New article mining method, device and storage medium

Abstract

The present disclosure provides a new item mining method, apparatus, and storage medium, relating to the field of computers. The method includes obtaining multiple item attribute combinations of a preset category; using an item multi-objective evaluation model to evaluate multiple preset objectives for each item attribute combination, where the item multi-objective evaluation model is learned from historical data of existing items of the preset category for multiple preset objectives; determining an optimal item attribute combination according to the evaluation results of the multiple preset objectives of the multiple item attribute combinations; and determining a new item of the preset category according to the optimal item attribute combination. Automatically mining new items through a machine learning model reduces the consumption of human, financial, and material resources in the design of new items, and the multi-objective learning model can output item attribute combinations that simultaneously meet multiple objectives. Through the interaction between multiple objectives, it helps to jump out of the local optimum of a single objective, and the sharing of underlying features reduces the model parameters to be adjusted and repeated calculations.

Description

Technical Field

[0001] The present disclosure relates to the field of computers, and particularly to a method and apparatus for mining new items and a storage medium. Background Art

[0002] With the rapid development of the e-commerce industry and the improvement of the efficiency of the supply chain production link, the iteration of items has been accelerating continuously, the trend of new item consumption has been rising continuously, and new items have become an important link for enterprises to attract consumers, seize the market opportunity, and enhance brand value.

[0003] Most of the current new item design methods are that product managers obtain user requirements through market research and market analysis according to the strategic needs of the enterprise. At the same time, through methods such as comment mining, the attention and requirements of users for each attribute of the item are mined, and the attributes of the item are changed based on experience to design new items, and then through trial production and sales, the design scheme of the new item is finally determined. Summary of the Invention

[0004] Embodiments of the present disclosure propose a solution for automatically mining new items through a machine learning model, which helps in the design of new items and reduces the consumption of human, financial, and material resources in the design of new items; and this model belongs to a multi-objective learning model, which can output a combination of item attributes that simultaneously meets multiple objectives, and during the learning process, through the interaction between multiple objectives, it helps to jump out of the local optimum of a single objective, and the underlying features of the multi-objective learning model are shared, which can reduce the model parameters to be adjusted and repeated calculations. In addition, through the heuristic search method, a more comprehensive combination of item attributes can be obtained, avoiding missing some combinations of item attributes, which is beneficial to mining more new items.

[0005] Some embodiments of the present disclosure propose a method for mining new items, including:

[0006] Obtaining multiple combinations of item attributes of a preset category;

[0007] Using an item multi-objective evaluation model to evaluate multiple preset objectives for each combination of item attributes, where the item multi-objective evaluation model is obtained by learning historical data of existing items of the preset category for the multiple preset objectives;

[0008] Determining an optimal combination of item attributes according to the evaluation results of multiple preset objectives of multiple combinations of item attributes;

[0009] Determining a new item of the preset category according to the optimal combination of item attributes.

[0010] In some embodiments, the obtaining of multiple combinations of item attributes of a preset category includes: using a heuristic search method to search for the item attributes of the preset category to obtain multiple combinations of item attributes of the preset category.

[0011] In some embodiments, the multi-objective evaluation model for items is obtained by learning the historical data of existing items of the preset category for the multiple preset objectives based on a multi-gated mixture-of-experts multi-objective learning model, a progressive hierarchical extraction multi-objective learning model, or a subnet routing multi-objective learning model.

[0012] In some embodiments, when the multi-objective evaluation model for items is obtained based on a multi-gated mixture-of-experts multi-objective learning model, the multi-objective evaluation model for items includes an input layer, a multi-gated mixture-of-experts layer, a tower layer, and an output layer connected in cascade, where:

[0013] The multi-gated mixture-of-experts layer includes a plurality of feedforward neural networks called experts connected to the input layer, and a plurality of gating networks connected to the input layer and the plurality of experts. Each gating network includes a gate connected to the input layer and a weighted sum calculator. The weighted sum calculator in each gating network performs a weighted sum operation on the output results of each expert according to the weight information provided by the gate in the gating network. Each gating network corresponds to a preset objective;

[0014] The tower layer includes a plurality of sub-tower layers. Each sub-tower layer is a fully connected neural network and is connected to a gating network;

[0015] The output layer includes a plurality of sub-output layers. Each sub-output layer is a fully connected neural network and is connected to a sub-tower layer.

[0016] In some embodiments, it further includes: performing one or more of the following operations before learning: initializing the weights and biases of each expert and each gate with a uniform distribution initializer; initializing the weights of the fully connected neural network of each sub-tower layer with a normal distribution initializer; initializing the weights of the fully connected neural network of each sub-output layer with a normal distribution initializer.

[0017] In some embodiments, learning the historical data of existing items of the preset category for the multiple preset objectives to obtain the multi-objective evaluation model for items includes: determining a training data set according to the historical data of existing items of the preset category, inputting the training feature data in the training data set into the multi-objective evaluation model for items, determining the total loss of the multiple preset objectives according to the gap information between the predicted value of each preset objective output by the multi-objective evaluation model for items and the true value of each preset objective in the training data set, and training the multi-objective evaluation model for items according to the total loss of the multiple preset objectives.

[0018] In some embodiments, learning the historical data of the existing items of the preset category for the multiple preset targets to obtain the multi-target evaluation model of the items includes: determining a validation data set according to the historical data of the existing items of the preset category, and after each round of training of the multi-target evaluation model of the items, using the validation data set to verify the multi-target evaluation model of the items, and evaluating the training effect using a preset evaluation function according to the verification result.

[0019] In some embodiments, learning the historical data of the existing items of the preset category for the multiple preset targets to obtain the multi-target evaluation model of the items includes: determining a test data set according to the historical data of the existing items of the preset category, testing the multi-target evaluation model of the items using the test data set, and using the multi-target evaluation model of the items that passes the test for new item mining.

[0020] In some embodiments, determining the preferred item attribute combination according to the evaluation results of the multiple preset targets of the multiple item attribute combinations includes:

[0021] Calculating the comprehensive evaluation value of the multiple preset targets of each item attribute combination according to the evaluation value of each preset target of each item attribute combination;

[0022] Determining the preferred item attribute combination from the multiple item attribute combinations according to the comprehensive evaluation value of each item attribute combination.

[0023] In some embodiments, the heuristic search methods include: genetic algorithm, particle swarm algorithm, ant colony algorithm, tabu search algorithm, simulated annealing algorithm.

[0024] In some embodiments, the multiple preset targets include: sales volume and unique visitor conversion rate.

[0025] In some embodiments, the historical data of the existing items includes: the historical sales data of the existing items, and the historical sales data of the existing items includes: the inherent attributes and sales information of the existing items.

[0026] Some embodiments of the present disclosure propose a new item mining device, including:

[0027] An attribute combination acquisition unit, configured to acquire multiple item attribute combinations of a preset category;

[0028] An evaluation unit, configured to evaluate multiple preset targets of each item attribute combination by using a multi-target evaluation model of the items, and the multi-target evaluation model of the items is obtained by learning the historical data of the existing items of the preset category for the multiple preset targets;

[0029] A preferred attribute combination determination unit, configured to determine a preferred item attribute combination according to the evaluation results of multiple preset goals of multiple item attribute combinations;

[0030] A new item determination unit, configured to determine a new item of the preset category according to the preferred item attribute combination.

[0031] In some embodiments, it further includes: a model learning unit, configured to learn the historical data of the existing items of the preset category for the multiple preset goals based on a multi-gated mixture of experts multi-objective learning model, a progressive hierarchical extraction multi-objective learning model, or a subnet routing multi-objective learning model to obtain the item multi-objective evaluation model.

[0032] In some embodiments, the attribute combination acquisition unit is configured to search for item attributes of a preset category using a heuristic search method to obtain multiple item attribute combinations of the preset category.

[0033] Some embodiments of the present disclosure propose a new item mining device, including: a memory; and a processor coupled to the memory, the processor being configured to execute the new item mining methods of each embodiment based on instructions stored in the memory.

[0034] Some embodiments of the present disclosure propose a non-transitory computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the new item mining methods of each embodiment are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. According to the following detailed description with reference to the drawings, the present disclosure can be more clearly understood.

[0036] Obviously, the drawings in the following description are only some embodiments of the present disclosure. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0037] Figure 1 A flowchart showing the new item mining method according to some embodiments of the present disclosure.

[0038] Figure 2A and 2B A schematic diagram showing the mutation and crossover of the chromosomes of each individual in the genetic algorithm according to some embodiments of the present disclosure.

[0039] Figure 3 A flowchart showing the new item mining method based on the genetic algorithm according to some embodiments of the present disclosure.

[0040] Figure 4Schematic diagram of an item multi-objective evaluation model obtained based on a multi-gated mixture-of-experts multi-objective learning model according to some embodiments of the present disclosure.

[0041] Figure 5 Schematic diagram of a learning method for an item multi-objective evaluation model according to some embodiments of the present disclosure.

[0042] Figure 6 Schematic diagram of a new item mining device according to some embodiments of the present disclosure.

[0043] Figure 7 Schematic diagram of a new item mining device according to some other embodiments of the present disclosure. Detailed implementation manners

[0044] Next, the technical solutions in the embodiments of the present disclosure will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present disclosure.

[0045] Figure 1 Schematic flowchart of a new item mining method according to some embodiments of the present disclosure.

[0046] As Figure 1 shown, the new item mining method of this embodiment includes the following steps.

[0047] In step 110, obtain multiple combinations of item attributes of a preset category.

[0048] In some embodiments, a heuristic search method is used to search for the item attributes of a preset category to obtain multiple combinations of item attributes of the preset category. The heuristic search method includes, for example, but is not limited to: genetic algorithm, particle swarm algorithm, ant colony algorithm, tabu search algorithm, simulated annealing algorithm. The main differences between various heuristic algorithms lie in the operation methods and stopping criteria during the search. Each heuristic algorithm is briefly described below, and the detailed description can refer to the related technologies.

[0049] The tabu search algorithm has a tabu list during the search, which stores the results (i.e., combinations of item attributes) that appeared in the previous search process, and these results will not be considered during the iteration to prevent search loops and getting stuck in local optima.

[0050] The simulated annealing algorithm is derived from the principle of solid annealing. The internal energy E is simulated as the objective function value f, and the temperature T evolves into the control parameter t. Starting from the initial solution i (the initial combination of item attributes) and the initial value of the control parameter t, the iteration of "generating a new solution → calculating the difference of the objective function → accepting or rejecting" is repeated for the current solution, and the value of t is gradually attenuated. The current solution at the end of the algorithm is the obtained approximate optimal solution (the optimal combination of item attributes). Among them, according to the level of temperature, when the result of the new search is worse than the result of the previous search, there is a certain probability of accepting this worse result, which helps to jump out of the local optimal solution. This probability is related to the temperature. The higher the temperature, the greater the probability. So when the temperature is high, it is equivalent to random search, and when the temperature is low, it is equivalent to local search. Moreover, after each round of search, the temperature will decrease at a certain ratio.

[0051] The ant colony algorithm is a probabilistic algorithm used to find the optimal path. The walking paths of ants represent the feasible solutions to the problem to be optimized, and all the paths of the entire ant colony constitute the solution space of the problem to be optimized (i.e., the combination of item attributes). Ants with shorter paths release more pheromone. As time goes by, the concentration of pheromone accumulated on the shorter paths gradually increases, and the number of ants choosing this path also increases. Eventually, the entire ant colony will converge to the best path under the action of positive feedback, and at this time, it corresponds to the optimal solution of the problem to be optimized (the optimal combination of item attributes).

[0052] The particle swarm algorithm simulates the behavior of a flock of birds randomly searching for food. In the particle swarm algorithm, each potential solution to the optimization problem is a bird in the search space, called a "particle". All particles have a fitness value determined by the function to be optimized, and each particle also has a velocity that determines the direction and distance of their "flight". The particle swarm algorithm is initialized as a group of random particles (the initial combination of item attributes), and then the optimal solution (the optimal combination of item attributes) is found through iteration. In each iteration, the particle updates itself by tracking two extreme values: the first is the optimal solution found by the particle itself, which is called the individual extreme value; the second is the optimal solution currently found by the entire population, which is called the global extreme value; alternatively, instead of using the entire population, a part of it can be used as the neighbors of the particle, which is called the local extreme value.

[0053] The genetic algorithm is inspired by the genetic laws of nature. Each population contains many individuals of different qualities (Individual, corresponding to the combination of item attributes in this embodiment). During the reproduction process, such as Figure 2A and 2BAs shown, the chromosomes of each individual (corresponding to the item attributes in this embodiment) are mutated, crossed over, etc. to generate a new population, and then according to the principle of survival of the fittest, excellent individuals are selected by calculating the fitness of the individuals (corresponding to evaluating and selecting the optimal item attribute combination by using the item multi-objective evaluation model in this embodiment).

[0054] As an example, hereinafter in combination with Figure 3 a new item mining method based on the genetic algorithm will be specifically described.

[0055] In step 120, the item multi-objective evaluation model is used to evaluate multiple preset objectives for each item attribute combination. The item multi-objective evaluation model is obtained by learning the historical data of the existing items of the preset category for the multiple preset objectives. The learning process will be specifically described hereinafter.

[0056] The item multi-objective evaluation model is a machine learning model and a multi-objective learning model. In some embodiments, the item multi-objective evaluation model is obtained by learning the historical data of the existing items of the preset category for the multiple preset objectives based on, for example, a multi-gate mixture-of-experts (MMoE) multi-objective learning model, a progressive layered extraction (PLE) multi-objective learning model, or a sub-network routing (SNR) multi-objective learning model.

[0057] As an example, hereinafter in combination with Figure 4 the item multi-objective evaluation model obtained based on the multi-gate mixture-of-experts multi-objective learning model will be specifically described.

[0058] In some embodiments, the multiple preset objectives include: sales volume and unique visitor (UV) conversion rate. The UV conversion rate refers to: within a statistical period, the ratio of the number of times of completed conversion behaviors to the total number of clicks on the promotion information. For example: 100 users saw the promotion information of the item, among which 10 users clicked on the information and jumped to the target website, and 3 users had conversion behaviors such as purchasing the item. Then the UV conversion rate of this promotion information is (3 / 10)*100% = 30%.

[0059] Sales volume and UV conversion rate are two correlated goals, and traditional single-goal optimization methods cannot reflect the mutual influence between different goals, and thus cannot well find the optimal combination that meets multiple goals simultaneously; while the new item mining method based on the multi-goal optimization method disclosed in this disclosure can simultaneously predict the two tasks related to new items, namely sales volume and UV conversion rate, and the local optimal solutions of different tasks are in different positions. Through the interaction between tasks, it helps to break out of the dilemma of local optimality.

[0060] In some embodiments, the historical data of the existing items includes: the historical sales data of the existing items, and the historical sales data of the existing items includes: the inherent attributes and sales information of the existing items. The inherent attributes of an item, for example, include the appearance attributes, functional attributes, mode attributes, etc. of the item. Taking a refrigerator as an example, a refrigerator has appearance attributes (such as length, width, height, panel material, color, door opening method, LCD screen, etc.), functional attributes (such as freshness preservation, multi-cycle, dry-wet separation storage, intelligent application, door-in-door, ice making, etc.), and mode attributes (such as refrigeration method, temperature control method, energy efficiency level, defrosting mode, fixed frequency / variable frequency, etc.). The sales information of an item, for example, includes price, shelf time, etc.

[0061] In step 130, according to the evaluation results of multiple preset goals for multiple item attribute combinations, determine the preferred item attribute combination.

[0062] In some embodiments, determining the preferred item attribute combination according to the evaluation results of multiple preset goals for multiple item attribute combinations includes: according to the evaluation value of each preset goal of each item attribute combination, for example, using a weighted summation calculation method, calculate the comprehensive evaluation value of multiple preset goals for each item attribute combination, where the weights of each preset goal can be set as needed. For example, set the weight of sales volume to be higher than the weight of UV conversion rate; according to the comprehensive evaluation value of each item attribute combination, determine the preferred item attribute combination from multiple item attribute combinations. For example, select several item attribute combinations with the largest comprehensive evaluation value as the preferred item attribute combination, and the selection quantity can be set as needed.

[0063] In step 140, according to the preferred item attribute combination, determine the new item of the preset category.

[0064] If none of the preferred item attribute combinations have appeared in the existing items, then all the preferred item attribute combinations can be used as the new items of the preset category. If some of the preferred item attribute combinations have appeared in the existing items, use the preferred item attribute combinations that have not appeared in the existing items as the new items of the preset category.

[0065] Embodiments of the present disclosure propose a solution for automatically mining new items through a machine learning model, which helps in the design of new items and reduces the consumption of human, financial, and material resources in the design of new items. Especially in the case where there are many item attributes and their combinations, it is basically impossible to complete such a huge design project manually, and it is also easy to miss some item attribute combinations. The machine mining solution of the present disclosure can well solve this problem; and this model belongs to a multi-objective learning model, which can output item attribute combinations that simultaneously meet multiple objectives. And during the learning process, through the interaction between multiple objectives, it helps to jump out of the local optimum of a single objective, and the underlying features of the multi-objective learning model are shared, which can reduce the model parameters to be adjusted and repeated calculations. In addition, through the heuristic search method, item attribute combinations can be obtained more comprehensively, avoiding missing some item attribute combinations, which is beneficial to mining more new items. In addition, the new item mining method of the embodiments of the present disclosure does not depend on time series factors. If new items are mined according to the time series prediction method, it is necessary to rely on the sales volume change characteristics, trends, and periodicity of items in the time dimension, etc. However, new items lack these characteristics, and even if predicted based on historical attribute combinations, errors will accumulate over time.

[0066] Figure 3 The flowchart showing the new item mining method based on genetic algorithm according to some embodiments of the present disclosure. As Figure 3 shown, the new item mining method based on genetic algorithm of this embodiment includes the following steps.

[0067] In step 310, set the maximization of the comprehensive evaluation value of multiple preset objectives as the mining objective, and set the value range of item attributes.

[0068] Let the item attribute set x = [x1, x2, …, x d , …, x D , where D is the number of item attributes, and x d is the d-th attribute of the item, d = 1, 2, …, D, and the value of each attribute x d is within its value range.

[0069] In step 320, use the genetic algorithm to first randomly search for several initial item attribute combinations as the initial population, and each item attribute combination is taken as an individual.

[0070] In step 330, at the beginning of each iteration, evaluate the fitness of individuals according to the principle of survival of the fittest and perform selection operations, select some better item attribute combinations for mutation and crossover operations, etc., and gradually make the item attribute combinations evolve towards the multi-objective optimum.

[0071] Among them, mutation means that some of the attribute values in an individual are randomly changed to other values of the attribute, as shown in Figure 2A shown. Crossover means that the attribute values between different individuals are exchanged, as shown in Figure 2B shown.

[0072] Among them, using the multi-objective evaluation model of items, the evaluation values of multiple preset objectives of an individual are obtained, and the weighted sum of the evaluation values of multiple preset objectives of the individual is calculated to obtain the comprehensive evaluation value of multiple preset objectives of the individual, which is used to represent the fitness of the individual. Among them, the larger the comprehensive evaluation value of the individual, the greater the fitness of the individual.

[0073] In step 340, it is judged whether the search stop condition is satisfied. If the search stop condition is not satisfied, return to step 320 to continue the iteration.

[0074] The search stop condition is, for example: the number of iterations of the population has reached the preset number of iterations, or no better item attribute combination has appeared within a certain number of iterations, etc.

[0075] In step 350, if the search stop condition is satisfied, output the top-k item attribute combinations with multiple optimal objectives as the design scheme of the newly mined item, where k is the number of design schemes of the newly mined item to be output and can be set.

[0076] Thus, the mining of new items is realized based on the genetic algorithm and the multi-objective evaluation model of items. Among them, some new types of item attribute combinations can be generated through operations such as mutation and crossover of the genetic algorithm, which is beneficial to providing the design scheme of new items.

[0077] Figure 4 The figure shows a schematic diagram of the item multi-objective evaluation model obtained based on the multi-gate mixture of experts multi-objective learning model in some embodiments of the present disclosure.

[0078] As Figure 4 shown, when the item multi-objective evaluation model is obtained based on the multi-gate mixture of experts multi-objective learning model, the item multi-objective evaluation model includes an input layer, a multi-gate mixture of experts layer, a tower layer, and an output layer cascaded in sequence.

[0079] The input layer (Input Layer) is used to instantiate the data input to this layer into a tensor, and the dimension is the number of features d of the input data. The input layer can be implemented, for example, based on the Input() function of a deep learning library (such as Keras).

[0080] The Multi-gate Mixture-of-Experts Layer (MMoE Layer) includes multiple feedforward neural networks (or expert feedforward networks) called experts that are connected to the input layer, and multiple gating networks that are connected to the input layer and the multiple experts. Different experts output different results based on the input features. The gating networks are used to perform weighted summation operations on the output results of each expert. The number of gating networks is the same as the number of preset targets, each gating network corresponds to a preset target, and different gating networks can perform weighted summation operations with different weights on the output results of each expert. Each gating network includes a gate connected to the input layer and a weighted summation calculator. The weighted summation calculator in each gating network performs a weighted summation operation on the output results of each expert according to the weight information provided by the gate in that gating network. Set the parameters of this layer. For example, set the number of hidden units h of the experts to 32, the number of experts n to 16, and the number of tasks k to 2. The parameters of this layer can be adjusted according to the actual situation and are not limited to the examples given. Each task corresponds to a preset target. Assuming there are 2 preset targets, then the number of tasks is 2.

[0081] The Tower Layer includes multiple sub-tower layers. The number of sub-tower layers is the same as the number of preset targets, and each sub-tower layer corresponds to a preset target. Each sub-tower layer is a fully connected neural network and is connected to a gating network. Each sub-tower layer takes the output of the corresponding gating network as input. Set the number of hidden units of this layer to 16, for example, which is not limited to the example given. Tower Layer 1 and Tower Layer 2 in the figure are two sub-tower layers.

[0082] The Output Layer includes multiple sub-output layers. The number of sub-output layers is the same as the number of preset targets, and each sub-output layer corresponds to a preset target. Each sub-output layer is a fully connected neural network and is connected to a sub-tower layer. Each sub-output layer takes the output of the corresponding sub-tower layer as input and outputs the prediction result of the corresponding preset target. Set the number of hidden units of this layer to 1, for example. Output Layer 1 and Output Layer 2 in the figure are two sub-output layers.

[0083] It can be seen from the item multi-objective evaluation model that the input layer and the expert feedforward network in the model are shared for multiple learning tasks, which can reduce the model parameters to be adjusted and repeated calculations.

[0084] Before learning, the activation functions of the item multi-objective evaluation model can be set. For example, set the activation function of each expert to the ReLU function, set the activation function of each gate to the Softmax function; set the activation function of each sub-tower layer to the ReLU function; set the activation function of each sub-output layer to the Linear function.

[0085] Before learning, the item multi-objective evaluation model can be initialized. For example, use a uniform distribution initializer to initialize the weights and biases of each expert and each gate; use a normal distribution initializer to initialize the weights of the fully connected neural network of each sub-tower layer; use a normal distribution initializer to initialize the weights of the fully connected neural network of each sub-output layer.

[0086] The working process of the item multi-objective evaluation model is described below.

[0087] In the input layer, the input layer instantiates the input data into a tensor, and the dimension of the tensor is the number of features d of the input data.

[0088] In the multi-gate mixture of experts layer, the multi-gate mixture of experts layer takes the tensor output by the input layer as the input x, and performs a weighted sum of the output results of each expert according to the weight information provided by each gate to obtain the output results f for each task k (x) = ∑ i g k (x) i f i (x), where f i (x) represents the output result of the i-th expert feed-forward network, and g k (x) i represents the weight of the k-th task corresponding gate to the output result of the i-th expert feed-forward network, x represents the input of the multi-gate mixture of experts layer, that is, the output of the input layer. For example, k = 1, 2, i = 1, 2, 3…16. The activation function of each expert feed-forward network is set to the ReLU function to make the calculation speed and convergence speed faster, then f i (x) = max{0, w i x + b}, w i is the weight information (matrix) of the i-th expert feed-forward network, which needs to be determined through training, and the matrix parameters are the number of features d of the input data × the number of hidden units h of the expert × the number of experts n, b is the bias information (matrix) of the expert feed-forward network, which needs to be determined through training, and the matrix parameters are the number of hidden units h of the expert × the number of experts n. The activation function of each gate is set to the Softmax function to convert the gate output result into non-negative numbers and the sum of each item is 1, then x iDenote the output of the gating for the $i$-th expert feed-forward network before the activation function processing. The output of the gating for each expert feed-forward network before the activation function processing can be represented by the combination $\{x i \} = w gk x + b$, where $w gk $ is the weight information (matrix) of the gating corresponding to the $k$-th task, which needs to be determined through training. The matrix parameter is the number of features $d$ of the input data $\times$ the number of experts $n \times$ the number of tasks $k$. $b$ is the bias information (matrix) of this gating, which needs to be determined through training. The matrix parameter is the number of experts $n \times$ the number of tasks $k$. In addition, before model training, the weights and biases of each expert feed-forward network and each gating can be initialized with a uniform distribution initializer, and no regularization term and constraint need to be set.

[0089] In the tower layer, for each task, the sub-tower layer takes the output result of the corresponding task of the upper layer multi-gating mixture of experts layer as the input $x$, and initializes the weights of the fully connected neural network of this layer with a normal distribution initializer. Set the activation function to the ReLU function. Then the output result $h k $ of the Tower Layer for the $k$-th task is $h k = \max\{0, wf

[0090] (x) + b\}$, where $w$ and $b$ represent the weight information (matrix) and bias information (matrix) of the sub-tower layer of the $k$-th task respectively. k In the output layer, for each task, the sub-output layer takes the output result of the corresponding task's sub-tower layer of the upper layer as the input, and initializes the weights of the fully connected neural network of this layer with a normal distribution initializer. The activation function is default to the Linear function, so that the final output result $y k = linear(h

[0091] ) outputs a single value through the only hidden layer unit of this layer.

[0092] Figure 5 It should be noted that the inputs of the multi-gating mixture of experts layer and the tower layer are both the outputs of their respective upper layers. Therefore, although the inputs of both the multi-gating mixture of experts layer and the tower layer are represented by $x$, the meaning of $x$ is different. The input $x$ of the multi-gating mixture of experts layer represents the output of its upper layer input layer, and the input $x$ of the tower layer represents the output of its upper layer multi-gating mixture of experts layer.

[0092] Figure 5 The figure shows a schematic diagram of the learning method of the item multi-objective evaluation model according to some embodiments of the present disclosure. As Figure 5 shown, the learning method of the item multi-objective evaluation model of this embodiment includes the following steps.

[0093] In step 510, obtain the historical data of existing items of a preset category, such as historical sales data.

[0094] Taking the refrigerator category as an example, obtain the appearance attributes (such as length, width, height, panel material, color, door opening method, LCD screen, etc.), functional attributes (such as freshness preservation, multi-cycle, dry-wet separation storage, intelligent application, door-in-door, ice making, etc.), mode attributes (such as refrigeration method, temperature control method, energy efficiency grade, defrosting mode, fixed frequency / variable frequency, etc.), and sales information (such as shelf price, shelf time, etc.) of various existing refrigerators.

[0095] Among them, in these historical data, for example, there are numerical discrete feature data such as depth, width, height, etc., numerical continuous feature data such as shelf price, etc., and non-numerical feature data such as brand, color, panel material, etc.

[0096] In step 520, preprocess the historical data, including data processing and feature engineering, etc., to convert the noisy original data set into a data set that can be used for model training.

[0097] In some embodiments, preprocessing the historical data includes, for example, but is not limited to the following processing:

[0098] 1) Remove the data that cannot be used to predict the preset goals of new items (such as sales volume, UV conversion rate, etc.), such as order volume, click-through rate, etc.

[0099] 2) Calculate the time difference based on the product shelf date and sold date.

[0100] 3) Based on the time difference, sum up the sales volume of each SKU (Stock Keeping Unit, the minimum inventory unit) within the number of days to be predicted in the future.

[0101] 4) Remove the sold date.

[0102] 5) Remove the duplicate rows of the data.

[0103] 6) Generate new features. For example, generate 3 new features of year, month, and day based on the shelf date, and generate the corresponding average sales volume and average price for each brand according to the historical sales volume and price of each brand of the item, etc., to represent the brand value and market share of the item, etc.

[0104] 7) Process the outliers of the data. For example, find the median of the corresponding UV conversion rate for the brand of the item and replace its outliers.

[0105] 8) Fill in the missing values of the corresponding features according to the values with the highest occurrence frequency of each feature of the item.

[0106] 9) Perform Min-Max normalization processing on the numerical discrete features to eliminate the influence of the unit and scale differences between each feature, making the process of finding the optimal solution smoother and easier to converge. Among them, the Min-Max normalization method is where min is the minimum value in the set where x is located, and max is the maximum value in the set where x is located.

[0107] 10) Perform Z-Score normalization on numerical continuous features to eliminate the influence of unit and scale differences between features, making the process of finding the optimal solution smoother and easier to converge. Among them, the Z-Score normalization method is where μ is the mean value in the set where x is located, and σ is the standard deviation in the set where x is located.

[0108] 11) Perform One-Hot encoding on non-numerical features to transform non-numerical features into a form that is beneficial for model learning. For example, for each feature, if it has m possible values, then after One-Hot encoding, it becomes m binary features, and these features are mutually exclusive, with only one activated at a time.

[0109] In step 530, the preprocessed dataset is divided to form a training dataset, a validation dataset, and a test dataset, for example, including the following steps.

[0110] 1) Shuffle the dataset in order to enhance the generalization ability of the model.

[0111] 2) Divide the shuffled dataset into a training set, a validation set, and a test set according to a certain ratio (such as 0.7:0.15:0.15).

[0112] 3) Split the feature data and target data in the three divided sets to form a training dataset (including training feature data and training target data), a validation dataset (including validation feature data and validation target data), and a test dataset (including test feature data and test target data).

[0113] Thus, a training dataset, a validation dataset, and a test dataset are determined according to the historical data of the existing items of the preset category.

[0114] In step 540, the item multi-objective evaluation model is trained using the training dataset.

[0115] In some embodiments, the training feature data in the training dataset is input into the item multi-objective evaluation model, and the total loss of multiple preset objectives is determined according to the gap information between the predicted value of each preset objective output by the item multi-objective evaluation model and the true value of each preset objective in the training dataset. The item multi-objective evaluation model is trained according to the total loss of the multiple preset objectives. Thus, the model parameters are continuously adjusted through training, making the total loss smaller and smaller.

[0116] For example, the total loss function of multiple preset targets is the Mean Absolute Error (MAE). where y mi is the true value of the i-th target of the m-th item data record, is the predicted value of the i-th target of the m-th item data record, 1 ≤ m ≤ M, and M is the number of item data records in the training dataset. Using MAE to determine the loss instead of the Mean Squared Error (MSE) can avoid the phenomenon of continuously increasing loss in the early stage of training and the model has better learning performance.

[0117] The optimizer of the model, for example, adopts the Adam optimizer, which can be applicable to scenarios with unstable objective functions and large-scale data and parameters.

[0118] Set the learning rate as an exponentially decaying learning rate: where lr is the learning rate, decay_rate is the decay rate of the learning rate within the range of (0, 1), global_step is the number of running epochs, and decay_steps is the number of epochs for decaying the learning rate once. The purpose is to enable the model to obtain a faster gradient descent speed in the early stage of training and gradually slow down with training to prevent the loss value from fluctuating up and down without convergence. Set some other training parameters. For example, the number of training epochs (epochs) is 200, and the batch size for training in each epoch is 256. After setting various parameters, the model can be trained by the automatic hyperparameter tuning tool Hyperopt.

[0119] In step 550, the item multi-objective evaluation model is verified using the validation dataset.

[0120] In some embodiments, after each epoch of training of the item multi-objective evaluation model, the item multi-objective evaluation model is verified using the validation dataset, and the training effect is evaluated using a preset evaluation function according to the verification result. If the evaluation result indicates that the item multi-objective evaluation model is optimized in this round of training, the item multi-objective evaluation model is trained for the next epoch. If the evaluation result indicates that the item multi-objective evaluation model has not been optimized in multiple epochs of preset number of epochs, the training is stopped.

[0121] Among them, the evaluation index can be set as the Weighted Mean Absolute Percentage Error (WMAPE) function, where y ni is the true value of the i-th target of the n-th item data record in the validation dataset, To verify the predicted value of the i-th target of the n-th item data record in the validation dataset. Using WMAPE for evaluation instead of MAPE reduces the impact caused by the difference in the order of magnitude of the true value y and also avoids the situation where the denominator is zero and cannot be calculated, making it more persuasive in predicting sales volume and UV conversion rate.

[0122] In addition, the evaluation metrics for the validation dataset can be set. If no optimization is achieved in multiple rounds of training in a preset number of rounds (such as 50 rounds), the training can be stopped in advance.

[0123] After the training is completed, the model and parameters under the optimal validation result are saved so that the model with the optimal validation result can be directly loaded during prediction.

[0124] In step 560, the item multi-objective evaluation model is tested using the test dataset, and the item multi-objective evaluation model that passes the test is used for new item mining.

[0125] Among them, during testing, the total loss function MAE of the multiple preset targets mentioned above can be used to evaluate the prediction loss of the model. If it is small enough and meets the business requirements, the item multi-objective evaluation model passes the test and can be used for new item mining.

[0126] Thus, an item multi-objective evaluation model is obtained by learning the historical data of existing items of a preset category for multiple preset targets.

[0127] Figure 6 The schematic diagram of a new item mining device showing some embodiments of the present disclosure is as follows. Figure 6 As shown, the new item mining device 600 of this embodiment includes: a memory 610 and a processor 620 coupled to the memory 610. The processor 620 is configured to execute the new item mining method in any of the foregoing embodiments based on the instructions stored in the memory 610.

[0128] For example, obtain multiple item attribute combinations of a preset category; use the item multi-objective evaluation model to evaluate multiple preset targets for each item attribute combination, where the item multi-objective evaluation model is obtained by learning the historical data of existing items of the preset category for the multiple preset targets; determine the preferred item attribute combination according to the evaluation results of the multiple preset targets of the multiple item attribute combinations; determine the new items of the preset category according to the preferred item attribute combination. Among them, the heuristic search method is used to search for the item attributes of the preset category to obtain multiple item attribute combinations of the preset category. Among them, the item multi-objective evaluation model is obtained by learning the historical data of existing items of the preset category for the multiple preset targets based on a multi-gate mixture of experts multi-objective learning model, a progressive hierarchical extraction multi-objective learning model, or a subnetwork routing multi-objective learning model.

[0129] Among them, the memory 610 can, for example, include a system memory, a fixed non-volatile storage medium, etc. The system memory stores, for example, an operating system, application programs, a boot loader, and other programs.

[0130] Among them, the processor 620 can be implemented in the form of a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete hardware components such as discrete gates or transistors.

[0131] The device 600 may further include an input / output interface 630, a network interface 640, a storage interface 650, etc. These interfaces 630, 640, 650 and the memory 610 and the processor 620 can be connected through a bus 660, for example. Among them, the input / output interface 630 provides a connection interface for input / output devices such as a display, a mouse, a keyboard, and a touch screen. The network interface 640 provides a connection interface for various networking devices. The storage interface 650 provides a connection interface for external storage devices such as an SD card and a USB flash drive. The bus 660 can use any bus structure in a variety of bus structures. For example, the bus structure includes but is not limited to an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, and a Peripheral Component Interconnect (PCI) bus.

[0132] Figure 7 Schematic diagram of a new item mining device showing some other embodiments of the present disclosure. As Figure 7 shown, the new item mining device 700 of this embodiment includes units 710-740 and may further include unit 750.

[0133] An attribute combination acquisition unit 710, configured to acquire a plurality of item attribute combinations of a preset category. For example, a plurality of item attribute combinations of the preset category are obtained by searching for item attributes of the preset category using a heuristic search method. The heuristic search method includes: genetic algorithm, particle swarm algorithm, ant colony algorithm, tabu search algorithm, simulated annealing algorithm.

[0134] An evaluation unit 720 is configured to evaluate multiple preset objectives for each combination of item attributes by using an item multi-objective evaluation model, where the item multi-objective evaluation model is obtained by learning from historical data of existing items of the preset category for the multiple preset objectives.

[0135] A preferred attribute combination determination unit 730 is configured to determine a preferred item attribute combination according to the evaluation results of the multiple preset objectives of multiple combinations of item attributes. For example, according to the evaluation values of each preset objective of each combination of item attributes, calculate the comprehensive evaluation value of the multiple preset objectives of each combination of item attributes; according to the comprehensive evaluation value of each combination of item attributes, determine the preferred item attribute combination from multiple combinations of item attributes.

[0136] A new item determination unit 740 is configured to determine a new item of the preset category according to the preferred item attribute combination.

[0137] A model learning unit 750 is configured to obtain the item multi-objective evaluation model by learning from historical data of existing items of the preset category for the multiple preset objectives based on a multi-gate mixture-of-experts multi-objective learning model, a progressive hierarchical extraction multi-objective learning model, or a subnet routing multi-objective learning model.

[0138] Some embodiments of the present disclosure propose a non-transitory computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the new item mining method of each embodiment are implemented.

[0139] Those skilled in the art should understand that the embodiments of the present disclosure can be provided as a method, a system, or a computer program product. Therefore, the present disclosure can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present disclosure can take the form of a computer program product implemented on one or more non-transitory computer-readable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer program code.

[0140] The present disclosure is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing in the process Figure 1 each process or multiple processes and / or blocks Figure 1a device for the functions specified in one or more boxes.

[0141] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device, and the instruction device implements the operations in the process Figure 1 one process or more processes and / or boxes Figure 1 the functions specified in one box or more boxes.

[0142] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one process or more processes and / or boxes Figure 1 the functions specified in one box or more boxes.

[0143] The above are only the preferred embodiments of the present disclosure, and are not intended to limit the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present disclosure shall be included within the protection scope of the present disclosure.

Claims

1. A method for excavating new items, characterized in that, Including: Obtaining multiple combinations of item attributes for a preset category, including: searching for item attributes of the preset category using a heuristic search method to obtain multiple combinations of item attributes of the preset category; Using an item multi-objective evaluation model to evaluate multiple preset objectives for each combination of item attributes, where the item multi-objective evaluation model is learned from historical data of existing items of the preset category for the multiple preset objectives; Determining an optimal combination of item attributes based on the evaluation results of the multiple preset objectives for the multiple combinations of item attributes; Determining a new item of the preset category according to the optimal combination of item attributes; Wherein, the item multi-objective evaluation model includes an input layer, a multi-gate mixture-of-experts layer, a tower layer, and an output layer connected in series in sequence, where: The multi-gate mixture-of-experts layer includes multiple feed-forward neural networks called experts connected to the input layer, and multiple gate networks connected to the input layer and the multiple experts. Each gate network includes a gate connected to the input layer and a weighted sum calculator. The weighted sum calculator in each gate network performs a weighted sum operation on the output results of each expert according to the weight information provided by the gate in this gate network. Each gate network corresponds to a preset objective, and the input layer and the feed-forward neural networks are shared for multiple learning tasks of the multiple preset objectives; The tower layer includes multiple sub-tower layers, and each sub-tower layer is a fully connected neural network and is connected to a gate network; The output layer includes multiple sub-output layers, and each sub-output layer is a fully connected neural network and is connected to a sub-tower layer.

2. The method according to claim 1, wherein It also includes: Performing one or more of the following operations before learning: Initializing the weights and biases of each expert and each gate using a uniform distribution initializer; initializing the weights of the fully connected neural network of each sub-tower layer using a normal distribution initializer; initializing the weights of the fully connected neural network of each sub-output layer using a normal distribution initializer.

3. The method according to claim 1, characterized in that Learning from the historical data of existing items of the preset category for the multiple preset objectives to obtain the item multi-objective evaluation model includes: Determining a training data set according to the historical data of existing items of the preset category, inputting the training feature data in the training data set into the item multi-objective evaluation model, determining the total loss of the multiple preset objectives according to the gap information between the predicted value of each preset objective output by the item multi-objective evaluation model and the true value of each preset objective in the training data set, and training the item multi-objective evaluation model according to the total loss of the multiple preset objectives.

4. The method according to claim 3, wherein Learning from the historical data of existing items of the preset category for the multiple preset objectives to obtain the item multi-objective evaluation model includes: Determining a validation data set according to the historical data of existing items of the preset category. After each round of training of the item multi-objective evaluation model, using the validation data set to validate the item multi-objective evaluation model, and evaluating the training effect using a preset evaluation function according to the validation result.

5. The method according to claim 4, wherein Learning from the historical data of existing items of the preset category for the multiple preset objectives to obtain the item multi-objective evaluation model includes: Determine a test data set based on the historical data of existing items of the preset category, use the test data set to test the multi-objective evaluation model of the items, and use the multi-objective evaluation model of the items that passes the test for new item mining.

6. The method according to claim 1, characterized in that Determining a preferred item attribute combination based on the evaluation results of multiple preset objectives for multiple item attribute combinations includes: Calculating the comprehensive evaluation value of multiple preset objectives for each item attribute combination according to the evaluation value of each preset objective of each item attribute combination; Determining a preferred item attribute combination from multiple item attribute combinations according to the comprehensive evaluation value of each item attribute combination.

7. The method according to claim 1, characterized in that, The heuristic search method includes: genetic algorithm, particle swarm algorithm, ant colony algorithm, tabu search algorithm, simulated annealing algorithm.

8. The method according to claim 1, characterized in that The multiple preset objectives include: sales volume and unique visitor conversion rate.

9. The method according to claim 1, wherein The historical data of the existing items includes: the historical sales data of the existing items, and the historical sales data of the existing items includes: the inherent attributes and sales information of the existing items.

10. A new item excavation device, characterized in that, Including: An attribute combination acquisition unit configured to acquire multiple item attribute combinations of a preset category, including: using a heuristic search method to search for item attributes of the preset category to obtain multiple item attribute combinations of the preset category; An evaluation unit configured to use a multi-objective evaluation model of items to evaluate multiple preset objectives of each item attribute combination, and the multi-objective evaluation model of items is learned based on the historical data of existing items of the preset category for the multiple preset objectives; A preferred attribute combination determination unit configured to determine a preferred item attribute combination according to the evaluation results of multiple preset objectives of multiple item attribute combinations; A new item determination unit configured to determine a new item of the preset category according to the preferred item attribute combination, wherein the multi-objective evaluation model of items includes an input layer, a multi-gate mixture of experts layer, a tower layer, and an output layer connected in series in sequence, wherein: The multi-gate mixture of experts layer includes multiple feed-forward neural networks called experts connected to the input layer, and multiple gate networks connected to the input layer and the multiple experts. Each gate network includes a gate connected to the input layer and a weighted sum calculator. The weighted sum calculator in each gate network performs a weighted sum operation on the output results of each expert according to the weight information provided by the gate in the gate network. Each gate network corresponds to a preset objective, and the input layer and the feed-forward neural networks are shared for multiple learning tasks of multiple preset objectives; The tower layer includes multiple sub-tower layers, and each sub-tower layer is a fully connected neural network and is connected to a gate network; The output layer includes multiple sub-output layers, and each sub-output layer is a fully connected neural network and is connected to a sub-tower layer.

11. A new item mining device, including: A memory; And a processor coupled to the memory, the processor being configured to execute the new item mining method according to any one of claims 1-9 based on instructions stored in the memory.

12. A non-transitory computer-readable storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the new item mining method according to any one of claims 1-9.

Images