Model training-based credible data space recommendation method and system

Abstract

The invention discloses a credible data space recommendation method and system based on model training. The method comprises the steps that S1, a data demander initiates a query request and submits the query request to a data market; s2, vectorizing the acquired query request by the data market; inputting the vectorized query request into an original recommendation model to obtain a recommendation result; s3, the data market feeds back a recommendation result to the data demand side; and S4, the data demander views the recommendation result, and feeds back behavior data when the recommendation result is viewed to the data market, so that the data market optimizes the original recommendation model according to the behavior data of the data demander. According to the method, an original recommendation model is updated and optimized in real time by using latest behavior data of a user, so that a recommendation result is ensured to be more accurate; by only inputting the vectorized query request into the original recommendation model to obtain the recommendation result, the recommendation service can be provided on the premise of not leaking user data, so that privacy leakage is effectively avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of data resource recommendation, and in particular to a method and system for recommending a trusted data space based on model training. Background Art

[0002] In the automotive field, a trusted data space for the automotive industry is created, which contains product data of various automotive suppliers, various types of automotive data, the latest developments in the industry, etc. If users want to quickly obtain the most accurate information, it is difficult to quickly obtain accurate data information through traditional keyword matching-based database queries. Traditional recommendation algorithms will also face various problems if they are directly transplanted. For example, the user's personal behavior information needs to be submitted to a centralized node for recommendation, so that the user's behavior information will be exposed in an untrustworthy environment.

[0003] There is currently no effective solution to the problems of inaccurate recommendation methods and difficulty in privacy protection in existing technologies. Summary of the invention

[0004] The embodiments of the present invention provide a recommendation method and system for a trusted data space based on model training to solve the problems of inaccurate recommendation methods and difficulty in privacy protection in the prior art.

[0005] To achieve the above-mentioned purpose, on the one hand, the present invention provides a recommendation method for a trusted data space based on model training, the method comprising: S1, the data demander initiates a query request and submits it to the data market; S2, the data market vectorizes the obtained query request; and inputs the vectorized query request into the original recommendation model to obtain the recommendation result; S3, the data market feeds back the recommendation result to the data demander; S4, the data demander views the recommendation result, and feeds back the behavior data when viewing the recommendation result to the data market, so that the data market optimizes the original recommendation model according to the behavior data of the data demander.

[0006] Optionally, S1 includes: the data demander initiates a query request; the connector of the data demander converts the query request into a keyword set, and submits the keyword set to the data market.

[0007] Optionally, before S1, the steps include: generating a unique UUID identifier for each query request; dividing all recommendation results corresponding to the current query request into multiple pages, and associating each page with the UUID identifier generated by the current query request.

[0008] Optionally, S3 includes: the data market feeds back the first page of recommendation results to the data demander; when the data demander requests to query more recommendation results, continue to query the remaining recommendation results in pages according to the UUID identifier generated by the current query request.

[0009] Optionally, the original recommendation model is obtained by the following method: the data market obtains the characteristic data and metadata description information of all users as a data set; and obtains the IDs and metadata IDs of all users and converts them into embedding vectors; the data set and the embedding vectors are input into the DeepFM model for multiple rounds of model training to obtain the original recommendation model.

[0010] On the other hand, the present invention provides a recommendation system for a trusted data space based on model training, the system comprising: an initiating unit, used by a data demander to initiate a query request and submit it to a data market; a recommendation unit, used by the data market to vectorize the query request obtained; and input the vectorized query request into the original recommendation model to obtain a recommendation result; a feedback unit, used by the data market to feed back the recommendation result to the data demander; an optimization unit, used by the data demander to view the recommendation result, and feed back the behavioral data when viewing the recommendation result to the data market, so that the data market optimizes the original recommendation model according to the behavioral data of the data demander.

[0011] Optionally, the initiating unit includes: an initiating subunit, used by the data demander to initiate a query request; a converting subunit, used by the connector of the data demander to convert the query request into a keyword set, and submit the keyword set to the data market.

[0012] Optionally, it also includes: a generation unit, used to generate a unique UUID identifier for each query request; divide all recommended results corresponding to the current query request into multiple pages, and associate each page with the UUID identifier generated by the current query request.

[0013] Optionally, the feedback unit includes: a feedback subunit, used for the data market to feed back the first page of recommended results to the data demander; and a paging query subunit, used for continuing to paging query the remaining recommended results according to the UUID identifier generated by the current query request when the data demander requests to query more recommended results.

[0014] Optionally, it also includes: a model training unit, which is used in the data market to obtain the characteristic data and metadata description information of all users as a data set; and obtain the IDs and metadata IDs of all users and convert them into embedding vectors; input the data set and the embedding vector into the DeepFM model for multiple rounds of model training to obtain the original recommendation model.

[0015] Beneficial effects of the present invention:

[0016] The present invention provides a recommendation method and system for a trusted data space based on model training, wherein the method trains the model based on the user's past behavior information after characterization and in combination with the data information. After the original recommendation model is trained, the user's latest behavior data is used to update and optimize the original recommendation model in real time to ensure that the recommendation results are more accurate. By only inputting vectorized query requests into the original recommendation model to obtain recommendation results, recommendation services can be provided without leaking user data, thereby effectively avoiding the risks of privacy leakage and data abuse and meeting the increasingly stringent requirements for data privacy protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of a method for recommending a trusted data space based on model training provided by an embodiment of the present invention;

[0018] Figure 2 It is a structural diagram of a recommendation system of a trusted data space based on model training provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] It is difficult to quickly obtain accurate data information using traditional recommendation algorithms (i.e., querying from the database based on keyword matching), and directly transplanting traditional recommendation algorithms into the trusted data space may cause difficulties in privacy protection.

[0021] A trusted data space is a data space used for data circulation, which includes data providers, data demanders, and a centralized metadata description center. In order to maximize the promotion of data circulation in the data space, we need to explore the interests of data demanders as much as possible. The more accurate the query and recommendation of data, the more conducive it is to data transactions between data providers and data demanders. Therefore, the present invention proposes a logical framework based on the data space and designs a recommendation system that conforms to the current framework. While data resources do not leave the domain in the trusted data space, the maximum circulation of data in the space can be achieved, and users of the trusted data space can be promoted to find data and recommend related content.

[0022] Taking the automobile trusted data space as an example, the present invention provides a specific query implementation example.

[0023] Figure 1 is a flow chart of a method for recommending a trusted data space based on model training provided by an embodiment of the present invention, such as Figure 1 As shown, the method includes:

[0024] S1. The data demander initiates a query request and submits it to the data market;

[0025] Specifically, the S1 includes:

[0026] S11. The data demander initiates a query request;

[0027] User A (data demander) wants to query the "2024 China New Energy Vehicle Market Report" and requires filtering by dimensions such as brand, model, and technology trends.

[0028] The original description of the request might be: "Query the report on China's new energy vehicle market in 2024, including market trends, major brands, sales data, and technological innovation."

[0029] S12: The connector of the data demander converts the query request into a keyword set, and submits the keyword set to the data market.

[0030] Keyword conversion: Convert user A's query request into a set of keywords for the system to search:

[0031] Keyword collection: ["2024", "China", "new energy vehicles", "market report", "brand", "model", "technology trend"]

[0032] The data demander submits the converted keyword set to the data market.

[0033] Furthermore, the data demander submits the converted keyword set and paging information (if any) to the data market, and also includes a new query identifier (UUID) to track the query.

[0034] In an optional implementation, a unique UUID identifier is generated for each query request; all recommendation results corresponding to the current query request are divided into multiple pages, and each page is associated with the UUID identifier generated by the current query request.

[0035] For example, a query request is assigned a unique UUID identifier in the data market: uuid-12345, so as to track subsequent paginated queries. Assuming that the current query request may correspond to 500 recommended results, the 500 recommended results corresponding to the current query request are divided into 5 pages, each with 100 recommended results. When requesting to view each page, the UUID identifier uuid-12345 is required.

[0036] The current query request data structure (including paging information):

[0037] {

[0038] "query_id":"uuid-12345", / / query identifier

[0039] "keywords":["2024","China","new energy vehicles","market report","brand","model","technological trends"], / / keyword collection

[0040] "pagination":{

[0041] "page":1, / / page number

[0042] "page_size":100 / / Number of recommended results per page

[0043] }

[0044] }

[0045] The above indicates a request to query the 100 recommended results on the first page.

[0046] S2. The data market vectorizes the query request obtained, and inputs the vectorized query request into the original recommendation model to obtain the recommendation result.

[0047] Specifically, after receiving a query request, the data market will vectorize the converted keyword set. Vectorization is the process of converting text into a digital format. The vectorized query request is input into the original recommendation model to obtain the recommendation results.

[0048] The original recommendation model is obtained by the following method:

[0049] The data market obtains the feature data and metadata description information of all users as a data set; and obtains the IDs and metadata IDs of all users and converts them into embedding vectors;

[0050] The dataset and embedding vectors are input into the DeepFM model for multiple rounds of model training to obtain the original recommendation model.

[0051] The data market only obtains the characteristic data and metadata description information of all users, and does not obtain the user's original data information. Therefore, the user's data will not be leaked, ensuring the privacy of the data.

[0052] The DeepFM model performs two calculations: one is to use the factorization machine (FM) to calculate the second-order interactions between features, and the other is to pass the output of the FM part and the embedding vector to the deep neural network (DNN) part, which is responsible for learning higher-order nonlinear interactions between features. The outputs of the FM part and the DNN part are combined to form the final recommendation result.

[0053] The original recommendation model contains all relevant data in the database and vectorizes this data (reports, articles, research results, etc.).

[0054] For example, the content of a report in the database may become:

[0055] {

[0056] "doc_id":"doc-001", / / Document ID

[0057] "title":"China's New Energy Vehicle Market Development Report 2024", / / Title

[0058] "vector":[0.21,0.56,0.43,0.67,0.89] / / vector

[0059] }

[0060] By matching the vectorized database with the vectorized keyword set, the original recommendation model calculates the most relevant results and generates output.

[0061] Output data (based on query results):

[0062]

[0063] The above represents all the recommended results output, in which two recommended results (document 1doc-001, document 2doc-002) are displayed.

[0064] S3. The data market feeds back the recommendation results to the data demander;

[0065] Specifically, S3 includes:

[0066] S31. The data market feeds back the first page of recommendation results to the data demander;

[0067] Display query results: The data market returns the recommended results to the data demander, and the data demander displays the results to user A.

[0068] Display content (first page of recommended results):

[0069]

[0070] The first page of 100 results will be displayed to user A.

[0071] S32: When the data demander requests to query more recommended results, the remaining recommended results are continuously queried in pages according to the UUID identifier generated by the current query request.

[0072] For example, user A viewed 100 results on page 1. To obtain more data, the same uuid-12345 is used for paging query, that is, the data demander initiates paging query. For example, user A now requests to query the second page of data (100 results per page).

[0073] Paginated query request:

[0074] {

[0075] "query_id":"uuid-12345", / / query identifier

[0076] "pagination":{

[0077] "page":2, / / Number of pages

[0078] "page_size":100 / / Number of recommended results per page

[0079] }}

[0080] S4. The data demander checks the recommendation results and feeds back the behavior data when checking the recommendation results to the data market, so that the data market optimizes the original recommendation model according to the behavior data of the data demander.

[0081] When viewing the recommendation results, user A clicks on the first article "China's New Energy Vehicle Market Development Report in 2024" and collects the second article "Analysis of New Energy Vehicle Technology Trends in 2024". These behavioral data will be fed back to the data market to optimize the recommendation model.

[0082] Specifically, user behavior data is represented as:

[0083] {

[0084] "query_id":"uuid-12345", / / query identifier

[0085] "user_behavior":[

[0086] {

[0087] "doc_id":"doc-001", / / Document ID

[0088] "action":"click", / / Action: click

[0089] "timestamp":"2024-12-05T10:30:00Z" / / timestamp

[0090] },

[0091] {

[0092] "doc_id":"doc-002", / / Document ID

[0093] "action":"collect", / / Action: Collect

[0094] "timestamp":"2024-12-05T10:32:00Z" / / timestamp

[0095] }

[0096] ]}

[0097] These behavior data (clicks, favorites, etc.) will be used to retrain the original recommendation model and optimize the recommendation results. For example, the original recommendation model will learn that user A prefers market trends and technical reports on new energy vehicles, and thus prioritize relevant content in subsequent queries.

[0098] Figure 2 A recommendation system based on a trusted data space of model training is provided in an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0099] The initiating unit 201 is used for the data demander to initiate a query request and submit it to the data market;

[0100] The initiating unit 201 includes:

[0101] The initiating subunit 2011 is used for the data demander to initiate a query request;

[0102] The conversion subunit 2012 is used for the connector of the data demander to convert the query request into a keyword set and submit the keyword set to the data market.

[0103] It also includes: a generation unit, which is used to generate a unique UUID identifier for each query request; divide all recommended results corresponding to the current query request into multiple pages, and associate each page with the UUID identifier generated by the current query request.

[0104] The recommendation unit 202 is used for the data market to vectorize the query request obtained; and input the vectorized query request into the original recommendation model to obtain the recommendation result;

[0105] It also includes: a model training unit, which is used in the data market to obtain the characteristic data and metadata description information of all users as a data set; and obtain the IDs and metadata IDs of all users and convert them into embedded vectors; input the data set and embedded vectors into the DeepFM model for multiple rounds of model training to obtain the original recommendation model.

[0106] Feedback unit 203, used for the data market to feed back the recommendation results to the data demander;

[0107] The feedback unit 203 includes:

[0108] Feedback subunit 2031, used for the data market to feed back the first page of recommendation results to the data demander;

[0109] The paging query subunit 2032 is used to continue paging querying the remaining recommendation results according to the UUID identifier generated by the current query request when the data demander requests to query more recommendation results.

[0110] The optimization unit 204 is used for the data demander to view the recommendation results and feed back the behavior data when viewing the recommendation results to the data market, so that the data market optimizes the original recommendation model according to the behavior data of the data demander.

[0111] The system provided by the present invention corresponds to the above method, and the specific implementation of the system will not be repeated here.

[0112] Beneficial effects of the present invention:

[0113] The present invention provides a recommendation method and system for a trusted data space based on model training, wherein the method trains the model based on the user's past behavior information after characterization and in combination with the data information. After the original recommendation model is trained, the user's latest behavior data is used to update and optimize the original recommendation model in real time to ensure that the recommendation results are more accurate. By only inputting vectorized query requests into the original recommendation model to obtain recommendation results, recommendation services can be provided without leaking user data, thereby effectively avoiding the risks of privacy leakage and data abuse and meeting the increasingly stringent requirements for data privacy protection.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A recommendation method for a trusted data space based on model training, characterized in that: include: S1. The data demander initiates a query request and submits it to the data market; S2. The data market vectorizes the obtained query requests; The vectorized query request is input into the original recommendation model to obtain the recommendation results; S3. The data market feeds back the recommendation results to the data demander; S4. The data demander checks the recommendation results and feeds back the behavior data when checking the recommendation results to the data market, so that the data market optimizes the original recommendation model according to the behavior data of the data demander.

2. The method according to claim 1, characterized in that The S1 includes: The data demander initiates a query request; The connector of the data demander converts the query request into a keyword set and submits the keyword set to the data market.

3. The method according to claim 1, characterized in that Before said S1 includes: Generate a unique UUID identifier for each query request; divide all recommended results corresponding to the current query request into multiple pages, and associate each page with the UUID identifier generated by the current query request.

4. The method according to claim 3, characterized in that The S3 includes: The data market will feed back the first page of recommended results to the data demander; When the data demander requests to query more recommended results, the remaining recommended results are continuously queried in pages according to the UUID identifier generated by the current query request.

5. The method according to claim 1, characterized in that The original recommendation model is obtained by the following method: The data market obtains the feature data and metadata description information of all users as a data set; and obtains the IDs and metadata IDs of all users and converts them into embedding vectors; The dataset and embedding vectors are input into the DeepFM model for multiple rounds of model training to obtain the original recommendation model.

6. A recommendation system based on a trusted data space for model training, characterized in that: include: The initiating unit is used by the data demander to initiate a query request and submit it to the data market; The recommendation unit is used by the data market to vectorize the query requests obtained; The vectorized query request is input into the original recommendation model to obtain the recommendation results; Feedback unit, used by the data market to feed back recommendation results to data demanders; The optimization unit is used for the data demander to view the recommendation results and feed back the behavior data when viewing the recommendation results to the data market, so that the data market optimizes the original recommendation model according to the behavior data of the data demander.

7. The system according to claim 6, characterized in that The initiating unit comprises: The initiating subunit is used by the data demander to initiate a query request; The conversion subunit is used for the connector of the data demander to convert the query request into a keyword set and submit the keyword set to the data market.

8. The system according to claim 6, characterized in that Also includes: The generating unit is used to generate a unique UUID identifier for each query request; divide all the recommendation results corresponding to the current query request into multiple pages, and associate each page with the UUID identifier generated by the current query request.

9. The system according to claim 8, characterized in that The feedback unit comprises: Feedback subunit, used by the data market to feed back the first page of recommendation results to the data demander; The paging query subunit is used to continue paging querying the remaining recommendation results according to the UUID identifier generated by the current query request when the data demander requests to query more recommendation results.

10. The system according to claim 6, characterized in that Also includes: The model training unit is used in the data market to obtain the feature data and metadata description information of all users and use them as data sets; And get all user IDs and metadata IDs and convert them into embedding vectors; The dataset and embedding vectors are input into the DeepFM model for multiple rounds of model training to obtain the original recommendation model.

Images