User tag automatic generation method based on multi-scene data

Abstract

The invention belongs to the technical field of e-commerce retail and industrial product purchasing, and particularly relates to a multi-scene data-based user tag automatic generation method, which comprises the steps of multi-scene data access and preprocessing, effective frequency statistics, data association feature recognition and user tag automatic generation. According to the invention, multi-scene data access is realized, such as purchase, browsing and collection data of articles of daily use and industrial products, and missing value filling, abnormal value elimination and format unified preprocessing are carried out; removing redundancy for browsing or collection; extracting single features, calculating associated features and assigning mutual information weights; triggering a label according to a scene threshold value, and storing the label through confidence verification; a high-precision fine particle label can be generated, and precision marketing, personalized recommendation and purchase demand matching are supported.

Description

Technical Field

[0001] This invention belongs to the field of e-commerce retail and industrial product procurement technology, specifically a method for automatically generating user tags based on multi-scenario data. Background Technology

[0002] In the fields of e-commerce retail and industrial product procurement, user tags are the core of accurately matching needs with goods / services. The number of purchases (reflecting actual needs), the number of views (reflecting initial interest), and the number of favorites (implying potential intentions) are three types of data that are crucial to the quality of tags.

[0003] However, current technology has significant limitations:

[0004] The data dimension is too narrow: relying solely on the number of views to determine "potential interests" or the number of purchases to determine "spending power" without integrating the number of favorites, which leads to misjudgments (e.g., frequently browsing industrial product X but not favorited or purchased, when it was actually just price comparison) or omissions (e.g. frequently favorited household product A but not purchased, potential demand is ignored).

[0005] Manually-driven rules: Tags are generated based on fixed thresholds (e.g., purchasing household item B more than 5 times is considered a high-frequency consumer), without exploring the correlation between the three types of data, and unable to identify segmented users such as "high browsing-high collection-low purchase" (e.g., household item C);

[0006] The statistical method is crude: the number of views and favorites is not deduplicated (e.g., 10 views of daily necessities B in 1 day are still counted), which leads to the distortion of features and mistakenly judges repeated operations as "high interest".

[0007] In summary, existing labeling technologies suffer from low accuracy and coarse granularity, necessitating a label generation technology that integrates three types of data, optimizes statistical rules, and mines related features. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies and solve the aforementioned technical problems, this invention proposes a method for automatically generating user tags based on multi-scenario data.

[0009] The technical solution adopted by this invention to solve its technical problem is as follows: This invention proposes a method for automatically generating user tags based on multi-scenario data, including the following steps:

[0010] S1. Multi-scenario data access and preprocessing: Access data sources according to the "user-category" dimension. The data sources include the purchase count, browsing count, and collection count of each of the daily necessities category A / B / C and the industrial products category X / Y / Z. Perform missing value filling, outlier removal, and format unification processing on the data sources.

[0011] S2. Valid Count Statistics: For the data processed by S1, perform redundancy removal statistics on the number of views, the number of favorites, and the number of purchases. The number of views and the number of favorites are redundant according to preset rules, and the number of purchases is counted based on the actual number of orders placed.

[0012] S3. Data association feature identification: Extract the number of purchases, valid views, and valid favorites for each category as single features, calculate the favorite-view ratio and purchase-view ratio for each category as association features, and assign weights to each feature based on mutual information.

[0013] S4. Automatic generation of user tags: Based on the sum of feature weights of each category and the scene threshold, candidate tags are triggered, the confidence level of the candidate tags is calculated and verified, and candidate tags that meet the confidence level requirements are added to the database as official user tags, while those that do not meet the requirements are discarded.

[0014] Furthermore, the specific rules for filling missing values ​​in S1 are as follows: if a user does not perform a certain type of behavior, the corresponding number of times is filled with 0; if data for a certain date is missing, it is filled with the average number of times for the corresponding category of the user in the past 7 days, and the filling error is allowed to be ±10%.

[0015] Furthermore, the specific rule for outlier removal in S1 is as follows: use three times the number of historical behaviors of the user in the corresponding category as a threshold, and remove the number of outliers that exceeds the threshold.

[0016] Furthermore, the specific rules for the unified format described in S1 are as follows: data is summarized according to "user ID-category ID-statistical period", where the statistical period for daily necessities categories A / B / C is set to "day", and the statistical period for industrial products categories X / Y / Z is set to "month".

[0017] Furthermore, the specific rules for redundancy removal statistics of browsing counts described in S2 are as follows: the number of times the same user browses the same category within the statistical period is counted as 1 valid browsing count; among them, the browsing statistical period for daily necessities categories A / B / C is 1 day, and the browsing statistical period for industrial products categories X / Y / Z is 3 days.

[0018] Furthermore, the specific rules for redundancy removal in the collection count described in S2 are as follows: the number of times the same user collects products of the same category is counted as 1 valid collection count; if a user cancels a collection of a certain category and then collects it again, if the interval is less than 7 days, it will not be counted repeatedly, and if the interval is ≥ 7 days, it will be counted as 1 valid collection count again.

[0019] Furthermore, the feature weight assignment based on mutual information described in S3 is as follows: the weight of the number of purchases is 0.9, the weight of the number of valid favorites is 0.6, the weight of the number of valid views is 0.3, and the weight of the associated features (favorites-views ratio, purchase-views ratio) is 0.7.

[0020] Furthermore, in S3, if the sum of the weights of purchase frequency, valid pageview frequency, valid collection frequency, and related features within the same category exceeds 2.0, then 2.0 is taken as the total weight of that category.

[0021] Furthermore, the specific settings of the scenario thresholds mentioned in S4 are as follows: the tag trigger threshold for the daily necessities scenario is 1.2, and the tag trigger threshold for the industrial products scenario is 1.5; if the total weight of a certain category is greater than or equal to the tag trigger threshold of the corresponding scenario, then the basic candidate tag is triggered; if the weight of the associated feature of a certain category is greater than or equal to 0.6, then the subdivision candidate tag is triggered.

[0022] Furthermore, the specific rules for confidence calculation and verification described in S4 are as follows: Confidence = (Number of valid features that trigger the label / Total number of features required for the label) × 100%; where the total number of features required for the basic candidate label is 2, and the confidence level must be ≥60%; the total number of features required for the subdivided candidate label is 3, and the confidence level must be ≥70%.

[0023] The beneficial effects of this invention are as follows:

[0024] 1. The user tag automatic generation method based on multi-scenario data described in this invention accesses three core data categories—purchase frequency, browsing frequency, and collection frequency—for daily necessities (A / B / C) and industrial products (X / Y / Z). It then performs missing value filling (filling with 0 for no activity, filling with the average of the last 7 days for missing dates with an error of ±10%), outlier removal (using 3 times the historical average as the threshold), and format standardization (daily necessities summarized by "day" and industrial products by "month") to achieve the completeness and standardization of multi-scenario, multi-dimensional behavioral data. This solves the problem of tag one-sidedness caused by existing technologies focusing on a single data dimension (such as misjudging price comparison users based solely on browsing frequency, or missing potential demand for high collection frequency based solely on purchase frequency).

[0025] 2. The user tag automatic generation method based on multi-scenario data described in this invention removes redundancy by setting differentiated cycles for browsing times according to scenarios (daily necessities are counted once for the same category within 1 day, and industrial products are counted once for the same category within 3 days), and removes duplicates for collection times by category (regardless of the number of collections within the same category, it is counted once) and cancels the re-collection interval of ≥7 days. This achieves the removal of redundancy and restoration of true intent in user behavior data, and solves the feature distortion problem caused by the coarse data statistics method of the existing technology (such as cumulative counting of repeated browsing and repeated collections) (such as mistakenly judging 10 repeated browsing times within 1 day as high interest).

[0026] 3. The user tag automatic generation method based on multi-scenario data described in this invention calculates the association features such as the collection-view ratio and purchase-view ratio of each category, assigns differentiated weights to the features based on mutual information (purchase frequency 0.9, association feature 0.7, effective collection 0.6, effective view 0.3), and sets scenario-based tag trigger thresholds (daily necessities 1.2, industrial products 1.5), thereby achieving accurate identification and tag adaptation of fine-grained user needs. This solves the problems of insufficient association feature mining and the "one-size-fits-all" thresholding in existing technologies, which result in coarse tag granularity and insufficient accuracy (such as the inability to identify high-intent hesitant users with "high views-high collections-low purchases"). Attached Figure Description

[0027] The invention will now be further described with reference to the accompanying drawings.

[0028] Figure 1 This is a flowchart of the method for automatically generating user tags based on multi-scenario data according to the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] like Figure 1 As shown, the present invention provides a method for automatically generating user tags based on multi-scenario data, which includes the following steps:

[0031] S1. Multi-scenario data access and preprocessing: Access data sources according to the "user-category" dimension. The data sources include the purchase count, browsing count, and collection count of each of the daily necessities category A / B / C and the industrial products category X / Y / Z. Perform missing value filling, outlier removal, and format unification processing on the data sources.

[0032] S2. Valid Count Statistics: For the data processed by S1, perform redundancy removal statistics on the number of views, the number of favorites, and the number of purchases. The number of views and the number of favorites are redundant according to preset rules, and the number of purchases is counted based on the actual number of orders placed.

[0033] S3. Data association feature identification: Extract the number of purchases, valid views, and valid favorites for each category as single features, calculate the favorite-view ratio and purchase-view ratio for each category as association features, and assign weights to each feature based on mutual information.

[0034] S4. Automatic generation of user tags: Based on the sum of feature weights of each category and the scene threshold, candidate tags are triggered, the confidence level of the candidate tags is calculated and verified, and candidate tags that meet the confidence level requirements are added to the database as official user tags, while those that do not meet the requirements are discarded.

[0035] Work Process: First, purchase, browsing, and collection data for daily necessities (A / B / C) and industrial products (X / Y / Z) are collected according to the "user-category" dimension. Preprocessing is performed by filling missing values ​​(filling in 0 for missing rows and using the average of the last 7 days for missing data), removing outliers (using 3 times the historical average as the threshold), and standardizing the format (user ID-category ID-period, daily for daily necessities and monthly for industrial products). Next, redundancy is removed from browsing (1 time within a 1-day period for daily necessities and 3 times within a 3-day period for industrial products) and collection (1 time within the same category, with a cancellation and re-collection interval of ≥7 days before re-collection). Purchases are counted based on actual orders. Then, single features (purchase, valid browsing, valid collection count) are extracted, correlation features (collection-browsing ratio, purchase-browsing ratio) are calculated, and weights are assigned according to mutual information (purchase 0.9, collection 0). 6. Browsing 0.3, Association 0.7); Finally, candidate tags are triggered according to the scenario threshold (daily necessities 1.2 / industrial products 1.5), and the confidence score (number of effective features / total number of required features × 100%) is calculated. If the basic tag is ≥60% and the sub-tag is ≥70%, it is added to the database; otherwise, it is discarded. This breaks the existing technology's "single-dimensional data dependence" and integrates three types of core behavioral data. It eliminates feature distortion caused by browsing and collection redundancy. It mines associated features to identify subdivided user types, solving the problems of one-sided tags and low accuracy. It can also cover the entire scenario of daily necessities and industrial products, providing a complete and real data source for subsequent tag generation. It builds a standardized data processing process to reduce manual intervention. It lays the foundation for fine-grained tag generation and supports business needs such as precision marketing and personalized recommendations.

[0036] As a specific implementation of the present invention: the specific rules for filling missing values ​​in S1 are as follows: if a user does not perform a certain type of behavior, the corresponding number of times is filled to 0; if data for a certain date is missing, it is filled according to the average number of times for the corresponding category of the user in the past 7 days, and the filling error is allowed to be ±10%.

[0037] Working Process: When a user does not make a purchase of household item A, the "Number of Purchases of Household Item A" is filled with 0. When a user's "Number of Browsing Household Item B" data is missing for a certain date, the average number of times household item B has been viewed by that user in the past 7 days is calculated (e.g., if the user viewed it 21 times in the past 7 days, the average is 3 times). This average is used to fill in the missing date data. If the filled value deviates from the theoretical true value within ±10% (e.g., if the true value is 3 times, the filled value is 2.7-3.3 times), then the filling is considered valid. This effectively solves the problem of "incomplete features due to missing data" and avoids the misjudgment of tags due to missing behavioral data (e.g., not filling with 0 might misjudge that the user has no need for this category). By filling with the average of the past 7 days and error control, the rationality of missing data is ensured, and the impact of filling deviation on subsequent statistics is reduced. The integrity of the data source is ensured, with no data missing loopholes. The filling rules are quantifiable and reproducible, requiring no subjective human judgment. Error control further improves data accuracy and provides a reliable basis for valid count statistics.

[0038] As a specific implementation of the present invention: the specific rule for outlier removal in S1 is: using 3 times the number of historical behaviors of the user in the corresponding category as a threshold, outlier counts exceeding the threshold are removed.

[0039] Working Process: If a user's historical average number of views for "Industrial Product X" is 5 (e.g., an average of 5 views per month over the past 10 months), the outlier threshold is set to 15 (5 x 3). When a user's number of views for "Industrial Product X" is 20 in a given statistical analysis, these 20 views are considered outliers and removed from the dataset. Only normal views that meet the threshold are retained (e.g., 4 or 6 views in subsequent statistics). This eliminates the problem of "feature distortion caused by failure to remove outliers" in existing technologies, and avoids overestimating user interest or demand intensity due to extreme values ​​(e.g., high frequency caused by misoperation or system statistical errors). It also filters out invalid and interfering data to ensure the authenticity of behavioral data. Furthermore, the threshold setting is based on the user's own historical data, which aligns with individual behavioral patterns and is more accurate than a fixed global threshold. It also reduces the negative impact of outliers on subsequent weight calculations and tag triggering.

[0040] As a specific implementation of the present invention: the specific rule for the unified format in S1 is: data is summarized according to "user ID-category ID-statistical period", wherein the statistical period for daily necessities categories A / B / C is set to "day", and the statistical period for industrial products categories X / Y / Z is set to "month".

[0041] Work Process: For user "U1's" daily necessities data, daily purchase, browsing, and collection counts are summarized in the format "U1-Category A-20240501" and "U1-Category B-20240502". For user "U2's" industrial products data, monthly purchase, browsing, and collection counts are summarized in the format "U2-Category X-202405" and "U2-Category Y-202406", ensuring that all data includes a three-dimensional identifier of "user-category-period". This effectively solves the problem of "disorganized data format and inability to correlate across periods" in existing technologies, achieving structured storage of daily necessities (daily cycle) and industrial products (monthly cycle) data, avoiding statistical discrepancies caused by inconsistent periods (e.g., daily statistics for industrial products may result in a large number of zero values ​​due to low-frequency purchases). It standardizes the data structure, allowing for direct access for subsequent effective count statistics and feature extraction. It also supports cross-period data comparison and analysis (e.g., the correlation between a user's industrial product purchases in a certain month and their daily daily necessities consumption in the same month). This reduces the format adaptation cost during data processing and improves algorithm efficiency.

[0042] As a specific implementation of the present invention: the specific rule for the redundancy removal statistics of browsing counts in S2 is as follows: the number of times the same user browses the same category within the statistical period is counted as 1 valid browsing count; wherein, the browsing statistical period for daily necessities categories A / B / C is 1 day, and the browsing statistical period for industrial products categories X / Y / Z is 3 days.

[0043] Working Process: User "U3" browses daily necessities A 12 times in 1 day, which is counted as only 1 "valid browsing count of daily necessities A" according to the rules; User "U4" browses industrial products X 5 times in 3 days (2 times on day 1, 1 time on day 2, and 2 times on day 3), which is counted as only 1 "valid browsing count of industrial products X" according to the rules; if the statistical period is exceeded (e.g., U3 browses category A again on day 2), the count is reset; This eliminates the problem of "overestimation of interest due to repeated counting of browsing counts" in the existing technology, avoids judging invalid browsing such as repeated refreshes and accidental touches (e.g., 12 browsings in 1 day) as high interest, and restores the true browsing intent; the valid browsing count is more in line with the user's true interest, improving the accuracy of feature data; a differentiated period is set according to the category scenario (daily for high-frequency daily necessities, 3 days for low-frequency industrial products), which conforms to the user behavior patterns of the two types of products; and the amount of redundant data is reduced, reducing the computing power consumption of subsequent feature calculations.

[0044] As a specific implementation of the present invention: the specific rules for redundancy removal statistics of collection counts in S2 are as follows: the number of times the same user collects the same category of goods is counted as 1 valid collection count; if a user cancels collection of a certain category and then collects it again, if the interval is less than 7 days, it will not be counted repeatedly, and if the interval is ≥ 7 days, it will be counted as 1 valid collection count again.

[0045] Working Process: User "U5" favorites 3 items under the "Household Goods C" category. According to the rules, this is counted as only one "valid favorite count for Household Goods C". If U5 cancels favorites of Household Goods C and then favorites it again after 5 days, the valid count is not counted again. If the user favorites it again after 8 days, it is counted as one "valid favorite count for Household Goods C". This effectively solves the problem of "exaggerated potential demand due to repeated counting of favorites" in existing technologies, avoids misjudging multiple favorites of the same category (e.g., 3 items) as high intent, and filters out short-term repeated operations (e.g., canceling and then favoriteing again after 5 days) through the interval rule, restoring the true intention of favorites. The valid favorite count accurately reflects the user's potential demand for the category, rather than the number of items. The interval rule distinguishes between "short-term hesitation" and "long-term re-intent", improving the accuracy of potential demand judgment. It provides real data support for subsequent calculation of related features (favorites-view ratio).

[0046] As a specific embodiment of the present invention: the feature weight assignment based on mutual information in S3 is as follows: the weight of the number of purchases is 0.9, the weight of the number of valid favorites is 0.6, the weight of the number of valid views is 0.3, and the weight of the associated features (favorites-view ratio, purchase-view ratio) is 0.7.

[0047] Work Process: For user "U6's" data on daily necessities (A), extract the single features "Number of Purchases = 2" (weight 0.9) and "Number of Valid Views = 3" (weight 0.3), and calculate the related feature "Purchase-View Ratio = 2 / 3 ≈ 0.67" (weight 0.7); For user "U7's" data on industrial products (Z), extract the single features "Number of Valid Favorites = 1" (weight 0.6) and "Number of Valid Views = 2" (weight 0.3), and calculate the related feature "Favorites-View Ratio = 1 / 2 = 0.5" (weight 0.7), assigning corresponding weight values ​​to each feature; breaking the existing... There is a problem with the technology of "equalizing feature weights, making it impossible to distinguish importance". By assigning mutual information values, feature priority is clearly defined - purchasing behavior (weight 0.9) best reflects the real demand, and related features (0.7) better reflect the demand type than single features (browsing 0.3), avoiding the wrong tag triggering due to chaotic weights; it realizes the quantification of feature importance, and the subsequent weight sum calculation is more in line with the demand logic; and the weight of related features is higher than that of basic browsing features, which helps to identify segmented users such as "high collection - low purchase"; at the same time, it provides a reasonable weight basis for setting scene thresholds and improves the accuracy of tag triggering.

[0048] As a specific embodiment of the present invention: In S3, if the sum of the weights of the number of purchases, the number of valid views, the number of valid favorites, and the associated features under the same category exceeds 2.0, then 2.0 is taken as the total weight of the category.

[0049] Work process: Calculate the total feature weight of user "U8" for industrial product X: purchase count (0.9) + valid collection count (0.6) + collection-view ratio (0.7) = 2.2. Since it exceeds 2.0, take 2.0 as the "total weight of industrial product X"; Calculate the total feature weight of user "U9" for household goods B: purchase count (0.9) + valid view count (0.3) + purchase-view ratio (0.7) = 1.9. Since it does not exceed 2.0, take 1.9 as the "total weight of household goods B"; Solve the problem of "excessive weight of single product category monopolizing labels". To address the issue of excessive label skew in a particular category due to the superposition of multiple features (e.g., industrial product X with a weight of 2.2), which could squeeze the label generation space for other categories (e.g., if a user has needs for both industrial product X and household goods B, avoid generating only industrial product labels), this approach balances the opportunities for generating labels across multiple categories, ensuring that users' multi-dimensional needs can be identified. Furthermore, the weight cap is set reasonably (2.0), avoiding both weight overflow and suppressing the total weight of normal categories (e.g., 1.9). At the same time, it ensures comprehensive label coverage, aligning with the actual scenario of users consuming / purchasing multiple categories.

[0050] As a specific embodiment of the present invention: the specific setting of the scene threshold in S4 is as follows: the tag trigger threshold for the daily necessities scene is 1.2, and the tag trigger threshold for the industrial products scene is 1.5; if the total weight of a certain category is ≥ the tag trigger threshold of the corresponding scene, then the basic candidate tag is triggered; if the weight of the associated feature of a certain category is ≥ 0.6, then the subdivision candidate tag is triggered.

[0051] Working process: User "U10" for daily necessities A: total weight = number of purchases (0.9) + purchase-view ratio (0.7) = 1.6 ≥ 1.2 (threshold for daily necessities), triggering the basic candidate label "high-frequency consumer of daily necessities A"; User "U11" for industrial products Y: total weight = number of purchases (0.9) + number of valid favorites (0.6) = 1.5 ≥ 1.5 (threshold for industrial products), triggering the basic candidate label "stable buyer of industrial products Y"; User "U12" for daily necessities C: associated feature (favorites-view ratio) weight = 0.7 ≥ 0.6, triggering the segmented candidate label "high-intent hesitant user of daily necessities C"; To address the issue of existing technologies having a "one-size-fits-all threshold that doesn't fit specific scenarios," this invention sets differentiated thresholds for daily necessities (high frequency, threshold 1.2) and industrial products (longer cycle, threshold 1.5). It also triggers subdivided tags through a related feature threshold (0.6), avoiding the problem of basic tags failing to cover specific needs. Furthermore, the scenario-based thresholds adapt to the user behavior patterns of both types of goods, improving the accuracy of basic tag triggering and allowing subdivided tag triggering rules to uncover implicit needs such as "high intent but no conversion," resulting in finer tag granularity. This also meets the needs of different business scenarios (e.g., daily necessities require high-frequency tags, while industrial products require stable procurement tags).

[0052] As a specific embodiment of the present invention: the specific rules for the confidence calculation and verification in S4 are as follows: confidence = (number of valid features that trigger the label / total number of features required for the label) × 100%; wherein, the total number of features required for the basic candidate label is 2, and the confidence level must be ≥ 60%; the total number of features required for the subdivided candidate label is 3, and the confidence level must be ≥ 70%.

[0053] Work process: "Industrial product Z high-intent hesitant user" (segmentation tag) requires 3 effective features (effective browsing, effective collection, collection-browsing ratio). If only effective browsing and collection-browsing ratio are triggered, the confidence level = (2 / 3) × 100% ≈ 66.7% < 70%, and the candidate tag is discarded; "Daily necessities A high-frequency consumer" (basic tag) requires 2 effective features (purchase, purchase-browsing ratio). If both are triggered, the confidence level = (2 / 2) × 100% = 100% ≥ 60%, and it is entered into the database as a formal tag; Solving existing technical problems. The problem of "low-quality tags being included in the database without confidence verification" is addressed by using dual verification of feature quantity and confidence threshold to filter out candidate tags with "incomplete features and low confidence" (such as 66.7% of sub-tags), ensuring the reliability of the tags included in the database. This improves the quality of the final tags and avoids low-confidence tags misleading business decisions (such as mistakenly pushing high-intent products to low-confidence users). Furthermore, the differentiated confidence requirements for basic and sub-tags align with the demand intensity of the two types of tags (sub-tags require higher confidence). At the same time, it forms a closed loop for tag quality control, ensuring the effectiveness of tag application.

[0054] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatically generating user tags based on multi-scene data, characterized in that, Comprise the following steps: S1, multi-scene data access and preprocessing: access data sources according to "user-category" dimensions, the data sources include the purchase times, the browsing times, the collection times of the life supplies category A / B / C and the industrial goods category X / Y / Z respectively corresponding; The data source is processed by missing value filling, outlier rejection and format unification; S2, effective times statistics: for the data processed in S1, the browsing times de-redundancy statistics, the collection times de-redundancy statistics and the purchase times statistics are carried out respectively, wherein the browsing times and the collection times are de-redundant according to the preset rule, and the purchase times are counted according to the actual order times; S3, data correlation feature identification: extract the purchase times, the effective browsing times and the effective collection times of each category as single features, calculate the collection-browsing ratio and the purchase-browsing ratio of each category as correlation features, and assign weights to each feature based on mutual information; S4, automatic generation of user label: according to the total weight of the features of each category and the scene threshold, the candidate label is triggered, the confidence of the candidate label is calculated and verified, and the candidate label meeting the confidence requirement is taken as the formal user label and stored, and the candidate label not meeting the confidence requirement is discarded. 2.The method of claim 1, wherein, The specific rules of the missing value filling in S1 are: if a user does not produce a certain behavior, the corresponding times are filled with 0; if the data of a certain date is missing, the average times of the corresponding category of the user in the last 7 days are filled, and the filling error is allowed to be ±10%. 3.The method of claim 1, wherein, The specific rules of the outlier rejection in S1 are: taking 3 times of the historical behavior times of the corresponding category of the user as the threshold, the abnormal times exceeding the threshold are rejected. 4.The method of claim 1, wherein, The specific rules of the format unification in S1 are: the data is summarized according to "user ID-category ID-statistical period", wherein the statistical period of the life supplies category A / B / C is set to "day", and the statistical period of the industrial goods category X / Y / Z is set to "month". 5.The method of claim 1, wherein, The specific rules of the browsing times de-redundancy statistics in S2 are: the times of the same user browsing the same category within the statistical period are all recorded as 1 effective browsing time; wherein the browsing statistical period of the life supplies category A / B / C is 1 day, and the browsing statistical period of the industrial goods category X / Y / Z is 3 days. 6.The method of claim 1, wherein, The specific rules of the collection times de-redundancy statistics in S2 are: the collection times of the same user to the same category are all recorded as 1 effective collection time; if a user cancels the collection of a category and then collects it again, the interval is <7 days, which is not repeated, and the interval is ≥7 days, which is recorded as 1 effective collection time. 7.The method of claim 1, wherein, The feature weight assignment based on mutual information in S3 is: the weight of the purchase times is 0.9, the weight of the effective collection times is 0.6, the weight of the effective browsing times is 0.3, and the weight of the correlation features (collection-browsing ratio, purchase-browsing ratio) is 0.

7. 8.The method of claim 7, wherein, In S3, under the same category, if the weight sum of the purchase times, the effective browsing times, the effective collection times and the correlation features exceeds 2.0, the total weight of the category is taken as 2.

0. 9.The method of claim 1, wherein, The specific setting of the scene threshold in S4 is that the label triggering threshold of the life supplies scene is 1.2, and the label triggering threshold of the industrial product scene is 1.5; if the total weight of a category is greater than or equal to the label triggering threshold of the corresponding scene, the basic candidate label is triggered; if the associated feature weight of a category is greater than or equal to 0.6, the subdivided candidate label is triggered. 10.The method of claim 1, wherein, The specific rules of the confidence degree calculation and verification in S4 are that the confidence degree=(the number of effective features triggering the label / the total number of features required by the label)×100%; wherein the total number of features required by the basic candidate label is 2, and the confidence degree needs to be greater than or equal to 60%; the total number of features required by the subdivided candidate label is 3, and the confidence degree needs to be greater than or equal to 70%.

Images