Product group recommendation method, device, electronic device and readable storage medium

Abstract

The embodiments of the present disclosure provide a method, device, electronic device and readable storage medium for recommending a product group, the method comprising: for each product in the target product group, generating an individual feature vector according to the individual features corresponding to the product; determining the weight of the product in the target product group according to the individual feature vector; generating a comprehensive feature vector of the target product group according to the individual feature vector of the product, the weight of the product in the target product group and the comprehensive features of the target product group; inputting the comprehensive feature vector into an activation function to predict the quality score of the target product group; and recommending the target product group to the user according to the quality score. The quality score of the product combination can be estimated according to the features of each product in the product combination, which helps to improve the accuracy of the recommendation and the flexibility of the product combination is better.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of personalized recommendation technology, and more particularly to a method, device, electronic device, and readable storage medium for recommending a product group. Background Art

[0002] Personalized recommendation determines the products that users may be interested in based on product features and user features, and recommends them to users. Personalized recommendation is widely used in various fields. For example, in the catering field, multiple dishes can be reasonably combined to form a dish combination to recommend to users.

[0003] In the prior art, the recommendation process of dish combinations mainly includes: first, according to the user's daily basic calorie requirement and the calories required to be consumed that day obtained in advance, the calories that the user needs to supplement when dining are calculated; then, according to the calorie value corresponding to each dish and the dish combination method, multiple alternative dish combinations for the current user are formulated; finally, according to the user's consumption behavior and taste preferences in the system, multiple dish combination plans are generated in the alternative dish combinations, so as to recommend the best dish combination to the user according to the ranking of the dish combination plans.

[0004] The inventor studied the above scheme and found that the accuracy of the estimated calories is poor, which leads to poor recommendation accuracy; in addition, the inherent combination method limits the flexibility of the combination. Summary of the invention

[0005] The embodiments of the present disclosure provide a method, device, electronic device and readable storage medium for recommending a product group, which can estimate the quality score of a product combination according to the characteristics of each product in the product combination, help improve the accuracy of recommendation, and provide better flexibility for product combinations.

[0006] According to a first aspect of an embodiment of the present disclosure, a method for recommending a commodity group is provided, the method comprising:

[0007] For each product in the target product group, generating an individual feature vector according to the individual features corresponding to the product;

[0008] Determining the weight of each commodity in the target commodity group according to the individual feature vector of each commodity;

[0009] Generate a comprehensive feature vector of the target product group according to the individual feature vector of each product, the weight of each product in the target product group, and the comprehensive features of the target product group;

[0010] The comprehensive feature vector is input into a pre-trained activation function to predict the quality score of the target product group, wherein the activation function includes a first activation sub-function for click-through rate and a second activation sub-function for conversion rate, and during the training process, the loss value is determined according to the click-through rate sample value pre-marked in the product combination sample, the click-through rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample value, and the conversion rate prediction value output by the second activation sub-function;

[0011] The target product group is recommended to the user according to the quality score.

[0012] According to a second aspect of an embodiment of the present disclosure, there is provided a device for recommending a commodity group, the device comprising:

[0013] An individual feature vector generation module, used to generate an individual feature vector for each commodity in the target commodity group according to the individual features corresponding to the commodity;

[0014] A commodity weight determination module, used to determine the weight of each commodity in the target commodity group according to the individual feature vector of each commodity;

[0015] A comprehensive feature vector generating module, used to generate a comprehensive feature vector of the target commodity group according to the individual feature vector of each commodity, the weight of each commodity in the target commodity group, and the comprehensive features of the target commodity group;

[0016] A quality score prediction module, used for inputting the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target product group, wherein the activation function includes a first activation sub-function for click-through rate and a second activation sub-function for conversion rate, and during the training process, determining the loss value according to the click-through rate sample value pre-marked in the product combination sample, the click-through rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample value, and the conversion rate prediction value output by the second activation sub-function;

[0017] A product group recommendation module is used to recommend the target product group to the user according to the quality score.

[0018] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0019] A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the aforementioned method for recommending a commodity group when executing the program.

[0020] According to a fourth aspect of an embodiment of the present disclosure, a readable storage medium is provided, characterized in that when instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the aforementioned method for recommending a product group.

[0021] The embodiment of the present disclosure provides a method and device for recommending a commodity group, the method comprising: for each commodity in the target commodity group, generating an individual feature vector according to the individual features corresponding to the commodity; determining the weight of each commodity in the target commodity group according to the individual feature vector of each commodity; generating a comprehensive feature vector of the target commodity group according to the individual feature vector of each commodity, the weight of each commodity in the target commodity group, and the comprehensive features of the target commodity group; inputting the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target commodity group, the activation function comprising a first activation sub-function for click-through rate and a second activation sub-function for conversion rate, during the training process, determining the loss value according to the click-through rate sample value pre-marked in the commodity combination sample, the click-through rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample value, and the conversion rate prediction value output by the second activation sub-function; recommending the target commodity group to the user according to the quality score. The embodiment of the present disclosure can estimate the quality score of the commodity combination according to the features of each commodity in the commodity combination, which helps to improve the recommendation accuracy, and the flexibility of the commodity combination is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the description of the embodiments of the present disclosure will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0023] Figure 1 A flowchart showing a method for recommending a commodity group in an embodiment of the present disclosure is shown;

[0024] Figure 2 A flowchart showing the steps of a method for recommending a commodity group in another embodiment of the present disclosure is shown;

[0025] Figure 3 A schematic diagram of a model structure for predicting quality scores based on individual feature vectors and comprehensive feature vectors in an embodiment of the present disclosure is shown;

[0026] Figure 4 A schematic diagram of a model structure for generating an individual feature vector for each commodity in an embodiment of the present disclosure is shown;

[0027] Figure 5A structural diagram showing a device for recommending a commodity group in an embodiment of the present disclosure;

[0028] Figure 6 A structural diagram showing a device for recommending a commodity group in another embodiment of the present disclosure;

[0029] Figure 7 A structural diagram of an electronic device in an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments in the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present disclosure.

[0031] Embodiment 1

[0032] Reference Figure 1 , which shows a step flow chart of a method for recommending a commodity group in an embodiment of the present disclosure, as follows.

[0033] Step 101 : for each commodity in the target commodity group, an individual feature vector is generated according to the individual features corresponding to the commodity.

[0034] The target product group is a product group whose quality score is to be determined, and is composed of multiple products. In the embodiment of the present disclosure, the target product group can be a product group set by the merchant based on experience, or a product group generated by other automated methods. The embodiment of the present disclosure does not limit the source of the target product group.

[0035] Individual features are various features that affect the success or failure of product recommendation, and may include but are not limited to: product name, product price, product sales volume, labels at various levels, and the number of the product in the target product group.

[0036] Individual feature vectors can be obtained by splicing individual features in a certain order. For example, product name, product price, product sales, labels at all levels, and the number of products in the target product group can be spliced ​​in order. Among them, product name and labels at all levels need to be converted into word vectors in advance, while product price, product sales, and the number of products in the target product group can be directly used as vector values.

[0037] Step 102: determining the weight of each commodity in the target commodity group according to the individual feature vector of each commodity.

[0038] The weight of the product in the target product group may be the influence of the product on the target product group. For example, the greater the weight, the greater the influence of the product on the success or failure of the recommendation of the target product group; the smaller the weight, the smaller the influence of the product on the success or failure of the recommendation of the target product group.

[0039] In the embodiments of the present disclosure, Figure 3 As shown, the weight of a product in the target product group can be predicted by Self-Attention. It can predict the weight of the product in the target product group based on the individual feature vector of each product input. Self-Attention is a common structure in deep learning technology, and the embodiments of this disclosure will not be described in detail.

[0040] Step 103 : generating a comprehensive feature vector of the target product group according to the individual feature vector of each product, the weight of each product in the target product group, and the comprehensive features of the target product group.

[0041] Specifically, first, the individual feature vector of each commodity is multiplied by the weight of the commodity in the target commodity group to obtain the weighted feature vector of each commodity; then, the weighted feature vectors of all commodities and the comprehensive features of the target commodity group are spliced ​​into the comprehensive feature vector of the target commodity group. Of course, it can be understood that the embodiments of the present disclosure do not limit the splicing order.

[0042] In practical applications, such as Figure 3 As shown, the weighted feature vectors of all commodities and the comprehensive features of the target commodity group can be directly input into the Concatenate layer, and the comprehensive feature vector of the target commodity group is output. Concatenate is a common structure in deep learning, and the embodiments of the present disclosure will not be described in detail.

[0043] In the embodiments of the present disclosure, the comprehensive feature vector of the target product group includes not only the individual features of the products, but also the comprehensive features of the target product group. Thus, the more information the comprehensive feature vector includes, the more accurate the description of the target product group is, which helps to improve the accuracy of the quality score.

[0044] Step 104: input the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target product group, wherein the activation function includes a first activation sub-function for click-through rate and a second activation sub-function for conversion rate, wherein the first activation sub-function and the second activation sub-function are pre-trained, and during the training process, the loss value is determined according to the click-through rate sample values ​​pre-marked in the product combination sample, the click-through rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample values, and the conversion rate prediction value output by the second activation sub-function.

[0045] In the embodiments disclosed herein, Figure 3 As shown, the quality score is calculated based on the click rate output by the first activation sub-function and the conversion rate output by the second activation sub-function.

[0046] The loss value takes into account both the loss of click-through rate and the loss of conversion rate, and combines the two losses as the loss value, so that the click-through rate and conversion rate corresponding to the trained model are better.

[0047] It is understood that the embodiments of the present disclosure do not limit the specific calculation formula of the loss value, as long as the loss value is composed of the loss value of the click rate and the loss value of the conversion rate. For example, the square loss value between the click rate prediction value and the click rate sample value and the square loss value between the conversion rate prediction value and the conversion rate sample value can be directly weighted to obtain the loss value.

[0048] Step 105: recommend the target product group to the user according to the quality score.

[0049] It can be understood that the higher the quality score, the higher the recommendation success rate of the target product group; and the lower the quality score, the lower the recommendation success rate of the target product group.

[0050] Specifically, multiple target product combinations may be arranged in descending order according to their quality scores, and target product combinations with top-ranked quality scores greater than or equal to a preset quality score threshold may be recommended to the user.

[0051] In summary, the embodiment of the present disclosure provides a method for recommending a commodity group, the method comprising: for each commodity in the target commodity group, generating an individual feature vector according to the individual features corresponding to the commodity; determining the weight of each commodity in the target commodity group according to the individual feature vector of each commodity; generating a comprehensive feature vector of the target commodity group according to the individual feature vector of each commodity, the weight of each commodity in the target commodity group, and the comprehensive features of the target commodity group; inputting the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target commodity group, the activation function comprising a first activation sub-function for click-through rate and a second activation sub-function for conversion rate, during the training process, determining the loss value according to the click-through rate sample value pre-marked in the commodity combination sample, the click-through rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample value, and the conversion rate prediction value output by the second activation sub-function; recommending the target commodity group to the user according to the quality score. The embodiment of the present disclosure can estimate the quality score of the commodity combination according to the features of each commodity in the commodity combination, which helps to improve the recommendation accuracy, and the flexibility of the commodity combination is better.

[0052] Embodiment 2

[0053] Reference Figure 2 , which shows a specific step flow chart of a method for recommending a commodity group in another embodiment of the present disclosure, as follows.

[0054] Step 201: for each product in the target product group, the product name is input into the convolutional neural network and the maximum pooling layer to predict the first sub-vector.

[0055] The product name here refers to the word vector corresponding to the product name. Since the product name usually includes multiple information, in addition to the main information, there is some redundant information, so it is necessary to remove the redundant information from the product name to obtain the first sub-vector containing only the main information.

[0056] In the embodiments of the present disclosure, Figure 4 As shown, the feature information in the product name can be extracted by CNN (Convolutional Neutral Network), and the main information in the feature information can be extracted by using the maximum pooling layer Max-Pooling, so that the first sub-vector can contain the main information in the product name. Convolutional neural network and Max-Pooling are both commonly used structures in deep learning, and the embodiments of the present disclosure will not be described in detail.

[0057] It should be noted that, in addition to the product name, if there are any customized features in the individual features of the product that contain redundant information, they all need to be processed in step 201 to obtain the first sub-vector.

[0058] Step 202: Input the label into the embedding layer to predict the second sub-vector.

[0059] The label may include labels of multiple levels. The greater the number of labels, the more accurate the description of the product by the obtained second sub-vector; the smaller the number of labels, the less accurate the description of the product by the obtained second sub-vector.

[0060] Among them, the label itself is a semantic information, which needs to be mapped into a vector representation through the embedding layer.

[0061] In the embodiments of the present disclosure, Figure 4 As shown, the Embeding layer is an embedding layer. Embedding is a commonly used structure in deep learning, and the embodiments of the present disclosure will not be described in detail.

[0062] It can be understood that since the tag information is usually set according to certain rules and does not contain redundant information, it does not need to be processed through step 201.

[0063] Step 203: Concatenate the first sub-vector, the second sub-vector, and the third vector composed of the continuous features to obtain an individual feature vector.

[0064] Among them, continuous features are numerical features, including but not limited to: product price, product sales, product ratings, etc.

[0065] It should be noted that since continuous features are numerical features, they do not need any processing and can be used as values ​​in the feature vector.

[0066] like Figure 4 As shown, the first sub-vector, the second sub-vector, and the third vector composed of the continuous features can be concatenated into an individual feature vector.

[0067] It can be understood that the embodiments of the present disclosure do not limit the splicing order.

[0068] Step 204: determine the weight of each commodity in the target commodity group according to the individual feature vector of each commodity.

[0069] This step can refer to the detailed description of step 102 and will not be described again here.

[0070] Step 205 : generating a comprehensive feature vector of the target product group according to the individual feature vector of each product, the weight of each product in the target product group, and the comprehensive features of the target product group.

[0071] This step can refer to the detailed description of step 103 and will not be described again here.

[0072] Step 206: input the comprehensive feature vector into the first activation subfunction to predict the click rate.

[0073] It can be understood that the first activation subfunction is used to predict the click rate of the target product group according to the comprehensive feature vector.

[0074] Step 207: input the comprehensive feature vector into the second activation subfunction to predict the conversion rate.

[0075] It can be understood that the second activation subfunction is used to predict the conversion rate of the target product group according to the comprehensive feature vector.

[0076] It should be noted that the first activation sub-function and the second activation sub-function may adopt existing activation functions, such as sigmoid function, tanh function, ReLU function, etc. The embodiments of the present disclosure do not limit the selection of activation sub-functions.

[0077] Step 208, a quality score is calculated based on the click rate and the conversion rate. The first activation sub-function and the second activation sub-function are pre-trained. During the training process, a loss value is determined based on the click rate sample values ​​pre-marked in the product combination sample, the click rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample values, and the conversion rate prediction value output by the second activation sub-function.

[0078] Specifically, the quality score can be obtained by performing a weighted operation based on the click-through rate and the conversion rate, including weighting by multiplication and weighting by sum. For example, weighting by sum means: adding the product of the click-through rate and the first weighting factor, and the product of the conversion rate and the second weighting factor to obtain the quality score, and weighting by multiplication means: multiplying the click-through rate and the conversion rate to obtain the quality score. The embodiments of the present disclosure do not limit the weighting algorithm, and can be calculated by various formulas of comprehensive click-through rate and conversion rate, as long as the following relationship remains unchanged: the greater the click-through rate and the greater the conversion rate, the greater the quality score; the smaller the click-through rate and the smaller the conversion rate, the smaller the quality score.

[0079] The embodiments of the present disclosure can not only take both the click-through rate and the conversion rate into consideration during training, but can also determine the quality score by combining the click-through rate and the conversion rate, thereby improving the accuracy of the quality score.

[0080] Optionally, in another embodiment of the present disclosure, step 208 includes sub-steps A1 to A2:

[0081] Sub-step A1, determining a visit-to-purchase rate score according to the product of the click rate and the conversion rate.

[0082] In the embodiments of the present disclosure, the product of the click rate and the conversion rate can be directly used as the visit-to-purchase rate score, and can also be transformed to obtain the visit-to-purchase rate score. The embodiments of the present disclosure do not limit it, as long as the following relationship is maintained: the larger the product of the click rate and the conversion rate, the larger the visit-to-purchase rate score; the smaller the product of the click rate and the conversion rate, the smaller the visit-to-purchase rate score.

[0083] Sub-step A2: using the click rate, the conversion rate, or the visit-to-purchase rate score as a quality score.

[0084] In the embodiments of the present disclosure, multiple types of quality scores can be supported, so that recommendations with different goals can be flexibly implemented according to actual applications. For example, recommendations with click-through rate as the goal can recommend target product combinations with higher click-through rates to users, recommendations with conversion rate as the goal can recommend target product combinations with higher conversion rates to users, and recommendations with visit-to-purchase rate score as the goal can take both click-through rate and visit-to-purchase rate into account.

[0085] In another embodiment of the present disclosure, the loss value is calculated according to the following formula:

[0086] The loss value is calculated according to the following formula:

[0087]

[0088] TrainLoss is the loss value, M is the number of product combination samples, w1 is the first preset weight, w2 is the second preset weight, y′ i,1 is the click rate prediction value of the first activation subfunction for the i-th product combination sample, y′ i,2 is the conversion rate prediction value output by the second activation subfunction for the i-th product combination sample, y i,1 ,y i,2 are the click rate sample value and conversion rate sample value pre-labeled for the i-th product combination sample, y i,1 ·y i,2 is the sample value of the visit-purchase rate corresponding to the i-th product combination sample, y′ i,1 y′ i,2 is the predicted value of the visit-to-purchase rate corresponding to the i-th product combination sample.

[0089] Optionally, in another embodiment of the present disclosure, the first preset weight and the second preset weight are determined by the following steps B1 to B3:

[0090] Step B1, based on multiple product combination samples, respectively calculating the variance of the click rate prediction value and the variance of the visit purchase rate prediction value to obtain a first variance and a second variance.

[0091] In the embodiment of the present disclosure, the click rate prediction value of the commodity combination sample can be used as a variable, so that the variance σ1 of the click rate prediction value can be calculated. 2 , you can refer to the following formula:

[0092]

[0093] Among them, M, i, y′ i,1 Please refer to the description in formula (1), which will not be repeated here. μ1 is the average value of the click rate prediction value.

[0094] In addition, the predicted conversion rate of the product combination sample can also be used as a variable, so that the variance of the predicted conversion rate value σ2 can be calculated 2 , you can refer to the following formula:

[0095]

[0096] Among them, M, i, y′ i,2Please refer to the description in formula (1), which will not be repeated here. μ2 is the average value of the predicted conversion rate.

[0097] Step B2: if the first variance is greater than the second variance, reducing the first preset weight and increasing the second preset weight.

[0098] Specifically, if the difference between the first variance and the second variance is greater than a certain threshold, the first preset weight is reduced and the second preset weight is increased; if the difference between the first variance and the second variance is less than or equal to a certain threshold, the first preset weight is not reduced and the second preset weight is not increased.

[0099] It can be understood that the degree of reduction and increase can be determined according to the difference between the first variance and the second variance. The larger the difference, the greater the degree of reduction and increase; the smaller the difference, the smaller the degree of reduction and increase.

[0100] Step B3: if the first variance is smaller than the second variance, increase the first preset weight and decrease the second preset weight.

[0101] Specifically, if the difference between the second variance and the first variance is greater than a certain threshold, the first preset weight is increased and the second preset weight is decreased; if the difference between the second variance and the first variance is less than or equal to a certain threshold, the first preset weight is not increased and the second preset weight is not decreased.

[0102] It can be understood that the degree of reduction and increase can be determined according to the difference between the second variance and the first variance. The larger the difference, the greater the degree of reduction and increase; the smaller the difference, the smaller the degree of reduction and increase.

[0103] In the embodiments of the present disclosure, when the variance of the click-through rate prediction value is large, that is, when the accuracy of the click-through rate prediction value is poor, the impact of the click-through rate prediction value on the quality score can be reduced; when the variance of the conversion rate prediction value is large, that is, when the accuracy of the conversion rate prediction value is poor, the impact of the conversion rate prediction value on the quality score can be reduced, which helps to further improve the accuracy of the quality score.

[0104] In another embodiment of the present disclosure, before training the first activation sub-function and the second activation sub-function, step C1 is further included:

[0105] Step C1, filtering the product combination sample according to preset rules, wherein the preset rules include one or more of the following rules: a filtering rule based on a conversion rate within a preset historical time period, and a filtering rule based on a click rate within a preset historical time period.

[0106] Specifically, the product combination samples with a conversion rate lower than a certain conversion rate threshold can be filtered out, and the product combination samples with a click rate threshold lower than a certain click rate threshold can be filtered out. In addition, the product combination samples with a low number of impressions can also be filtered out.

[0107] The embodiments of the present disclosure can filter the commodity combination samples to ensure the quality of the commodity combination samples, which helps to improve the accuracy of the training results.

[0108] In another embodiment of the present disclosure, before training the first activation sub-function and the second activation sub-function, steps C2 to C3 are further included:

[0109] Step C2: expanding the sample of the commodity combination by using a preset blacklist.

[0110] The blacklist may be a commodity combination sample that does not meet the preset conditions, wherein the preset conditions may include but are not limited to: conditions based on price, conditions based on commodity types in the commodity combination sample, conditions based on user-defined conditions, etc. Commodity combination samples whose prices do not meet the conventional prices may be added to the blacklist, commodity combination samples whose commodity types in the commodity combination samples are all beverage types may be added to the blacklist, and commodity combination samples with poor user satisfaction may be added to the blacklist.

[0111] Of course, in order to reduce the space occupied by the blacklist, the storage object of the blacklist can be determined as a preset blacklist condition, so that the product combination that meets the preset blacklist condition can be used as a product combination sample and marked as a negative sample, for example, the click rate and conversion rate are both 0. Among them, the blacklist condition is opposite to the above preset condition. For example, the preset condition is that the price meets the regular price, and the blacklist condition is that the price does not meet the regular price; the preset condition is that the product type is not all beverage type, and the blacklist condition is that the product type is all beverage type; the preset condition is that the user satisfaction is good, and the blacklist condition is that the user satisfaction is poor.

[0112] Step C3: Expand the sample of the commodity combination through a preset whitelist.

[0113] The preset whitelist may include commodity combinations that meet preset conditions. Specifically, the whitelist may be generated by acquiring commodity combinations that meet preset conditions from a database of the commodity sales platform.

[0114] The preset conditions may refer to the description in step C2 and will not be described in detail here.

[0115] Similarly, in order to reduce the space occupied by the whitelist, the storage objects of the whitelist can be determined as preset conditions, and then when the product combination samples are expanded, the product combinations that meet the preset conditions are obtained from the database of the product sales platform as product combination samples, and marked as positive samples. For example, the click-through rate and conversion rate of the product combination sample on the product sales platform are used as marking information.

[0116] In the embodiments of the present disclosure, the commodity combination samples may be expanded through a blacklist and a whitelist, so that the commodity combination samples have diversity, which helps to improve the accuracy of the training results.

[0117] Step 209: recommend the target product group to the user according to the quality score.

[0118] This step can refer to the following relationship description of step 105, which will not be repeated here.

[0119] In summary, based on Example 1, the embodiment of the present disclosure provides another method for recommending product groups. In addition to the beneficial effects of Example 1, the embodiment of the present disclosure can also extract main information from product names containing redundant information, and splice product names, labels, and other continuous features into individual feature vectors, so that the individual feature vectors contain information of various dimensions, which helps to improve the accuracy of the quality score; it can also take into account both click-through rate and conversion rate during training, and determine the quality score based on the comprehensive click-through rate and conversion rate, so that the accuracy of the quality score is better; it can also achieve recommendations for different goals; it can also dynamically adjust the first preset weight and the second preset weight, which helps to improve the accuracy of the quality score; it can also filter the product combination samples to ensure the quality of the product combination samples, which helps to improve the accuracy of the training results; it can also expand the product combination samples through blacklists and whitelists, so that the product combination samples have diversity, which helps to improve the accuracy of the training results.

[0120] Embodiment 3

[0121] Reference Figure 4 , which shows a structural diagram of a device for recommending a commodity group in another embodiment of the present disclosure, as follows.

[0122] The individual feature vector generating module 301 is used to generate an individual feature vector for each commodity in the target commodity group according to the individual features corresponding to the commodity.

[0123] The commodity weight determination module 302 is used to determine the weight of each commodity in the target commodity group according to the individual feature vector of each commodity.

[0124] The comprehensive feature vector generating module 303 is used to generate a comprehensive feature vector of the target product group according to the individual feature vector of each product, the weight of each product in the target product group, and the comprehensive features of the target product group.

[0125] The quality score prediction module 304 is used to input the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target product group. The activation function includes a first activation sub-function for click-through rate and a second activation sub-function for conversion rate. During the training process, the loss value is determined based on the click-through rate sample values ​​pre-marked in the product combination sample, the click-through rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample values, and the conversion rate prediction value output by the second activation sub-function.

[0126] The product group recommendation module 305 is used to recommend the target product group to the user according to the quality score.

[0127] In summary, an embodiment of the present disclosure provides a device for recommending a product group, the device comprising: an individual feature vector generation module, for generating an individual feature vector for each product in a target product group according to the individual features corresponding to the product; a product weight determination module, for determining the weight of each product in the target product group according to the individual feature vector of each product; a comprehensive feature vector generation module, for generating a comprehensive feature vector of the target product group according to the individual feature vector of each product, the weight of each product in the target product group, and the comprehensive features of the target product group; a quality score prediction module, for inputting the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target product group, the activation function comprising a first activation sub-function for click-through rate and a second activation sub-function for conversion rate, and during the training process, determining the loss value according to the click-through rate sample value pre-marked in the product combination sample, the click-through rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample value, and the conversion rate prediction value output by the second activation sub-function; and a product group recommendation module, for recommending the target product group to a user according to the quality score. The embodiments of the present disclosure can estimate the quality score of a product combination according to the characteristics of each product in the product combination, which helps to improve the accuracy of recommendation and provide better flexibility for the product combination.

[0128] Embodiment 3 is a device embodiment corresponding to Embodiment 1. The detailed description may refer to Embodiment 1 and will not be repeated here.

[0129] Embodiment 4

[0130] Reference Figure 5 , which shows a structural diagram of a device for recommending a commodity group in an embodiment of the present disclosure, as follows.

[0131] The individual feature vector generation module 401 is used to generate an individual feature vector for each commodity in the target commodity group according to the individual features corresponding to the commodity; optionally, in an embodiment of the present disclosure, the individual feature vector generation module 401 includes:

[0132] The convolution and maximization submodule 4011 is used to input the product name into the convolutional neural network and the maximum pooling layer for each product in the target product group to predict a first sub-vector.

[0133] The embedding prediction submodule 4012 is used to input the label into the embedding layer to predict the second sub-vector.

[0134] The individual feature vector concatenation submodule 4013 is used to concatenate the first subvector, the second subvector, and the third vector composed of the continuous features to obtain an individual feature vector.

[0135] The commodity weight determination module 402 is used to determine the weight of each commodity in the target commodity group according to the individual feature vector of each commodity.

[0136] The comprehensive feature vector generating module 403 is used to generate a comprehensive feature vector of the target product group according to the individual feature vector of each product, the weight of each product in the target product group, and the comprehensive features of the target product group.

[0137] The quality score prediction module 404 is used to input the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target product group, wherein the activation function includes a first activation sub-function for click-through rate and a second activation sub-function for conversion rate. During the training process, the loss value is determined according to the click-through rate sample value pre-marked in the product combination sample, the click-through rate prediction value output by the first activation sub-function, the pre-marked conversion rate sample value, and the conversion rate prediction value output by the second activation sub-function; optionally, in an embodiment of the present disclosure, the quality score prediction module 404 includes:

[0138] The click rate prediction submodule 4041 is used to input the comprehensive feature vector into the first activation subfunction to predict the click rate.

[0139] The conversion rate prediction submodule 4042 is used to input the comprehensive feature vector into the second activation subfunction to predict the conversion rate.

[0140] The quality score calculation submodule 4043 is used to calculate the quality score according to the click rate and conversion rate.

[0141] The product group recommendation module 405 is used to recommend the target product group to the user according to the quality score.

[0142] Optionally, in another embodiment of the present disclosure, the quality score calculation submodule 4043 includes:

[0143] The visit-to-purchase rate score prediction unit is used to determine the visit-to-purchase rate score according to the product of the click rate and the conversion rate.

[0144] The quality score prediction unit is used to use the click rate, the conversion rate or the visit-to-purchase rate score as the quality score.

[0145] In another embodiment of the present disclosure, the loss value is calculated according to the following formula:

[0146] The loss value is calculated according to the following formula:

[0147]

[0148] TrainLoss is the loss value, M is the number of product combination samples, w1 is the first preset weight, w2 is the second preset weight, y′ i,1 is the click rate prediction value of the first activation subfunction for the i-th product combination sample, y′ i,2 is the conversion rate prediction value output by the second activation subfunction for the i-th product combination sample, y i,1 ,y i,2 are the click rate sample value and conversion rate sample value pre-labeled for the i-th product combination sample, y i,1 ·y i,2 is the sample value of the visit-purchase rate corresponding to the i-th product combination sample, y′ i,1 y′ i,2 is the predicted value of the visit-to-purchase rate corresponding to the i-th product combination sample.

[0149] Optionally, in another embodiment of the present disclosure, the first preset weight and the second preset weight are determined by the following modules:

[0150] The variance calculation module is used to calculate the variance of the click rate prediction value and the variance of the visit-to-purchase rate prediction value based on multiple commodity combination samples, and obtain a first variance and a second variance.

[0151] The first weight adjustment module is configured to reduce the first preset weight and increase the second preset weight if the first variance is greater than the second variance.

[0152] The second weight adjustment module is configured to increase the first preset weight and decrease the second preset weight if the first variance is smaller than the second variance.

[0153] In another embodiment of the present disclosure, before training the first activation sub-function and the second activation sub-function, the following modules are also included:

[0154] The sample filtering module is used to filter the product combination samples according to preset rules, wherein the preset rules include one or more of the following rules: filtering rules based on conversion rates within a preset historical time period, filtering rules based on click rates within a preset historical time period.

[0155] In another embodiment of the present disclosure, before training the first activation sub-function and the second activation sub-function, the following modules are also included:

[0156] The first sample expansion module is used to expand the sample of the commodity combination through a preset blacklist.

[0157] The second sample expansion module is used to expand the sample of the commodity combination through a preset white list.

[0158] In summary, based on the third embodiment, the embodiment of the present disclosure provides another device for recommending a product group. In addition to the beneficial effects of the third embodiment, the embodiment of the present disclosure can also extract main information from the product name containing redundant information, and splice the product name, label, and other continuous features into an individual feature vector, so that the individual feature vector contains information of various dimensions, which helps to improve the accuracy of the quality score; splice the individual feature vector and the comprehensive features of the target product group into a comprehensive feature vector of the target product group, so that the comprehensive feature vector contains information of various dimensions, which helps to improve the accuracy of the quality score; it can also take into account both the click-through rate and the conversion rate during training, and determine the quality score by combining the click-through rate and the conversion rate, so that the accuracy of the quality score is better; it can also achieve recommendations for different goals; it can also dynamically adjust the first preset weight and the second preset weight, which helps to improve the accuracy of the quality score; it can also filter the product combination samples to ensure the quality of the product combination samples, which helps to improve the accuracy of the training results; it can also expand the product combination samples through blacklists and whitelists, so that the product combination samples have diversity, which helps to improve the accuracy of the training results.

[0159] Embodiment 4 is a device embodiment corresponding to Embodiment 2. The detailed description may refer to Embodiment 2 and will not be repeated here.

[0160] The embodiment of the present disclosure also provides an electronic device, referring to Figure 6 , including: a processor 501, a memory 502, and a computer program 5021 stored in the memory 502 and executable on the processor, wherein when the processor 501 executes the program, the method for recommending a commodity group of the aforementioned embodiment is implemented.

[0161] An embodiment of the present disclosure further provides a readable storage medium. When instructions in the storage medium are executed by a processor of an electronic device, the electronic device can execute the method for recommending a commodity group of the aforementioned embodiment.

[0162] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0163] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system or other device. Various general purpose systems may also be used together with the teachings based thereon. According to the above description, it is apparent that the structure required for constructing such systems is. In addition, the embodiments of the present disclosure are not directed to any particular programming language either. It should be understood that the contents of the embodiments of the present disclosure described herein may be realized using various programming languages, and the description of the specific languages ​​above is intended to disclose the best mode of implementation of the embodiments of the present disclosure.

[0164] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the embodiments of the present disclosure can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0165] Similarly, it should be understood that in order to streamline the present disclosure and aid in understanding one or more of the various inventive aspects, in the above description of the exemplary embodiments of the present disclosure, the various features of the embodiments of the present disclosure are sometimes grouped together into a single embodiment, figure, or description thereof. However, the disclosed method should not be interpreted as reflecting the following intention: that the claimed embodiments of the present disclosure require more features than the features explicitly recited in each claim. More specifically, as reflected in the claims below, the inventive aspects lie in less than all the features of the single embodiment disclosed above. Therefore, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, with each claim itself as a separate embodiment of the embodiments of the present disclosure.

[0166] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and in addition they may be divided into a plurality of submodules or subunits or subcomponents. Except that at least some of such features and / or processes or units are mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed in this manner may be combined in any combination. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0167] The various component embodiments of the embodiments of the present disclosure may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components in the recommendation device of the commodity group according to the embodiments of the present disclosure. The embodiments of the present disclosure may also be implemented as a device or apparatus program for executing part or all of the methods described herein. Such a program implementing an embodiment of the present disclosure may be stored on a computer-readable medium, or may be in the form of one or more signals. Such a signal may be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0168] It should be noted that the above embodiments illustrate rather than limit the embodiments of the present disclosure, and that those skilled in the art may design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets should not be constructed as a limitation on the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of multiple such elements. The embodiments of the present disclosure may be implemented by means of hardware including several different elements and by means of appropriately programmed computers. In a unit claim that lists several devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order. These words may be interpreted as names.

[0169] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0170] The above description is only a preferred embodiment of the embodiments of the present disclosure and is not intended to limit the embodiments of the present disclosure. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the embodiments of the present disclosure should be included in the protection scope of the embodiments of the present disclosure.

[0171] The above is only a specific implementation of the embodiment of the present disclosure, but the protection scope of the embodiment of the present disclosure is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed by the embodiment of the present disclosure, which should be included in the protection scope of the embodiment of the present disclosure. Therefore, the protection scope of the embodiment of the present disclosure should be based on the protection scope of the claims.

Claims

1. A method for recommending a product group, characterized in that: The method comprises: For each product in the target product group, generating an individual feature vector according to the individual features corresponding to the product; Determining the weight of each commodity in the target commodity group according to the individual feature vector of each commodity; Generate a comprehensive feature vector of the target product group according to the individual feature vector of each product, the weight of each product in the target product group, and the comprehensive features of the target product group; The comprehensive feature vector is input into a pre-trained activation function to predict the quality score of the target product group, wherein the activation function includes a first activation sub-function for click-through rate and a second activation sub-function for conversion rate. During the training process, a loss value is determined according to a click-through rate sample value pre-marked in a product combination sample, a click-through rate prediction value output by the first activation sub-function, a pre-marked conversion rate sample value, and a conversion rate prediction value output by the second activation sub-function. Specifically, a squared loss value between a click-through rate prediction value output by the first activation sub-function and the pre-marked click-through rate sample value is weighted with a squared loss value between a conversion rate prediction value output by the second activation sub-function and the pre-marked conversion rate sample value to determine the loss value. The target product group is recommended to the user according to the quality score.

2. The method according to claim 1, characterized in that The individual features include a product name, a label corresponding to the product, and a continuous feature. The step of generating an individual feature vector according to the individual features corresponding to the product includes: Input the product name into a convolutional neural network and a maximum pooling layer to predict a first sub-vector; Input the label into the embedding layer to predict a second sub-vector; The first sub-vector, the second sub-vector, and a third vector composed of the continuous features are concatenated to obtain an individual feature vector.

3. The method according to claim 1, characterized in that The step of inputting the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target product group includes: Inputting the comprehensive feature vector into the first activation subfunction to predict the click rate; Inputting the comprehensive feature vector into the second activation subfunction to predict the conversion rate; A quality score is calculated based on the click rate and conversion rate.

4. The method according to claim 3, characterized in that The step of calculating the quality score according to the click rate and the conversion rate includes: Determine a visit-to-purchase rate score based on the product of the click rate and the conversion rate; The click rate, the conversion rate, or the visit-to-purchase rate score is used as a quality score.

5. The method according to any one of claims 1 to 4, characterized in that: The step of determining the loss value according to the click rate sample value pre-marked in the product combination sample, the click rate prediction value output by the first activation sub-function, the conversion rate sample value pre-marked, and the conversion rate prediction value output by the second activation sub-function during the training process includes: The loss value is calculated according to the following formula: TrainLoss is the loss value, M is the number of product combination samples, w1 is the first preset weight, w2 is the second preset weight, y' i,1 is the click rate prediction value of the first activation subfunction for the i-th product combination sample, y' i,2 is the conversion rate prediction value output by the second activation subfunction for the i-th product combination sample, y i,1 ,y i,2 are the click rate sample value and conversion rate sample value pre-labeled for the i-th product combination sample, y i,1 ·y i,2 is the sample value of the visit-purchase rate corresponding to the i-th product combination sample, y' i,1 y' i,2 is the predicted value of the visit-to-purchase rate corresponding to the i-th product combination sample.

6. The method according to claim 5, characterized in that The first preset weight and the second preset weight are determined by the following steps: Based on multiple product combination samples, respectively calculating the variance of the click rate prediction value and the variance of the visit purchase rate prediction value to obtain a first variance and a second variance; If the first variance is greater than the second variance, reducing the first preset weight and increasing the second preset weight; If the first variance is smaller than the second variance, the first preset weight is increased and the second preset weight is decreased.

7. The method according to claim 5, characterized in that Before training the first activation sub-function and the second activation sub-function, the method further includes: The commodity combination sample is filtered according to preset rules, wherein the preset rules include one or more of the following rules: a filtering rule based on a conversion rate within a preset historical time period, and a filtering rule based on a click rate within a preset historical time period.

8. The method according to claim 5, characterized in that Before training the first activation sub-function and the second activation sub-function, the method further includes: Expanding the sample of the product combination by using a preset blacklist; The commodity combination sample is expanded by using a preset whitelist.

9. A device for recommending a product group, characterized in that: The device comprises: An individual feature vector generation module, used to generate an individual feature vector for each commodity in the target commodity group according to the individual features corresponding to the commodity; A commodity weight determination module, used to determine the weight of each commodity in the target commodity group according to the individual feature vector of each commodity; A comprehensive feature vector generating module, used to generate a comprehensive feature vector of the target commodity group according to the individual feature vector of each commodity, the weight of each commodity in the target commodity group, and the comprehensive features of the target commodity group; A quality score prediction module, used for inputting the comprehensive feature vector into a pre-trained activation function to predict the quality score of the target product group, wherein the activation function includes a first activation sub-function for click-through rate and a second activation sub-function for conversion rate. During the training process, a loss value is determined according to a click-through rate sample value pre-marked in a product combination sample, a click-through rate prediction value output by the first activation sub-function, a pre-marked conversion rate sample value, and a conversion rate prediction value output by the second activation sub-function. Specifically, the loss value is determined by weighting a square loss value between a click-through rate prediction value output by the first activation sub-function and the pre-marked click-through rate sample value and a square loss value between a conversion rate prediction value output by the second activation sub-function and the pre-marked conversion rate sample value; A product group recommendation module is used to recommend the target product group to the user according to the quality score.

10. An electronic device, characterized in that: include: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for recommending a commodity group according to one or more of claims 1 to 8 is implemented.

11. A readable storage medium, characterized in that: When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method for recommending a commodity group as described in one or more of method claims 1-8.

Images