A feature extraction method and device

Abstract

The present invention discloses a feature extraction method and device, and relates to the field of computer technology. A specific implementation of the method includes: determining the item category in response to the input of the item category range, obtaining the basic features corresponding to the item category, and determining the positive example items and the positive example feature vector corresponding to the item category; using a generator to sample basic features and combine them to generate a combined item, combining the feature values ​​of each basic feature to generate a feature vector of the combined item; inputting the feature vector into a discriminator to obtain a discrimination result; if the discrimination result is a preset value, inputting the feature vector and the positive example feature vector into an objective function to obtain a target parameter, and if the target parameter meets the preset conditions, the sampled basic feature is determined as the target feature. This implementation is based on basic feature sampling and feature vector discrimination to quickly capture and extract the optimal basic features that best represent the most competitive characteristics at the moment.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a feature extraction method and device. Background Art

[0002] Predicting competitive trends for market segments (consumers who will respond immediately to a given marketing campaign) is the basis for market item design decisions and an important issue in the field of market demand forecasting. Its main purpose is to predict the needs of specific user groups in the market, thereby capturing the most competitive item features and avoiding vicious competition (such as price wars) caused by similar items.

[0003] In the process of implementing the present invention, the inventors found that the prior art has at least the following problems:

[0004] There are a wide variety of existing items. The basic characteristics of items in each category are different, and the basic characteristics of each item may change dynamically, which makes it difficult to discover the representative optimal characteristics of each category. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a feature extraction method and device, which can at least solve the problem that optimal features are difficult to mine in the prior art.

[0006] To achieve the above object, according to one aspect of an embodiment of the present invention, a feature extraction method is provided, comprising:

[0007] In response to the input of the item category range, determine the item category, obtain basic features corresponding to the item category, and determine positive example items and positive example feature vectors corresponding to the item category;

[0008] Using a generator to sample basic features and combine them to generate a combined object, and combining the feature values ​​of each basic feature to generate a feature vector of the combined object;

[0009] Inputting the feature vector into a discriminator to obtain a discrimination result; wherein the generator and the discriminator are located in a generative adversarial network;

[0010] If the discrimination result is a preset value, the feature vector and the positive example feature vector are input into an objective function to obtain a target parameter. If the target parameter meets a preset condition, the sampled basic feature is determined as the target feature.

[0011] Optionally, if the target parameter meets a preset condition, determining the sampled basic feature as the target feature includes:

[0012] Determine whether the target parameter is within a preset value range, and if so, determine the sampled basic feature as the target feature;

[0013] If not, the target parameters are used to update the parameters of the generator and the discriminator to obtain updated generator and discriminator;

[0014] Using the updated generator and discriminator and the objective function, the above basic feature sampling, feature vector discrimination and target parameter generation steps are repeated until the obtained target parameter is within the preset numerical range, and the sampled basic features are determined as target features.

[0015] Optionally, the determining whether the target parameter is within a preset value range further includes:

[0016] Determine whether the current number of iterations reaches the preset number. If so, determine the sampled basic features as target features; if not, determine whether the target parameters are within the preset value range.

[0017] Optionally, before determining the item category in response to the input of the item category range, the method further includes:

[0018] According to the physical properties of each item in the item information database, determine the category to which each item belongs in each category level, so as to construct a category range table;

[0019] The determining of the item category in response to the input of the item category range includes:

[0020] In response to the input of the item category range, the item categories within the item category range are determined using the category range table.

[0021] Optionally, before using the category range table to determine the item category within the item category range, the method further includes:

[0022] A field corresponding to the item category range is obtained from a preset query field table, and then the item category corresponding to the field is extracted from the category range table.

[0023] Optionally, the positive example items are stored in an item set, and the item set corresponds to the item category;

[0024] Before determining the positive example items and positive example feature vectors corresponding to the item category, the method further includes:

[0025] Determining a first set of items in response to input of a market segment scope and a trend measurement indicator, wherein the trend measurement indicator is used to measure competitive trends between items;

[0026] The intersection of the item set and the first item set is obtained, and the items in the item set that are in the intersection are marked as positive examples, and the remaining items are marked as negative examples.

[0027] Optionally, the method further includes: determining a subcategory belonging to the item category based on an item category tree structure corresponding to the item category, and then acquiring an item set corresponding to the subcategory.

[0028] Optionally, the determining of the first set of items in response to the input of the market segment scope and the trend measurement indicator further includes:

[0029] Acquire a first field corresponding to the market segment range from a preset query field table, and then extract a second set of items corresponding to the first field from an item interaction attribute basic table;

[0030] Acquire a second field corresponding to the trend measurement indicator from the preset query field table, and then extract a third set of items corresponding to the second field from the item interaction attribute basic table;

[0031] The intersection of the second item set and the third item set is taken to obtain the first item set.

[0032] Optionally, it also includes: based on the user's selection records of each item, counting the selection quantity corresponding to each item, combining the information data of each user, and generating the basic table of item interaction attributes.

[0033] Optionally, after marking the items in the item set that are in the intersection as positive examples and marking the remaining items as negative examples, the method further includes:

[0034] Determine an associated item category that has an associated relationship with the item category, and determine a competing item that has a competitive relationship with the positive example item in a fourth item set corresponding to the associated item category;

[0035] If the competing product is not in the third item set or the associated item category is not within the item category range, the competing product is marked as a negative example, otherwise it is marked as a positive example.

[0036] Optionally, the using a generator to sample basic features and combine them to generate a combined item includes:

[0037] For a single basic feature, count the number of first items in the item set that contain both the single basic feature and other basic features, and the number of second items that contain the other basic features;

[0038] Calculating a ratio of the first number of items to the second number of items, and using the ratio as a sampling distribution input of the generator to obtain a feature value for the single basic feature;

[0039] In response to the selection of at least one basic feature, the combined item is generated by combining the orderings between the basic features.

[0040] Optionally, the inputting the feature vector into a discriminator to obtain a discrimination result includes: if the feature vector is within a vector domain composed of the positive example feature vectors, generating a discrimination result of the preset value.

[0041] Optionally, inputting the feature vector into a discriminator to obtain a discrimination result includes:

[0042] The discriminator counts the total number of items and the total number of positive example items in the item set, and calculates the proportion of the total number of positive example items in the total number of items;

[0043] For the basic features of the combined feature vectors, counting the total amount of first items and the total amount of first positive example items in the item set having a single basic feature;

[0044] Determine a first proportion of the first positive example total amount of items in the total amount of positive example items, and then calculate a first product of the first proportions of all basic features of the feature vector;

[0045] Determine a second proportion of the first total amount of items in the total amount of items, and then calculate a second product of the second proportions of all basic features of the feature vector;

[0046] Based on the proportion, the first product, and the second product, a probability value that the feature vector is a positive example item is determined, and the probability value is used as a discrimination result.

[0047] To achieve the above object, according to another aspect of an embodiment of the present invention, a feature extraction device is provided, comprising:

[0048] an acquisition module, configured to determine the item category in response to an input of an item category range, acquire basic features corresponding to the item category, and determine positive example items and positive example feature vectors corresponding to the item category;

[0049] A sampling module, used to use the generator to sample basic features and combine them to generate a combined object, and combine the feature values ​​of each basic feature to generate a feature vector of the combined object;

[0050] A discriminant module, used for inputting the feature vector into a discriminator to obtain a discriminant result; wherein the generator and the discriminator are located in a generative adversarial network;

[0051] A determination module is used to input the feature vector and the positive example feature vector into an objective function to obtain a target parameter if the discrimination result is a preset value, and to determine the sampled basic feature as a target feature if the target parameter meets a preset condition.

[0052] Optionally, the determining module is used to:

[0053] Determine whether the target parameter is within a preset value range, and if so, determine the sampled basic feature as the target feature;

[0054] If not, the target parameters are used to update the parameters of the generator and the discriminator to obtain updated generator and discriminator;

[0055] Using the updated generator and discriminator and the objective function, the above basic feature sampling, feature vector discrimination and target parameter generation steps are repeated until the obtained target parameter is within the preset numerical range, and the sampled basic features are determined as target features.

[0056] Optionally, the determination module is further used to: determine whether the current number of iterations reaches a preset number, and if so, determine the sampled basic features as target features; if not, determine whether the target parameters are within the preset value range.

[0057] Optionally, a table building module is also included to:

[0058] According to the physical properties of each item in the item information database, determine the category to which each item belongs in each category level, so as to construct a category range table;

[0059] The acquisition module is used to:

[0060] In response to the input of the item category range, the item categories within the item category range are determined using the category range table.

[0061] Optionally, the acquisition module is used to: acquire a field corresponding to the item category range from a preset query field table, and then extract the item category corresponding to the field from the category range table.

[0062] Optionally, the positive example items are stored in an item set, and the item set corresponds to the item category;

[0063] The acquisition module is further used for:

[0064] Determining a first set of items in response to input of a market segment scope and a trend measurement indicator, wherein the trend measurement indicator is used to measure competitive trends between items;

[0065] The intersection of the item set and the first item set is obtained, and the items in the item set that are in the intersection are marked as positive examples, and the remaining items are marked as negative examples.

[0066] Optionally, the acquisition module is further used to: determine the sub-category belonging to the item category based on the item category tree structure corresponding to the item category, and then acquire the item set corresponding to the sub-category.

[0067] Optionally, the acquisition module is further used to:

[0068] Acquire a first field corresponding to the market segment range from a preset query field table, and then extract a second set of items corresponding to the first field from an item interaction attribute basic table;

[0069] Acquire a second field corresponding to the trend measurement indicator from the preset query field table, and then extract a third set of items corresponding to the second field from the item interaction attribute basic table;

[0070] The intersection of the second item set and the third item set is taken to obtain the first item set.

[0071] Optionally, the acquisition module is further used to: based on the user's selection records of each item, count the selection quantity corresponding to each item, and generate the basic table of item interaction attributes in combination with the information data of each user.

[0072] Optionally, the acquisition module is further used to:

[0073] Determine an associated item category that has an associated relationship with the item category, and determine a competing item that has a competitive relationship with the positive example item in a fourth item set corresponding to the associated item category;

[0074] If the competing product is not in the third item set or the associated item category is not within the item category range, the competing product is marked as a negative example, otherwise it is marked as a positive example.

[0075] Optionally, the sampling module is used to:

[0076] For a single basic feature, count the number of first items in the item set that contain both the single basic feature and other basic features, and the number of second items that contain the other basic features;

[0077] Calculating a ratio of the first number of items to the second number of items, and using the ratio as a sampling distribution input of the generator to obtain a feature value for the single basic feature;

[0078] In response to the selection of at least one basic feature, the combined item is generated by combining the orderings between the basic features.

[0079] Optionally, the discrimination module is used to: if the feature vector is within the vector domain composed of the positive example feature vectors, generate a discrimination result of the preset value.

[0080] Optionally, the discrimination module is used to:

[0081] The discriminator counts the total number of items and the total number of positive example items in the item set, and calculates the proportion of the total number of positive example items in the total number of items;

[0082] For the basic features of the combined feature vectors, counting the total amount of first items and the total amount of first positive example items in the item set having a single basic feature;

[0083] Determine a first proportion of the first positive example total amount of items in the total amount of positive example items, and then calculate a first product of the first proportions of all basic features of the feature vector;

[0084] Determine a second proportion of the first total amount of items in the total amount of items, and then calculate a second product of the second proportions of all basic features of the feature vector;

[0085] Based on the proportion, the first product, and the second product, a probability value that the feature vector is a positive example item is determined, and the probability value is used as a discrimination result.

[0086] To achieve the above objective, according to another aspect of an embodiment of the present invention, a feature extraction electronic device is provided.

[0087] The electronic device of an embodiment of the present invention includes: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement any of the feature extraction methods described above.

[0088] To achieve the above objective, according to another aspect of an embodiment of the present invention, a computer-readable medium is provided, on which a computer program is stored, and when the program is executed by a processor, any of the feature extraction methods described above is implemented.

[0089] According to the solution provided by the present invention, one embodiment of the above invention has the following advantages or beneficial effects: innovatively transforming the problem of predicting the demand of specific user groups in a segmented market into the problem of intelligent identification of optimal item features based on market competition trends, quickly and accurately mining the optimal features of items, and providing a reference for item design and sales, which is a key step for enterprises to expand their markets and enhance their competitiveness in the future.

[0090] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0092] Figure 1 is a schematic diagram of the main process of a feature extraction method according to an embodiment of the present invention;

[0093] Figure 2 It is the item category tree structure provided by the embodiment of the present invention;

[0094] Figure 3 is a flow chart of an optional feature extraction method according to an embodiment of the present invention;

[0095] Figure 4 is a flow chart of another optional feature extraction method according to an embodiment of the present invention;

[0096] Figure 5 is a schematic diagram of a feature extraction method process framework according to an embodiment of the present invention;

[0097] Figure 6 is a schematic diagram of main modules of a feature extraction device according to an embodiment of the present invention;

[0098] Figure 7 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0099] Figure 8 It is a schematic diagram of the structure of a computer system of a mobile device or a server suitable for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0100] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0101] Based on the basic features of each item category that have been acquired, the embodiments of the present invention propose how to reduce the amount of basic features and quickly capture and extract the optimal features that best represent the most competitive characteristics at the moment.

[0102] See also Figure 1, which shows a main flow chart of a feature extraction method provided by an embodiment of the present invention, comprising the following steps:

[0103] S101: determining an item category in response to an input of an item category range, acquiring basic features corresponding to the item category, and determining positive example items and positive example feature vectors corresponding to the item category;

[0104] S102: using a generator to sample basic features and combine them to generate a combined object, and combining the feature values ​​of each basic feature to generate a feature vector of the combined object;

[0105] S103: Inputting the feature vector into a discriminator to obtain a discrimination result; wherein the generator and the discriminator are located in a generative adversarial network;

[0106] S104: If the discrimination result is a preset value, the feature vector and the positive example feature vector are input into an objective function to obtain a target parameter. If the target parameter meets a preset condition, the sampled basic feature is determined as a target feature.

[0107] In the above implementation, for step S101, the present invention provides a demand input interface for users (such as design departments) to define as query input conditions for calling the "intelligent optimal object feature mining device", and uses the query input conditions as service request content.

[0108] The present invention mainly describes the basic features of the item category, and the input condition is mainly "query item category range", as shown in Table 1:

[0109] Table 1:

[0110]

[0111] The search conditions for the item category range include:

[0112] ① The category of goods being targeted;

[0113] ② The brand of the item being targeted;

[0114] ③ Scope of items.

[0115] After entering the query conditions, the "intelligent optimal item feature mining device" is called to complete the target feature mining operation under the basic characteristics of the subsequent item category.

[0116] The "intelligent optimal item feature mining device" determines the item category based on the item category range input by the user. The specific determination method is described in the following Figure 3 The description shown is not repeated here. Search by item category range IFor example, query to obtain item category I, and combine the items belonging to item category I to form an item set SKU I .

[0117] If the item category contains multiple subcategories, we can determine the set of items that meet the requirements of each subcategory, and then start from each subcategory to extract the most features of each subcategory. Figure 2 As shown in the tree structure of item categories, under the “large household appliances” category, there are 6 subcategories: “washing machines - drum washing and drying machines”, “washing machines - pulsator type”, “washing machines - drum type”, “dryers”, “air conditioners”, and “refrigerators”. Then, the target features of each subcategory are mined under each subcategory.

[0118] It should be noted that the item category tree structure can be traversed based on the depth / breadth first search method. For example, when starting from the root node, the leftmost branch of household appliances is selected to connect to the "large appliances" node. The "large appliances" category is called "item category I". Similarly, the classification under the "large appliances" category is also obtained through layer-by-layer screening.

[0119] Each item category has basic characteristics and is pre-stored in the item information database; the basic characteristics are attributes, which include not only the common attributes common to most items in this category (such as price, size, weight, etc.), but also special attributes that can distinguish it from other item categories (such as washing machines have a wash ratio attribute, refrigerators have a door opening method attribute, and mobile phones have a memory size attribute).

[0120] It should be noted that the embodiment of the present invention determines the basic characteristics of the item category in advance based on the item attributes (for example, calculating the chi-square value of a single attribute to the item category), and then the basic characteristics can be directly queried based on the table. The present invention does not describe the method for determining the basic characteristics.

[0121] Some items in the item set are positive examples, that is, the best items, and some are negative examples, that is, items that do not meet the search requirements. For details on whether an item is a positive example, please refer to the following Figure 4 The description shown is not repeated here.

[0122] For a single item under item category I, its information data is obtained from the item information database, and the feature quantization value of each basic feature is combined to generate its own feature vector, and combined with its category label (which can be determined according to the item category tree structure), its basic data is obtained, as shown in Table 2 (the present invention mainly uses the feature vector of the positive example item):

[0123] Table 2:

[0124]

[0125] The present invention defines a feature vector as follows: the vector dimension is determined by the basic features of the item category, each dimension is a category after the basic features are quantified, and the content of each dimension is a dictionary in the form of (key, value), where the key is the domain value under the basic feature, and the value is a positive number. Usually, the feature value has only two values: 0 and 1. If a certain attribute exists in the item, it is 1, otherwise it is 0.

[0126] For step S102, the item set SKU I Although it contains a large number of items, the basic feature combinations of each item may not be fully covered. For example, there are 3 attribute values ​​under each of the 3 basic features, so there are a total of 27 combinations, but there are only 15 in reality.

[0127] Considering this problem, the present invention utilizes a generator to randomly sample basic features and generate new items, namely, combined items.

[0128] There are two main sampling methods (with z representing the basic features):

[0129] 1) Randomly select a value in the value range of the basic characteristics of the item category, such as z1, z2, and z3 represent three basic characteristics, forming a feature vector form SKU (z1, z2, z3). For example, select brand, color, and size as three basic characteristics.

[0130] 2) Combining basic features based on sampling algorithms, such as the Gibbs sampling algorithm; the core idea of ​​the sampling algorithm is to restore the original distribution of the variable as much as possible so that the expectation of the variable distribution is not affected by the variable taking a large value but a small variable.

[0131] ① Select at least one basic feature and determine at least one attribute value in each basic feature; for example, brand A and brand B in the brand attribute, white and red in the color attribute;

[0132] ② Establish joint probability, which is mainly used for subsequent objective functions. For example, in this category of items, the number of items with brand A and white color is a, and the number of items with brand A and red color is b, so the probabilities are a / (a+b) and b / (a+b) respectively. Based on this joint probability, some attribute values ​​or attributes (basic features) with lower probability values ​​can be eliminated.

[0133] ③ Set the number of iterations τ and start the loop iteration;

[0134] ④ Establish conditional probability for each basic feature, that is, the conditional probability of one attribute under all other attributes; for example, for the first brand feature, count the number of first items with both brand characteristics and color characteristics under the item category I, divide the number of first items by the number of second items with color characteristics, and calculate the conditional probability of the brand feature;

[0135] ⑤Use the conditional probability of the basic feature as the sampling distribution input to obtain a new feature value, such as

[0136] ⑥When the number of iterations reaches τ, the current new feature vector is returned

[0137] It should be noted that this step is a process of random combination of basic features, similar to exhausting all combinations of basic features. If the number of selections is not considered, method 1) can include all combinations, and method 2) the exhaustive range is limited by the number of iterations.

[0138] Furthermore, the number of iterations set here is used to re-collect basic features when the discrimination result of the feature vector of the combined object does not meet the requirements or the target parameters after inputting the target function do not meet the preset conditions.

[0139] Taking this a step further, sampling is performed by transforming a difficult problem into one that can be solved through sampling, a process performed using the generator in a generative adversarial network.

[0140] For step S103, the generative adversarial network includes, in addition to the generator, a discriminator, which is mainly used in the present invention to determine whether the feature vector of the combined item is a positive item (optimal item).

[0141] The input to the discriminant model is the basic features of the item category and its value. The model can be implemented in the following three ways:

[0142] 1) Methods based on geometric spatial position are used to determine whether the feature vector of the combined object is within the domain of the positive feature vector of a certain category, such as comparing whether the cosine angle of two feature vectors is greater than a preset value, or comparing whether the distance between two objects is within a predetermined range (such as Euclidean distance);

[0143] 2) Based on the naive Bayesian method, it mainly borrows Bayesian theory, product rule of probability theory and conditional independence theory. On the basis of assuming that the basic characteristics of items are independent of each other, it calculates the probability that the characteristics of items belong to the positive example, compares the preset selection probability value, and judges whether it is close to the positive example;

[0144] The following are some specific examples:

[0145] It is known that all items under item category I contain only three basic features z1, z2 and z3, and there are 400 items in total (85 positive examples and 315 negative examples).

[0146] The total number of items with feature z1 is 100, of which there are 80 positive examples;

[0147] The total number of items with feature z2 is 200, of which there are 2 positive examples;

[0148] The total number of items with feature z3 is 300, of which there are 3 positive examples.

[0149] Now there is a new combination of item A consisting of z1 and z3. The method to judge whether the new item A is a positive example item is:

[0150] ① Calculate the probability that the positive example contains z1 p1_best = 80 / 85, the probability that the positive example contains z3 p3_best = 3 / 85, and calculate the proportion of positive examples p_best = 85 / 400;

[0151] ② Calculate the proportion of items containing feature z1, p1 = 100 / 400, and calculate the proportion of items containing feature z3, p3 = 300 / 400;

[0152] ③Calculate the probability that item A composed of z1 and z3 is a positive item:

[0153] p_a=(p1_best*p3_best*p_best) / (p1*p3)=0.037

[0154] ④ Determine whether p_a is greater than 0.01 (0.01 is an artificially set threshold and can be adjusted). If it is greater than 0.01, then item A is the best item. For example, if p_a>0.01 in ③, then item A is a positive item.

[0155] Among them, the condition independence theory is that there are three variables a, b, and c. A depends on variable c but does not depend on variable b. Then the conditional probability that a depends on b and c is p(a|b,c)=p(a|c).

[0156] 3) Based on the compression algorithm, mainly borrowing the definition of Kolmogorov complexity, setting the hyperparameter (normalized compression distance), combining the basic features of the items with the hierarchical structure of the item category classification tree, using a hierarchical clustering-based method, calculating the normalized compression distance between the two items and comparing it with the hyperparameter, and then making a judgment.

[0157] For step S104, in the generative adversarial algorithm, the value of each basic feature of the object is randomly generated, but under the guidance of a goal to increase the pass rate of the deception discriminator, it is also necessary to increase the success rate of identifying positive items. In the confrontation between the two units (i.e., the generator and the discriminator), the value of each basic feature is continuously corrected to ultimately obtain the optimal feature.

[0158] In the data model, G is used to express the generator and D is used to express the discriminator. The optimization objective function is as follows:

[0159]

[0160] Among them, V(G,D) is the binary cross entropy function, for example:

[0161] J(w)=-[ylog(z)+(1-y)log(1-z)],z=h(x)

[0162] Among them, mini-max is: for example, when a feature vector formed by a certain combination of values ​​under the basic features of an item is within the domain of the positive feature vector, a reward score is obtained; if it is within the domain of the negative item feature, a penalty score is obtained; that is, while maximizing the reward score, the penalty score is minimized. The optimization process of a basic feature of an item is actually a set of behaviors for how a feature of an item chooses the next value. The optimal feature is found through continuous trial and error and game.

[0163] If the combined item is a positive item, a positive item is selected from the item set, and the positive item feature vector of the selected positive item and the feature vector of the combined item are input into the objective function together to obtain the target parameter.

[0164] If the target parameter is within the preset value range, the iteration step is exited and the basic features of the combined items are taken as the most characteristic features of the item category. Otherwise, the generator and discriminator need to be updated based on the target parameter, and steps S102 to S104 are repeated based on the updated generator and discriminator until the target parameter is within the preset value range.

[0165] Furthermore, the number of iterations is set to be limited. If the number of iterations is reached or the target parameter is within a preset value range, the iteration is stopped. Otherwise, the generator and the discriminator need to be updated based on the target parameter and the iterative operation is repeated until the number of iterations is reached or the target parameter is within a preset value range.

[0166] To summarize the optimization problem of basic features under the above item categories, for a given item category, from the different values ​​of the basic features of the items, unsupervised learning is used to mine the optimal value to form the target feature.

[0167] The input consists of two parts:

[0168] 1) Specify the item category and the corresponding basic features to use the generator to sample the basic features to form a new combined item;

[0169] 2) Under the specified item category, the items and their positive and negative label data table are used to combine the feature vectors of the combined items.

[0170] The output is:

[0171] 1) Target features under specified item categories;

[0172] 2) The value of each target feature, such as sorting out the types and contents of the optimal item features under each item category, as shown in Table 3.

[0173] Table 3:

[0174]

[0175] The user receives the results returned by the "intelligent optimal item feature mining device". When designing a certain category of items in the future, he can refer to the target features under the category, as shown in Table 4, to learn the current market demand focus for dryers for users in the suburbs of first-tier cities, thereby capturing the current needs of specific user groups and the corresponding optimal item features that meet the competition trend.

[0176] Table 4:

[0177]

[0178] The method provided in the above embodiment, based on basic feature sampling and feature vector discrimination, innovatively defines the problem of item category target feature mining as how to select a part of the basic feature value set that is closest to the positive example item set from the value set under the basic feature type of the item category, so as to quickly capture and extract the optimal basic features that best represent the most competitive characteristics at the moment.

[0179] See also Figure 3 , shows a schematic flow chart of an optional feature extraction method according to an embodiment of the present invention, comprising the following steps:

[0180] S301: Determine the category to which each item belongs in each category level according to the physical properties of each item in the item information database, thereby constructing a category range table;

[0181] S302: In response to the input of the item category range, using the category range table, determining the item categories within the item category range;

[0182] S303: Acquire basic features corresponding to the item category, and determine positive example items and positive example feature vectors corresponding to the item category;

[0183] S304: using a generator to sample basic features and combine them to generate a combined object, and combining the feature values ​​of each basic feature to generate a feature vector of the combined object;

[0184] S305: Inputting the feature vector into a discriminator to obtain a discrimination result; wherein the generator and the discriminator are located in a generative adversarial network;

[0185] S306: If the discrimination result is a preset value, the feature vector and the positive example feature vector are input into an objective function to obtain a target parameter. If the target parameter meets a preset condition, the sampled basic feature is determined as a target feature.

[0186] In the above implementation, for steps S303 to S306, see Figure 1 The description of steps S101 to S103 is not repeated here.

[0187] In the above implementation, for steps S301 and S302, the "intelligent item category adaptive feature mining device" disassembles the query conditions input by the user and finds the relevant fields of the corresponding item category range from the storage system to facilitate subsequent calibration of the items.

[0188] The present invention pre-constructs a query field table (or called an extended table) construction unit to parse query input conditions and form a data table that can be used by subsequent modules, during which the data table fields and contents in the storage system need to be queried.

[0189] The stored data is in various forms, but the text in the query conditions needs to be mapped to the field names in the data table. For example, if you want to find the "category" related fields for "which categories of goods", you can directly find the "category" related fields. For another example, if you want to find the "top three sales of the same brand within a unit time period", you need to parse out at least the "time", "sales" and "brand" fields. Ideally, the storage system contains all the fields in the query conditions. For example, see Table 5:

[0190] Table 5:

[0191]

[0192] In actual operation, some fields in the query conditions may not exist in the storage system. In view of this situation, a selection button is set in one or more input items of the user's query condition input interface, and the query field is pre-customized for each optional option, so as to achieve the purpose of field queryability.

[0193] Furthermore, if the number of query condition fields is too small, you can expand the query fields according to the pre-defined content that meets the query conditions to form a query field expansion table. For example, to parse "the top three sales of the same brand within a unit time period", you need to parse at least the three fields of "time", "sales" and "brand". The parsing process can be regarded as expansion. Figure 6 As shown:

[0194] Table 6:

[0195]

[0196] The scope of item categories may include primary categories, secondary categories, tertiary categories, etc., as shown in Table 7. The query field may also involve other "non-category" information fields, such as sales volume, GMV (Gross Merchandise Volume), etc.:

[0197] Table 7:

[0198]

[0199]

[0200] In the above implementation, when the item category is too broad or the item range is too large, the item category division can be prioritized, and the item sub-category division can be automatically performed by establishing an item category division tree to form a category range table that can be queried, thereby achieving the purpose of denoising the basic features of the item category; and the query field can be expanded to obtain a wider range of items.

[0201] See also Figure 4 , shows a schematic flow chart of another optional feature extraction method according to an embodiment of the present invention, comprising the following steps:

[0202] S401: determining an item category in response to an input of an item category range, and acquiring basic features and an item set corresponding to the item category;

[0203] S402: Determine a first set of items in response to input of a market segment range and a trend measurement index; wherein the trend measurement index is used to measure competitive trends between items;

[0204] S403: Taking the intersection of the item set and the first item set, marking the items in the item set that are in the intersection as positive examples, and marking the remaining items as negative examples;

[0205] S404: using a generator to sample basic features and combine them to generate a combined object, and combining the feature values ​​of each basic feature to generate a feature vector of the combined object;

[0206] S405: Inputting the feature vector into a discriminator to obtain a discrimination result; wherein the generator and the discriminator are located in a generative adversarial network;

[0207] S406: If the discrimination result is a preset value, the feature vector and the positive example feature vector are input into the objective function to obtain a target parameter. If the target parameter meets the preset conditions, the sampled basic feature is determined as the target feature; wherein the positive example feature vector corresponds to the positive example item.

[0208] In the above embodiment, for steps S401, S404 to S406, see Figure 1 The description of steps S101 to S103 is not repeated here.

[0209] This embodiment mainly describes the item set SKU under item category I I , how to determine which items meet the criteria of market segment scope and competitive trend, and if so, mark them as positive examples; among them, positive examples represent the best.

[0210] In the above implementation, for step S402, the input of the "intelligent optimal item feature mining device" may include, in addition to the "item category range", the following:

[0211] 1) Market segmentation, including:

[0212] ① The targeted regions, such as first-tier, second-tier, and third-tier cities;

[0213] ② Targeted groups, such as gender groups, age groups, user level groups, etc.;

[0214] 2) The standard indicators for measuring market competitiveness trends (i.e. trend measurement indicators) can be based on economic perspectives such as sales volume and profit, such as the top three sales volume of the same brand in a unit time period, the top three total profits of the same category, etc.

[0215] After the user has entered the query conditions, he can call the "intelligent optimal item feature mining device" to perform the most characteristic mining operation under the item category.

[0216] It should be noted that for the above "market segment range", "item category range" and "trend measurement indicators", a single interface can be set for simultaneous input, or two or three interfaces can be set for users to input separately, but it must be ensured that the input interface for "item category range" is located first.

[0217] The present invention pre-establishes a basic table of item interaction (interaction between people and items) attributes to store the user's selection behavior data for each item and the user's information data, including one piece of data for each item sold every day and in each regional dimension (regional dimension can be selected in terms of province, city, district, county, village), which contains four categories of information, namely:

[0218] a) Economic statistics of goods: such as average cost price, average transaction price, total sales volume, total sales, average profit, average profit margin;

[0219] b) Item marketing statistics: total exposure, total purchase, total order, average conversion rate, number of items on sale, and item promotion discount ratio;

[0220] c) Item operation text information: item title, item search and click keywords (set), item review set;

[0221] d) Item audience statistics: number of users reached by the item, distribution of levels of reached users (the cumulative number of people at each user level), age distribution of reached users, and gender distribution of reached users, as shown in Table 8:

[0222] Table 8:

[0223]

[0224]

[0225] According to the table, the second set of items that meet the "market segment range" and the third set of items that meet the "trend measurement indicators" can be queried respectively, and the intersection can be calculated to obtain the first set of items that meet both the "market segment range" and the "trend measurement indicators".

[0226] Furthermore, Figure 3 Similarly, the first field corresponding to the "market segment range" can be determined first, and then based on the first field, the corresponding items can be found from the basic table of item interaction attributes to obtain the second item set.

[0227] Taking the "trend measurement indicator" of the top three sales of the same brand within a unit time period as an example, it is necessary to at least parse out the three fields of "time", "sales" and "brand", and then through logical processing, to get a field "is the top three sales of the same brand within a unit time period".

[0228] For step S403, for the item set SKU I For each item in , determine whether it is in the first item set. If so, mark it as a positive example, otherwise mark it as a negative example. Specifically, the item set SKU I Intersect with the first item set and add SKUI The items in the first item set are marked as positive examples, and the remaining items are marked as negative examples.

[0229] In addition to the above methods, other methods can also be used:

[0230] 1) Group items into SKUs I Intersect the second set of items that meet the "market segment range" and convert the item set SKU I The items in the intersection are marked with t, t∈{0,1}, which is 1 if they are in the intersection, and 0 if they are not;

[0231] 2) Group items into SKUs I Intersect the third set of items that meet the "trend measurement indicator" and convert the item set SKU I The items in m i Mark, m i ∈{0,1};

[0232] 3) For item set SKU I The items that meet both the “market segment scope” and “trend measurement indicators” are marked as positive examples, that is, m i If t is 1 at the same time, otherwise it is a negative example. This process is equivalent to a logical operation.

[0233] SKU for item collection I After labeling each item, the final output is an "item and item positive and negative label data table", which uses items, market segments and measurement indicator types as dimensions, and records whether the item is a positive example under these dimensions.

[0234] Furthermore, in order to expand the positive and negative examples, the competitive products can be determined through the positive example items. Specifically, the competitive products of each positive example item are obtained by associating the fourth item set of the item category cluster. However, if the competitive product is not within the scope of the queried item category or does not meet the measurement criteria, the competitive product is marked as a negative example. It should be noted that the sample dimension at this time is not fixed, and can be the item dimension or the order dimension.

[0235] It should be noted that the purpose of this implementation process is to expand negative examples to other categories (or subcategories). For example, in the "dryer" subcategory, positive and negative examples are marked. Dryer X is a positive example, and its competitor is dryer Y. In other subcategories, "washing machines - drum washing and drying machines", the two are not in the same subcategory, but they are competitors. It was later discovered that dryer Y could not meet the measurement criteria. The reason may be that the attributes such as price and drying weight are not competitive. Therefore, in order to extract reasonable prices and specific values ​​of drying weight, dryer Y is used as a negative example and compared with the positive example dryer X for subsequent optimal item comparison.

[0236] The method provided in the above embodiment distinguishes positive and negative examples of items in an item set under an item category based on the market segment scope and trend measurement indicators, and uses the expansion of related item clusters to expand positive and negative example items, and establishes a mapping between items and measurement standards.

[0237] See also Figure 5 , which shows a schematic diagram of a feature extraction process framework provided by an embodiment of the present invention:

[0238] 1) Demand module, which is used to propose requirements for the item design part, including specifying the scope of market segments, defining market competitive trend measurement standards (i.e. trend measurement standards), and defining the scope of item categories.

[0239] 2) Raw data extraction and processing module, including a query field (expansion) table construction unit, an item attribute basic table construction unit, an item category basic feature extraction unit, and a related item category cluster basic table construction unit;

[0240] ① Query field (expansion) table construction unit: used to disassemble the query conditions input by the user and determine the fields corresponding to the query conditions according to the query field expansion table, such as the field corresponding to the item category, the first field of the market segment range and the second field of the trend measurement indicator;

[0241] ②Item attribute basic table construction unit: used to construct according to the attributes of each item in the item information database;

[0242] ③ Extracting basic characteristic units of item categories: for a certain item category, mining its general and individual attributes as its basic characteristics;

[0243] ④ Extract the basic table construction unit of the related item category cluster: used to determine the relationship between item categories, mainly used to expand the positive and negative examples of the specified item category.

[0244] 3) Intelligent item analysis module: input a specified item category to mine the best items that meet the measurement criteria under the item category, or mine the worst items that do not meet the measurement criteria, and finally output a set of item category labels to serve the intelligent item target feature mining module;

[0245] 4) Intelligent item target feature mining module, including item model mathematical expression establishment unit, item feature quality judgment unit and intelligent optimization unit;

[0246] ① Item model mathematical expression establishment unit: used to determine the feature vector of each item (including combined items) according to the feature value after the basic features are quantified;

[0247] ②Item feature quality judgment unit: Use the item attribute basic table to extract item economic attribute information, and according to the defined market competitiveness trend measurement standard, dig out the best items that meet the measurement standard;

[0248] ③ Intelligent optimization unit: Using the common and representative basic features of the item category, applying intelligent technology, and automatically mining the target features of items under each category based on the item-measurement standard mapping set.

[0249] The method provided in the embodiment of the present invention innovatively transforms the problem of predicting the demand of specific user groups in a segmented market into the problem of intelligently identifying the optimal item features based on market competition trends. It quickly and accurately mines the target features of items and provides a reference for item design and sales. It is a key step for enterprises to expand their markets and enhance their competitiveness in the future.

[0250] See also Figure 6 , which shows a schematic diagram of main modules of a feature extraction device 600 provided in an embodiment of the present invention, including:

[0251] An acquisition module 601 is used to determine the item category in response to the input of the item category range, obtain the basic features corresponding to the item category, and determine the positive example items and the positive example feature vector corresponding to the item category;

[0252] The sampling module 602 is used to use the generator to sample basic features and combine them to generate a combined object, and combine the feature values ​​of each basic feature to generate a feature vector of the combined object;

[0253] A discriminator module 603 is used to input the feature vector into a discriminator to obtain a discriminant result; wherein the generator and the discriminator are located in a generative adversarial network;

[0254] The determination module 604 is used to input the feature vector and the positive example feature vector into the objective function to obtain the target parameter if the discrimination result is a preset value, and to determine the sampled basic feature as the target feature if the target parameter meets the preset condition.

[0255] Optionally, the determining module 604 is used to:

[0256] Determine whether the target parameter is within a preset value range, and if so, determine the sampled basic feature as the target feature;

[0257] If not, the target parameters are used to update the parameters of the generator and the discriminator to obtain updated generator and discriminator;

[0258] Using the updated generator and discriminator and the objective function, the above basic feature sampling, feature vector discrimination and target parameter generation steps are repeated until the obtained target parameter is within the preset numerical range, and the sampled basic features are determined as target features.

[0259] Optionally, the determination module 604 is further used to: determine whether the current number of iterations reaches a preset number, and if so, determine the sampled basic features as target features; if not, determine whether the target parameters are within the preset value range.

[0260] Optionally, a table building module 605 (not shown in the figure) is further included, which is used to:

[0261] According to the physical properties of each item in the item information database, determine the category to which each item belongs in each category level, so as to construct a category range table;

[0262] The acquisition module 601 is used to:

[0263] In response to the input of the item category range, the item categories within the item category range are determined using the category range table.

[0264] Optionally, the acquisition module 601 is used to: acquire a field corresponding to the item category range from a preset query field table, and then extract the item category corresponding to the field from the category range table.

[0265] Optionally, the positive example items are stored in an item set, and the item set corresponds to the item category;

[0266] The acquisition module 601 is further used for:

[0267] Determining a first set of items in response to input of a market segment scope and a trend measurement indicator, wherein the trend measurement indicator is used to measure competitive trends between items;

[0268] The intersection of the item set and the first item set is obtained, and the items in the item set that are in the intersection are marked as positive examples, and the remaining items are marked as negative examples.

[0269] Optionally, the acquisition module 601 is further used to: determine a subcategory belonging to the item category based on an item category tree structure corresponding to the item category, and then acquire an item set corresponding to the subcategory.

[0270] Optionally, the acquisition module 601 is further used to:

[0271] Acquire a first field corresponding to the market segment range from a preset query field table, and then extract a second set of items corresponding to the first field from an item interaction attribute basic table;

[0272] Acquire a second field corresponding to the trend measurement indicator from the preset query field table, and then extract a third set of items corresponding to the second field from the item interaction attribute basic table;

[0273] The intersection of the second item set and the third item set is taken to obtain the first item set.

[0274] Optionally, the acquisition module 601 is further used to: based on the user's selection records of each item, count the selection quantity corresponding to each item, and generate the basic table of item interaction attributes in combination with the information data of each user.

[0275] Optionally, the acquisition module 601 is further used to:

[0276] Determine an associated item category that has an associated relationship with the item category, and determine a competing item that has a competitive relationship with the positive example item in a fourth item set corresponding to the associated item category;

[0277] If the competing product is not in the third item set or the associated item category is not within the item category range, the competing product is marked as a negative example, otherwise it is marked as a positive example.

[0278] Optionally, the sampling module 602 is used to:

[0279] For a single basic feature, count the number of first items in the item set that contain both the single basic feature and other basic features, and the number of second items that contain the other basic features;

[0280] Calculating a ratio of the first number of items to the second number of items, and using the ratio as a sampling distribution input of the generator to obtain a feature value for the single basic feature;

[0281] In response to the selection of at least one basic feature, the combined item is generated by combining the orderings between the basic features.

[0282] Optionally, the discrimination module 603 is configured to generate a discrimination result of the preset value if the feature vector is within the vector domain formed by the positive example feature vectors.

[0283] Optionally, the identification module 603 is used to:

[0284] The discriminator counts the total number of items and the total number of positive example items in the item set, and calculates the proportion of the total number of positive example items in the total number of items;

[0285] For the basic features of the combined feature vectors, counting the total amount of first items and the total amount of first positive example items in the item set having a single basic feature;

[0286] Determine a first proportion of the first positive example total amount of items in the total amount of positive example items, and then calculate a first product of the first proportions of all basic features of the feature vector;

[0287] Determine a second proportion of the first total amount of items in the total amount of items, and then calculate a second product of the second proportions of all basic features of the feature vector;

[0288] Based on the proportion, the first product, and the second product, a probability value that the feature vector is a positive example item is determined, and the probability value is used as a discrimination result.

[0289] In addition, the specific implementation content of the device described in the embodiment of the present invention has been described in detail in the method described above, so the repeated content will not be described again here.

[0290] The device provided in the embodiment of the present invention innovatively transforms the problem of predicting the demand of specific user groups in a segmented market into the problem of intelligently identifying the optimal item features based on market competition trends, quickly and accurately mines the target features of items, and provides a reference for item design and sales. This is a key step for enterprises to expand their markets and enhance their competitiveness in the future.

[0291] Figure 7 An exemplary system architecture 700 is shown in which embodiments of the present invention may be applied.

[0292] like Figure 7 As shown, system architecture 700 may include terminal devices 701, 702, 703, network 704 and server 705 (just an example). Network 704 is used to provide a medium for communication links between terminal devices 701, 702, 703 and server 705. Network 704 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0293] The user can use the terminal devices 701, 702, 703 to interact with the server 705 through the network 704 to receive or send messages, etc. Various communication client applications can be installed on the terminal devices 701, 702, 703.

[0294] The terminal devices 701 , 702 , and 703 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers.

[0295] The server 705 may be a server that provides various services, such as a background management server that provides support for shopping websites browsed by users using the terminal devices 701 , 702 , and 703 (only as an example).

[0296] It should be noted that the method provided in the embodiment of the present invention is generally executed by the server 705 , and accordingly, the device is generally set in the server 705 .

[0297] It should be understood that Figure 7 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0298] Reference below Figure 8 , which shows a schematic diagram of the structure of a computer system 800 of a terminal device suitable for implementing an embodiment of the present invention. Figure 8 The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0299] like Figure 8 As shown, the computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage part 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the system 800 are also stored. The CPU 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0300] The following components are connected to the I / O interface 805: an input section 806 including a keyboard, a mouse, etc.; an output section 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, a modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as needed. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 810 as needed, so that a computer program read therefrom is installed into the storage section 808 as needed.

[0301] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 809, and / or installed from the removable medium 811. When the computer program is executed by the central processing unit (CPU) 801, the above-mentioned functions defined in the system of the present invention are executed.

[0302] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0303] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0304] The modules involved in the embodiments of the present invention may be implemented by software or hardware. The modules described may also be set in a processor, for example, it may be described as: a processor includes an acquisition module, a sampling module, a discrimination module and a determination module. The names of these modules do not constitute a limitation on the modules themselves in some cases, for example, the determination module may also be described as a "target feature determination module".

[0305] As another aspect, the present invention further provides a computer-readable medium, which may be included in the device described in the above embodiment; or may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device includes:

[0306] In response to the input of the item category range, determine the item category, obtain basic features corresponding to the item category, and determine positive example items and positive example feature vectors corresponding to the item category;

[0307] Using a generator to sample basic features and combine them to generate a combined object, and combining the feature values ​​of each basic feature to generate a feature vector of the combined object;

[0308] Inputting the feature vector into a discriminator to obtain a discrimination result; wherein the generator and the discriminator are located in a generative adversarial network;

[0309] If the discrimination result is a preset value, the feature vector and the positive example feature vector are input into an objective function to obtain a target parameter. If the target parameter meets a preset condition, the sampled basic feature is determined as the target feature.

[0310] According to the technical solution of the embodiment of the present invention, the problem of demand prediction for specific user groups in a segmented market is innovatively transformed into the problem of intelligent identification of optimal item features based on market competition trends. The target feature mining of items is quickly and accurately performed, providing a reference for item design and sales. This is a key step for enterprises to expand their markets and enhance their competitiveness in the future.

[0311] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A feature extraction method, characterized in that: include: In response to the input of the item category range, determine the item category, obtain basic features corresponding to the item category, and determine positive example items and positive example feature vectors corresponding to the item category; Using a generator to sample basic features and combine them to generate a combined object, and combining the feature values ​​of each basic feature to generate a feature vector of the combined object; Inputting the feature vector into a discriminator to obtain a discrimination result; wherein the generator and the discriminator are located in a generative adversarial network; If the discrimination result is a preset value, the feature vector and the positive example feature vector are input into an objective function to obtain a target parameter. If the target parameter meets a preset condition, the sampled basic feature is determined as the target feature.

2. The method according to claim 1, characterized in that If the target parameter meets the preset condition, the sampled basic feature is determined as the target feature, including: Determine whether the target parameter is within a preset value range, and if so, determine the sampled basic feature as the target feature; If not, the target parameters are used to update the parameters of the generator and the discriminator to obtain updated generator and discriminator; Using the updated generator and discriminator and the objective function, the above basic feature sampling, feature vector discrimination and target parameter generation steps are repeated until the obtained target parameter is within the preset numerical range, and the sampled basic features are determined as target features.

3. The method according to claim 2, characterized in that The determining whether the target parameter is within a preset value range further includes: Determine whether the current number of iterations reaches the preset number. If so, determine the sampled basic features as target features; if not, determine whether the target parameters are within the preset value range.

4. The method according to claim 1, characterized in that: Before determining the item category in response to the input of the item category range, the method further includes: According to the physical properties of each item in the item information database, determine the category to which each item belongs in each category level, so as to construct a category range table; The determining of the item category in response to the input of the item category range includes: In response to the input of the item category range, the item categories within the item category range are determined using the category range table.

5. The method according to claim 4, characterized in that Before using the category range table to determine the item categories within the item category range, the method further includes: A field corresponding to the item category range is obtained from a preset query field table, and then the item category corresponding to the field is extracted from the category range table.

6. The method according to claim 1, characterized in that The positive example items are stored in an item set, and the item set corresponds to the item category; Before determining the positive example items and positive example feature vectors corresponding to the item category, the method further includes: Determining a first set of items in response to input of a market segment scope and a trend measurement indicator, wherein the trend measurement indicator is used to measure competitive trends between items; The intersection of the item set and the first item set is obtained, and the items in the item set that are in the intersection are marked as positive examples, and the remaining items are marked as negative examples.

7. The method according to claim 6, characterized in that Also includes: Based on the item category tree structure corresponding to the item category, a subcategory belonging to the item category is determined, and then an item set corresponding to the subcategory is obtained.

8. The method according to claim 6, characterized in that The determining of the first set of items in response to the input of the market segment scope and the trend measurement indicator further includes: Acquire a first field corresponding to the market segment range from a preset query field table, and then extract a second set of items corresponding to the first field from an item interaction attribute basic table; Acquire a second field corresponding to the trend measurement indicator from the preset query field table, and then extract a third set of items corresponding to the second field from the item interaction attribute basic table; The intersection of the second item set and the third item set is taken to obtain the first item set.

9. The method according to claim 8, characterized in that Also includes: Based on the user's selection records of each item, the selection quantity corresponding to each item is counted, and combined with the information data of each user, the basic table of item interaction attributes is generated.

10. The method according to claim 6, characterized in that After marking the items in the intersection of the item set as positive examples and marking the remaining items as negative examples, the method further includes: Determine an associated item category that has an associated relationship with the item category, and determine a competing item that has a competitive relationship with the positive example item in a fourth item set corresponding to the associated item category; If the competing product is not in the third item set or the associated item category is not within the item category range, the competing product is marked as a negative example, otherwise it is marked as a positive example.

11. The method according to claim 1, characterized in that The method of using a generator to sample basic features and combine them to generate a combined item includes: For a single basic feature, count the number of first items in the item set that contain both the single basic feature and other basic features, and the number of second items that contain the other basic features; Calculating a ratio of the first number of items to the second number of items, and using the ratio as a sampling distribution input of the generator to obtain a feature value for the single basic feature; In response to the selection of at least one basic feature, the combined item is generated by combining the orderings between the basic features.

12. The method according to claim 1, characterized in that The step of inputting the feature vector into a discriminator to obtain a discrimination result includes: If the feature vector is within the vector domain formed by the positive example feature vectors, a discrimination result of the preset value is generated.

13. The method according to claim 1, characterized in that The step of inputting the feature vector into a discriminator to obtain a discrimination result includes: The discriminator counts the total number of items and the total number of positive example items in the item set, and calculates the proportion of the total number of positive example items in the total number of items; For the basic features of the combined feature vectors, counting the total amount of first items and the total amount of first positive example items in the item set having a single basic feature; Determine a first proportion of the first positive example total amount of items in the total amount of positive example items, and then calculate a first product of the first proportions of all basic features of the feature vector; Determine a second proportion of the first total amount of items in the total amount of items, and then calculate a second product of the second proportions of all basic features of the feature vector; Based on the proportion, the first product, and the second product, a probability value that the feature vector is a positive example item is determined, and the probability value is used as a discrimination result.

14. A feature extraction device, characterized in that: include: an acquisition module, configured to determine the item category in response to an input of an item category range, acquire basic features corresponding to the item category, and determine positive example items and positive example feature vectors corresponding to the item category; A sampling module, used to use the generator to sample basic features and combine them to generate a combined object, and combine the feature values ​​of each basic feature to generate a feature vector of the combined object; A discriminant module, used for inputting the feature vector into a discriminator to obtain a discriminant result; wherein the generator and the discriminator are located in a generative adversarial network; A determination module is used to input the feature vector and the positive example feature vector into an objective function to obtain a target parameter if the discrimination result is a preset value, and to determine the sampled basic feature as a target feature if the target parameter meets a preset condition.

15. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 13.

16. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 13 is implemented.

Images