Resource allocation method, electronic device, storage medium, and program product

Abstract

The application provides a resource allocation method, an electronic device, a storage medium and a program product, relates to the technical field of computers, and is used for predicting the probability of passenger cancellation of resources to determine whether to allocate resources to passengers. The method comprises the following steps: determining initial features based on passenger data of a target passenger and resource data of to-be-issued resources; inputting the initial features into each expert network respectively to obtain multiple expert extraction features; each expert network corresponds to a resource allocation target; the expert network is used for extracting the correlation between the multiple initial features related to the resource allocation target to obtain the expert extraction features; weighting and fusing the multiple expert extraction features based on the weight corresponding to each resource allocation target in the multiple resource allocation targets to obtain fused features; predicting the prediction probability of the target passenger using the to-be-issued resources based on the fused features; and if the prediction probability is greater than or equal to a preset value, allocating the to-be-issued resources to the target passenger.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a resource distribution method, electronic device, storage medium, and program product. Background Technology

[0002] Currently, the air transport industry is accelerating its transformation towards intelligent and refined operations. With the full online transformation of passenger booking, check-in, frequent flyer programs, and ancillary services, airlines have accumulated massive amounts of multi-dimensional operational and customer behavior data, which is increasingly being recognized as a key strategic asset. At the same time, the rapid development of artificial intelligence, machine learning, and big data analytics technologies has provided powerful tools for unlocking the value of this data.

[0003] Traditional airline marketing relies heavily on experience and judgment, resulting in low resource allocation efficiency. In recent years, with changes in revenue management concepts, precision marketing has gradually become an industry consensus. However, related research is still in its early stages: most findings focus on macro-level input-output analysis, lacking in-depth research on customer needs.

[0004] In current airline marketing practices, the allocation of marketing resources generally relies on the experience and judgment of sales personnel or pre-set rules. For example, static indicators such as frequent flyer status, historical spending, and flight frequency are often used as the main basis for resource allocation. However, this extensive resource allocation method, dominated by human experience, has significant limitations. Summary of the Invention

[0005] This application provides a resource allocation method, electronic device, storage medium, and program product, which are used to extract features of passenger data and resources to be allocated, extract the correlation between features and resource allocation targets through an expert network, determine fused features, and then determine the probability of passengers canceling resources to be allocated based on the fused features, and decide whether to allocate resources based on the comparison between the predicted probability and the preset value.

[0006] Firstly, this application provides a resource allocation method, comprising: determining initial features based on passenger data of a target passenger and resource data of resources to be allocated; inputting the initial features into each of multiple expert networks to obtain multiple expert-extracted features; each expert network corresponding to a resource allocation target; the expert networks being used to extract the correlation between multiple initial features related to the resource allocation target to obtain expert-extracted features; the resource allocation target including at least one of the following: resource redemption rate, resource cost, and passenger long-term value; weighted fusion of multiple expert-extracted features based on the weights corresponding to each resource allocation target to obtain fused features; predicting the probability of a target passenger using the resources to be allocated based on the fused features; and if the predicted probability is greater than or equal to a preset value, allocating the resources to be allocated to the target passenger.

[0007] The technical solution provided in this application offers at least the following benefits: By constructing an expert network targeting multiple resource allocation objectives (such as resource redemption rate, resource cost, and long-term passenger value), each expert network learns the correlation between passenger and resource characteristics related to a specific objective. Then, based on the business weights of each objective, the features extracted by the experts are weighted and fused to generate a fused feature that balances multi-objective optimization. This predicts the probability that a target passenger will use the resources to be allocated, and allocation is only executed when the predicted probability reaches a preset threshold. This effectively avoids resource waste or loss of user value, and improves the accuracy of resource allocation, the flexibility of business adaptation, and the overall input-output efficiency.

[0008] In one possible implementation, initial features are determined based on passenger data of the target passengers and resource data of the resources to be distributed, including: extracting passenger features based on passenger data and extracting resource features based on resource data of the resources to be distributed; concatenating passenger features and resource features, and performing dimensionality reduction processing on the concatenated features based on principal component analysis to obtain initial features; the number of dimensions of the initial features is less than the sum of the number of dimensions of passenger features and resource features.

[0009] Based on the above implementation method, by extracting passenger features and resource features from passenger data and resource data respectively, and then concatenating the two and performing dimensionality reduction using principal component analysis, not only is the key information of passenger feature and resource feature matching preserved, but also the redundancy and noise interference of the high-dimensional feature space are effectively reduced, resulting in lower initial feature dimension and higher information density, providing an input basis for the efficient training and stable inference of the subsequent expert network.

[0010] In one possible implementation, before constructing multiple feature pairs based on passenger features and resource features, the method further includes: preprocessing the passenger features and resource features respectively, wherein the preprocessing includes at least one of the following: denoising, missing value imputation, normalization, embedding, or one-hot encoding.

[0011] Based on the above implementation method, preprocessing operations such as denoising, missing value filling, normalization, embedding or one-hot encoding are performed on passenger features and resource features before feature splicing. This can unify the representation form and scale of features from different sources and of different types, eliminate the bias caused by missing data and anomalies, thereby enhancing the model's ability to understand real user behavior and resource attributes, and providing more standardized data input for subsequent feature fusion and deep modeling.

[0012] In one possible implementation, after determining the initial features, the method further includes: inputting the initial features into a multilayer perceptron to determine the attention weight of each dimension of the initial features; determining the product of each dimension of the initial features and its corresponding attention weight to obtain weighted features; and inputting the initial features into each of the multiple expert networks to obtain multiple expert-extracted features, including: inputting the weighted features into each of the multiple expert networks to obtain multiple expert-extracted features.

[0013] Based on the above implementation, before the initial features are input into the expert network, a multilayer perceptron is introduced to calculate the attention weights of each dimension of features and generate a weighted feature representation. This enables the model to automatically identify and strengthen the feature dimensions that are more discriminative to the current resource allocation target, and suppress the influence of irrelevant or noisy features. Using this weighted feature as the input to the expert network further enhances the ability of each expert to capture key correlation patterns under their respective objectives, thereby enhancing the pertinence and effectiveness of the features extracted by the experts, and ultimately improving the quality of the fused features and the accuracy of prediction.

[0014] In one possible implementation, based on the weights corresponding to each resource allocation objective among multiple resource allocation objectives, the features extracted by multiple experts are weighted and fused to obtain fused features. This includes: transforming each expert's extracted features based on a preset transformation matrix to obtain transformed features corresponding to each expert network; and weighting and fusing multiple transformed features based on the weights corresponding to each resource allocation objective among multiple resource allocation objectives to obtain fused features.

[0015] Based on the above implementation method, before weighted fusion of features extracted from multiple experts, the features output by each expert are first transformed by a preset transformation matrix to remove redundant information from different experts' outputs in each feature dimension and extract the difference information. Then, weighted fusion is performed by combining the weights of each resource allocation goal to ensure that the fusion process retains the professionalism of each expert and achieves a dynamic balance of goal orientation, thereby generating fusion features that better reflect the comprehensive intent of multiple goals and improving the rationality of the final decision.

[0016] In one possible implementation, predicting the probability of a target passenger using the resources to be allocated based on fused features includes: inputting the fused features into a prediction model to determine the predicted value; using an activation function to convert the predicted value into a predicted probability; and the predicted probability being positively correlated with the predicted value.

[0017] Based on the above implementation method, the fused features are input into the prediction model to obtain the predicted value, and then converted into the predicted probability through the activation function. This probability is positively correlated with the predicted value, which intuitively reflects the likelihood of passengers using the resources to be allocated. This probabilistic output not only facilitates comparison with preset thresholds to make allocation decisions, but also supports sorting and prioritizing the usage preferences of different passengers, thereby enhancing the fineness of resource allocation.

[0018] In one possible implementation, the method further includes: determining a preset value based on a preset resource allocation target; the preset resource allocation target is any one of a plurality of resource allocation targets, and each preset resource allocation target has a corresponding preset value range.

[0019] Based on the above implementation method, by binding the preset value with the specific resource allocation target and configuring a corresponding preset value range for each target, the threshold for the issuance decision is no longer a fixed constant, but a strategy parameter that can be dynamically adjusted according to the business target, thereby enhancing the system's adaptability to different operational scenarios.

[0020] Secondly, embodiments of this application provide a resource distribution device, comprising: The processing module is used to determine initial characteristics based on the passenger data of the target passengers and the resource data of the resources to be distributed; The input module is used to input initial features into each of the multiple expert networks to obtain multiple expert-extracted features; each expert network corresponds to a resource allocation objective; the expert network is used to extract the correlation between multiple initial features related to the resource allocation objective to obtain expert-extracted features; the resource allocation objective includes at least one of the following: resource redemption rate, resource cost, and passenger long-term value; The processing module is used to perform weighted fusion of features extracted by multiple experts based on the weight corresponding to each resource allocation target among multiple resource allocation targets, so as to obtain fused features; The prediction module is used to predict the probability of a target passenger using the resources to be allocated based on the fused features. The processing module, if the predicted probability is greater than or equal to the preset value, is used to distribute resources to the target passengers.

[0021] Thirdly, this application provides an electronic device, including: a processor and a memory, the processor being coupled to the memory; the memory being used to store computer instructions, which are loaded and executed by the processor to enable the computer device to implement the method described in the first aspect.

[0022] Fourthly, this application provides a computer-readable storage medium comprising: computer software instructions; which, when executed in an electronic device, cause the electronic device to implement the method described in the first aspect.

[0023] Fifthly, this application provides a computer program product comprising a computer program; when the computer program is run in an electronic device, it causes the electronic device to implement the method described in the first aspect.

[0024] The beneficial effects of the second to fifth aspects mentioned above are described in the corresponding description of the first aspect and will not be repeated here. Attached Figure Description

[0025] Figure 1 A flowchart illustrating a resource allocation method provided in an embodiment of this application; Figure 2 A flowchart illustrating yet another resource allocation method provided in this application embodiment; Figure 3 This is a schematic diagram illustrating the composition of a resource distribution device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0026] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0027] The terms "first" and "second," etc., used in the specification and drawings of this application are used to distinguish different objects or to distinguish different treatments of the same object, rather than to describe a specific order of objects.

[0028] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0029] It should be noted that in the embodiments of this application, the words "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0030] To facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish the same or similar items with essentially the same function and effect. Those skilled in the art can understand that the terms "first" and "second" are not intended to limit the quantity or execution order.

[0031] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0032] In the past, airlines relied heavily on human experience in marketing decisions, often allocating resources based on intuition or convention, resulting in low efficiency. Today, with the continuous evolution of revenue management concepts, the industry is increasingly inclined to adopt precision marketing strategies. Nevertheless, related research is still in its early stages, with most work focusing on macro-level input-output assessments and less on in-depth exploration of the subtle differences in needs among different passenger groups.

[0033] In practice, airlines typically allocate marketing resources based on fixed rules, such as static labels like frequent flyer status, past spending, or number of flights. While this method is simple and easy to implement, it is essentially a one-size-fits-all, crude approach that fails to reflect the real, dynamic, and diverse needs of passengers. This limits the potential for further improvement in marketing effectiveness and also results in a waste of marketing resources.

[0034] Based on this, this application provides a resource allocation method. It constructs an expert network targeting multiple resource allocation objectives (such as resource redemption rate, resource cost, and long-term passenger value), enabling each expert network to learn the correlation between passenger and resource characteristics related to a specific objective. Then, based on the business weights of each objective, the features extracted by the experts are weighted and fused to generate a fused feature that balances multi-objective optimization. This predicts the probability that a target passenger will use the resources to be allocated, and allocation is only performed when the predicted probability reaches a preset threshold. This effectively avoids resource waste or loss of user value, improving the accuracy of resource allocation, the flexibility of business adaptation, and the overall input-output efficiency.

[0035] The embodiments provided in this application will now be described in detail with reference to the accompanying drawings.

[0036] The resource distribution method provided in this application can be applied to electronic devices. These electronic devices can be servers, such as a server cluster consisting of multiple servers, a single server, a computer, or a processor or processing chip within a server or computer. For example, electronic devices can also be terminals, such as mobile phones, tablets, wearable devices, in-vehicle devices, laptops, and ultra-mobile personal computers (Ultra). Mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), etc. It should be noted that the embodiments of this application do not limit the specific form of the electronic device.

[0037] Figure 1 This is a flowchart illustrating a resource allocation method provided in an embodiment of this application. Figure 1 As shown, the resource distribution method provided in this application can be implemented through the aforementioned electronic device, specifically including S101~S105.

[0038] S101. Determine initial characteristics based on passenger data of the target passengers and resource data of resources to be allocated.

[0039] When resources need to be allocated to passengers, it is necessary to obtain passenger data of the target passengers and resource data of the resources to be allocated as a basis for decision-making, so as to provide data support for whether to allocate resources. In this process, passenger data and resource data of the resources to be allocated need to be transformed into features that can be identified and quantified by the model.

[0040] Among them, target passengers refer to specific individual passengers for whom a decision is currently being made regarding the distribution of resources. In some cases, target passengers may also refer to a category of passengers. Passenger data includes static passenger data, such as frequent flyer class, age, and region, as well as dynamic behavioral data, such as historical ticket purchase frequency, recent travel time, and number of order changes. Resources to be distributed are marketing resources planned to be distributed to target passengers, such as discount coupons, free baggage allowances, and VIP lounge privileges. Initial features are the raw input variables extracted or concatenated from passenger data and resource data. By extracting features, effective content can be extracted from a large amount of data for subsequent model processing.

[0041] In one implementation, S101 can be achieved through S1011-S1012: S1011. Extract passenger features based on passenger data, and extract resource features based on resource data to be issued.

[0042] In resource allocation scenarios, raw data typically exists in unstructured or semi-structured forms, such as user behavior logs, order records, and resource metadata. These raw fields cannot be directly input into machine learning models for effective learning. Passenger data contains a wealth of information related to user intent, preferences, and capabilities, while resource data describes the attributes of the resources to be allocated. By extracting passenger features and resource features separately, multi-resource reuse can be supported, allowing the same user feature to be combined with different resource features.

[0043] Specifically, passenger data refers to the historical and real-time behavioral and attribute data of target passengers, including but not limited to: static attributes, such as age, gender, city of residence, membership level, etc.; dynamic behaviors, such as the number of times viewed in the past 7 / 30 days, order frequency, average order value, coupon redemption history, etc.; and contextual information, such as current session behavior, device type, access channel, etc. Resource data refers to the metadata of resources to be issued, such as a discount coupon or a free upgrade benefit, including but not limited to: resource type, such as discount coupon, cash coupon, service benefit; face value / intensity, such as "20 yuan off" or "20% off", etc. Usage parameters, such as "valid for purchases of 100 or more"; validity period, such as "valid for 7 days"; cost, such as the cost borne by the platform, such as 15 yuan; scope of application, such as flight routes, product categories, and time period restrictions.

[0044] During feature extraction, the extracted features can be preprocessed. Preprocessing methods include, but are not limited to, denoising, missing value imputation, normalization, embedding, or one-hot encoding. For continuous features, Z-score normalization can be used, while for discrete features, one-hot encoding can be used for discretization mapping.

[0045] In one implementation, embedding can be used to directly map discrete passenger features into low-dimensional dense vectors, replacing one-hot encoding.

[0046] S1012. Passenger features and resource features are concatenated, and the concatenated features are dimensionality reduced based on principal component analysis to obtain the initial features.

[0047] Concatenating passenger and resource features aims to construct a joint representation of users and resources, capturing the correlation between them. However, direct concatenation leads to a dramatic increase in feature dimensionality, generating a large amount of redundant information and noise, resulting in a sharp increase in the model's computational cost and the risk of overfitting. Therefore, principal component analysis (PCA) is used for dimensionality reduction, which compresses the feature space while retaining most of the information, improving model efficiency and stability. The initial feature dimension obtained after dimensionality reduction is less than the sum of the dimensions of passenger and resource features.

[0048] Specifically, after the features are preprocessed by S1011, the preprocessed passenger features and resource features are concatenated in the last dimension. The process of performing principal component analysis on the concatenated features can be found in the implementation methods of related technologies, and will not be elaborated here.

[0049] After obtaining the initial features, the method also includes S201-S202: S201. Input the initial features into the multilayer perceptron and determine the attention weight of each dimension of the initial features.

[0050] Although the initial features are dimensionality-reduced through principal component analysis, the importance of each dimension is not equal. Some dimensions may be highly correlated with whether a user will use the resource, while others may be noisy or weakly correlated features. Inputting all dimensions into the subsequent model with the same importance would dilute the key signals and reduce prediction accuracy. By introducing an attention mechanism, the model dynamically evaluates the contribution of each feature dimension to the current prediction task, thereby enhancing the influence of key features, suppressing interference from irrelevant or redundant features, and improving the model's adaptability to different combinations of users and resources.

[0051] Specifically, a two-layer multilayer perceptron is used to process the initial features. The first layer performs dimensionality reduction on the initial features and extracts key features. The second layer performs dimensionality increase on the initial features to ensure feature integrity. Finally, the sigmoid activation function is used to generate attention weights.

[0052] In one implementation, a Transformer encoder can be used to capture long-range feature dependencies as an alternative to a multilayer perceptron.

[0053] S202. Determine the product of each dimension feature and its corresponding attention weight in the initial features to obtain the weighted features.

[0054] After obtaining the attention weights, they need to be applied to the initial features. By multiplying element-wise, features with high weights are amplified, while features with low weights are weakened, thereby generating a new, goal-oriented feature that focuses more on key information.

[0055] S102. Input the initial features into each of the multiple expert networks to obtain multiple expert-extracted features.

[0056] After the initial feature construction is completed, the feature is automatically fed into all pre-trained expert networks in parallel. When performing resource allocation, multiple resource allocation objectives often need to be considered simultaneously, such as resource redemption rate, resource cost, and long-term passenger value. Different resource allocation objectives focus on different data models, and a single model is difficult to take into account the optimization of multiple objectives. By adopting a multi-expert network structure, each expert network can focus on learning the feature associations most relevant to its objective.

[0057] In this context, an expert network can be a submodule of a neural network. Each expert network typically consists of 2-3 fully connected layers, and the activation function can be ReLU or GELU. Each expert network is specifically trained and optimized for a particular resource allocation objective. For example, one expert network might learn the correlation between features and the objective to improve the resource redemption rate, thereby extracting the corresponding features. The resource redemption rate refers to the proportion of resources actually used after allocation; resource cost refers to the financial or operational costs required to allocate the resource; and long-term passenger value refers to the potential future revenue generated by the user.

[0058] Specifically, each expert network is assigned a resource allocation target. All expert networks share the same initial feature input, but each has independent parameters. After the initial features are input into the expert network, each expert outputs a vector of fixed dimensions, which serves as the expert-extracted feature under that resource allocation target.

[0059] In one implementation, after obtaining the weighted features through S202 above, the weighted features are input into each expert network to obtain multiple expert-extracted features.

[0060] S103. Based on the weight of each resource allocation objective among multiple resource allocation objectives, the features extracted by multiple experts are weighted and fused to obtain the fused features.

[0061] In real-world business scenarios, different objectives have varying degrees of importance. By introducing configurable or learnable weights, the priorities among multiple objectives can be dynamically balanced.

[0062] Among them, the target weight refers to the relative importance of each resource allocation target in the current scenario. It can be set manually or learned automatically through a gating mechanism. The weighted fusion refers to the linear combination of multiple expert-extracted features according to weights or fusion through other methods to generate a comprehensive representation. The fusion feature refers to the feature representation that integrates the intentions of multiple resource allocation targets and is used for the final prediction.

[0063] Specifically, if the weights are static, the expert-extracted features output by each expert are directly weighted and summed to obtain the fused features; if the weights are dynamic, the initial features are used as input through a gated network to output the attention weights of each expert, and the expert-extracted features are fused based on the attention weights of each expert. The fused features can then be further processed through a fusion layer, such as a multilayer perceptron, to perform nonlinear transformations and enhance expressive power.

[0064] In one implementation, S103 can be implemented through S1031-S1032: S1031. Based on the preset transformation matrix, transform the features extracted from each expert network to obtain the transformed features corresponding to each expert network.

[0065] In a multi-expert architecture, each expert network learns user-resource matching patterns for a specific resource allocation objective. Despite differing objectives, certain underlying signals may be of common concern to multiple experts, leading to potentially similar opinions. This means that expert opinions under different resource allocation objectives contain both consensus and disagreement. Simple weighted averaging can obscure these fine-grained differences. Therefore, a pre-defined transformation matrix is ​​used to linearly map the features extracted from each expert. The core purpose is not only dimensional alignment but also to remove redundant information (i.e., consensus) from the outputs of multiple experts and extract the points of disagreement (disagreements) between their opinions.

[0066] S1032. Based on the weights corresponding to each resource allocation target among multiple resource allocation targets, multiple transformation features are weighted and fused to obtain fused features.

[0067] The core of feature fusion lies in the weighted fusion of the outputs of multiple experts along each feature dimension. The process involves dynamically determining which experts are more credible in each dimension, automatically suppressing redundant or conflicting expert responses, and retaining the most discriminative signals. Specifically, linear fusion can be performed along the expert dimensions using 1D convolution, by weighting and summing the outputs of different experts in each dimension to obtain the fused features.

[0068] S104. Based on the fusion features, predict the probability of target passengers using the resources to be allocated.

[0069] The core issue in resource allocation is "whether the user will use this resource." By predicting the probability of use, the effectiveness of resource allocation can be quantified, the waste caused by ineffective allocation can be avoided, and a basis for decision-making regarding subsequent allocations can be provided.

[0070] The predicted probability refers to a value between 0 and 1 in the model output, representing the likelihood that the target passenger will actually use the resource after receiving it.

[0071] Specifically, the fused features are input into a pre-trained model, which outputs predicted probabilities. The model can be a perceptron with multiple layers, and the last layer uses an activation function to output probability values.

[0072] In one implementation, S104 can be implemented using S1041-S1042: S1041. Input the fused features into the prediction model to determine the predicted value.

[0073] While fused features are highly abstract and encompass the multidimensional intent of matching users with resources, they are still an intermediate representation and cannot be directly used for decision-making. To address the core business question of whether users will redeem resources, a specialized predictive model is needed to map fused features into an interpretable and comparable numerical output.

[0074] Specifically, when the fused features are input into the prediction model, the model outputs a predicted value. However, the predicted value only reflects the likelihood of a user redeeming resources; that is, the higher the value, the more likely the user is to redeem resources. But the predicted value cannot be directly interpreted as a "probability." Therefore, it is necessary to transform the predicted value to obtain a "predicted probability" that can be used for decision-making.

[0075] S1042. Use an activation function to convert the predicted values ​​into predicted probabilities.

[0076] Since the predicted values ​​cannot be used directly for decision-making, an activation function is needed to convert the predicted values ​​into predicted probabilities between 0 and 1, so that the decision on whether to "issue" or "not issue" can be made based on the predicted probabilities.

[0077] The prediction probability is positively correlated with the prediction value; that is, the larger the prediction value, the greater the prediction probability.

[0078] S105. If the predicted probability is greater than or equal to the preset value, distribute the resources to be issued to the target passengers.

[0079] The predicted probability is a value, and a decision threshold needs to be set to translate it into a clear "distribute" or "do not distribute" action. This threshold can be adjusted according to business objectives to achieve precise and efficient resource allocation.

[0080] The preset value refers to the set probability threshold, such as 0.7. If the predicted probability is higher than this value, it is considered worthwhile to distribute resources. Distributing resources refers to pushing coupons, benefits, etc. to users. The push methods include, but are not limited to, application push, SMS, account credit, etc.

[0081] In one implementation, preset values ​​can be determined based on preset resource allocation goals. These preset resource allocation goals can be any one of multiple goals, each with a corresponding preset value range. For example, if the current goal is to maximize the redemption rate, a lower predicted probability, such as a preset value ≥ 0.2, is acceptable to broaden the coverage. If the core objective is to strictly control resource costs, a higher confidence level is needed, such as a preset value ≥ 0.6, to avoid waste caused by ineffective distribution. If the focus is on cultivating high-value users, a more lenient threshold may be set only for high-value users, while a higher preset value may be applied to ordinary users. By binding preset values ​​to specific preset resource allocation goals and configuring reasonable value ranges for them, dynamic decision-making driven by resource distribution goals can be achieved, enabling the system to flexibly respond to the needs of different operational stages and different customer services.

[0082] Specifically, the electronic device calculates the predicted probability of each target passenger for the resource in real time. If the predicted probability is greater than or equal to a preset value, the resource allocation action is triggered.

[0083] Based on the above embodiments, by constructing expert networks for multiple resource allocation objectives, each expert network learns the correlation between passenger and resource characteristics related to a specific objective. Then, based on the business weights of each objective, the features extracted by the experts are weighted and fused to generate a fused feature that takes into account multi-objective optimization. This predicts the probability that a target passenger will use the resources to be allocated, and allocation is only executed when the predicted probability reaches a preset threshold. This effectively avoids resource waste or loss of user value, and improves the accuracy of resource allocation, the flexibility of business adaptation, and the overall input-output efficiency.

[0084] The resource allocation method of this application embodiment is described below with reference to a specific example, such as... Figure 2 As shown, the resource allocation method includes S301-S309: S301. Data preparation.

[0085] The raw passenger and resource data are transformed into structured feature pairs, which contain both continuous and discrete features. For continuous features, Z-score normalization is used to perform a linear transformation on the variables. This process preserves the magnitude relationship of numerical features and ensures the comparability of features with different units, such as "cumulative mileage in the past year" and "number of check-in days in the past month". For example, for variable x, the normalized representation is... in, The normalized features are represented by μ, where μ is the data mean and σ is the data standard deviation. For discrete features, one-hot encoding is used to discretize the variables. If there are n possible values ​​for a category, an n-dimensional binary vector is generated. The feature vector pairs are obtained through the above process. , where p is the passenger feature dimension and r is the resource feature dimension.

[0086] S302, Feature Dimensionality Reduction.

[0087] After obtaining the structured feature pairs passenger characteristics and The joint feature vector of variable x can be obtained by directly concatenating the features. in, Represent the real number field. Then, principal component analysis is used to obtain the dimensionality-reduced vector. PCA stands for Principal Component Analysis. k This represents the dimension after dimensionality reduction. The resulting vector... This is equivalent to the initial feature in the above embodiments.

[0088] S303, Key Feature Extraction.

[0089] Will The input is fed into a two-layer multilayer perceptron, and in the first layer... Perform dimensionality reduction processing, Reduced from K-dimensional to M-dimensional. Where ReLU is the activation function. For attention weight vectors, For bias terms M and K For dimensions.

[0090] Features after dimensionality reduction u The input is fed into the second layer to obtain the upgraded features. in, This is the attention weight vector.

[0091] Attention weights are generated, and finally, the Sigmoid function is used to generate the attention weights. in Let z be the Sigmoid activation function, and z be the input to the function. v Attention weights are obtained as input to the function. w .

[0092] The obtained attention weights and Multiplying each element individually yields the weighted feature. .

[0093] S304, Expert Network Feature Extraction.

[0094] The obtained weighted features are input into multiple expert networks. Through a shared expert network and dynamic gating mechanism, independent weight assignments are learned for different resource allocation objectives. The shared expert network consists of multiple fully connected layers, and the k-dimensional features are further transformed into e-dimensional features using the nonlinear activation function ReLU to extract potential nonlinear patterns, providing a basis for multi-objective decision-making. For each expert, the output is: in Let represent the weight matrix of the i-th expert network.

[0095] S305, Feature Weighting.

[0096] Use the Column-Conv1d fusion module to combine the outputs of n expert networks. Concatenate them into a three-dimensional tensor Where B is the batch size, n is the number of experts, and e is the output dimension of each expert. Then, a 1×1 convolution is used for linear fusion along the expert dimensions, weighted summing of the outputs of different experts at each dimension to remove redundant information. The fused features are: in The attention weights represent the weighted features. The bias term represents the weighted feature. It is the expert output after redundancy removal. .

[0097] S306, Probability Distribution Calculation.

[0098] The dynamic gating mechanism is used to dynamically allocate weights to each expert network for each resource allocation target t. The input to the dynamic gating mechanism is the features after MLP-Attention weighting. The scores are converted into a probability distribution using the softmax function. in, Let the weight matrix be independent for resource allocation target t. This is the bias term for the resource allocation target t.

[0099] S307, Feature Fusion.

[0100] For each resource allocation target t, dynamic gating weights are used to adjust the expert output after Column-Conv1d fusion. By performing a weighted summation, the fusion feature is obtained: .

[0101] S308, Probability Prediction.

[0102] Fusion features Input the data into the model to obtain the predicted value output by the model. Then Convert to variable score , This reflects the tendency of resource allocation target t. Finally, the Sigmoid activation function is used to suppress the score. Shrink to [0,1] and convert to interpretable predictive probabilities: ,in Let represent the resource write-off probability matrix under resource allocation target t.

[0103] S309, Resource Decision-Making.

[0104] During the strategy generation phase, the decision to allocate a resource to a passenger can be based on the passenger's resource redemption probability. The strategy generation engine can be represented as follows: ,in This represents the preset value. It can be dynamically adjusted according to business needs. For example, under a resource allocation target t, resource settings with stricter inventory control are possible. =0.7 to avoid wasting resources; different objectives can be set independently. value.

[0105] This application embodiment can divide the following device into functional modules based on the above method example. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. Optionally, the module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0106] In some embodiments, this application also provides a resource allocation apparatus. The resource allocation apparatus may include one or more functional modules for implementing the methods described in the above embodiments. For example, Figure 3 This is a schematic diagram illustrating the composition of a resource distribution device 300 provided in an embodiment of this application. Figure 3 As shown, the resource distribution device 300 includes: Processing module 301 is used to determine initial characteristics based on passenger data of the target passenger and resource data of the resources to be distributed; Input module 302 is used to input initial features into each of the multiple expert networks to obtain multiple expert-extracted features; each expert network corresponds to a resource allocation objective; the expert network is used to extract the correlation between multiple initial features related to the resource allocation objective to obtain expert-extracted features; the resource allocation objective includes at least one of the following: resource redemption rate, resource cost, and passenger long-term value; The processing module 301 is also used to perform weighted fusion of the features extracted by multiple experts based on the weight corresponding to each resource allocation target among multiple resource allocation targets, so as to obtain fused features; The prediction module 303 is used to predict the probability of a target passenger using the resources to be allocated based on the fused features; The processing module 301, if the predicted probability is greater than or equal to a preset value, is used to issue resources to be distributed to the target passenger.

[0107] In one implementation, the processing module 301 is further configured to: extract passenger features based on passenger data, and extract resource features based on resource data of resources to be issued; concatenate the passenger features and resource features, and perform dimensionality reduction processing on the concatenated features based on principal component analysis to obtain initial features; the number of dimensions of the initial features is less than the sum of the number of dimensions of the passenger features and resource features.

[0108] In one implementation, the processing module 301 is further used to preprocess the passenger features and resource features respectively, and the preprocessing includes at least one of the following: denoising, missing value filling, normalization processing, embedding processing or one-hot encoding.

[0109] In one implementation, the input module 302 is further configured to input the initial features into a multilayer perceptron, determine the attention weight of each dimension of the initial features, and determine the product of each dimension of the initial features and its corresponding attention weight to obtain weighted features. The weighted features are then input into each of the multiple expert networks to obtain multiple expert-extracted features.

[0110] In one implementation, the processing module 301 is further configured to: transform the features extracted from each expert based on a preset transformation matrix to obtain the transformed features corresponding to each expert network; and perform weighted fusion of multiple transformed features based on the weights corresponding to each resource allocation target among multiple resource allocation targets to obtain fused features.

[0111] In one implementation, the prediction module 303 is further configured to input the fused features into the prediction model to determine the predicted value; use an activation function to convert the predicted value into a predicted probability; and the predicted probability is positively correlated with the predicted value.

[0112] In one implementation, the prediction module 303 is further configured to determine a preset value based on a preset resource allocation target; the preset resource allocation target is any one of a plurality of resource allocation targets, and each preset resource allocation target has a corresponding preset value range.

[0113] When the methods of the above embodiments are implemented in hardware, this embodiment of the invention provides a possible structural schematic diagram of the electronic device involved in the above embodiments. For example... Figure 4As shown, the electronic device 400 includes: a processor 402, a communication interface 403, and a bus 404. Optionally, the electronic device 400 may also include a memory 401.

[0114] Processor 402 may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 402 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 402 may also be a combination that implements computing functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0115] Communication interface 403 is used to connect to other devices via a communication network. This communication network can be Ethernet, wireless access network, wireless local area network (WLAN), etc.

[0116] The memory 401 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0117] In one possible implementation, the memory 401 can exist independently of the processor 402. The memory 401 can be connected to the processor 402 via a bus 404 and is used to store instructions or program code. When the processor 402 calls and executes the instructions or program code stored in the memory 401, it can implement the method provided in the embodiments of the present invention.

[0118] In another possible implementation, the memory 401 can also be integrated with the processor 402.

[0119] Bus 404 can be an Extended Industry Standard Architecture (EISA) bus, etc. Bus 404 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The symbol is represented by only one line, but this does not mean that there is only one bus or one type of bus.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above device can be divided into different functional modules to complete all or part of the functions described above.

[0121] This application also provides a computer-readable storage medium. All or part of the processes in the above method embodiments can be executed by computer instructions instructing related hardware. The program can be stored in the aforementioned computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. The computer-readable storage medium can be any of the foregoing embodiments or memory. The aforementioned computer-readable storage medium can also be an external storage device of the aforementioned device, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the aforementioned device. Further, the aforementioned computer-readable storage medium can include both internal storage units of the aforementioned service call device and external storage devices. The aforementioned computer-readable storage medium is used to store the aforementioned computer program and other programs and data required by the aforementioned device. The aforementioned computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0122] This application also provides a computer program product comprising a computer program that, when run on a computer, causes the computer to perform the methods provided in the above embodiments.

[0123] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A resource distribution method, characterized in that, The method includes: Based on the passenger data of the target passengers and the resource data of the resources to be allocated, the initial characteristics are determined; The initial features are input into each of the multiple expert networks to obtain multiple expert-extracted features; each expert network corresponds to a resource allocation objective; the expert network is used to extract the correlation between multiple initial features related to the resource allocation objective to obtain expert-extracted features; the resource allocation objective includes at least one of the following: resource redemption rate, resource cost, and passenger long-term value. Based on the weight corresponding to each resource allocation objective among multiple resource allocation objectives, the features extracted by the multiple experts are weighted and fused to obtain the fused features; Based on the fusion features, predict the probability that the target passenger will use the resources to be allocated; If the predicted probability is greater than or equal to a preset value, the resources to be distributed are sent to the target passenger.

2. The method according to claim 1, characterized in that, The initial characteristics determined based on passenger data of the target passengers and resource data of resources to be allocated include: Passenger features are extracted based on the passenger data, and resource features are extracted based on the resource data of the resources to be issued. The passenger features and resource features are concatenated, and the concatenated features are then subjected to dimensionality reduction processing based on principal component analysis to obtain the initial features; the number of dimensions of the initial features is less than the sum of the number of dimensions of the passenger features and the resource features.

3. The method according to claim 2, characterized in that, Before concatenating the passenger features and the resource features, the method further includes: The passenger features and the resource features are preprocessed respectively, and the preprocessing includes at least one of the following: denoising, missing value imputation, normalization, embedding, or one-hot encoding.

4. The method according to claim 1, characterized in that, After determining the initial features, the method further includes: The initial features are input into a multilayer perceptron to determine the attention weight of each dimension of the initial features; The weighted features are obtained by multiplying each dimension of the initial features with its corresponding attention weight. The step of inputting the initial features into each of the multiple expert networks to obtain multiple expert-extracted features includes: The weighted features are input into each of the multiple expert networks to obtain the multiple expert-extracted features.

5. The method according to claim 1, characterized in that, The weighted fusion of the features extracted by the multiple experts, based on the weight corresponding to each resource allocation objective among multiple resource allocation objectives, yields the fused features, including: Based on the preset transformation matrix, the features extracted from each expert are transformed to obtain the transformed features corresponding to each expert network. Based on the weights corresponding to each resource allocation objective among multiple resource allocation objectives, multiple transformation features are weighted and fused to obtain fused features.

6. The method according to claim 1, characterized in that, The prediction of the probability that the target passenger will use the resources to be allocated based on the fused features includes: The fused features are input into the prediction model to determine the predicted value; An activation function is used to convert the predicted value into a predicted probability; the predicted probability is positively correlated with the predicted value.

7. The method according to claim 1, characterized in that, The method further includes: Based on a preset resource allocation target, the preset value is determined; the preset resource allocation target is any one of the plurality of resource allocation targets, and each preset resource allocation target has a corresponding preset value range.

8. An electronic device, characterized in that, It includes a processor and a memory, the processor being coupled to the memory; the memory is used to store computer instructions, which are loaded and executed by the processor to enable the computer device to perform the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes computer-executable instructions that, when executed on a computer, cause the computer to perform the method of any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when run on an electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 7.

Images