Model training method and device, equipment and storage medium

Abstract

The invention provides a model training method and device, equipment and a storage medium, and relates to the technical field of computers, in particular to the technical fields of data processing, large language models, intelligent e-commerce and the like. According to the specific implementation scheme, a plurality of commodity conversion samples are extracted from a plurality of information samples; inputting the plurality of commodity conversion samples into a to-be-trained conversion amount prediction model to obtain a plurality of first prediction values, the first prediction values being related to a commodity conversion prediction amount and a commodity minimum amount; determining a loss function according to the plurality of first predicted values; and training a conversion amount prediction model based on the loss function. The accuracy of predicting the commodity conversion amount can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to technical fields such as data processing, large language models, and smart e-commerce. Background Art

[0002] In recent years, the rapid development of the e-commerce industry has led to an increase in the variety of goods and an expansion of the price range of various goods. This phenomenon has led to a decrease in the accuracy of product conversion amount prediction and the prediction results sometimes deviate from reality, which in turn affects the results of sorting product information. Therefore, how to accurately predict the product conversion amount and make the prediction results meet actual requirements is an urgent problem to be solved. Summary of the invention

[0003] The present invention provides a model training method, apparatus, device and storage medium.

[0004] According to one aspect of the present disclosure, a model training method is provided, comprising:

[0005] Extract multiple commodity conversion samples from multiple information samples;

[0006] Inputting a plurality of commodity conversion samples into a conversion amount prediction model to be trained to obtain a plurality of first prediction values, wherein the first prediction values ​​are related to the commodity conversion prediction amount and the commodity minimum amount;

[0007] Determining a loss function based on the plurality of first prediction values;

[0008] The conversion amount prediction model is trained based on the loss function.

[0009] According to another aspect of the present disclosure, there is provided an information sorting method, comprising:

[0010] obtaining a plurality of first information;

[0011] Determine a first prediction value of each piece of first information by using a pre-trained conversion amount prediction model;

[0012] Determine, according to the first prediction value of each piece of first information, a predicted amount of commodity conversion corresponding to each piece of first information;

[0013] sorting the plurality of first information based on the predicted commodity conversion amount corresponding to each piece of first information;

[0014] The conversion amount prediction model is obtained by training using any training method provided in the present disclosure.

[0015] According to another aspect of the present disclosure, a model training device is provided, comprising:

[0016] A sample extraction module, used to extract multiple commodity conversion samples from multiple information samples;

[0017] An input module, used for inputting a plurality of commodity conversion samples into a conversion amount prediction model to be trained to obtain a plurality of first prediction values, wherein the first prediction values ​​are related to the commodity conversion prediction amount and the commodity minimum amount;

[0018] A loss function determination module, used to determine a loss function according to a plurality of first prediction values;

[0019] The training module is used to train the conversion amount prediction model based on the loss function.

[0020] According to another aspect of the present disclosure, there is provided an information sorting device, comprising:

[0021] An acquisition module, used for acquiring a plurality of first information;

[0022] A first determination module, used to determine a first prediction value of each first information by using a pre-trained conversion amount prediction model;

[0023] A second determination module is used to determine the commodity conversion prediction amount corresponding to each first information according to the first prediction value of each first information;

[0024] A sorting module, used to sort the multiple first information based on the predicted commodity conversion amount corresponding to each first information;

[0025] The conversion amount prediction model is obtained by training using any training device provided in the present disclosure.

[0026] According to another aspect of the present disclosure, there is provided an electronic device, comprising:

[0027] at least one processor; and

[0028] a memory communicatively connected to the at least one processor; wherein,

[0029] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute any method in the embodiments of the present disclosure.

[0030] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute any method according to the embodiments of the present disclosure.

[0031] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which implements any method according to the embodiments of the present disclosure when executed by a processor.

[0032] The present invention extracts commodity conversion samples from multiple information samples, and uses the conversion amount prediction model to be trained to predict multiple first prediction values ​​corresponding to the commodity conversion samples. The first prediction value is related to the commodity conversion prediction amount and the commodity minimum amount. Since the first prediction value is related to the commodity minimum amount, by limiting the value range of the first prediction value, the lower limit of the commodity conversion prediction amount corresponding to each commodity conversion sample can be indirectly limited, so that the commodity conversion prediction amount is not less than the commodity minimum amount, so that the prediction result meets the reasonable lower limit requirement, thereby improving the reliability of the prediction result.

[0033] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0035] Figure 1 is a schematic diagram of an application scenario according to an embodiment of the present disclosure;

[0036] Figure 2 is a flow chart of an implementation of a model training method according to an embodiment of the present disclosure;

[0037] Figure 3 is a schematic diagram of an overall process according to an embodiment of the present disclosure;

[0038] Figure 4 is a comparison diagram of a first Poisson distribution and a second Poisson distribution according to an embodiment of the present disclosure;

[0039] Figure 5 is a flowchart of an implementation method of an information sorting method according to an embodiment of the present disclosure;

[0040] Figure 6 is a structural schematic diagram of a model training device 600 according to an embodiment of the present disclosure;

[0041] Figure 7 is a structural diagram of an information sorting device 700 according to an embodiment of the present disclosure;

[0042] Figure 8 A schematic block diagram of an example electronic device 800 that may be used to implement embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0043] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure 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 of the present disclosure. Similarly, for the sake of clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0044] The “and / or” in the embodiments of the present disclosure indicates that three relationships may exist. For example, A and / or B may indicate the following three situations: A exists alone, A and B exist at the same time, and B exists alone. The term “at least one” herein indicates any combination of at least two of any one or more of a plurality of types. For example, at least one of A, B, and C may indicate any one or more elements selected from the set consisting of A, B, and C. The terms “first” and “second” herein refer to and distinguish between multiple similar technical terms, and do not mean to limit the order or to limit the meaning to only two. For example, the first feature and the second feature refer to two types / two features. The first feature may be one or more, and the second feature may also be one or more.

[0045] In recent years, the rapid development of the e-commerce industry has increased the number of commodity categories and the price range between commodities in different categories has also become larger and larger. From high-end expensive jewelry and luxury goods to low-end affordable daily consumer goods, the distribution of commodity transaction amounts presents a multi-layered feature. At the same time, consumer purchasing behavior is becoming increasingly diversified, and the frequent holding of promotional activities has made a single commodity often presented in the form of a variety of package combinations. For example, a certain shampoo may have a variety of packages with different prices, such as two bottles and three bottles, for consumers to choose from. The above phenomena have increased the difficulty of predicting the commodity conversion amount.

[0046] Currently, existing commodity conversion amount prediction methods can include:

[0047] A method for predicting the conversion amount of goods based on a regression model. In the application scenario of the regression model, in order to make the prediction results meet the lower limit constraint requirements of the amount in the actual business, the truncation method is used to adjust the output of the model. The basic idea of ​​this method is to truncate the prediction results of the model, that is, when the prediction value is lower than the preset lower limit, it is directly corrected to the lower limit. On the surface, this processing method can make all prediction results meet the conditions of the lower limit constraint.

[0048] However, there is a significant problem in the actual application of the truncation method, which is that a large number of predicted values ​​are forced to be pulled down to near the lower limit. This trend not only weakens the model's ability to distinguish the actual conversion price differences of different commodities, but also affects the accuracy and practicality of the prediction. Because in the actual business environment, the conversion prices of commodities often show diverse characteristics, and the truncation method cannot accurately capture and reflect this difference.

[0049] In order to overcome the limitations of the truncation method, another idea is to constrain the output of the model by selecting a suitable activation function. As an important component of the neural network, the selection of the activation function has a crucial impact on the performance of the model. By selecting and designing the activation function, the output of the model can be kept above the specified lower limit value, thereby meeting the requirements of the lower limit constraint.

[0050] However, the actual conversion price lower limits of different commodities are different. If a prediction model is built separately for each commodity and a corresponding activation function is set, it will not only greatly increase the complexity of model development and maintenance, but also reduce the flexibility and applicability of the model.

[0051] In addition, models based on deep learning, such as deep neural network (DNN), convolutional neural network (CNN), graph neural network (GNN), or models using reinforcement learning technology, can directly predict the conversion amount of goods, but cannot effectively meet the lower limit constraint.

[0052] Therefore, the current existing product conversion amount prediction methods cannot meet the lower limit constraint while making accurate predictions, which in turn affects the application of the product conversion amount. For example, the display order of the corresponding information of each product is determined by the product conversion amount corresponding to each product.

[0053] In order to solve the above problems, the embodiments of the present disclosure propose a model training method and an information sorting method. Figure 1 is a schematic diagram of an application scenario according to an embodiment of the present disclosure, such as Figure 1As shown, the application scenario schematic diagram of the embodiment of the present disclosure may include but is not limited to a model training device 110 and an information sorting device 120, and the model training device 110 and the information sorting device 120 can communicate with each other through any type of wired or wireless network. Specifically, the model training device 110 can be used to train the conversion amount prediction model, and send the trained conversion amount prediction model to the information sorting device 120; the information sorting device 120 can load and install the conversion amount prediction model, and after receiving the information corresponding to each commodity input by the user, each information is input into the conversion amount prediction model, and the conversion amount prediction model outputs the commodity conversion prediction amount corresponding to each information, and sorts each information based on the commodity conversion prediction amount corresponding to each information. Among them, the information sorting device 120 proposed in the embodiment of the present disclosure includes but is not limited to electronic devices such as mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, game consoles, e-book readers, multimedia playback devices, wearable devices, etc.; the model training device 110 may include an electronic device or server for providing background management for the information sorting device 120. In addition, the embodiments of the present disclosure do not impose specific restrictions on the number of model training devices 110 and information sorting devices 120. For example, the application scenario diagram of the embodiments of the present disclosure may include one or more model training devices 110 and one or more information sorting devices 120.

[0054] Figure 2 is a flow chart of an implementation of a model training method according to an embodiment of the present disclosure, including:

[0055] S210, extracting multiple commodity conversion samples from multiple information samples;

[0056] S220, inputting a plurality of commodity conversion samples into a conversion amount prediction model to be trained to obtain a plurality of first prediction values, wherein the first prediction values ​​are related to the commodity conversion prediction amount and the commodity minimum amount;

[0057] S230, determining a loss function according to a plurality of first prediction values;

[0058] S240: training the conversion amount prediction model based on the loss function.

[0059] It can be seen from the above training method that during the training process of the conversion amount prediction model, the model can output the first prediction value corresponding to each commodity conversion sample, and the first prediction value is related to the commodity conversion prediction amount and the commodity minimum amount. Since the first prediction value is related to the commodity minimum amount, by limiting the value range of the first prediction value, the lower limit of the commodity conversion prediction amount corresponding to each commodity conversion sample can be indirectly limited, so that the commodity conversion prediction amount is not less than the commodity minimum amount, so that the prediction result meets the reasonable lower limit requirement, thereby improving the reliability of the prediction result.

[0060] Figure 3 It is a schematic diagram of the overall process according to an embodiment of the present disclosure. In the overall process of the embodiment of the present disclosure, the goal is to calculate the effective cost per mille (eCPM) corresponding to each product information. In this process, the conversion rate (Central Video Recorder, CVR) prediction module (also known as the CVR tower) is mainly used to predict the CVR of the product click sample, and the CVR tower is trained based on the prediction results; and the conversion amount prediction model is used to predict the gross merchandise volume (GMV) of the product conversion sample, and the conversion amount prediction model is trained based on the prediction results. In the embodiment of the present disclosure, the conversion amount prediction model can also be called the GMV tower.

[0061] In some implementations, extracting multiple commodity conversion samples from multiple information samples includes:

[0062] Obtain multiple information samples and determine the user behavior corresponding to each information sample;

[0063] When the user behavior corresponding to the information sample is a conversion behavior, the information sample is determined as a commodity conversion sample;

[0064] Extract multiple product conversion samples.

[0065] In some embodiments, it further comprises:

[0066] Record the actual amount of product conversion corresponding to each product conversion sample.

[0067] First, it is necessary to collect information samples from various channels and sources. These information samples may come from users' online behavior records, transaction records, browsing records, search records, click records, etc., and each information sample contains user data, which is helpful for subsequent analysis of user behavior and then determining product conversion samples.

[0068] After obtaining multiple information samples, the sample screening module 311 in the sample processing module 310 analyzes each sample to determine its corresponding user behavior. User behavior includes browsing, clicking, adding to cart, conversion, etc. In this process, data analysis techniques and algorithms such as machine learning and deep learning can be used. These technologies can help us extract valuable information from massive data and accurately determine the user behavior corresponding to each information sample.

[0069] After determining the user behavior corresponding to each information sample, it is necessary to further mark the user behavior of the product sample using the sample marking module 312. Click behavior refers to the behavior of a user clicking on a certain product information; conversion behavior refers to the process of a user changing from initial behavior such as browsing and clicking to final behavior such as actual purchase and payment.

[0070] In order to identify product conversion samples, clear conversion behavior definitions and identification criteria can be set. For example, the behavior of a user submitting an order and successfully paying according to an information sample corresponding to a product can be defined as a conversion behavior, and the corresponding information sample can be identified as a product conversion sample.

[0071] After determining the commodity conversion samples, it is necessary to extract these samples from the original information samples and record the commodity conversion amount corresponding to each commodity conversion sample for subsequent analysis and processing.

[0072] By using the above method, by identifying user behavior, it is possible to distinguish information samples with click behavior, and further extract information samples with conversion behavior from click behavior samples, thereby providing a basis for screening product conversion samples. At the same time, product click samples can be used to train the CVR tower, and product conversion samples can be used to train the GMV tower, thereby improving the prediction accuracy of each prediction model.

[0073] In some embodiments, the conversion amount prediction model includes an embedding module, a neural gate module, and a prediction module;

[0074] Inputting multiple commodity conversion samples into the conversion amount prediction model to be trained to obtain multiple first prediction values, including:

[0075] For each product conversion sample, perform the following operations:

[0076] The commodity conversion sample is input into the embedding module, and the feature vector of the commodity conversion sample is extracted by using the embedding module; the weight corresponding to each element in the feature vector is set by using the neural gate module; and the first prediction value is determined by using the prediction module based on the feature vector and the weight corresponding to each element.

[0077] In the embodiment of the present disclosure, the product click samples and product conversion samples are the outputs of the previous steps, and include user purchase behavior data, which are the basis for subsequent analysis and prediction. In order to extract the features of each product click sample and product conversion sample from these data, a sample mapping unit 320 is required.

[0078] The sample mapping unit 320 includes an embedding module 321 and a neural gate module 322 .

[0079] The core function of the embedding module 321 is to convert high-dimensional and sparse commodity sample data into a feature vector with high position density. In this process, not only the dimension of the data is reduced, but also the key information and internal structure of the original data are preserved. Through the processing of the embedding module 321, each commodity sample is mapped to a feature vector of fixed length.

[0080] After obtaining the feature vector, it is necessary to assign weights to the elements in the feature vector, which can be achieved through the neural gate module 322 in the model.

[0081] The neural gate module 322 is responsible for dynamically adjusting the weights of each element in the feature vector. The weight distribution is not fixed, but is determined based on the feature vector itself, the requirements of the prediction task, and the model parameters. Through the processing of the neural gate module 322, those feature elements that have an important impact on the prediction results will be assigned higher weights, while those feature elements with less impact will be assigned lower weights. At the same time, as the model is iterated, the weights corresponding to the feature elements will also change dynamically.

[0082] After obtaining the weighted feature vector, the multi-task prediction unit 330 can be used to predict the CVR of the product click sample and the GMV of the product conversion sample respectively. In order to predict the CVR of each product click sample, the CVR tower prediction module 331 can be used, which makes predictions based on the feature vector of the product click sample and the weights corresponding to its internal elements; in order to predict the GMV of each product conversion sample, the GMV tower prediction module 332 can be used, which makes predictions based on the conversion amount, feature vector and weights corresponding to its internal elements of each product conversion sample. Both the CVR tower prediction module 331 and the GMV tower prediction module 332 contain multiple neural network structures such as fully connected layers and activation functions. These structures work together to convert feature vectors and weights into final prediction values.

[0083] For the CVR tower, the CVR prediction values ​​output by the CVR tower prediction module 331 can be used to construct a cross entropy loss, and the CVR tower can be trained based on the cross entropy loss, thereby improving the accuracy of the CVR tower's prediction of the CVR of the product. Correspondingly, for the GMV tower, the first prediction values ​​output by the GMV tower prediction module 332 can be used to construct a loss function, and the GMV tower can be trained based on the loss function, thereby improving the accuracy of the prediction of the GMV of the product. Through continuous training and optimization, the CVR tower prediction module 331 and the GMV tower prediction module 332 can learn how to extract useful information from the feature vector and accurately predict the target value.

[0084] After completing the training of the CVR tower and the GMV tower, the product information can be input into the CVR tower and the GMV tower respectively, and the CVR and GMV corresponding to the product information can be output to calculate the eCPM of the product.

[0085] By using the above method, by converting product samples into feature vectors, not only the dimension of the data is reduced, but also the key information and internal structure of the original data are retained, thereby improving the efficiency of feature extraction. In addition, by setting weights for each element in the feature vector, the model can focus on the features that affect the prediction results, thereby improving the flexibility of the model.

[0086] In the disclosed embodiment, the loss function construction and training process of the GMV tower are introduced in detail.

[0087] In some implementations, the first prediction value is related to the predicted amount of commodity conversion and the minimum amount of commodity, including:

[0088] The first predicted value satisfies a first functional relationship with the predicted amount of commodity conversion and the minimum amount of commodity.

[0089] In some implementations, the first predicted value obtained by the conversion amount prediction model belongs to a first value range, and the first value range is related to the first functional relationship.

[0090] In some implementations, the first functional relationship includes: the first prediction value is equal to the first ratio minus 1, the first ratio being the ratio of the predicted amount of commodity conversion to the minimum amount of the commodity;

[0091] The first value range is greater than or equal to 0.

[0092] In the embodiment of the present disclosure, in order to make the prediction result of the conversion amount prediction model reach a reasonable minimum standard, a first functional relationship is defined to specifically illustrate the correlation between the first prediction value, the commodity conversion prediction amount and the commodity minimum amount. The first functional relationship is shown in formula (1):

[0093]

[0094] Among them, PV represents the first predicted value, GMV represents the total transaction amount of the product determined by the predicted amount of product conversion, and min_sku_price is the minimum amount of the product. If the product is sold in single form, min_sku_price is the product price itself. If the product is sold in the form of a package or combination, min_sku_price is the minimum amount of the package or combination. When the value range of the first predicted value is greater than or equal to 0, the predicted amount of product conversion is greater than or equal to the minimum amount of the product.

[0095] Through the above method, the product conversion prediction amount predicted by the model can be reasonable, logical and practical, and this method can narrow the model's prediction range for the first prediction value, thereby improving the accuracy of the product conversion prediction amount and providing a data basis for the subsequent GMV tower training.

[0096] In some implementations, determining a loss function according to a plurality of first prediction values ​​includes:

[0097] Fitting the distribution of the plurality of first prediction values ​​to a first Poisson distribution;

[0098] Determine the real amount of commodity conversion of each commodity conversion sample, and determine the first real value of each commodity conversion sample according to the real amount of commodity conversion, the minimum amount of commodities and the first functional relationship;

[0099] Fitting the distribution of the plurality of first real values ​​to a second Poisson distribution;

[0100] A loss function is determined according to the first Poisson distribution and the second Poisson distribution.

[0101] The conversion amount prediction model is used to predict the first prediction values ​​for multiple commodity conversion samples. In order to deeply understand the distribution characteristics of these first prediction values, the Poisson distribution in statistics is used for fitting.

[0102] Poisson distribution is a discrete probability model that is suitable for describing the number of random events that occur in a fixed time and space. In the embodiments of the present disclosure, it can be used to describe the distribution of multiple first prediction values. Through fitting, a first Poisson distribution corresponding to multiple first prediction values ​​can be obtained, which reveals the distribution law of the first prediction values ​​in the population.

[0103] Next, it is necessary to determine the real amount of commodity conversion for each commodity conversion sample. These real amounts represent the actual economic value of commodity conversion. At the same time, it is also necessary to determine the minimum amount of each commodity. Based on the real amount of commodity conversion and the minimum amount of each commodity, combined with the defined first functional relationship, the first real value of each commodity conversion sample can be calculated. With these first real values, the Poisson distribution is used for fitting again to obtain the second Poisson distribution, which describes the distribution of the first real value in the population.

[0104] According to the first Poisson distribution and the second Poisson distribution, a loss function can be constructed. The loss function is a function used in fields such as machine learning to evaluate the difference between the model prediction result and the actual result. In the embodiment of the present disclosure, the first Poisson distribution can be used as the prediction result, and the second Poisson distribution can be used as the actual result. By comparing the difference between the first Poisson distribution and the second Poisson distribution, the Poisson distribution negative log-likelihood loss (PoissonNLLLoss) between the two Poisson distributions is used as the loss function, and a loss function that can reflect the prediction accuracy and error can be constructed.

[0105] In some embodiments, determining a loss function according to a first Poisson distribution and a second Poisson distribution includes:

[0106] Calculate the Poisson distribution negative log-likelihood loss between the first Poisson distribution and the second Poisson distribution;

[0107] Poisson distribution negative log-likelihood loss is used as the loss function.

[0108] Figure 4 is a comparison diagram of the first Poisson distribution and the second Poisson distribution according to an embodiment of the present disclosure. Figure 4 As shown, the first Poisson distribution is X1~λ1, and the second Poisson distribution is X2~λ2, where X1 and X2 are random variables in the first predicted value and the first true value, respectively, λ1 and λ2 represent the parameters of the first Poisson distribution and the second Poisson distribution, respectively, the horizontal axis in the figure represents the predicted target, and the vertical axis represents the probability of the target occurring. For example, since the predicted amount of commodity conversion is similar to the minimum amount of commodity, and the actual amount of commodity conversion is similar to the minimum amount of commodity, after being constrained by the first functional relationship, the probability of the first predicted value and the first true value both appearing around 0 is relatively high.

[0109] Correspondingly, the probability density function of the first Poisson distribution is:

[0110]

[0111] The probability density function of the second Poisson distribution is:

[0112]

[0113] The Poisson distribution negative log-likelihood loss is calculated based on the first Poisson distribution and the second Poisson distribution. If λ1 is the probability distribution parameter fitted by the model output result, and no normalization and logarithmization are performed. The loss L corresponding to the kth sample n for:

[0114] L n =―logP(X1=k)=λ1―klogλ1+log(k!) (4)

[0115] After normalization and logarithmization, the loss L corresponding to the kth sample is n for:

[0116] L n =―logP(X1=k)=exp(λ1)―kλ1+log(k!) (5)

[0117] Based on the above calculation results, the Poisson distribution negative log-likelihood loss is used as the loss function, and the conversion amount prediction model is iteratively trained based on this function.

[0118] Through this method, the difference between the model's predicted probability distribution and the true distribution can be reflected. Based on this difference, the negative log-likelihood loss of the Poisson distribution is determined to guide the model training.

[0119] The disclosed embodiment also provides an information sorting method. Figure 5 is a flowchart of an implementation method of information sorting according to an embodiment of the present disclosure, including:

[0120] S510, obtaining a plurality of first information;

[0121] S520: Determine a first prediction value of each piece of first information by using a pre-trained conversion amount prediction model;

[0122] S530: Determine the predicted conversion amount of the commodity corresponding to each first information according to the first predicted value of each first information;

[0123] S540, sorting the plurality of first information based on the predicted commodity conversion amount corresponding to each first information;

[0124] Among them, the conversion amount prediction model is trained using the above training method.

[0125] By using this method, each first information is predicted through a pre-trained conversion amount prediction model to determine the corresponding first prediction value. Since the first prediction value is related to the minimum amount of the product, by limiting the value range of the first prediction value, the product conversion prediction amount corresponding to each first information can be indirectly controlled not to be lower than the minimum amount of the product. Finally, multiple first information are sorted based on the product conversion prediction amount, thereby improving the accuracy and reliability of the sorting of the first information. Among them, the first information of the embodiment of the present disclosure includes advertisements, introduction information, etc. of the product.

[0126] In some implementations, the first predicted value satisfies a first functional relationship with the predicted amount of commodity conversion and the minimum amount of commodity;

[0127] Determining the commodity conversion forecast amount corresponding to each first information according to the first forecast value of each first information includes: determining the commodity conversion forecast amount according to the first forecast value, the first functional relationship and the minimum amount of the commodity.

[0128] In some implementations, the first functional relationship includes: the first predicted value is equal to the first ratio minus 1, and the first ratio is the ratio of the predicted amount of commodity conversion to the minimum amount of the commodity.

[0129] In the embodiment of the present disclosure, in order to set a reasonable lower limit of the prediction result of the conversion amount prediction model, a first functional relationship is defined, which specifically describes the internal relationship between the first prediction value, the commodity conversion prediction amount and the commodity minimum amount. Among them, the commodity conversion prediction amount is:

[0130] GMV = (PV + 1) min_sku_price (6)

[0131] Among them, GMV represents the total transaction amount of the product determined by the predicted conversion amount of the product, PV represents the first predicted value, min_sku_price is the minimum amount of the product. If the product is sold as a single piece, min_sku_price is the product price itself. If the product is sold as a package or combination, min_sku_price is the minimum amount of the package or combination.

[0132] By adopting the above method, by establishing the relationship between GMV and the minimum amount of goods, GMV can be made greater than the minimum amount of goods, thereby improving the reliability of the model predicting GMV, thereby providing data support for subsequent information sorting.

[0133] The above introduces the process of determining the predicted amount of commodity conversion corresponding to each first information. For each first information, the corresponding predicted revenue needs to be determined. Finally, the first information is sorted with reference to the predicted revenue corresponding to each first information.

[0134] The predicted return of First Information is determined by the following values:

[0135] (1) The predicted amount of commodity conversion, i.e., GMV, can be determined using the conversion amount prediction model proposed in this disclosure;

[0136] (2) Click-Through-Rate (CTR) of the first information, which can be obtained by an external CTR prediction model;

[0137] (3) Return on Investment (ROI) is the net benefit obtained by users through First Information, which can be determined based on the parameters set by the user;

[0138] (4) The CVR of a product is the proportion of users who have visited or browsed the first information and have purchased the product corresponding to the first information, which can be determined based on the CVR tower mentioned above.

[0139] In the disclosed embodiment, the predicted revenue can be determined by eCPM. The revenue of one effective exposure is:

[0140] M=CTR×CVR×GMV / ROI (7)

[0141] Based on the revenue from one effective exposure, eCPM can be determined in two ways:

[0142] (1) Multiply the revenue from one effective exposure by one thousand to directly get the eCPM;

[0143] (2) The conversion amount prediction model is used to predict the conversion amount of the first information multiple times, and the effective exposure income is calculated once using each prediction result, and the eCPM is determined based on the multiple effective exposure incomes.

[0144] Finally, the first information can be ranked according to the eCPM corresponding to each first information. Generally speaking, the larger the eCPM value, the higher the corresponding first information is ranked.

[0145] The present disclosure also provides a model training device. Figure 6 is a schematic diagram of the structure of a model training device 600 according to an embodiment of the present disclosure, comprising:

[0146] A sample extraction module 610 is used to extract multiple commodity conversion samples from multiple information samples;

[0147] An input module 620, for inputting a plurality of commodity conversion samples into a conversion amount prediction model to be trained to obtain a plurality of first prediction values, wherein the first prediction values ​​are related to the commodity conversion prediction amount and the commodity minimum amount;

[0148] A loss function determination module 630, configured to determine a loss function according to a plurality of first prediction values;

[0149] The training module 640 is used to train the conversion amount prediction model based on the loss function.

[0150] In some embodiments, the sample extraction module 610 is used to:

[0151] Obtain multiple information samples and determine the user behavior corresponding to each information sample;

[0152] When the user behavior corresponding to the information sample is a conversion behavior, the information sample is determined as a commodity conversion sample;

[0153] Extract multiple product conversion samples.

[0154] In some implementations, the sample extraction module 610 is further configured to:

[0155] Record the actual amount of product conversion corresponding to each product conversion sample.

[0156] In some embodiments, the conversion amount prediction model includes an embedding module, a neural gate module, and a prediction module;

[0157] The input module 620 is used to:

[0158] For each product conversion sample, perform the following operations:

[0159] The commodity conversion sample is input into the embedding module, and the feature vector of the commodity conversion sample is extracted by using the embedding module; the weight corresponding to each element in the feature vector is set by using the neural gate module; and the first prediction value is determined by using the prediction module based on the feature vector and the weight corresponding to each element.

[0160] In some implementations, the first prediction value is related to the predicted amount of commodity conversion and the minimum amount of commodity, including:

[0161] The first predicted value satisfies a first functional relationship with the predicted amount of commodity conversion and the minimum amount of commodity.

[0162] In some implementations, the first predicted value obtained by the conversion amount prediction model belongs to a first value range, and the first value range is related to the first functional relationship.

[0163] In some implementations, the first functional relationship includes: the first predicted value is equal to the first ratio minus 1, and the first ratio is the ratio of the predicted amount of commodity conversion to the minimum amount of the commodity;

[0164] The first value range is greater than or equal to 0.

[0165] In some implementations, the loss function determination module 630 is configured to:

[0166] Fitting the distribution of the plurality of first prediction values ​​to a first Poisson distribution;

[0167] Determine the real amount of commodity conversion of each commodity conversion sample, and determine the first real value of each commodity conversion sample according to the real amount of commodity conversion, the minimum amount of commodities and the first functional relationship;

[0168] Fitting the distribution of the plurality of first real values ​​to a second Poisson distribution;

[0169] A loss function is determined according to the first Poisson distribution and the second Poisson distribution.

[0170] In some implementations, the loss function determination module 630 is configured to:

[0171] Calculate the Poisson distribution negative log-likelihood loss between the first Poisson distribution and the second Poisson distribution;

[0172] The Poisson distribution negative log-likelihood loss is used as the loss function.

[0173] The disclosed embodiment also provides an information sorting device, Figure 7 is a schematic diagram of the structure of an information sorting device 700 according to an embodiment of the present disclosure, comprising:

[0174] An acquisition module 710, configured to acquire a plurality of first information;

[0175] A first determination module 720 is used to determine a first prediction value of each first information by using a pre-trained conversion amount prediction model;

[0176] The second determination module 730 is used to determine the predicted conversion amount of the commodity corresponding to each first information according to the first predicted value of each first information;

[0177] A sorting module 740, configured to sort the plurality of first information based on the predicted commodity conversion amount corresponding to each first information;

[0178] The conversion amount prediction model is obtained by training using the above-mentioned training device.

[0179] In some implementations, the first predicted value satisfies a first functional relationship with the predicted amount of commodity conversion and the minimum amount of commodity;

[0180] The second determination module 730 is used to determine the predicted amount of commodity conversion according to the first predicted value, the first functional relationship and the minimum amount of the commodity.

[0181] In some implementations, the first functional relationship includes: the first predicted value is equal to the first ratio minus 1, and the first ratio is the ratio of the predicted amount of commodity conversion to the minimum amount of the commodity.

[0182] For the description of specific functions and examples of each module and submodule of the device in the embodiment of the present disclosure, reference can be made to the relevant description of the corresponding steps in the above method embodiment, which will not be repeated here.

[0183] In the technical solution disclosed in the present invention, the acquisition, storage and application of personal information of users involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0184] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0185] Figure 8 A schematic block diagram of an example electronic device 800 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0186] like Figure 8 As shown, the device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the device 800 can also be stored. The computing unit 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.

[0187] A number of components in the device 800 are connected to the I / O interface 805, including: an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a disk, an optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0188] The computing unit 801 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 801 performs the various methods and processes described above, such as detection methods. For example, in some embodiments, the detection method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 800 via ROM 802 and / or communication unit 809. When the computer program is loaded into RAM 803 and executed by the computing unit 801, one or more steps of the detection method described above may be performed. Alternatively, in other embodiments, the computing unit 801 may be configured to perform the detection method in any other appropriate manner (e.g., by means of firmware).

[0189] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0190] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0191] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, 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 foregoing.

[0192] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0193] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0194] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0195] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0196] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A model training method, comprising: Extract multiple commodity conversion samples from multiple information samples; Inputting the plurality of commodity conversion samples into a conversion amount prediction model to be trained to obtain a plurality of first prediction values, wherein the first prediction values ​​are related to the commodity conversion prediction amount and the commodity minimum amount; Determining a loss function according to the plurality of first prediction values; The conversion amount prediction model is trained based on the loss function.

2. The method according to claim 1, wherein: The step of extracting a plurality of commodity conversion samples from a plurality of information samples includes: Acquire a plurality of the information samples, and determine a user behavior corresponding to each of the information samples; In a case where the user behavior corresponding to the information sample is a conversion behavior, determining the information sample as the commodity conversion sample; Extract a plurality of conversion samples of the commodity.

3. The method according to claim 2, further comprising: The actual amount of commodity conversion corresponding to each of the commodity conversion samples is recorded.

4. The method according to claim 3, wherein: The conversion amount prediction model includes an embedding module, a neural gate module and a prediction module; The step of inputting the plurality of commodity conversion samples into a conversion amount prediction model to be trained to obtain a plurality of first prediction values ​​includes: For each of the product conversion samples, perform the following operations: The commodity conversion sample is input into the embedding module, and the feature vector of the commodity conversion sample is extracted by using the embedding module; the weight corresponding to each element in the feature vector is set by using the neural gate module; and the first prediction value is determined by using the prediction module based on the feature vector and the weight corresponding to each element.

5. The method according to any one of claims 1 to 4, wherein: The first prediction value is related to the predicted amount of commodity conversion and the minimum amount of commodity, including: The first predicted value satisfies a first functional relationship with the predicted conversion amount of the commodity and the minimum amount of the commodity.

6. The method according to claim 5, wherein: The first prediction value obtained by the conversion amount prediction model belongs to a first value range, and the first value range is related to the first functional relationship.

7. The method according to claim 6, wherein: The first functional relationship includes: the first prediction value is equal to the first ratio minus 1, and the first ratio is the ratio of the predicted amount of the commodity conversion to the minimum amount of the commodity; The first value range is greater than or equal to 0.

8. The method according to any one of claims 5 to 7, wherein: The determining of the loss function according to the plurality of first prediction values ​​comprises: Fitting the distribution of the plurality of first prediction values ​​to a first Poisson distribution; Determine the real amount of commodity conversion of each of the commodity conversion samples, and determine the first real value of each of the commodity conversion samples according to the real amount of commodity conversion, the minimum amount of the commodity and the first functional relationship; Fitting the distribution of the plurality of first true values ​​to a second Poisson distribution; The loss function is determined according to the first Poisson distribution and the second Poisson distribution.

9. The method according to claim 8, wherein: The determining the loss function according to the first Poisson distribution and the second Poisson distribution includes: Calculating a Poisson distribution negative log-likelihood loss between the first Poisson distribution and the second Poisson distribution; The Poisson distribution negative log-likelihood loss is used as the loss function.

10. A method for sorting information, comprising: obtaining a plurality of first information; Determine a first prediction value of each of the first information by using a pre-trained conversion amount prediction model; Determine, according to the first predicted value of each of the first information, a predicted amount of commodity conversion corresponding to each of the first information; sorting the plurality of first information based on the predicted commodity conversion amounts corresponding to the respective first information; Wherein, the conversion amount prediction model is trained using any training method described in claims 1-9.

11. The method according to claim 10, wherein: The first predicted value satisfies a first functional relationship with the predicted conversion amount of the commodity and the minimum amount of the commodity; Determining the commodity conversion forecast amount corresponding to each of the first information according to the first forecast value of each of the first information includes: determining the commodity conversion forecast amount according to the first forecast value, the first functional relationship and the minimum amount of the commodity.

12. The method according to claim 10 or 11, wherein: The first functional relationship includes: the first prediction value is equal to a first ratio minus 1, and the first ratio is a ratio of the predicted conversion amount of the commodity to the minimum amount of the commodity.

13. A model training device, comprising: A sample extraction module, used to extract multiple commodity conversion samples from multiple information samples; An input module, used for inputting the plurality of commodity conversion samples into a conversion amount prediction model to be trained to obtain a plurality of first prediction values, wherein the first prediction values ​​are related to the commodity conversion prediction amount and the commodity minimum amount; A loss function determination module, used to determine a loss function according to the plurality of first prediction values; A training module is used to train the conversion amount prediction model based on the loss function.

14. The device according to claim 13, wherein: The sample extraction module is used to: Acquire a plurality of the information samples, and determine a user behavior corresponding to each of the information samples; In a case where the user behavior corresponding to the information sample is a conversion behavior, determining the information sample as the commodity conversion sample; Extract a plurality of conversion samples of the commodity.

15. The device according to claim 14, wherein the sample extraction module is further configured to: The actual amount of commodity conversion corresponding to each of the commodity conversion samples is recorded.

16. The device according to claim 15, wherein: The conversion amount prediction model includes an embedding module, a neural gate module and a prediction module; The input module is used for: For each of the product conversion samples, perform the following operations: The commodity conversion sample is input into the embedding module, and the feature vector of the commodity conversion sample is extracted by using the embedding module; the weight corresponding to each element in the feature vector is set by using the neural gate module; and the first prediction value is determined by using the prediction module based on the feature vector and the weight corresponding to each element.

17. The device according to any one of claims 13 to 16, wherein: The first prediction value is related to the predicted amount of commodity conversion and the minimum amount of commodity, including: The first predicted value satisfies a first functional relationship with the predicted conversion amount of the commodity and the minimum amount of the commodity.

18. The device according to claim 17, wherein: The first prediction value obtained by the conversion amount prediction model belongs to a first value range, and the first value range is related to the first functional relationship.

19. The device according to claim 18, wherein: The first functional relationship includes: the first prediction value is equal to the first ratio minus 1, and the first ratio is the ratio of the predicted amount of the commodity conversion to the minimum amount of the commodity; The first value range is greater than or equal to 0.

20. The device according to any one of claims 17 to 19, wherein: The loss function determination module is used to: Fitting the distribution of the plurality of first prediction values ​​to a first Poisson distribution; Determine the real amount of commodity conversion of each of the commodity conversion samples, and determine the first real value of each of the commodity conversion samples according to the real amount of commodity conversion, the minimum amount of the commodity and the first functional relationship; Fitting the distribution of the plurality of first true values ​​to a second Poisson distribution; The loss function is determined according to the first Poisson distribution and the second Poisson distribution.

21. The device according to claim 20, wherein: The loss function determination module is used to: Calculating a Poisson distribution negative log-likelihood loss between the first Poisson distribution and the second Poisson distribution; The Poisson distribution negative log-likelihood loss is used as the loss function.

22. An information sorting device, comprising: An acquisition module, used for acquiring a plurality of first information; A first determination module, configured to determine a first prediction value of each of the first information by using a pre-trained conversion amount prediction model; A second determination module, configured to determine a predicted commodity conversion amount corresponding to each of the first information according to the first predicted value of each of the first information; A sorting module, configured to sort the plurality of first information based on the predicted commodity conversion amounts corresponding to the respective first information; Wherein, the conversion amount prediction model is obtained by training using any training device described in any one of claims 13-21.

23. The device according to claim 22, wherein: The first predicted value satisfies a first functional relationship with the predicted conversion amount of the commodity and the minimum amount of the commodity; The second determination module is used to determine the predicted conversion amount of the commodity according to the first predicted value, the first functional relationship and the minimum amount of the commodity.

24. The device according to claim 22 or 23, wherein: The first functional relationship includes: the first prediction value is equal to a first ratio minus 1, and the first ratio is a ratio of the predicted conversion amount of the commodity to the minimum amount of the commodity.

25. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 12.

26. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-12.

27. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 12.

Images