A method and system for mining features of second-hand goods
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING ZHUANZHUAN SPIRIT TECH CO LTD
- Filing Date
- 2021-11-26
- Publication Date
- 2026-06-05
Smart Images
Figure CN116188028B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Internet applications, and in particular to a method and system for mining the characteristics of second-hand goods based on commodity quality inspection reports. Background Technology
[0002] With the development of internet applications and the logistics industry, more and more merchants and consumers tend to complete transactions on e-commerce platforms. Among these, secondhand e-commerce platforms provide a convenient place for people to dispose of their unwanted items. Due to varying usage conditions, the condition of secondhand goods differs greatly. For example, even mobile phones of the same brand and model may differ significantly: one may be unopened and unused, while the other's casing may be damaged due to use. Using only common product characteristics such as brand and model number is clearly insufficient to distinguish between these two items. Therefore, to more accurately recommend products to users, a more detailed feature mining of secondhand goods is needed. Summary of the Invention
[0003] In view of the technical problems existing in the prior art, the present invention proposes a method and system for mining features of second-hand goods, which can be used to obtain valuable product features from quality inspection reports to enrich the description of second-hand goods.
[0004] To address the aforementioned technical problems, according to one aspect of the present invention, a method for mining features of second-hand goods is provided, comprising the following steps: reading a product quality inspection report and obtaining text information therefrom; extracting one or more undetermined product features from the text information; adding the undetermined product features to a first information feature set of a functional model to form a second information feature set; the functional model obtaining a first predicted evaluation value and a second predicted evaluation value respectively using the first information feature set and the second information feature set as inputs; comparing the first predicted evaluation value and the second predicted evaluation value; and confirming that the undetermined product features are usable in response to the second predicted evaluation value being greater than the first predicted evaluation value.
[0005] According to another aspect of the present invention, the present invention also provides a second-hand goods feature mining system, including a text acquisition module, a feature extraction module, an input feature determination module, a model evaluation module, and a determination module, wherein the text acquisition module is configured to read a goods quality inspection report and obtain text information therefrom; the feature extraction module is connected to the text acquisition module and is configured to extract one or more undetermined goods features from the text information; the input feature determination module is connected to the feature extraction module and is configured to add the undetermined goods features to a first information feature set of a functional model to form a second information feature set; the model evaluation module is connected to the input feature determination module and is configured to use the first information feature set and the second information feature set as inputs to the functional model, respectively, to obtain a first predicted evaluation value and a second predicted evaluation value; the determination module is connected to the model evaluation module and is configured to compare the first predicted evaluation value and the second predicted evaluation value, and confirm that the undetermined goods feature is usable when the second predicted evaluation value is greater than the first predicted evaluation value.
[0006] This invention extracts usable and valid product features from product quality inspection reports, thereby refining the description of second-hand goods. This enables downstream storage services such as search and recommendation to better distinguish second-hand goods and provide more accurate recall products. Attached Figure Description
[0007] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings, wherein:
[0008] Figure 1 This is a flowchart of a second-hand goods feature mining method according to an embodiment of the present invention;
[0009] Figure 2 This is a flowchart of extracting the characteristics of a commodity to be determined from the contents of quality inspection items according to an embodiment of the present invention;
[0010] Figure 3 This is a flowchart illustrating the extraction of features of a product from description information according to an embodiment of the present invention;
[0011] Figure 4 This is a flowchart illustrating the process of obtaining two predicted evaluation values using an online model, according to one embodiment of the present invention.
[0012] Figure 5 This is a flowchart illustrating the process of training a model and obtaining two prediction evaluation values according to one embodiment of the present invention.
[0013] Figure 6 This is a block diagram illustrating the principle of a second-hand goods feature mining system according to an embodiment of the present invention.
[0014] Figure 7 This is a block diagram illustrating the principle of a text acquisition module according to an embodiment of the present invention;
[0015] Figure 8 This is a block diagram illustrating the principle of a feature extraction module according to an embodiment of the present invention;
[0016] Figure 9 This is a block diagram illustrating the principle of a secondhand goods feature mining system according to another embodiment of the present invention; and
[0017] Figure 10 This is a block diagram of a second-hand goods feature mining system according to another embodiment of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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.
[0019] In the following detailed description, reference can be made to the accompanying drawings, which form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Specific embodiments of the present application are described in sufficient detail below to enable those skilled in the art to implement the technical solutions of the present application. It should be understood that other embodiments may also be utilized, or structural, logical, or electrical changes may be made to the embodiments of the present application.
[0020] In this invention, to obtain detailed information about secondhand goods and increase user trust in the products, some secondhand e-commerce platforms conduct quality inspections on the secondhand goods sold on their platforms and issue quality inspection reports based on the results. The inspection items vary depending on the product and its category. For example, for computers, the inspection items include multiple major items and sub-items within those major items. Major items may include CPU testing, hard drive testing, memory testing, screen testing, and appearance testing, while sub-items may include "CPU series" and "CPU model" in the CPU testing item, and "hard drive 1," "hard drive 2," etc., in the hard drive testing item. In one embodiment, when generating a quality inspection report, the system provides multiple options for each inspection item. For example, the "CPU" series may include Intel's "Ryzen," "Pentium," and "Celeron" series, and AMD's "Ryzen," "Athlon," and "Sempron" series. "CPU models" may include "i9-10900K," "i7-9700KF," and "R9-3900X," etc. When generating the quality inspection report, the system selects the corresponding content from the quality inspection items based on the product. If no corresponding option is available, authorized personnel can add content to the quality inspection item options based on the actual situation of the product. Therefore, in the resulting quality inspection report, each quality inspection item corresponds to one quality inspection content. Since quality inspection items cannot cover all possible product conditions, quality inspection reports usually include explanatory information to explain situations not covered in the inspection items. For example, a mobile phone quality inspection report might include the following explanatory information: "Upon inspection, the device's back cover is detached from its adhesive, there is sealant preventing the detection of internal disassembly / repair liquid immersion, the front camera glass is cracked, the screen has dents, and SIM card 1 exhibits intermittent card reading problems." Analysis of the various sections of the above quality inspection report shows that the information included in the report provides a more detailed description of the product, making the inspected product more easily identifiable among similar products. Therefore, this invention provides a product feature mining method and system that extracts usable and effective product features from quality inspection reports issued for secondhand goods, thereby enriching the features of secondhand goods.
[0021] Figure 1 This is a flowchart of a secondhand goods feature mining method according to an embodiment of the present invention. In this embodiment, the goods feature mining method includes the following steps:
[0022] Step S1: Read the product quality inspection report and obtain its text information. In one embodiment, the product quality inspection report corresponding to the product is stored in a database. The product quality inspection report has an independent identifier, or it is identified using the product identifier corresponding to the product. The product quality inspection report is read from the database, including the quality inspection items, their specific contents, and any additional explanatory information.
[0023] Step S2: Extract one or more product features to be determined from the text information. Since the text information includes quality inspection items and additional explanatory information, product features to be determined are extracted from both the quality inspection items and the additional explanatory information.
[0024] In one embodiment, such as Figure 2 The diagram shows a flowchart for extracting the characteristics of goods to be determined from the quality inspection items. The process includes the following steps:
[0025] Step S21a: Extract the names of the quality inspection items, such as "CPU model", "color", "material", "size", etc.
[0026] Step S22a: Check if there is already a product feature with the same name for this type of product. If there is, discard it in step S23a. If not, record it as a pending product feature in step S24a and save the content corresponding to the quality inspection item as the feature value.
[0027] In another embodiment, such as Figure 3 The diagram shows a flowchart for extracting features of a product from its description. The process includes the following steps:
[0028] Step S21b involves segmenting the explanatory information into words and calculating word frequencies to determine the first vocabulary with a frequency greater than a threshold. For example, using a natural language semantic model, the explanatory information is segmented and labeled to obtain multiple words. Then, the frequency of each word is calculated. The word frequency refers to the number of times a word appears in the file divided by the total number of words in the file. In this embodiment, it refers to the number of times a word appears in all quality inspection reports divided by the number of quality inspection reports. Then, high-frequency words are determined based on the word frequency threshold. To distinguish words with different functions, high-frequency words are referred to as the first vocabulary in this embodiment.
[0029] Step S22b: Extract the context information of the first word. The context information may include, for example, a complete sentence containing the first word, or a complete sentence containing the first word plus the two sentences preceding and following it.
[0030] Step S23b: Extract one or more second words from the context information; for example, extract two similar words from a complete sentence that includes the first word as the second word.
[0031] Step S24b: Combine the second word with the first word into a phrase.
[0032] Step S25b: Take a phrase and count the number of times the second word and the first word of the phrase co-occur, that is, count the number of times it appears in the quality inspection report.
[0033] Step S26b: Compare the co-occurrence counts of the second word and the first word with the threshold value, and determine whether the co-occurrence counts of the second word and the first word are greater than or equal to the threshold. If the co-occurrence counts of the second word and the first word that make up the word group are greater than or equal to the threshold, then in step S27b, it is determined that the second word and the first word constitute a pending product feature; otherwise, in step S28b, it is determined that the word group cannot be used as a pending product feature.
[0034] Step S29b: Determine if there are any other phrases that include the first word. If so, return to step S25b; otherwise, end the mining based on the first word.
[0035] The phrase consisting of the first and second words obtained by the aforementioned method can be used as a product feature, such as "screen scratches" or "dust in the camera", and its feature value can be set to "yes / no" or "no / no".
[0036] In another embodiment, when a first word and multiple second words constitute multiple pending product features, the first word can be used as the pending product feature, and the second words as feature values. For example, when obtaining "screen broken," "screen scratched," "screen shattered," etc., where "screen" is the first word, and other words such as "broken," "scratched," and "shattered" are the second words, then "screen" can be used as the pending product feature, and "broken," "scratched," and "shattered" can be used as feature values. Of course, "screen broken," "screen scratched," and "screen shattered" can also be used as a single product feature, with "present / yes" or "absent / no" as the feature value.
[0037] After identifying potential product features from the quality inspection report, they need to be verified to determine their usability. Since product features are used in downstream services such as recommendation and search, which employ machine learning models to predict product recall based on product and user features, this invention applies the mined product features to the models used in these services. The usability of the product feature is determined by comparing the model performance before and after its use. Specifically, the following steps are included:
[0038] Step S3: The undetermined product feature is added to the first information feature set of the functional model to form a second information feature set. The functional model can be any one or more models used in various services, such as a click-through rate model, conversion rate model, etc. The model uses the feature set as input to make predictions to obtain predicted values. The features include, for example, user features, product features, and scenario features, etc. For distinction, this invention refers to the original feature set of the model as the first information feature set, and the information feature set formed by adding an undetermined product feature to the first information feature set as the second information feature set.
[0039] Step S4: The functional model obtains a first prediction evaluation value and a second prediction evaluation value by using the first information feature set and the second information feature set as inputs, respectively. In this step, a pre-trained model, such as one currently being used online, can be used, or a model specifically designed for use can be trained. In one embodiment, such as... Figure 4 As shown, the process of obtaining two predicted evaluation values using the model currently being applied online is as follows:
[0040] Step S41a: Obtain the sample set of the online model to form the first test set. The sample set includes multiple first information feature sets used by the online model for prediction. Each first information feature set is a sample. The sample includes positive samples and negative samples. For example, for the click-through rate model, the positive sample is the set of all features corresponding to when the user clicks at a specific location (such as the homepage), and the negative sample is the set of all features corresponding to when the user does not click at a specific location (such as the homepage).
[0041] Step S42a: Add undetermined product features to the first information feature set corresponding to each sample in the sample set to form a second information feature set, thereby obtaining the second test set.
[0042] Step S43a: Input the corresponding samples from the two test sets into the online model to obtain the corresponding first and second predicted values.
[0043] Step S44a: According to the selected evaluation parameters, the first predicted evaluation value and the second predicted evaluation value are calculated based on the first predicted value and the second predicted value, respectively. In one embodiment, the evaluation parameter can be AUC (Area Under Curve), which refers to the probability that the model outputs a positive sample as positive and a negative sample as positive when randomly selecting a positive sample and a negative sample, respectively. For this invention, if the model predicts using a feature set that includes the mined features of the undetermined product, and its AUC value (hereinafter referred to as the second AUC value) is larger than the AUC value when predicting using a feature set that does not include the mined features of the undetermined product (hereinafter referred to as the first AUC value), it indicates that the mined features of the undetermined product can improve the model performance. The greater the difference between the second AUC value and the first AUC value, the greater the improvement of the model performance by the mined features of the undetermined product, thus determining that the mined features of the undetermined product are more useful. In one embodiment, the AUC value can be calculated using Formula 1-1:
[0044] in,
[0045] Among them, (P positive, P 负) represents the predicted value of a pair of positive and negative samples, where M is the number of positive samples and N is the number of negative samples.
[0046] In another embodiment, AUC can be calculated using formula 1-2:
[0047]
[0048] Where, r j The positive samples are sorted, where M is the number of positive samples and N is the number of negative samples.
[0049] GAUC can also be used as an evaluation parameter. GAUC is the weighted average of the AUC values when making predictions for a single user. Its weight can be set to the number of impressions or the number of clicks, depending on the model. The number of impressions is the number of products viewed by the user, and the number of clicks is the number of products clicked by the user.
[0050] In another embodiment, such as Figure 5 As shown, a model is first trained, and then the model's prediction evaluation value is calculated. The specific steps include:
[0051] Step S41b: Obtain the necessary data to form the first sample set. For example, use all data from the past sixty days to the present as the sample set. Of course, depending on the function and category of the model, determine the feature values and labels of each sample used by the model based on the obtained data, such as click or no click, conversion or no conversion, etc.
[0052] Step S42b: Divide the sample set into a first training set and a first test set. For example, use all data from the past sixty days to the previous three days as the training set, and use the data from the past three days as the test set.
[0053] Step S43b: Train the model using the first training set samples, and obtain the first model after training.
[0054] Step S44b: Using the first test set samples, input them into the trained model to obtain a first prediction evaluation value. For example, first input the samples in the first test set into the model to obtain multiple first prediction values, and then calculate the first prediction evaluation value of the model based on the first prediction values and according to the evaluation parameters.
[0055] Step S45b: Add the mined features of the undetermined goods to the samples of the first training set and the first test set to form the second training set and the second test set.
[0056] Step S46b: Train the model using the second training set samples, and obtain the second model after training.
[0057] Step S47b: Input the second test set samples into the trained model to obtain the second prediction evaluation value.
[0058] Of course, during the aforementioned model training process, it is necessary to determine whether the trained model is usable. For example, after training, the model can be evaluated using its own test set. The AUC value can be calculated based on the predicted values obtained from the test set and compared with a threshold, such as 0.5. If the AUC value is greater than or equal to the threshold, the model is considered usable. Other metrics can also be used to evaluate model usability, such as recall, accuracy, precision, and F1 score. Once the model is determined to be usable, the AUC value obtained from the test set samples is used as the prediction evaluation value required in this invention.
[0059] Step S5: Compare the first predicted evaluation value and the second predicted evaluation value.
[0060] Step S6: Determine whether the second predicted evaluation value is greater than the first predicted evaluation value. If the second predicted evaluation value is greater than the first predicted evaluation value, then in step S7, confirm that the pending product feature is available. If the second predicted evaluation value is less than the first predicted evaluation value, then in step S8, confirm that the pending product feature is unavailable and discard the pending product feature.
[0061] Through the above embodiments, after extracting product features from the quality inspection report, a model is used to verify them to determine whether the product features can improve the model's performance. If so, it indicates that the extracted product features are usable.
[0062] In another embodiment, the available undetermined product features are further added to the input feature set of the online functional model for online A / B testing. For example, the online model using the original features is called the base model, or version A functional model, and the model with the added features is called the exp model, or version B functional model. Both are run online with 50% traffic. If the click-through rate / conversion rate of the exp model is higher than that of the base model, the features are considered effective. In other words, this embodiment further verifies the effectiveness of the mined features through actual use.
[0063] In the foregoing embodiments, the trained model is, for example, a logistic regression model or a wide & deep model, etc., and depending on its function, it can be a click-through rate model or a conversion rate model, etc.
[0064] In one embodiment, when the phrase "camera dust" was detected, an AUC value greater than 0.55 based on a logistic regression model indicates that the model trained with this feature is usable. When this feature was added to the online model, the AUC value obtained with the feature increased by more than three thousandths compared to the AUC value obtained without the feature, confirming that the feature met the deployment criteria. After deploying this feature, the user order completion rate (number of orders / number of users) increased by 2.41%. Analysis shows that since camera dust has a relatively small impact on photo quality but offers a relatively low price, the "camera dust" feature helps the model better identify purchase intentions.
[0065] Figure 6 This is a block diagram of a second-hand commodity feature mining system according to an embodiment of the present invention. The system includes a text acquisition module 1, a feature extraction module 2, an input feature determination module 3, a model evaluation module 4, and a determination module 5. The text acquisition module 1 is used to read commodity quality inspection reports from the database and obtain text information from them.
[0066] In one embodiment, such as Figure 7 As shown, the text acquisition module 1 includes a quality inspection item information extraction unit 11 and a description information extraction unit 12. The quality inspection item information extraction unit 11 extracts multiple quality inspection item information from the product quality inspection report as text information for extracting product characteristics, such as the quality inspection items and corresponding quality inspection content in the quality inspection report, such as reading the quality inspection item "CPU model" and its detected content "9-10900K", "body color" and its corresponding content "silver gray", etc. The description information extraction unit 12 extracts additional description information from the product quality inspection report. For example, the description text in the additional document column of the quality inspection report is "Upon inspection, the back cover of the device is delaminated, there is sealant that cannot detect internal disassembly and repair liquid immersion, the front camera glass is cracked, the screen has dents, and SIM card 1 has intermittent card reading problems. It is recommended that buyers purchase according to their actual needs. For detailed test results, please refer to the quality inspection report."
[0067] The feature extraction module 2 is connected to the text acquisition module 1 and is used to extract one or more features of the product to be determined from the text information. In one embodiment, such as... Figure 8As shown, the feature extraction module 2 includes a first feature extraction unit 21 and a second feature extraction unit 22. The first feature extraction unit 21 is used to extract the features of the pending product from the quality inspection item information in the product quality inspection report. For example, the field name of the quality inspection item is extracted as a feature, and the corresponding content of the quality inspection item is used as the feature value. For example, "CPU model" is used as a feature, and the corresponding content "9-10900K" is used as the feature value. Then, the product feature database is queried to check if the same feature already exists. If it does, the feature and feature value are discarded. If the product feature database does not contain the feature or the feature value, it is treated as a pending product feature. The second feature extraction unit 22 is used to extract the features of the pending product from the description information in the product quality inspection report. The second feature extraction unit 22 further includes a first vocabulary unit 221, a context information unit 222, a second vocabulary unit 223, and a feature determination unit 334. The first vocabulary unit 221 segments the description information, counts word frequencies, and determines the first vocabulary with a word frequency greater than a threshold. For example, a trained natural language semantic model is used. The explanatory information is input into the model, and after processing, multiple word segments are obtained. Then, the frequency of each word segment is counted in all quality inspection reports. The frequency of each word segment is divided by the number of quality inspection reports to obtain its word frequency. High-frequency words are then determined based on a word frequency threshold. In this embodiment, high-frequency words are determined as the first vocabulary. The context information unit 222 is connected to the first vocabulary unit 221 and extracts the context information of the first vocabulary from the explanatory information. For example, it extracts the entire sentence containing the first vocabulary, or the entire sentence containing the first vocabulary along with its preceding and following sentences. The second vocabulary unit 223 is connected to the context information unit 222 and extracts one or more second vocabulary words from the context information. For example, the two most similar content words before and after the first vocabulary are used as the second vocabulary words. The feature determination unit 334 is connected to the first vocabulary unit 221 and the second vocabulary unit 223 respectively, and is configured to count the co-occurrence frequency of each second vocabulary and the first vocabulary; when the co-occurrence frequency of the second vocabulary and the first vocabulary is greater than a threshold, it is determined that the second vocabulary and the first vocabulary constitute a pending product feature. In another embodiment, when multiple second vocabularys constitute multiple pending product features with the same first vocabulary, the feature determination unit 224 determines the first vocabulary as a pending product feature and the second vocabulary as a feature value.
[0068] The input feature determination module 3 is connected to the feature extraction module 2 and is used to add the undetermined product feature to the first information feature set of the functional model to form a second information feature set. Different functional models have corresponding feature sets to complete the prediction and calculation for their respective functions. In this invention, the feature set originally used by the functional model is called the first information feature set. The input feature determination module 3 adds a mined undetermined feature to the first information feature set to form the second information feature set. Based on the model used by the model evaluation module 4, its corresponding sample set is obtained. Each sample in the sample set consists of a series of features, called a feature set. If the currently used model is an online model, the feature set currently used to constitute the sample is called the first information feature set. A certain number of samples from the online model are obtained to form a first test set. The undetermined product feature is added to the first information feature set corresponding to the samples in the first test set to form the second information feature set, thus obtaining the second test set.
[0069] The model evaluation module 4 is connected to the input feature determination module 3. It takes the first information feature set and the second information feature set as inputs to the functional model to obtain a first predicted evaluation value and a second predicted evaluation value. Specifically, samples from the first test set are input to the online model to obtain multiple predicted values. According to selected evaluation parameters, such as AUC values, the AUC value is calculated based on the predicted values and recorded as the first AUC value. The calculation process is described in the method description and will not be repeated here. The model evaluation module 4 inputs samples from the second test set to the online model to obtain multiple predicted values. According to selected evaluation parameters, such as AUC values, the AUC value is calculated based on the predicted values and recorded as the second AUC value.
[0070] The determination module 5 is connected to the model evaluation module 4 and is used to compare the first predicted evaluation value and the second predicted evaluation value. When the second predicted evaluation value is greater than the first predicted evaluation value, the undetermined product feature is confirmed to be usable. For example, when the second AUC value is greater than the first AUC value, it can be determined that the currently evaluated undetermined product feature can improve the model's performance, and it is determined to be a usable product feature.
[0071] Figure 9 This is a block diagram illustrating the principle of a secondhand goods feature mining system according to another embodiment of the present invention. In this embodiment, in addition to including... Figure 6In addition to the text acquisition module 1, feature extraction module 2, input feature determination module 3, model evaluation module 4, and determination module 5 in the illustrated embodiment, it also includes a sample set module 6 and a model training module 7. In this embodiment, the model training module 7 is used for model training. The sample set module 6 acquires corresponding sample data according to the needs of model training. In one embodiment, the sample set module 6 processes the corresponding data into samples composed of feature sets according to the features described in the model, and divides them into two sample sets. The input feature determination module 3 adds the features of the product to be determined to one of the sample sets (called the second sample set), and divides the sample set into a training set and a test set, thus obtaining the first training set and the first test set in the first sample set, and the second training set and the first test set in the second sample set. The model training module 7 trains the first functional model using the first training set samples and the second functional model using the training set samples. The model then undergoes testing on its respective test set, evaluation according to evaluation parameters, adjustment of model hyperparameters, and retraining until the trained model meets the requirements. After model training is complete, model training module 7 sends a notification to model evaluation module 4, which tests the first functional model using samples from the first test set in the first sample set to obtain a first predicted evaluation value, and tests the second functional model using samples from the second test set in the second sample set to obtain a second predicted evaluation value. Determination module 5 compares the first and second predicted evaluation values, and confirms the availability of the undetermined product feature if the second predicted evaluation value is greater than the first predicted evaluation value.
[0072] Figure 10This is a schematic diagram of a second-hand goods feature mining system according to another embodiment of the present invention. In this embodiment, based on the aforementioned embodiment, an AB experiment module 8 is further included, which includes a feature addition unit 81, a functional model feedback unit 82, and an effectiveness judgment module 83. The feature addition unit 81 is connected to the determination module 5, and adds available goods features to the input feature set of the online B-version functional model to form a B-version feature set. The online functional model is divided into two versions, A and B, each accounting for 50% of the traffic. The B-version functional model uses the B-version feature set to complete its function, and the A-version functional model uses the A-version feature set to complete its function. The effect analysis system in the platform analyzes the effects of these two versions of the functional model to obtain their respective functional effect data. For example, for a model predicting user order rate, the functional effect data can be the user order completion rate (number of orders / number of users). The functional model feedback unit 82 obtains the first functional effect data of online functional model version A and online functional model version B from the effect analysis system, and sends them to the validity judgment module 83. The validity judgment module 83 compares the first functional effect data and the second functional effect data; if the second functional effect data is better than the first functional effect data, the available pending product feature is determined to be valid. For example, when the feature "camera dust" is added to the feature set of version B, the user order completion rate obtained after running the functional model of version B increases by 2.41% compared with the user order completion rate obtained after running the functional model of version A without adding this feature. That is, this feature can help the model better identify purchase intentions, and therefore it is a very effective product feature.
[0073] After discovering useful and effective product features, this invention can add them to the quality inspection items, thereby improving the comprehensiveness of the quality inspection. Furthermore, by adding the discovered useful and effective product features to the product feature database for application in online models, the functionality of each model can be effectively improved, better serving users of the second-hand e-commerce platform and providing a better user experience.
[0074] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the scope of the invention. Therefore, all equivalent technical solutions should also fall within the scope of the invention.
Claims
1. A method for feature mining of second-hand goods, comprising: Read the product quality inspection report, obtain the quality inspection item information and description information, and extract multiple characteristics of the products to be determined; The step of extracting the characteristics of pending commodities from the quality inspection item information includes: extracting the name of the quality inspection item; querying whether there is already a commodity feature with the same name for the corresponding commodity; if there is, discarding it; if not, recording it as a characteristic of pending commodities, and using the content detected in the quality inspection item as the feature value. The steps for extracting features of products to be determined from the description information include: segmenting the description information into words, counting word frequencies, and determining a first word whose word frequency is greater than a threshold; extracting context information of the first word; extracting one or more second words from the context information; counting the number of times each second word co-occurs with the first word; and when the number of times the second word co-occurs with the first word is greater than a threshold, determining that the second word and the first word constitute features of products to be determined, and when the first word constitutes multiple features of products to be determined with multiple second words respectively, using the first word as the feature of products to be determined and the second word as the feature value. The features of the undetermined goods are added to the first set of information features of the functional model to form the second set of information features; The functional model is trained, and the steps include: A first set of information features is obtained to form a first sample set, and the feature value and label of each sample used by the model are determined; the first sample set is divided into a first training set and a first test set; the first model is obtained by training with the first training set. A second set of information features is obtained to form a second sample set, which is then divided into a second training set and a second test set. The second model is obtained by training with the second training set. Input the first / second test sets into the first / second model respectively to obtain the first / second prediction evaluation values; compare the first prediction evaluation value and the second prediction evaluation value; and If the second predicted evaluation value is greater than the first predicted evaluation value, the undetermined product feature is confirmed to be available.
2. A second-hand goods feature mining system, comprising: The text acquisition module is configured to read product quality inspection reports and extract text information from them. A feature extraction module, which is connected to the text acquisition module, is configured to extract multiple features of the products to be determined from the text information; An input feature determination module, which is connected to the feature extraction module, is configured to add the features of the product to be determined to the first information feature set of the functional model to form a second information feature set; The model evaluation module is connected to the input feature determination module and is configured to take the first information feature set and the second information feature set as inputs to the functional model, respectively, to obtain a first prediction evaluation value and a second prediction evaluation value. as well as A determination module, which is connected to the model evaluation module, is configured to compare the first predicted evaluation value and the second predicted evaluation value, and confirm that the feature of the product to be determined is available when the second predicted evaluation value is greater than the first predicted evaluation value. The text acquisition module includes a quality inspection item information extraction unit, configured to extract multiple quality inspection item information from the product quality inspection report as text information for extracting product features; wherein the feature extraction module includes a first feature extraction unit, configured to extract the undetermined product features from the quality inspection item information in the product quality inspection report; The text acquisition module includes a description information extraction unit, configured to extract additional description information from the product quality inspection report; the feature extraction module includes a second feature extraction unit, configured to extract the features of the product to be determined from the description information in the product quality inspection report. The second feature extraction unit further includes: a first vocabulary unit, configured to segment the description information, count word frequencies, and determine first vocabulary words with a frequency greater than a threshold; a context information unit, connected to the first vocabulary unit, configured to extract context information of the first vocabulary words from the description information; a second vocabulary unit, connected to the context information unit, configured to extract one or more second vocabulary words from the context information; and a feature determination unit, connected to the first vocabulary unit and the second vocabulary unit, configured to count the co-occurrence frequency of each second vocabulary word with the first vocabulary word; and when the co-occurrence frequency of the second vocabulary word with the first vocabulary word is greater than a threshold, determining that the second vocabulary word and the first vocabulary word constitute a pending product feature; when the first vocabulary word constitutes multiple pending product features with multiple second vocabulary words respectively, the feature determination unit uses the first vocabulary word as the pending product feature and the second vocabulary word as the feature value; A sample set module, connected to the input feature determination module, is configured to provide a first sample set and a second sample set for the functional model. The first / second sample set includes corresponding training and test sets. Samples in the first sample set are composed of a first set of information features, and samples in the second sample set are composed of a second set of information features. A model training module, which is connected to the sample set module, is configured to train a first functional model and a second functional model respectively based on training set samples from the first sample set and the second sample set. Correspondingly, the model evaluation module is connected to the model training module, and the configuration is to test the first functional model with the first test set samples in the first sample set to obtain the first predicted evaluation value, and to test the second functional model with the second test set samples in the second sample set to obtain the second predicted evaluation value.