to this, automated machine learning (AutoML) is gaining attention. AutoML is a concept that automates part or all of the development of AI models. By using AutoML, even people without specialized knowledge of machine learning (non-experts in machine learning) are increasingly able to create models with simple operations. There are three basic steps that AutoML users must take before training a model. The first is to select the machine learning task, i.e., what kind of analysis they want to perform before performing machine learning. The second is to process (preprocess) the data before training. The third is to execute model training.

[0004] As such, AutoML allows users to select a machine learning task appropriate for the analysis they want to perform, upload their data, and, if necessary, perform preprocessing to train the model. However, if users cannot select a task appropriate for the data to be processed and their purpose, they will not be able to obtain the desired AI model. To avoid this situation, it is important to have a method to support the selection of machine learning tasks. For example, a known method supports machine learning task selection by providing recipes that describe machine learning preprocessing and algorithms and allowing users to search for recipes using an interactive interface.

[0005] However, according to the inventor's research, when this method only provides recipes, non-experts in machine learning have difficulty selecting an appropriate machine learning task that matches the data to be processed and the purpose, making it difficult to arrive at the appropriate task. In particular, when there are too many recipes, it may be difficult to select one that suits the data from the vast number of recipes. Furthermore, even if the analytical purpose is determined, the machine learning task is not necessarily uniquely determined. For example, when using a model for image anomaly detection, there are supervised learning and unsupervised learning approaches, but non-experts in machine learning do not understand the difference between the two. Therefore, the above method does not allow for appropriate machine learning task selection, which hinders non-experts from utilizing machine learning.

[0006] Therefore, it is desirable to enable even non-experts in machine learning to select an appropriate machine learning task based on the data to be processed. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Patent No. 6419859 Summary of the Invention [Problem to be solved by the invention]

[0008] The problem to be solved by the present invention is to provide a machine learning utilization support device, method, and program that enables even a non-expert in machine learning to select an appropriate machine learning task based on the data to be processed. [Means for solving the problem]

[0009] A machine learning utilization support device according to an embodiment includes a receiving unit, a conversion unit, a database, a search unit, and a suggestion unit. The receiving unit receives a dataset to be processed as a machine learning task. The conversion unit converts each sample in the dataset into first data in a common data format. The database previously stores machine learning task candidates in association with second data having the common data format and relating to descriptions of the candidates. The search unit searches the database for machine learning task candidates associated with the second data corresponding to the first data, based on the first data. The suggestion unit generates suggested data including the machine learning task candidates and descriptions of the candidates, based on the search results for the candidates. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a block diagram showing an example of the configuration of a machine learning utilization support device according to a first embodiment. [Figure 2] 4 is a flowchart for explaining the operation in the first embodiment. [Figure 3] FIG. 10 is a block diagram showing an example of the configuration of a machine learning application support device according to a second embodiment. [Figure 4] FIG. 11 is a schematic diagram showing an example of the configuration of a machine learning task database according to the second embodiment. [Figure 5] FIG. 10 is a sequence diagram for explaining the operation in the second embodiment. [Figure 6] FIG. 6 is a schematic diagram for explaining the operation in the second embodiment. [Figure 7] FIG. 10 is a schematic diagram showing a display example of a data set and an explanation according to the second embodiment. [Figure 8]FIG. 13 is a schematic diagram showing an example of the configuration of a machine learning task database according to the third embodiment. [Figure 9] FIG. 11 is a sequence diagram for explaining the operation in the third embodiment. [Figure 10] FIG. 10 is a schematic diagram for explaining the operation in the third embodiment. [Figure 11] FIG. 10 is a schematic diagram for explaining the operation in the fourth embodiment. [Figure 12] FIG. 13 is a schematic diagram for explaining the operation in the fifth embodiment. [Figure 13] FIG. 20 is a schematic diagram showing an example of the configuration of a machine learning task database according to the sixth embodiment. [Figure 14] FIG. 20 is a sequence diagram for explaining the operation in the sixth embodiment. [Figure 15] FIG. 13 is a schematic diagram for explaining the operation in the sixth embodiment. [Figure 16] FIG. 23 is a schematic diagram showing a display example of a data set and an explanation according to the sixth embodiment. [Figure 17] FIG. 20 is a schematic diagram showing an example of the configuration of a machine learning task database according to the seventh embodiment. [Figure 18] FIG. 13 is a sequence diagram for explaining the operation in the seventh embodiment. [Figure 19] FIG. 23 is a schematic diagram showing an example of a search prompt according to the seventh embodiment. [Figure 20] FIG. 23 is a schematic diagram showing an example of a search prompt according to the seventh embodiment. [Figure 21] FIG. 23 is a schematic diagram showing an example of a response to a search prompt in the seventh embodiment. [Figure 22] FIG. 23 is a schematic diagram showing an example of the configuration of a machine learning task database according to the eighth embodiment. [Figure 23] FIG. 20 is a sequence diagram for explaining the operation in the eighth embodiment. [Figure 24] FIG. 23 is a schematic diagram showing an example of a summary prompt according to the eighth embodiment. [Figure 25] FIG. 23 is a schematic diagram showing an example of a response to a summary prompt in the eighth embodiment. [Figure 26] FIG. 13 is a sequence diagram for explaining the operation in the ninth embodiment. [Figure 27] FIG. 23 is a schematic diagram showing an example of a question prompt according to the ninth embodiment. [Figure 28] FIG. 23 is a schematic diagram showing an example of a question prompt according to the ninth embodiment. [Figure 29] FIG. 23 is a schematic diagram showing an example of a response to a question prompt in the ninth embodiment. [Figure 30] FIG. 20 is a block diagram showing an example of the configuration of a machine learning utilization system according to a tenth embodiment. [Figure 31] FIG. 22 is a sequence diagram for explaining the operation in the tenth embodiment. [Figure 32] FIG. 23 is a schematic diagram showing an example of displaying evaluation results in the tenth embodiment. [Figure 33] FIG. 23 is a block diagram showing an example of the configuration of a machine learning utilization system according to a modified example of the tenth embodiment. [Figure 34] FIG. 22 is a block diagram showing an example of the hardware configuration of a machine learning utilization system according to the eleventh embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, each embodiment will be described by way of example with reference to the drawings.

[0012] First Embodiment 1 is a block diagram showing an example of the configuration of a machine learning utilization support device according to embodiment 1. The machine learning utilization support device 10 includes a data receiving unit 11, a data conversion unit 12, a machine learning task database 13, a search unit 14, and a machine learning task proposal unit 15.

[0013] Here, the data accepting unit 11 accepts a user dataset 200 uploaded by a user 1. Here, the user dataset 200 is a dataset to be processed in a machine learning task. The user dataset 200 may be in any format, such as image data, tabular (structured) data, time-series data, or text data, or may be in a format combining these. Furthermore, in conventional technology, when uploading a user dataset 200, the directory structure of the user dataset 200 had to be appropriately structured depending on the algorithm and learning method (such as supervised learning or unsupervised learning) used for learning. In this embodiment, the directory structure of the user dataset 200 does not need to be a specific structure. Note that the data accepting unit 11 is an example of a accepting unit.

[0014] The data conversion unit 12 converts each sample of the user data set 200 into first data in a common data format. Examples of common data formats include text, vector, image, and tabular data, and these may be combined to ultimately form a common data format. The data conversion unit 12 is an example of a conversion unit.

[0015] The machine learning task database 13 stores information about machine learning tasks. For example, the machine learning task database 13 may be a relational database (RDB) that stores information in a table format consisting of rows and columns, or a vector database that stores and manages data in a vector format. The search results may be in JSON format or may be received in list format. For example, the machine learning task database 13 stores machine learning task candidates in advance in association with second data having a common data format and relating to explanations of the candidates. It is preferable that the explanations of the candidates include, for example, reasons why the candidates are inferred to be valid, from the perspective of facilitating the user's selection of candidates. The machine learning task database 13 is an example of a database.

[0016] The search unit 14 selects highly relevant machine learning tasks from the machine learning task database 13, which stores information on machine learning tasks, based on the samples (first data) converted into a common data format. For example, the search unit 14 searches the machine learning task database 13 for machine learning task candidates associated with second data corresponding to the first data, based on the first data, and obtains search results 300. The search results 300 include at least the machine learning task candidates out of the machine learning task candidates and second data related to explanations of the candidates.

[0017] The machine learning task proposal unit 15 may present the search result 300 on the screen as is, or, if the search result 300 includes multiple machine learning task candidates, may sort and present the multiple candidates. In either case, the machine learning task proposal unit 15 generates proposal data including machine learning task candidates and descriptions of the candidates based on the candidate search result 300. Here, the proposal data is screen data, and may be displayed on a display unit (not shown) or transmitted to a user terminal (not shown). The machine learning task proposal unit 15 is an example of a proposal unit.

[0018] Next, the operation of the information processing device configured as above will be described with reference to the flowchart of FIG.

[0019] Now, in step ST10, the data accepting unit 11 accepts the user data set 200 uploaded by the user 1.

[0020] After step ST10, in step ST20, the data conversion section 12 converts each sample of the user data set 200 into first data in a common data format.

[0021] After step ST20, in step ST30, the search unit 14 searches the machine learning task database 13 for machine learning task candidates associated with second data corresponding to the first data, based on the first data in a common data format.

[0022] After step ST30, in step ST40, the machine learning task proposal unit 15 generates proposal data including machine learning task candidates and explanations of the candidates based on the candidate search results 300. Thereafter, the machine learning task proposal unit 15 proposes the machine learning task candidates together with their explanations, for example, by displaying the proposal data on a display unit (not shown).

[0023] This allows User 1 to select an appropriate machine learning task based on the proposed content.

[0024] As described above, according to the first embodiment, the data accepting unit 11 accepts a dataset to be processed by a machine learning task. The data converting unit 12 converts each sample in the dataset into first data in a common data format. The machine learning task database 13 previously stores machine learning task candidates associated with second data having a common data format and relating to descriptions of the candidates. The searching unit 14 searches the machine learning task database 13 for machine learning task candidates associated with the second data corresponding to the first data, based on the first data. The machine learning task proposing unit 15 generates proposed data including machine learning task candidates and descriptions of the candidates, based on the search results. In this way, by converting the original data into a common data format to make it searchable and proposing machine learning task candidates and their descriptions based on the search results, even non-experts in machine learning can select an appropriate machine learning task based on the data to be processed.

[0025] <Second embodiment> Next, a machine learning utilization support device according to a second embodiment will be described.

[0026] The second embodiment is a specific example of the first embodiment, and is an example of a case where the common data format is text format and the proposed data is displayed on the display unit.

[0027] Figure 3 is a block diagram showing an example of the configuration of a machine learning utilization support device according to a second embodiment. Parts that are substantially the same as those in Figure 1 are assigned the same reference numerals and detailed descriptions thereof will be omitted. Here, differences will be mainly described. Similarly, descriptions of overlapping parts will be omitted in each of the following embodiments. Compared to the configuration shown in Figure 1, this machine learning utilization support device 10 further includes a display unit 16 that displays a screen including a task selection support display unit 16a.

[0028] Here, the display unit 16 displays the machine learning task candidates included in the proposal data and explanations of the candidates.

[0029] The task selection support display unit 16a is a display unit for machine learning task candidates included in the proposal data and explanations of the candidates. The task selection support display unit 16a may display only highly relevant machine learning tasks, sort the tasks by relevance, or highlight highly relevant machine learning tasks. Furthermore, the task selection support display unit 16a may include a GUI (graphical user interface) that accepts user operations in addition to a display area for explanations of the candidates. Here, user operations can include any operation related to task selection, such as selecting a machine learning task or asking a question.

[0030] Furthermore, since the common data format is text, the data conversion unit 12 has the functions of generating captions (captioning) and counting words in addition to the functions described above. Captioning can be performed using a self-developed caption generation AI or a publicly available caption generation AI model. For example, a large-scale language model (LLM) distributed as open source or a language model available via an API in a cloud service can be used. Fine-tuning can be performed to fine-tune the model to suit the task. The captioning language can be Japanese or another language. To perform captioning, the language input by the user can be converted to a common language different from the input, and the output can be retranslated to match the user's language. Word counting can be performed by determining keywords in advance and counting the number of occurrences of words (keywords) included in captions according to the keywords. Alternatively, when performing word counting, technical terms can be extracted using natural language processing and the number of occurrences of words (keywords) included in captions can be counted.

[0031] In either case, the data conversion unit 12 generates captions that explain each sample of the user data set 200 in text format, tallies the number of times that words contained in the captions appear, and sets the words that appear relatively frequently as the first data.

[0032] As shown in Figure 4, the machine learning task database 13 stores machine learning task candidates in advance in association with keywords (second data) that have a common data format and are related to the explanation of the candidates. Here, the second data is the keywords of the candidates.

[0033] The search unit 14 searches for machine learning task candidates from the machine learning task database 13 by keyword search based on the words that are the first data. However, this is not limited to this, and the search unit 14 may also count the number of matches between the words that are the first data and keywords, and use keywords with a relatively large number of matches as second data to search for machine learning task candidates associated with the second data.

[0034] The other configurations are the same as those in the first embodiment.

[0035] Next, the operation of the machine learning utilization support device configured as above will be described using the sequence diagram in Fig. 5 and the schematic diagrams in Fig. 6 and 7. The following description will discuss the above-mentioned steps ST10 to ST40 using more detailed steps.

[0036] (Step ST10) Now, the data accepting unit 11 accepts the user data set 200 uploaded by the user 1 (step ST11). The data accepting unit 11 stores the accepted user data set 200 in the data store 11a (step ST12). Thereafter, the data accepting unit 11 sends a data conversion start command including a data set path to the data conversion unit 12 (step ST13). This completes step ST10.

[0037] (Step ST20) Upon receiving the data conversion start command, the data conversion unit 12 acquires the user data set 200 from the data store 11a based on the data set path extracted from the data conversion start command (steps ST21 to ST22).

[0038] After step ST22, the data conversion unit 12 generates a caption that explains each sample of the acquired user data set 200 in a text format (step ST23). For example, as shown in Fig. 6, the data conversion unit 12 generates a caption "Two people are looking at a person who has fallen" from image 201 of each sample, and generates a caption "A person is walking" from image 202 of each sample.

[0039] After step ST23, the data conversion unit 12 counts the number of times that words included in the caption appear, and sets the words that appear relatively frequently as first data (step ST24). For example, the data conversion unit 12 counts the number of times that words "fall", "person", and "walk" appear in the caption, and sets the words "fall", "person", and "walk" that appear relatively frequently as first data. This first data is used as keyword information for search.

[0040] After step ST24, the data conversion unit 12 sends a search start command including the keyword information "fall", "person", and "walk" to the search unit 14 (step ST25), thereby completing step ST20.

[0041] (Step ST30) When the search unit 14 receives a search start command, it performs a keyword search by sending a machine learning information query including the keyword information "fall," "person," and "walking" extracted from the search start command to the machine learning task database 13 (step ST31).

[0042] After step ST31, the search unit 14 receives machine learning information including search results 300 of machine learning task candidates and keyword information from the machine learning task database 13 (step ST32). Note that the machine learning information received in step ST32 does not necessarily have to include keyword information. In this case, the search unit 14 can acquire machine learning information including search results 300 of candidates and keyword information by including keyword information extracted from the search start command in the machine learning information.

[0043] In either case, after step ST32, the search unit 14 sends a proposal start command including the candidate search result 300 and machine learning information including keyword information to the machine learning task proposal unit 15 (step ST33), thereby completing step ST30.

[0044] (Step ST40) Upon receiving the proposal start command, the machine learning task proposal unit 15 executes a proposal process based on the machine learning information extracted from the proposal start command (step ST41). The proposal process is a process of generating proposal data including machine learning task candidates and descriptions of the candidates based on, for example, search results 300 of machine learning task candidates included in the machine learning information and keyword information.

[0045] After step ST41, the machine learning task proposing unit 15 proposes candidate machine learning tasks together with explanations by displaying a task selection support display unit 16a based on the proposal data and a user dataset 200 on the display unit 16, as shown in Fig. 7. This allows the user 1 to select an appropriate machine learning task based on the proposal content.

[0046] As described above, according to the second embodiment, the display unit 16 displays the machine learning task candidates included in the proposal data and explanations of the candidates. Therefore, in addition to the effects described above, the display by the display unit 16 can prompt the user 1 to select a machine learning task candidate.

[0047] Furthermore, according to the second embodiment, the common data format is text format, and the second data is candidate keywords. The data conversion unit 12 generates captions that explain each sample in the user dataset 200 in text format, tallies the number of times each word in the caption appears, and defines the words that appear relatively frequently as the first data. The search unit 14 searches the machine learning task database 13 for candidate machine learning tasks by keyword search based on the words that are the first data. Therefore, in addition to the effects described above, it is possible to suggest machine learning tasks to the user based on keywords registered in advance in the machine learning task database 13.

[0048] <Third embodiment> Next, a machine learning utilization support device according to a third embodiment will be described.

[0049] The third embodiment is a modification of the first or second embodiment, and is an example of a case where the common data format is a vector format.

[0050] Accordingly, in addition to the above-mentioned functions, the data conversion unit 12 converts each sample of the user data set 200 into a first fixed-length vector, which is first data. More specifically, the user data set 200 is a set of image data. The data conversion unit 12 generates captions that explain each sample of the user data set 200 in text format, and converts the captions into first fixed-length vectors.

[0051] The machine learning task database 13 stores machine learning task candidates in advance in association with vectors (second data) that have a common data format and relate to the descriptions of the candidates, as shown in Figure 8. Here, the second data is a second fixed-length vector into which the descriptions of the candidates have been converted.

[0052] The search unit 14 searches for candidates for machine learning tasks by performing a vector similarity search between the first fixed-length vector and the second fixed-length vector.

[0053] The other configurations are the same as those of the second embodiment.

[0054] Next, the operation of the machine learning utilization support device configured as above will be described using the sequence diagram in Fig. 9 and the schematic diagram in Fig. 10. In the following explanation, among the above-mentioned steps ST10 to ST40, steps ST20 to ST30 that are different from those in the second embodiment will be described in more detail.

[0055] Now, step ST10 is executed in the same manner as described above.

[0056] (Step ST20) Steps ST21 to ST23 are executed in the same manner as described above, and captions that explain each sample of the user data set 200 in text form are generated.

[0057] After step ST23, the data conversion unit 12 vectorizes the caption (step ST24A) and converts the caption into a first fixed-length vector (first data), as shown in Fig. 10. This first data is used as vector information for search.

[0058] The vectorization method may use a technique for converting the importance of words in a sentence into vectors or a deep learning-based technique. Each sample of the user data set 200 in the data conversion unit 12 may use all of the images in the user data set 200, or may use only a portion of the data.

[0059] After step ST24A, the data conversion unit 12 sends a search start command including vector information [1.4, 1.5, 2.3] to the search unit 14 (step ST25A), thereby completing step ST20.

[0060] (Step ST30) Upon receiving a search start command, the search unit 14 performs a vector search by sending a machine learning information query including the vector information [1.4, 1.5, 2.3] extracted from the search start command to the machine learning task database 13 (step ST31A). The vector search may be performed using cosine similarity, which measures similarity by calculating the cosine of the angle between vectors, Euclidean distance (L2 distance) between vectors, Manhattan distance (L1 distance) between vectors, or Mahalanobis distance, which is a distance measure that takes into account covariance and the correlation coefficient of features. The machine learning task database 13 may generate captions for a dataset prepared in advance in an RDB and register the converted vectors, or it may use a vector database.

[0061] After step ST31A, the search unit 14 receives machine learning information including search results 300 of machine learning task candidates and explanations of the candidates from the machine learning task database 13 (step ST32A).

[0062] After step ST32A, the search unit 14 sends a proposal start command including the search result 300 of the candidate and machine learning information including an explanation of the candidate to the machine learning task proposal unit 15 (step ST33), thereby completing step ST30.

[0063] Thereafter, step ST40 is executed in the same manner as described above.

[0064] As described above, according to the third embodiment, the common data format is a vector format. The data conversion unit 12 converts each sample of the user data set 200 into a first fixed-length vector, which is the first data. The second data is a second fixed-length vector into which the candidate explanation is converted. The search unit 14 searches for candidates for the machine learning task by performing a vector similarity search between the first fixed-length vector and the second fixed-length vector. Therefore, even when the common data format is a vector format, the above-described effects can be obtained.

[0065] Furthermore, according to the third embodiment, the user dataset 200 is a set of image data. The data conversion unit 12 generates captions that explain each sample in the user dataset 200 in text format and converts the captions into first fixed-length vectors. Therefore, in addition to the effects described above, vector search in the image dataset enables high-speed search and makes it possible to propose machine information task candidates from more information than in the second embodiment. In other words, it becomes possible to propose machine learning tasks to the user that are closer to the thinking of machine learning experts.

[0066] <Fourth embodiment> Next, a machine learning utilization support device according to a fourth embodiment will be described.

[0067] The fourth embodiment is a modification of the third embodiment, and is an example of a case where the user data set 200 is a set of table data.

[0068] Accordingly, in addition to the above-mentioned functions, the data conversion unit 12 generates captions that explain the characteristics of the user data set 200 in text format based on the file name, header string, and statistics of each column of each sample in the user data set 200, and converts the captions into first fixed-length vectors.

[0069] The other configurations are the same as those of the third embodiment.

[0070] According to the above configuration, in step ST23 of step ST20 described above, the data conversion unit 12 generates captions in a text format that explain the table data 210, which is each sample of the user data set 200, as shown in Fig. 11. For example, the data conversion unit 12 generates captions in a text format that explain the characteristics of the user data set 200 based on the file name of the table data 210, the header character strings "temperature", "pressure", and "determination result", and the statistics of each column, "2 normal, 1 abnormal".

[0071] The caption generation performed by the data conversion unit 12 may be performed using a rule-based algorithm, or the characteristics of the data set may be explained using LLM. Statistical quantities for each column may include the mean, median, maximum / minimum value, variance, standard deviation, etc. The header string may use the column name as is, or may include a string similar to the header string. If the table data has classification items, the number of data items belonging to the classification items may be counted.

[0072] After step ST23, the data conversion unit 12 vectorizes the caption (step ST24A) and converts the caption into a first fixed-length vector (first data). This first data is used as vector information for search.

[0073] Thereafter, the processes from step ST25A onwards are executed in the same manner as described above.

[0074] As described above, according to the fourth embodiment, the user dataset 200 is a set of table data. The data conversion unit 12 generates captions that explain the characteristics of the user dataset 200 in text format based on the file name, header string, and column statistics of each sample in the user dataset 200, and converts the captions into first fixed-length vectors. Therefore, in addition to the effects described above, vector search in the table dataset enables high-speed search and makes it possible to propose machine learning task candidates from more information than in the second embodiment. In other words, it becomes possible to propose machine learning tasks to the user that are closer to the thinking of machine learning experts.

[0075] <Fifth embodiment> Next, a machine learning utilization support device according to a fifth embodiment will be described.

[0076] The fifth embodiment is a modification of the third embodiment, and is an example of a case where the user data set 200 is a set of time-series data.

[0077] Accordingly, in addition to the above-mentioned functions, the data conversion unit 12 generates captions that explain the characteristics of the user data set 200 in text format based on the file name, series name, and statistics for each series of each sample in the user data set 200, and converts the captions into first fixed-length vectors.

[0078] The other configurations are the same as those of the third embodiment.

[0079] According to the above configuration, in step ST23 of step ST20 described above, the data conversion unit 12 generates captions in a text format that explain the time-series data 220, which is each sample of the user data set 200, as shown in Fig. 12. For example, the data conversion unit 12 generates captions in a text format that explain the characteristics of the user data set 200 based on the file name of the time-series data 220, the series names "temperature" and "humidity," and the statistics for each series, "maximum temperature 44.5°C, minimum temperature 20.4°C."

[0080] The caption generation performed by the data conversion unit 12 may be performed using a rule-based algorithm, or the characteristics of the data set may be explained using LLM. The statistics for each column may be the mean, median, maximum / minimum value, variance, standard deviation, or the amount of change over a certain period of time.

[0081] After step ST23, the data conversion unit 12 vectorizes the caption (step ST24A) and converts the caption into a first fixed-length vector (first data). This first data is used as vector information for search.

[0082] Thereafter, the processes from step ST25A onwards are executed in the same manner as described above.

[0083] As described above, according to the fourth embodiment, the user dataset 200 is a set of time-series data. The data conversion unit 12 generates captions that explain the characteristics of the user dataset 200 in text format based on the file name, series name, and statistics for each series of each sample in the user dataset 200, and converts the captions into first fixed-length vectors. Therefore, in addition to the effects described above, vector search in the time-series dataset enables high-speed search and makes it possible to propose machine learning task candidates from more information than in the second embodiment. In other words, it becomes possible to propose machine learning tasks to the user that are closer to the thinking of machine learning experts.

[0084] <Modifications of the Fourth and Fifth Embodiments> In the fourth embodiment, the user data set 200 is described as a set of table data, but is not limited to this. Similarly, in the fifth embodiment, the user data set 200 is described as a set of time-series data, but is not limited to this. That is, the user data set 200 may be a set of multimodal data including table data and time-series data. In this case, the data conversion unit 12 generates captions that explain the characteristics of the user data set 200 in text format based on the file name, data format, and header character string included in the data set, and converts the captions into first fixed-length vectors. According to this modification, the effects of the fourth embodiment and the fifth embodiment can be simultaneously obtained for one user data set 200.

[0085] Sixth Embodiment Next, a machine learning utilization support device according to a sixth embodiment will be described.

[0086] The sixth embodiment is a modification of the first or second embodiment, and is an example of a case where the common data format is a vector format.

[0087] Accordingly, in addition to the above-described functions, the data conversion unit 12 converts each sample of the user data set 200 into a first fixed-length vector, which is first data, and estimates a first domain of the user data set 200 based on the first fixed-length vector. In domain estimation, a model trained in advance to estimate a domain from a vector may be used, or a publicly available model may be used. The model for estimating the domain may be an LLM, a model learned through supervised deep learning, or a model that learns vector distances through deep metric learning. When displaying candidates to the user, the selected domain information may or may not be displayed in text.

[0088] As shown in Figure 13, the machine learning task database 13 stores machine learning task candidates in advance, in association with vectors (second data) that have a common data format and relate to the candidate's explanation, and the candidate's second domain. Here, the second data is a second fixed-length vector into which the candidate's explanation (not shown) is converted. The candidate's second domain also functions as the candidate's explanation.

[0089] The search unit 14 searches for machine learning task candidates based on the first domain and the first fixed-length vector.

[0090] The other configurations are the same as those of the second embodiment.

[0091] Next, the operation of the machine learning utilization support device configured as above will be described using the sequence diagram in Fig. 14 and the schematic diagrams in Fig. 15 and 16. In the following explanation, among the above-mentioned steps ST10 to ST40, steps ST20 to ST30 that are different from those in the second embodiment will be described using more detailed steps.

[0092] Now, step ST10 is executed in the same manner as described above.

[0093] (Step ST20) Steps ST21 to ST23 are executed in the same manner as described above, and captions that explain each sample of the user data set 200 in text form are generated.

[0094] After step ST23, the data conversion unit 12 vectorizes the caption (step ST24A) and converts the caption into a first fixed-length vector (first data), as shown in Fig. 15. This first data is used for domain estimation and search as vector information [1.5, 1.4, 2.2].

[0095] After step ST24A, the data conversion unit 12 estimates the first domain "person-related" of the user data set 200 using the trained model based on the vector information [1.5, 1.4, 2.2] (step ST24B). This first domain is used for search as domain information "person-related".

[0096] After step ST24B, the data conversion unit 12 sends a search start command including the vector information [1.4, 1.5, 2.3] and the domain information [human system] to the search unit 14 (step ST25B), thereby completing step ST20.

[0097] (Step ST30) When the search unit 14 receives a search start command, it sends a machine learning information query including the vector information [1.4, 1.5, 2.3] extracted from the search start command and the domain information [person-related] to the machine learning task database 13, thereby performing a vector search filtered by person-related (step ST31B).

[0098] After step ST31B, the search unit 14 receives machine learning information including search results 300 of machine learning task candidates from the machine learning task database 13 (step ST32A).

[0099] After step ST32A, the search unit 14 sends a proposal start command including the candidate search result 300 and machine learning information including domain information "person-related" that explains the candidate to the machine learning task proposal unit 15 (step ST33). This ends step ST30.

[0100] Thereafter, step ST40 is executed in the same manner as described above. For example, as shown in Fig. 16, the machine learning task proposing unit 15 causes the display unit 16 to display candidate search results 300, a task selection support display unit 16a based on proposal data including domain information "people" that explains the candidates, and a user dataset 200. User 1 selects an appropriate machine learning task based on the displayed proposal content.

[0101] As described above, according to the sixth embodiment, the common data format is vector data. The data conversion unit 12 converts each sample in the user dataset 200 into a first fixed-length vector, which is first data, and estimates a first domain of the user dataset 200 based on the first fixed-length vector. The machine learning task database 13 associates and stores machine learning task candidates, the second domains of the candidates, and second data. The second data is a second fixed-length vector into which a description of the candidate is converted. The search unit 14 searches for machine learning task candidates based on the first domain and the first fixed-length vector. Therefore, in addition to the effects described above, it is possible to narrow the user's selection range by narrowing the domain. In other words, a user with a narrower selection range can select a machine learning task more easily than in the first to fifth embodiments.

[0102] Seventh Embodiment Next, a machine learning utilization support device according to a seventh embodiment will be described.

[0103] The seventh embodiment is a modification of the second embodiment, and is an example of a case where a search is performed using a search prompt.

[0104] Here, the common data format is text format.

[0105] The data conversion unit 12 generates captions that explain each sample of the user data set 200 in text format as first data.

[0106] The machine learning task database 13 stores machine learning task candidates in advance in association with explanatory text (second data) of the candidates that has a common data format, as shown in Fig. 17. The search unit 14 generates a search prompt including a caption, and searches for machine learning task candidates based on the search prompt using a large-scale language model (LLM).

[0107] The other configurations are the same as those of the second embodiment.

[0108] Next, the operation of the machine learning utilization support device configured as above will be described using the sequence diagram in Fig. 18 and the schematic diagrams in Fig. 19 to Fig. 21. In the following explanation, among the above-mentioned steps ST10 to ST40, steps ST20 to ST30 that are different from those in the second embodiment will be described using more detailed steps.

[0109] Now, step ST10 is executed in the same manner as described above.

[0110] (Step ST20) Steps ST21 to ST23 are executed in the same manner as described above, and captions that explain each sample of the user data set 200 in text form are generated.

[0111] After step ST23, the data conversion unit 12 sends a search start command including the caption to the search unit 14 (step ST25C), thereby completing step ST20.

[0112] (Step ST30) Upon receiving the search start command, the search unit 14 sends a machine learning information query to the machine learning task database 13 to acquire pairs of machine information tasks and explanatory text (step ST31C). Thereafter, the search unit 14 acquires machine learning information including pairs of machine learning tasks and explanatory text from the machine learning task database 13 (step ST32C).

[0113] After step ST32C, the search unit 14 generates a search prompt 310 including the caption extracted from the search start command, as shown in FIGS. 19 and 20 (step ST33C).

[0114] The search prompt 310 may be generated by combining a caption and a description of the machine learning task. The description of the machine learning task may be one that is pre-registered in the machine learning task database 13 and included in the search prompt 310, or the machine learning task may be described by an LLM, which may then be used to search for the machine learning task. Examples of LLMs include GPT (OpenAI API) provided by OpenAI, Inc., and PaLM (Google AI PaLM 2 - Google AI) provided by Google, Inc. These LLMs may be used, or may be fine-tuned before use, or an LLM model trained from scratch may be used without using these LLMs.

[0115] After step ST33C, the search unit 14 searches for machine learning task candidates from the machine learning information acquired in step ST32C using a large-scale language model (LLM) based on the search prompt 310 (step ST34C), thereby acquiring a response 320 as shown in FIG.

[0116] After step ST34C, the search unit 14 sends a proposal start command to the machine learning task proposing unit 15, which includes the search result 300 of the candidate and a response 320 including an explanation of the candidate as the search result (step ST35C), thereby completing step ST30.

[0117] Thereafter, step ST40 is executed in the same manner as described above.

[0118] As described above, according to the seventh embodiment, the common data format is text format. The data conversion unit 12 generates, as first data, captions that explain each sample in the user dataset 200 in text format. The search unit 14 generates a search prompt 310 including the caption, and searches for machine learning task candidates using a large-scale language model based on the search prompt 310. Therefore, in addition to the effects described above, it is possible to improve generalizability when an unknown dataset is uploaded. In other words, it is possible to propose to the user machine learning tasks that are closer to the thinking of machine learning experts than the vector search of the third to sixth embodiments.

[0119] Eighth Embodiment A machine learning utilization support device according to the eighth embodiment will be described.

[0120] The eighth embodiment is a modification of the seventh embodiment, and is an example of a case where a search is performed using a summary prompt.

[0121] Here, the common data format is text format.

[0122] The data conversion unit 12 generates captions that explain each sample of the user data set 200 in text format as first data.

[0123] As shown in FIG. 22, the machine learning task database 13 stores machine learning task candidates in advance in association with second data having a common data format and relating to explanations of the candidates. The second data includes explanatory text relating to the candidate explanations and fixed-length vectors converted from the explanatory text. Note that the machine learning task database 13 may store explanatory text for machine learning tasks that has been prepared in advance and vectorized using text-to-vec, which vectorizes text. The explanatory text may be prepared in PDF, Word, or HTML format, and large pieces of text may be divided into small chunks (paragraphs or sentences) using text-to-vec.

[0124] The search unit 14 generates a summary prompt including a caption, summarizes the user dataset 200 using a large-scale language model based on the summary prompt, and searches for machine learning task candidates from the machine learning task database 13 using a vector similarity search using the large-scale language model based on the summary results.

[0125] The other configurations are the same as those of the seventh embodiment.

[0126] Next, the operation of the machine learning utilization support device configured as above will be described using the sequence diagram in Fig. 23 and the schematic diagrams in Fig. 24 and 25. In the following explanation, among the above-mentioned steps ST10 to ST40, steps ST20 to ST30 that are different from the seventh embodiment will be described using more detailed steps.

[0127] Now, step ST10 is executed in the same manner as described above.

[0128] (Step ST20) Steps ST21 to ST23 are executed in the same manner as described above, and captions that explain each sample of the user data set 200 in text form are generated.

[0129] After step ST23, the data conversion unit 12 sends a search start command including the caption to the search unit 14 (step ST25C), thereby completing step ST20.

[0130] (Step ST30) The search unit 14 generates a summary prompt 330 including the caption extracted from the search start command (step ST31D), as shown in Fig. 24. The summary prompt 330 may explain the characteristics of the entire data set, or may explain the characteristics using only a portion of the data set. The summary prompt 330 may or may not include a statement limiting the number of characters to be output.

[0131] After step ST31D, the search unit 14 acquires a summary of the user data set 200 using a large-scale language model (LLM) based on the summary prompt 330 (step ST32D). As a result, the search unit 14 acquires a response 340, which is a result of summarizing the user data set 200, as shown in FIG.

[0132] After step ST32D, the search unit 14 searches for machine learning task candidates from the machine learning task database 13 using a large-scale language model based on the response 340 by vector similarity search, and obtains search results 300 of the candidates and search results including descriptions of the candidates (steps ST33D to ST34D).

[0133] After step ST34D, the search unit 14 sends a proposal start command including the search results 300 of the candidates and the explanations of the candidates to the machine learning task proposal unit 15 (step ST35D), thereby completing step ST30.

[0134] Thereafter, step ST40 is executed in the same manner as described above.

[0135] As described above, according to the eighth embodiment, the common data format is text format. The data conversion unit 12 generates, as first data, captions that describe each sample in the user data set 200 in text format. The search unit 14 generates a summary prompt 330 including the caption, summarizes the user data set 200 using a large-scale language model based on the summary prompt 330, and searches the machine learning task database 13 for machine learning task candidates using a vector similarity search based on the summary result. Therefore, in addition to the effects described above, performing a vector search makes it possible to obtain machine learning task information more quickly than in the seventh embodiment. In other words, it is possible to shorten the time it takes for the user 1 to accept a proposed machine learning task candidate.

[0136] Furthermore, according to the eighth embodiment, the machine learning task database 13 stores machine learning task candidates in advance in association with second data having a common data format and relating to explanations of the candidates. The second data includes explanatory text relating to the candidate and fixed-length vectors into which the explanatory text has been converted. In other words, by including fixed-length vectors into which explanatory text has been converted in advance, vector search can be performed without vectorizing the explanatory text during search.

[0137] <Ninth embodiment> Next, a machine learning utilization support device according to a ninth embodiment will be described.

[0138] The ninth embodiment is a modification of the seventh embodiment, and is an example of a case where, when the user cannot select a searched candidate, the candidates are narrowed down using a question prompt.

[0139] Here, in addition to the above-mentioned functions, if the user is unable to select a searched candidate, the machine learning task suggestion unit 15 generates a question prompt including an instruction to generate a question to narrow down the searched candidates, generates a question using a large-scale language model based on the question prompt, and upon receiving an answer from user 1 to the question, generates new suggestion data based on the answer including candidates narrowed down from the searched candidates and an explanation of the narrowed down candidate.

[0140] Furthermore, the machine learning task proposing unit 15 may sequentially narrow down the machine learning task candidates by repeatedly generating questions to narrow down the candidates and generating proposed data based on the answers to the questions.

[0141] The question prompt may be used when User 1 reselects the machine learning task, or if reselection is not possible, a different question may be prepared. The question prompt may also include the contents of the search results. The search results may include a description of the task stored in the machine learning task database 13, or the description of the machine learning task may not be included in the prompt and the LLM may be left to consider it. The number of tokens sent to the LLM may or may not be limited.

[0142] The question prompt may be generated based on the search results. The question prompt may also include an instruction to generate a question that selects an affirmative or negative answer. That is, the question prompt may or may not specify an output format. If an output format is specified, it may be in JSON format, list format, or the like.

[0143] The question prompt may include an instruction to generate a question necessary for selecting a candidate. In other words, the question prompt may or may not generate a question necessary for selecting a machine learning task. However, from the perspective of reducing the burden on the user, it is preferable that the question prompt generate a question necessary for selecting a machine learning task (avoiding meaningless questions). In this case, for example, if the search results list "human anomaly detection," "crack detection," and "visual inspection image classification" as machine learning task candidates, a question necessary for selecting a machine learning task (such as "Do you need to detect whether there are any anomalies?") is generated.

[0144] The question prompt may also include an instruction for generating a reason for generating the question, such as "By asking whether it is necessary to detect whether there are any abnormalities in the image data, the possibility of detecting human anomalies and cracks can be narrowed down."

[0145] The other configurations are the same as those of the seventh embodiment.

[0146] Next, the operation of the machine learning utilization support device configured as above will be described using the sequence diagram in Fig. 26 and the schematic diagrams in Fig. 27 to Fig. 29. The following description will use more detailed steps for step ST50, which is different from the seventh embodiment, after the above-mentioned steps ST10 to ST40.

[0147] Steps ST10 to ST40 are now executed in the same manner as described above. As a result, machine learning task proposal unit 15 proposes machine learning task candidates together with explanations by displaying task selection support display unit 16a based on the proposal data and user dataset 200 on display unit 16. Here, if user 1 is unable to select a machine learning task, step ST50 is initiated.

[0148] (Step ST50) When the question button is pressed by the user 1, the GUI of the task selection support display unit 16a sends an instruction to generate a question prompt to the machine learning task proposing unit 15 (step ST51).

[0149] When the machine learning task proposing unit 15 receives the instruction to generate a question prompt, it generates a question prompt 350 including an instruction to generate a question for narrowing down the searched candidates, as shown in FIGS. 27 and 28 (step ST52).

[0150] After step ST52, the machine learning task proposing unit 15 generates a question 360 using a large-scale language model based on the question prompt 350 (step ST53), as shown in Fig. 29. The machine learning task proposing unit 15 also causes the display unit 16 to display the question 360 (step ST54).

[0151] After step ST54, when the machine learning task proposal unit 15 receives an answer from user 1 to question 360, it narrows down the machine learning task candidates based on the answer (step ST56). The machine learning task proposal unit 15 also generates new proposal data including candidates narrowed down from the searched candidates and explanations of the narrowed down candidates. The machine learning task proposal unit 15 then displays a task selection support display unit 16a based on the proposal data and a user dataset 200 on the display unit 16, thereby proposing machine learning task candidates together with explanations. This allows user 1 to select an appropriate machine learning task based on the proposal content.

[0152] Furthermore, if the user is again unable to select a candidate, the machine learning task proposal unit 15 may sequentially narrow down the machine learning task candidates by repeatedly generating questions to narrow down the candidates and generating proposal data based on the answers to the questions.

[0153] As described above, according to the ninth embodiment, when a user is unable to select a searched candidate, the machine learning task proposal unit 15 generates a question prompt 350 including an instruction to generate a question for narrowing down the searched candidates. Based on the question prompt 350, the machine learning task proposal unit 15 generates a question 360 using a large-scale language model. Upon receiving an answer from the user 1 to the question 360, the machine learning task proposal unit 15 generates new proposal data including candidates narrowed down from the searched candidates and an explanation of the narrowed down candidate based on the answer. Therefore, in addition to the aforementioned candidates, it is possible to reduce the possibility that the user 1 will be unable to select a machine learning task. Further, advantages of having the LLM prepare questions include improved generalizability compared to rule-based questions, and the ease of adding tasks with the LLM, whereas with rule-based questions, it is necessary to think up a question when adding a machine learning task.

[0154] Furthermore, according to the ninth embodiment, the machine learning task proposing unit 15 may sequentially narrow down the machine learning task candidates by repeatedly generating questions to narrow down the candidates and generating proposal data based on the answers to the questions. In this case, it is possible to further reduce the possibility that the user 1 will be unable to select a machine learning task.

[0155] Additionally, according to a ninth embodiment, the question prompt 350 may include instructions for generating a question 360 with a choice of affirmative or negative answers. For example, the question 360 may be formatted as a "yes" or "no" answer, allowing even non-technical users to select a machine learning task.

[0156] According to the ninth embodiment, the question prompt 350 may include an instruction to generate a question 360 necessary for selecting a candidate. In this case, the generated questions are limited to those necessary for selecting a machine learning task, and the user will not answer meaningless questions.

[0157] According to the ninth embodiment, the question prompt may include an instruction to generate the reason for generating the question. In this case, it is possible to not only ask a question but also to convey to the user why the question was asked, which has the effect of making it easier for the user to understand the intent of the question.

[0158] <Tenth embodiment> Next, a machine learning utilization system according to a tenth embodiment will be described.

[0159] The tenth embodiment is a modification of the first to ninth embodiments, and is an example in which the machine learning utilization support device 10 further includes a machine learning model evaluation device capable of communication.

[0160] 30 is a block diagram showing an example of the configuration of a machine learning utilization system according to the tenth embodiment. The machine learning utilization system 30 includes the machine learning utilization support device 10 and a machine learning model evaluation device 20.

[0161] Here, the machine learning model evaluation device 20 includes a search result receiving unit 21, an evaluation data receiving unit 22, a machine learning model 23 stored in memory, a machine learning model evaluation unit 24, and a display unit 25.

[0162] The search result receiving unit 21 receives the search result 300 from the search unit 14 .

[0163] The evaluation data receiving unit 22 receives a user data set 200 for evaluating machine learning.

[0164] The machine learning model 23 is a trained model that has been trained in advance. The trained model may be a model prepared for each machine learning task, or may be a basic model that can handle multiple tasks.

[0165] If there is a trained model for the machine learning task corresponding to the received search result, the machine learning model evaluation unit 24 evaluates the trained model based on the user dataset 200 to be evaluated and the received search result 300.

[0166] The display unit 25 displays a condition input section 25a, which is a GUI, and accepts input of conditions for evaluating machine learning. The display unit 25 displays the evaluation results of the trained model on a result display section 25b.

[0167] Next, the operation of the machine learning utilization system configured as above will be explained using the sequence diagram in Fig. 31 and the schematic diagram in Fig. 32. The following explanation will use more detailed steps from step ST30 onwards, which are different from the seventh embodiment, among the above-mentioned steps ST10 to ST40.

[0168] Now, steps ST10 to ST20 are executed in the same manner as described above.

[0169] (Step ST30) Furthermore, steps ST31C to ST34C are executed in the same manner as described above, and the search unit 14 obtains search results 300 of machine learning task candidates.

[0170] After step ST34C, the search unit 14 sends a model evaluation start command including the candidate search result 300 to the machine learning model evaluation unit 24 via the search result receiving unit 21 (step ST35E), thereby completing step ST30.

[0171] (Step ST60) If there is a trained model for the machine learning task corresponding to the received search result, the machine learning model evaluation unit 24 evaluates the trained model based on the user dataset 200 to be evaluated and the received search result 300 (step ST61).

[0172] After step ST61, the machine learning model evaluation unit 24 performs a display process to generate display data including the evaluation results (step ST62).

[0173] After step ST62, the machine learning model evaluation unit 24 displays the evaluation result on the result display unit 25b of the display unit 25 based on the display data, as shown in Fig. 32 (step ST63). Here, the result display unit 25b displays sample output images 201a and 201b when the image 201 in the user dataset 200 is input to a trained model related to the machine learning task. This makes it possible to show the evaluation result to the user.

[0174] As described above, according to the tenth embodiment, the machine learning model is a trained model that has been trained in advance. If there is a trained model for the machine learning task corresponding to the received search results, the machine learning model evaluation unit 24 evaluates the trained model based on the user dataset 200 to be evaluated and the received search results 300. The display unit 25 displays the evaluation results of the trained model. Therefore, the actual evaluation results can be automatically obtained and can be used as an indicator when the user selects a task. In other words, since the user can predict the learning results to a certain extent, they can identify tasks that are not expected to have good accuracy and easily remove those tasks from their options.

[0175] (Variation) Although the tenth embodiment describes a case where a trained model exists for the machine learning task corresponding to the received search results, this is not limiting. For example, a configuration may be adopted in which learning is possible even when there is no machine learning model for the machine learning task corresponding to the received search results. As shown in FIG. 33 , this modification further includes a data division unit 26 and a machine learning model training unit 27 compared to the configuration shown in FIG. 31 . Here, when there is no machine learning model for the machine learning task corresponding to the received search results, the data division unit 26 divides the user dataset 200 to be evaluated into a first dataset and a second dataset. The machine learning model training unit 27 trains a machine learning model for the machine learning task based on the first dataset. The machine learning model evaluation unit 24 evaluates the machine learning model obtained by training based on the second dataset. The display unit 25 displays the evaluation results of the machine learning model. Therefore, according to this modification, whereas conventional AutoML systems require the selection of a machine learning task to perform learning, it is possible to search for a trainable task and automatically perform learning simply by uploading a dataset. In other words, even non-experts in machine learning can easily perform learning.

[0176] <Eleventh embodiment> 34 is a block diagram showing an example of the hardware configuration of a machine learning utilization support system according to an eleventh embodiment. The eleventh embodiment is a specific example of the first to tenth embodiments, in which the machine learning utilization support device 10 and the machine learning model evaluation device 20 are each realized by a computer.

[0177] The machine learning utilization support device 10 and the machine learning model evaluation device 20 each include a CPU (Central Processing Unit) 2, 2a, a RAM (Random Access Memory) 3, 3a, a ROM (Read Only Memory) 4, 4a, storage 5, 5a, a display device 6, 6a, an input device 7, 7a, and a communication device 8, 8a, all of which are connected by a bus.

[0178] The CPUs 2 and 2a are processors that execute arithmetic processing and control processing according to programs. The CPUs 2 and 2a use predetermined areas of the RAMs 3 and 3a as work areas and execute the processing of each part of the machine learning utilization support device 10 and the machine learning model evaluation device 20 in cooperation with programs stored in the ROMs 4 and 4a and storages 5 and 5a. Each of the CPUs 2 and 2a and the processors may be referred to as a processing circuit.

[0179] The RAMs 3 and 3a are memories such as SDRAMs (Synchronous Dynamic Random Access Memories), and function as work areas for the CPUs 2 and 2a.

[0180] The ROMs 3 and 3a are memories that store programs and various information in a non-rewritable manner.

[0181] The storages 5, 5a are devices that write and read data to magnetic recording media such as hard disk drives (HDDs), semiconductor storage media such as flash memories, magnetically recordable storage media such as HDDs, optically recordable storage media, etc. The storages 5, 5a write and read data to the storage media in response to control from the CPUs 2, 2a. The storages 5, 5a are an example of computer memory.

[0182] The display devices 6 and 6a are displays such as LCDs (Liquid Crystal Displays), etc. The display devices 6 and 6a display various information based on display signals from the CPUs 2 and 2a.

[0183] The input devices 7, 7a are input devices such as a mouse, a keyboard, etc. The input devices 7, 7a receive information input by a user as an instruction signal, and output the instruction signal to the CPUs 2, 2a.

[0184] The communication devices 8 and 8a communicate with external devices via a network under the control of the CPUs 2 and 2a. The communication devices 8 and 8a may also be called communication circuits.

[0185] The instructions shown in the processing procedures described in the above-described embodiments can be executed based on a software program. A general-purpose computer system can store this program in advance and, by loading this program, achieve effects similar to those achieved by the operations of the machine learning utilization support device 10 and the machine learning model evaluation device 20 described above. The instructions described in the above-described embodiments are recorded as a computer-executable program on a magnetic disk (such as a flexible disk or hard disk), an optical disk (such as a CD-ROM, CD-R, CD-RW, DVD-ROM, DVD±R, DVD±RW, or Blu-ray (registered trademark) Disc), a semiconductor memory, or a similar non-transitory recording medium. The recording medium may take any storage format as long as it is readable by a computer or an embedded system. The storage medium may also be referred to as a non-transitory computer-readable storage medium. By loading the program from the recording medium and causing a CPU to execute the instructions described in the program based on the program, a machine learning utilization support method can be realized, which involves operations similar to those of the machine learning utilization support device 10 and the machine learning model evaluation device 20 described in the above-described embodiments. Of course, when a computer acquires or reads a program, it may acquire or read the program through a network.

[0186] In addition, an OS (operating system), database management software, network middleware, etc. running on a computer may execute some of the processes required to realize this embodiment based on instructions from a program installed on the computer or embedded system from a recording medium.

[0187] Furthermore, the recording medium in this embodiment is not limited to a medium independent of a computer or an embedded system, but also includes a recording medium that stores or temporarily stores a program downloaded via a LAN, the Internet, or the like.

[0188] Furthermore, the number of recording media is not limited to one, and cases where the processing in this embodiment is executed from multiple media are also included in the recording media in this embodiment, and the media may have any configuration.

[0189] The computer or embedded system in this embodiment is for executing each process in this embodiment based on a program stored on a recording medium, and may be configured as either a device consisting of a single device such as a personal computer or a microcomputer, or a system in which multiple devices are connected to a network.

[0190] Furthermore, the computer in this embodiment is not limited to a personal computer, but also includes an arithmetic processing unit, a microcomputer, etc. included in information processing equipment, and is a general term for equipment or devices that can realize the functions in this embodiment by a program.

[0191] (Modifications of each embodiment) Each embodiment and each modified example may be expressed as a method or a program including each step of the machine learning utilization support device 10 and the machine learning model evaluation device 20 described above.

[0192] According to at least one of the embodiments described above, even a non-expert in machine learning can select an appropriate machine learning task based on the data to be processed. This also applies to at least one of the modified examples described above.

[0193] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, as well as within the scope of the invention described in the claims and their equivalents. [Explanation of symbols]

[0194] 10...machine learning utilization support device, 11...data acceptance unit, 12...data conversion unit, 13...machine learning task database, 14...search unit, 15...machine learning task proposal unit, 16, 25...display unit, 20...machine learning model evaluation device, 21...search result acceptance unit, 22...evaluation data acceptance unit, 23...machine learning model, 24...machine learning model evaluation unit, 26...data division unit, 27...machine learning model learning unit, 200...user dataset, 300...search results, 310...search prompts, 330…summary prompts, 350…question prompts.

Claims

1. a reception unit that receives a dataset to be processed in a machine learning task; a conversion unit that converts each sample of the data set into first data in a common data format; a database that stores in advance machine learning task candidates and second data having the common data format and relating to explanations of the candidates in association with each other; a search unit that searches the database for machine learning task candidates associated with the second data corresponding to the first data, based on the first data; a suggestion unit that generates suggestion data including candidates for the machine learning task and descriptions of the candidates based on the search results of the candidates; A machine learning utilization support device equipped with

2. The machine learning utilization support device according to claim 1 , further comprising a display unit that displays machine learning task candidates included in the proposal data and explanations of the candidates.

3. the common data format is a text format, the second data is the candidate keyword; the conversion unit generates captions that describe each sample in the data set in the text format, counts the number of times that words included in the captions appear, and defines the words that appear relatively frequently as the first data; The machine learning utilization support device according to claim 1 , wherein the search unit searches the database for the machine learning task candidates by keyword search based on the words that are the first data.

4. the common data format is a vector format, the conversion unit converts each sample of the data set into a first fixed-length vector that is the first data; the second data is a second fixed-length vector into which the candidate description is transformed; The machine learning utilization support device according to claim 1 , wherein the search unit searches for the machine learning task candidates by performing a vector similarity search between the first fixed-length vector and the second fixed-length vector.

5. the dataset is a set of image data; The machine learning utilization support device according to claim 4 , wherein the conversion unit generates a caption that describes each sample in the dataset in a text format and converts the caption into the first fixed-length vector.

6. the data set is a set of table data, 5. The machine learning utilization support device according to claim 4, wherein the conversion unit generates captions that explain characteristics of the dataset in text format based on file names, header strings, and statistics of each column of each sample in the dataset, and converts the captions into the first fixed-length vectors.

7. the dataset is a set of time series data; 5. The machine learning utilization support device according to claim 4, wherein the conversion unit generates captions that explain characteristics of the dataset in text format based on file names, series names, and statistics for each series of each sample in the dataset, and converts the captions into the first fixed-length vectors.

8. the dataset is a set of multimodal data including table data and time series data; 5. The machine learning utilization support device according to claim 4, wherein the conversion unit generates captions that explain characteristics of the dataset in text format based on file names, data formats, and header character strings included in the dataset, and converts the captions into the first fixed-length vectors.

9. the common data format is vector data, the conversion unit converts each sample of the dataset into a first fixed-length vector that is the first data, and estimates a first domain of the dataset based on the first fixed-length vector; the database stores the machine learning task candidates, second domains of the candidates, and the second data in association with each other; the second data is a second fixed-length vector into which the candidate description is transformed; The machine learning utilization support device according to claim 1 , wherein the search unit searches for the machine learning task candidates based on the first domain and the first fixed-length vector.

10. the common data format is a text format, the conversion unit generates, as the first data, a caption that describes each sample in the data set in a text format; The machine learning utilization assistance device according to claim 1 , wherein the search unit generates a search prompt including the caption, and searches for candidates for the machine learning task using a large-scale language model based on the search prompt.

11. the common data format is a text format, the conversion unit generates, as the first data, a caption that describes each sample in the data set in a text format; 3. The machine learning utilization support device according to claim 1, wherein the search unit generates a summary prompt including the caption, summarizes the dataset based on the summary prompt using a large-scale language model, and searches the database for candidates for the machine learning task by vector similarity search using the large-scale language model based on the summary result.

12. the database stores in advance machine learning task candidates in association with second data having the common data format and relating to explanations of the candidates; The machine learning utilization support device according to claim 11 , wherein the second data includes an explanatory sentence regarding the explanation of the candidate, and a fixed-length vector obtained by converting the explanatory sentence.

13. 11. The machine learning utilization support device according to claim 10, wherein, when a user is unable to select one of the searched candidates, the suggestion unit generates a question prompt including an instruction to generate a question for narrowing down the searched candidates, generates the question based on the question prompt using a large-scale language model, and, upon receiving an answer from the user to the question, generates new suggestion data including candidates narrowed down from the searched candidates and descriptions of the narrowed down candidates based on the answer.

14. 14. The machine learning utilization support device according to claim 13, wherein the suggestion unit sequentially narrows down the machine learning task candidates by repeatedly generating questions for narrowing down the candidates and generating proposal data based on answers to the questions.

15. The machine learning utilization assistance device according to claim 14 , wherein the question prompt includes an instruction to generate a question for which an affirmative answer or a negative answer can be selected.

16. The machine learning utilization assistance device according to claim 14 , wherein the question prompt includes an instruction to generate a question necessary for selecting the candidate.

17. The machine learning utilization assistance device according to claim 14 , wherein the question prompt includes an instruction for generating a reason for generating the question.

18. The machine learning utilization support device according to claim 1 ; a machine learning model evaluation device capable of communicating with the machine learning utilization support device; It is equipped with The machine learning model evaluation device, a search result receiving unit that receives the search results; an evaluation data receiving unit that receives a dataset for evaluating machine learning; a division unit that divides the dataset to be evaluated into a first dataset and a second dataset when there is no machine learning model for the machine learning task corresponding to the received search result; a learning unit that learns a machine learning model for the machine learning task based on the first dataset; an evaluation unit that evaluates the machine learning model obtained by the learning based on the second dataset; a display unit that displays an evaluation result of the machine learning model; A machine learning utilization system equipped with

19. The machine learning model is a trained model that has been trained in advance, when there is a trained model for the machine learning task corresponding to the received search results, the evaluation unit evaluates the trained model based on the dataset to be evaluated and the received search results; The display unit displays the evaluation result of the trained model. The machine learning utilization system according to claim 18.

20. a receiving unit receiving a dataset to be processed in a machine learning task; a conversion unit converting each sample of the data set into first data in a common data format; a database storing in advance machine learning task candidates and second data having the common data format and relating to explanations of the candidates in association with each other; a search unit searching, based on the first data, for machine learning task candidates associated with the second data corresponding to the first data in the database; a suggestion unit generating suggestion data including candidates for the machine learning task and descriptions of the candidates based on the search results of the candidates; A method for supporting the use of machine learning.

21. A computer with memory The ability to accept a dataset to process machine learning tasks; converting each sample of the data set into first data in a common data format; a function of previously storing in the memory, in association with a candidate machine learning task, second data having the common data format and relating to an explanation of the candidate; a function of searching, based on the first data, in the memory for candidate machine learning tasks associated with the second data corresponding to the first data; generating, based on the search results of the candidates, proposal data including candidates for the machine learning task and descriptions of the candidates; A program to support the use of machine learning to achieve this.

Images