Menu category optimization method and system based on historical meal ordering data

Abstract

The invention belongs to the technical field of meal ordering platform management, and discloses a menu category optimization method and system based on historical meal ordering data. The method comprises the following steps: collecting historical meal ordering data of a meal ordering platform, and extracting real-time multi-dimensional combination features about the historical meal ordering data at the current moment according to a preset multi-dimensional feature engineering space; according to the real-time multi-dimensional combination features, a pre-trained user menu category demand analysis model is used for analysis, and a real-time user menu category demand analysis result is obtained; according to the real-time user dish demand analysis result, generating a scheme by using a pre-trained menu category optimization model to obtain a real-time menu category optimization scheme; and according to the real-time menu category optimization scheme, optimizing preset menu category data of the meal ordering platform to obtain optimized menu category data. According to the invention, the problems of poor comprehensiveness, low accuracy and low intelligent degree in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of meal ordering platform management, and in particular relates to a menu category optimization method and system based on historical meal ordering data. Background Art

[0002] With the rapid development of online food ordering platforms, more and more users are choosing online ordering as a convenient way to spend their daily lives. On these platforms, users can choose dishes based on their preferences. The menu selection has become a key factor in determining whether a platform is popular with users. This means that the menu selection of these platforms needs to be continuously updated and optimized to meet users' actual needs. Therefore, how to optimize the menu selection of these platforms based on historical ordering data, adapting it to evolving user needs and improving user satisfaction and platform revenue, has become an important research topic.

[0003] The existing technology has many defects, including:

[0004] 1) Poor comprehensiveness: Existing technologies only consider partial information from historical ordering data when optimizing menu categories, such as dish sales or user ratings, while ignoring the comprehensive impact of multi-dimensional information such as time, space, and behavior. This fails to fully capture users' complex needs, resulting in significant deviations between optimization results and actual needs.

[0005] 2) Low accuracy: Existing technologies use static optimization rules or fixed optimization strategies, that is, one-time optimization based on historical ordering data, while ignoring real-time changing needs. As a result, the optimization results may not be able to adapt to the ever-changing user needs and market changes, and the accuracy cannot meet the requirements.

[0006] 3) Low level of intelligence: Existing technologies require a lot of manual intervention and lack an intelligent and automated menu category optimization system, resulting in optimization plans lagging behind actual needs and making it impossible to adjust menu categories in a timely manner to adapt to market changes. Summary of the Invention

[0007] In order to solve the problems of poor comprehensiveness, low accuracy and low intelligence in the existing technology, the purpose of the present invention is to provide a menu category optimization method and system based on historical ordering data.

[0008] The technical solution adopted in the present invention is:

[0009] A method for optimizing menu categories based on historical meal ordering data includes the following steps:

[0010] Collect historical ordering data from the ordering platform and extract real-time multi-dimensional combined features of the historical ordering data at the current moment based on the preset multi-dimensional feature engineering space;

[0011] Based on the real-time multi-dimensional combination features, the pre-trained user menu category demand analysis model is used to perform analysis and obtain the real-time user menu category demand analysis results;

[0012] Based on the real-time user dish demand analysis results, a pre-trained menu category optimization model is used to generate a real-time menu category optimization solution.

[0013] According to the real-time menu category optimization plan, the preset menu category data of the ordering platform is optimized to obtain the optimized menu category data.

[0014] Furthermore, the multidimensional feature engineering space includes a time feature engineering architecture, a space feature engineering architecture, a user behavior feature engineering architecture, and a multidimensional fusion feature engineering architecture. The multidimensional fusion feature engineering architecture is connected to the time feature engineering architecture, the space feature engineering architecture, and the user behavior feature engineering architecture respectively.

[0015] Furthermore, the time feature engineering is equipped with a periodic consumption feature extraction model, the spatial feature engineering is equipped with a consumption circle feature extraction model, the user behavior feature engineering architecture is equipped with a user portrait feature extraction model, and the multi-dimensional combination feature engineering architecture is equipped with a tensor combination model.

[0016] Furthermore, historical ordering data from the ordering platform is collected, and based on a preset multidimensional feature engineering space, real-time multidimensional combined features of the historical ordering data at the current moment are extracted, including the following steps:

[0017] Collect historical ordering data from the ordering platform and pre-process the historical ordering data to obtain pre-processed historical ordering data;

[0018] Use the periodic consumption feature extraction model of time feature engineering in the preset multi-dimensional feature engineering space to extract the real-time time features of historical ordering data at the current moment;

[0019] Use the consumer circle feature extraction model of spatial feature engineering in the multidimensional feature engineering space to extract the real-time spatial features of historical ordering data at the current moment;

[0020] Use the user profile feature extraction model of user behavior feature engineering in the multi-dimensional feature engineering space to extract real-time user behavior features of historical ordering data at the current moment;

[0021] The tensor combination model of the multidimensional combination feature engineering architecture in the multidimensional feature engineering space combines the real-time time features, the real-time spatial features and the real-time user behavior features to obtain real-time multidimensional combination features.

[0022] Furthermore, the user menu category demand analysis model is constructed based on the LSTM-DAGN algorithm, and the user menu category demand analysis model includes a time feature channel, a spatial feature channel, a user behavior feature channel and a dynamic gated attention module. The time feature channel, the spatial feature channel and the user behavior feature channel are set in parallel, and the time feature channel, the spatial feature channel and the user behavior feature channel are all connected to the dynamic gated attention module.

[0023] Furthermore, based on the real-time multi-dimensional combination features, a pre-trained user menu category demand analysis model is used to perform analysis to obtain real-time user menu category demand analysis results, including the following steps:

[0024] Use the time feature channel in the pre-trained user menu category demand analysis model to perform additional feature enhancement on the real-time time feature in the real-time multi-dimensional combination feature to obtain the enhanced real-time time feature;

[0025] Use the spatial feature channel in the user menu category demand analysis model to embed the real-time spatial features in the real-time multi-dimensional combination features into graph features to obtain the embedded real-time spatial features;

[0026] The user behavior feature channel in the user menu category demand analysis model is used to perform preference evolution on the real-time user behavior features in the real-time multi-dimensional combination features, and the real-time user behavior features after preference evolution are obtained;

[0027] According to the dynamic gated attention weight, the dynamic gated attention module in the user menu category demand analysis model is used to fuse the enhanced real-time temporal features, the embedded real-time spatial features, and the real-time user behavior features after preference evolution to obtain real-time fused features.

[0028] Based on the real-time fusion features, user menu category demand analysis is performed to obtain real-time user menu category demand analysis results.

[0029] Furthermore, the holiday vector is fused with the real-time time feature in the real-time multi-dimensional combined feature to obtain the enhanced real-time time feature.

[0030] Embed the merchant distribution map feature into the real-time spatial feature in the real-time multi-dimensional combination feature to obtain the embedded real-time spatial feature;

[0031] According to historical user behavior characteristics, a contrastive learning algorithm is used to perform preference evolution on the real-time user behavior characteristics in the real-time multi-dimensional combination characteristics, and the real-time user behavior characteristics after preference evolution are obtained.

[0032] Furthermore, the menu category optimization model is constructed based on the MOCPO algorithm, and the menu category optimization model includes an influence factor generation module, an optimization target update module, an initialization module, an iterative optimization module and a vector decoding module which are connected in sequence.

[0033] Furthermore, based on the real-time user dish demand analysis results, a pre-trained menu category optimization model is used to generate a solution to obtain a real-time menu category optimization solution, which includes the following steps:

[0034] Based on the real-time user menu demand analysis results, the influence factor generation module of the pre-trained menu category optimization model is used to generate the corresponding real-time influence factors;

[0035] Based on the real-time influencing factors, the optimization target update module of the menu category optimization model is used to update the predicted optimization target, obtain the real-time optimization target, and set the real-time fitness function based on the real-time optimization target;

[0036] Encode the initial real-time menu category optimization plan into individual vectors of the initialization module, and use the initialization module of the menu category optimization model to generate several initial solutions based on the individual vectors;

[0037] Based on the real-time fitness function, the iterative optimization module of the menu category optimization model is used to iteratively optimize several initial solutions to obtain the optimal solution;

[0038] Use the vector decoding module of the menu category optimization model to decode the individual vectors of the optimal solution and obtain the optimal real-time menu category optimization plan.

[0039] A menu category optimization system based on historical meal ordering data is used to implement a menu category optimization method, comprising a multi-dimensional combination feature extraction unit, a menu category demand analysis unit, a menu category optimization unit, and an optimization solution execution unit connected in sequence.

[0040] The beneficial effects of the present invention are:

[0041] The present invention provides a menu category optimization method and system based on historical meal ordering data, which constructs a three-dimensional feature engineering space including time, space, and user behavior, and realizes the comprehensive capture of complex user needs. This multi-dimensional feature fusion method effectively overcomes the defect of poor comprehensiveness of the existing technology, making the optimization result closer to actual needs; uses a user menu category demand analysis model for analysis, explores the deep relationship between user needs and menu categories in historical meal ordering data, uses a menu category optimization model to dynamically generate menu category optimization solutions, and dynamically adjusts and continuously optimizes according to real-time user needs, which can respond to market changes in a timely manner and effectively improve the market adaptability of menu categories and the accuracy of optimization solutions; adopts automated feature extraction, intelligent analysis and dynamic optimization to realize the construction of an automated and intelligent system for menu category optimization. Through this automated decision-making mechanism, human intervention is reduced and the degree of intelligence is improved.

[0042] Other beneficial effects of the present invention will be further described in the specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flowchart of the menu category optimization method based on historical meal ordering data in the present invention.

[0044] Figure 2 This is a structural block diagram of the menu category optimization system based on historical ordering data in the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0046] Example 1:

[0047] like Figure 1 As shown, this embodiment provides a menu category optimization method based on historical ordering data, including the following steps:

[0048] S1: Collect historical ordering data from the ordering platform and extract real-time multi-dimensional combined features of the historical ordering data at the current moment based on the preset multi-dimensional feature engineering space;

[0049] Historical ordering data includes historical order basic information, historical order execution data, historical consumption behavior data, historical operation management data, and historical extended data;

[0050] Basic information of historical orders includes:

[0051] Order ID: This includes the order number, platform source (such as Meituan, Ele.me, etc.), order time, payment time, etc., used to uniquely identify each transaction;

[0052] User information: This includes user ID, contact information, address, membership level, etc. Some systems support associating user profile data (such as consumption preferences and historical behavior);

[0053] Dish details: This includes dish name, category (e.g., staple food / drink), unit price, specifications (e.g., spiciness / portion), notes (e.g., "less oil"), etc. Some systems also record information about dish combination packages.

[0054] Historical order execution data includes:

[0055] Order status: including status records such as pending payment, accepted, in production, in delivery, completed, canceled, and abnormal status (such as timeout, refund);

[0056] Delivery information: including rider ID, delivery time (order receipt - delivery time), delivery distance, delivery fee, real-time track, etc.

[0057] Payment information: record payment method (online / offline), actual payment amount, discounts (such as coupons, platform subsidies), and account details (such as platform commission);

[0058] Historical consumption behavior data includes:

[0059] Consumption frequency: daily / weekly / monthly order volume, repurchase rate, and consumption time distribution (e.g., lunch / dinner market share);

[0060] Consumption amount: average order value, total order amount, price fluctuations of dishes, and user consumption levels (e.g., high frequency low consumption / low frequency high consumption);

[0061] Related consumption: dish pairing patterns (e.g., the correlation between "hamburger + fries"), combo purchase preferences, and cross-category consumption ratios;

[0062] Historical operations management data includes:

[0063] Production data: kitchen order response time, dish preparation time, food delivery efficiency (such as average food delivery per hour), and waste rate;

[0064] Service quality: delivery on-time rate, negative review rate, complaint type (such as food quality issues / delivery delays), customer satisfaction score;

[0065] Supply chain data: food inventory and order matching, stocking warning (such as out-of-stock rate of popular products), cold chain logistics timeliness;

[0066] Historical extended data dimensions include:

[0067] Marketing effectiveness: coupon redemption rate, event participation (such as discounts / flash sales), and fission communication effect;

[0068] Environmental data: weather impact (such as changes in order volume on rainy days), holiday fluctuations, and monitoring of competitive product activities;

[0069] Equipment data: card machine transaction records, smart food cabinet pickup data, and unmanned delivery equipment status;

[0070] The multidimensional feature engineering space includes the temporal feature engineering architecture, the spatial feature engineering architecture, the user behavior feature engineering architecture, and the multidimensional fusion feature engineering architecture. The multidimensional fusion feature engineering architecture is connected to the temporal feature engineering architecture, the spatial feature engineering architecture, and the user behavior feature engineering architecture respectively.

[0071] The time feature engineering setting includes a periodic consumption feature extraction model, the spatial feature engineering setting includes a consumption circle feature extraction model, the user behavior feature engineering architecture setting includes a user portrait feature extraction model, and the multi-dimensional combination feature engineering architecture setting includes a tensor combination model;

[0072] The formula is:

[0073] X k =Σx(n)×e -j2πkn / N ,k=0,1,...,N-1

[0074] Where, X k is the Fourier coefficient, a complex number, representing the signal component at frequency k, where k ranges from 0 to N-1. It is used to analyze the periodicity of different frequencies; x(n) is a discrete time series (order quantity / customer price series); N is the number of sampling points (7*24=168 hours of granularity is recommended); j is the imaginary unit; k is the frequency indicator; n is the time indicator; e -j2πkn / N It is a complex exponential kernel used to convert time domain signals to frequency domain, involving the calculation of frequency and phase;

[0075] X time =RF(X k )

[0076] Where, X time is the time feature; X(*) is the time feature extraction function;

[0077] The temporal feature engineering architecture is primarily responsible for extracting time-related feature information from historical ordering data. This architecture typically includes a periodic consumption feature extraction model to identify and extract user consumption patterns and cyclical patterns across different time dimensions (such as hours, days, weeks, and seasons). For example, this model can identify peak lunch and dinner times, differences in weekend and weekday consumption, and holiday consumption patterns.

[0078] By identifying cyclical patterns in mealtimes, the framework can accurately capture users' potential demand for specific dishes during specific time periods. For example, during lunch, users may prefer quick and healthy dishes, while during dinner, they may prefer hearty and diverse dishes.

[0079] The cyclical consumption patterns extracted by the temporal feature engineering framework provide important temporal information for the demand forecasting model, significantly improving the accuracy of the model's predictions of user menu category demand over the next period of time. This allows the platform to formulate more refined operational strategies, such as launching discounted dishes or set meals that meet user needs during specific time periods, thereby increasing user satisfaction and platform revenue.

[0080] The spatial feature engineering architecture is primarily responsible for extracting spatially relevant feature information from historical ordering data. This architecture typically includes a consumer circle feature extraction model to identify and extract consumer spending patterns and preferences across different geographic locations. For example, this model can identify consumption levels, consumption habits, and popular dishes in different regions.

[0081] The formula is:

[0082]

[0083] θ k′ =(μ k′ ,∈ k′ ,π k′ )

[0084] Where, is the probability density, used for user attribution calculation, The larger the value, the higher the probability that the user belongs to this circle; is the observation data, including longitude and latitude coordinates; θ k′ is a complete parameter set, expressing the complete parameter set of the k′th consumer circle; k′ is the indicator of the consumer circle; μ k′ is the geographic centroid coordinate (latitude + longitude); ∈ k′ is the spatial distribution characteristics of consumer circles; k′ The market weight of the consumer circle; is the Gaussian distribution function in three-dimensional space (longitude, latitude, time);

[0085]

[0086] Where, X space is the spatial feature; Rd(*) is the spatial feature extraction function;

[0087] By identifying consumption characteristics in different regions, the framework can accurately identify differences in consumer preferences and needs within specific areas. For example, in economically developed areas, users may be more inclined to choose high-quality, high-priced dishes; whereas in economically underdeveloped areas, users may be more inclined to choose dishes with high cost-effectiveness.

[0088] Based on the analysis results of the spatial feature engineering architecture, the platform can formulate regionally differentiated operational strategies. For example, it can launch special dishes or services that meet the needs of local users in different regions, thereby improving user satisfaction and market share.

[0089] The spatial feature engineering architecture can help the platform optimize resource allocation. For example, it can rationally allocate inventory and distribution resources based on consumption demand in different regions, thereby improving operational efficiency.

[0090] The user behavior feature engineering architecture is primarily responsible for extracting user behavior-related feature information from historical ordering data. This architecture typically includes a user profile feature extraction model, which is used to construct user profiles and identify and extract features such as user consumption habits, preferences, and attributes. For example, this model can identify user taste preferences, spending power, and dining scenarios.

[0091] The formula is:

[0092] Document-topic distribution: θ d ~Dir((α)

[0093] Topic-behavior distribution: φ z ~Dir((β)

[0094] User behavior sequence: w n’ ~Multinomial(φ zn’ )

[0095] Where θ d is the topic distribution of the document (user behavior sequence), that is, the proportion of different topics in each user's behavior; Dir(α) is the Dirichlet distribution; α is the hyperparameter of the Dirichlet distribution, which controls the distribution of topics. Here, α is set to 36 1 / 36 vectors, indicating that the prior probability of each topic is the same, all 1 / 36. At the time of initialization, the model assumes that the probability of each topic appearing in the document is equal and there is no bias; φ zis the behavior distribution under the zth topic, that is, the probability distribution of different user behaviors under a certain topic; z is the topic indicator; Dir(β) is the Dirichlet distribution; β is also a hyperparameter of the Dirichlet distribution, which is set to 0.1 here. A smaller β value will lead to a more concentrated distribution of behaviors within the topic, that is, the probability of some behaviors appearing under each topic is higher, while other behaviors are less likely to appear. This may be related to the characteristics of user behavior, such as some behaviors (such as clicks) are more common under specific topics; z n ' is the hidden variable of the subject; Multinomial(*) is the multinomial distribution function; φ zn’ The distribution of user behaviors according to themes; n’ It is the user behavior sequence;

[0096] X behavior =Rx(θ d ,φ z ,w n′ )

[0097] Where, X space is the user behavior feature; Rd(*) is the user behavior feature extraction function;

[0098] By building user portraits, the architecture can accurately depict users' consumption characteristics and needs, providing a foundation for personalized recommendations and precision marketing;

[0099] Based on the analysis results of the user behavior feature engineering architecture, the platform can make more accurate personalized recommendations. For example, it can recommend dishes that suit the user's taste preferences, thereby improving user satisfaction and repurchase rate.

[0100] User behavior feature engineering architecture can help platforms conduct targeted marketing, such as launching different promotions for users with different spending power, thereby improving marketing efficiency and conversion rates;

[0101] The multi-dimensional fusion feature engineering architecture is mainly responsible for combining the feature information extracted by the temporal feature engineering architecture, the spatial feature engineering architecture, and the user behavior feature engineering architecture to generate a more comprehensive and in-depth feature representation. This architecture usually includes a tensor combination model to combine features of different dimensions and capture the interactive relationships between features.

[0102] χ=ZK[X time ,X space ,X behavior ]

[0103] Where ZK[*] is the tensor combination function; χ is the multidimensional fusion feature;

[0104] By integrating feature information from different dimensions, this architecture can more comprehensively capture the complex needs of users and avoid the limitations of single-dimensional features.

[0105] The feature representation generated by the multi-dimensional combined feature engineering architecture contains richer information and can significantly improve the predictive and generalization capabilities of subsequent models.

[0106] Collect historical ordering data from the ordering platform and extract real-time multi-dimensional combined features of the historical ordering data at the current moment based on the preset multi-dimensional feature engineering space. This includes the following steps:

[0107] S1-1: Collect historical ordering data from the ordering platform and pre-process the historical ordering data to obtain pre-processed historical ordering data;

[0108] S1-2: Use the periodic consumption feature extraction model of time feature engineering in the preset multi-dimensional feature engineering space to extract the real-time time features of the historical ordering data at the current moment;

[0109] S1-3: Use the consumer circle feature extraction model of spatial feature engineering in the multidimensional feature engineering space to extract the real-time spatial features of historical ordering data at the current moment;

[0110] S1-4: Use the user profile feature extraction model of user behavior feature engineering in the multidimensional feature engineering space to extract real-time user behavior features of historical ordering data at the current moment;

[0111] S1-5: The tensor combination model of the multidimensional combination feature engineering architecture in the multidimensional feature engineering space combines the real-time time features, the real-time spatial features, and the real-time user behavior features to obtain real-time multidimensional combination features;

[0112] S2: Based on the real-time multi-dimensional combination features, the pre-trained user menu category demand analysis model is used to perform analysis to obtain the real-time user menu category demand analysis results;

[0113] The user menu category demand analysis model is built based on the Long Short Term Memory (LSTM)-Dynamic Gated Attention Network (DAGN) algorithm. The model includes a temporal feature channel, a spatial feature channel, a user behavior feature channel, and a dynamic gated attention module. The temporal feature channel, spatial feature channel, and user behavior feature channel are set in parallel and are all connected to the dynamic gated attention module.

[0114] The input dimensions of the user menu category demand analysis model include: 6-dimensional features (temporal features) such as period parameter, amplitude, and time zone offset; 8-dimensional features (spatial features) such as latitude and longitude centroid, circle affiliation, and circle density; Latent Dirichlet Allocation (LDA) topic distribution; and label distribution, totaling 34-dimensional features (user behavior features). The activation function is the improved Swish function (which improves gradient stability compared to ReLU).

[0115] The time feature channel processes and analyzes real-time input temporal features (such as period parameters, amplitude, and time zone offset). This channel leverages the recurrent structure of LSTM (Long Short-Term Memory) networks to capture long-term dependencies and cyclical patterns in time series data. Through LSTM, the model can learn about changes in user consumption habits over different timescales (such as hours, days, weeks, and months), such as preferences for specific dishes during specific time periods.

[0116] LSTM can effectively capture long-term dependencies in time series data, enabling the model to understand long-term trends in user spending habits. Through LSTM, the model can adapt to and predict cyclical changes in user spending habits, such as differences in spending between weekdays and weekends. LSTM's ability to process time series data enhances the model's robustness to temporal features, ensuring stable performance across different time scales.

[0117] The spatial feature channel is responsible for processing and analyzing the input real-time spatial features (such as latitude and longitude centroid, circle affiliation, circle density, etc.). This channel uses graph feature embedding technology to embed spatial features into a low-dimensional vector space, thereby capturing the complex relationships between spatial features. For example, through graph embedding, the model can understand the differences in consumption habits of users in different regions.

[0118] Graph feature embedding technology can capture complex relationships between spatial features, such as differences in user consumption habits across different regions. Through graph embedding, high-dimensional spatial features are compressed into a low-dimensional vector space, reducing the computational complexity of the model. Graph feature embedding technology also enhances the model's ability to generalize spatial dimensional features, enabling the model to better adapt to data distribution across different spatial regions.

[0119] The user behavior feature channel is responsible for processing and analyzing real-time user behavior features (such as LDA topic distribution and tag distribution) as input. This channel dynamically updates and optimizes user behavior features using preference evolution techniques. For example, through comparative learning algorithms, the model can learn the latest changes in user behavior features, thereby more accurately reflecting the user's current preferences.

[0120] Preference evolution technology can dynamically update user behavior characteristics, allowing the model to promptly reflect the latest changes in user preferences; through comparative learning algorithms, the model can learn subtle changes in user behavior characteristics, enhancing the model's adaptability to user behavior characteristics; dynamically updated user behavior characteristics improve the model's prediction accuracy for user menu category needs

[0121] The dynamic gated attention module is built based on the dynamic gated attention mechanism, which includes: a multi-head self-attention structure, a gated control structure, and a cross-modal feature alignment loss function. It is the core of the entire model and is responsible for integrating the outputs of the temporal feature channel, the spatial feature channel, and the user behavior feature channel. The module includes a multi-head self-attention structure, a gated control structure, and a cross-modal feature alignment loss function. The multi-head self-attention structure can capture the complex interactions between different features. The gated control structure can dynamically adjust the weights of different feature channels. The cross-modal feature alignment loss function can ensure that the outputs of different feature channels are semantically aligned. The multi-head self-attention structure can capture the complex interactions between different features, such as the interaction between temporal, spatial, and user behavior features. The gated control structure can dynamically adjust the weights of different feature channels, so that the model can automatically optimize the feature fusion strategy according to the characteristics of the input data. The cross-modal feature alignment loss function ensures that the outputs of different feature channels are semantically aligned, enhancing the overall consistency of the model. Through the dynamic gated attention module, the model can more effectively integrate multi-dimensional feature information, significantly improving the performance of the model.

[0122] Based on the real-time multi-dimensional combination features, a pre-trained user menu category demand analysis model is used to perform analysis to obtain the real-time user menu category demand analysis results, including the following steps:

[0123] S2-1: Use the time feature channel in the pre-trained user menu category demand analysis model to perform additional feature enhancement on the real-time time feature in the real-time multi-dimensional combination feature to obtain the enhanced real-time time feature;

[0124] In this embodiment, the holiday vector is fused with the real-time time feature in the real-time multi-dimensional combined feature to obtain an enhanced real-time time feature.

[0125] S2-2: Use the spatial feature channel in the user menu category demand analysis model to perform graph feature embedding on the real-time spatial features in the real-time multi-dimensional combination features to obtain the embedded real-time spatial features;

[0126] In this embodiment, the merchant distribution map feature is embedded in the real-time spatial feature in the real-time multi-dimensional combination feature to obtain the embedded real-time spatial feature;

[0127] S2-3: Using the user behavior feature channel in the user menu category demand analysis model to perform preference evolution on the real-time user behavior features in the real-time multi-dimensional combination features, and obtaining the real-time user behavior features after preference evolution;

[0128] In this embodiment, based on historical user behavior characteristics, a contrastive learning algorithm is used to perform preference evolution on the real-time user behavior characteristics in the real-time multi-dimensional combination characteristics to obtain the real-time user behavior characteristics after preference evolution;

[0129] S2-4: Based on the dynamic gated attention weights, the dynamic gated attention module in the user menu category demand analysis model performs feature fusion on the enhanced real-time temporal features, the embedded real-time spatial features, and the real-time user behavior features after preference evolution to obtain real-time fused features.

[0130] S2-5: Based on the real-time fusion features, perform user menu category demand analysis to obtain real-time user menu category demand analysis results;

[0131] S3: Based on the real-time user menu demand analysis results, a pre-trained menu category optimization model is used to generate a real-time menu category optimization solution.

[0132] The menu category optimization model is built based on the Multi-objective Crested Porcupine Optimizer (MOCPO) algorithm, and includes an influence factor generation module, an optimization target update module, an initialization module, an iterative optimization module, and a vector decoding module connected in sequence.

[0133] Based on the real-time user menu demand analysis results, a pre-trained menu category optimization model is used to generate a real-time menu category optimization solution, which includes the following steps:

[0134] S3-1: Based on the real-time user menu demand analysis results, use the influence factor generation module of the pre-trained menu category optimization model to generate the corresponding real-time influence factors;

[0135] By simultaneously considering multiple influencing factors, such as menu diversity, customer experience, and cost control, through an iterative optimization process, it is possible to find a balance between multiple objectives and generate menu category optimization solutions that meet multiple needs, adapting to the diverse needs of different user groups. In this embodiment, the real-time influencing factors include user satisfaction, supply chain constraints, and menu diversity.

[0136] S3-2: Based on the real-time influencing factors, use the optimization target update module of the menu category optimization model to update the predicted optimization target, obtain the real-time optimization target, and set the real-time fitness function based on the real-time optimization target;

[0137] The formula is:

[0138] fit(x)=min[W1×AX(x)+W2×AC(x)+W3×AD(x)]

[0139] Where fit(x) is the real-time fitness function; AX(x) is the user complaint function; AC(x) is the supply chain constraint function; AD(x) is the dish uniqueness function; x is the MOCPO individual; W1, W2, and W3 are the first, second, and third weight values;

[0140] S3-3: Encode the initial real-time menu category optimization solution into an individual vector of the initialization module. Based on the individual vector, use the initialization module of the menu category optimization model to generate several initial solutions (initial MOCPO individuals), and obtain an initial MOCPO population consisting of several initial MOCPO individuals.

[0141] The formula is:

[0142]

[0143] Where, is the initial MOCPO individual of the Circle chaos map, i.e. the initial solution; is the randomly generated initial MOCPO individual; i' is the MOCPO individual indicator; compared with the randomly distributed population, the initial position distribution of the improved MOCPO individuals generated by the Circle chaotic mapping sequence is more uniform, which expands the search range of the algorithm in space, increases the diversity of group positions, and to a certain extent improves the defect of the algorithm easily falling into local extreme values, thereby improving the optimization efficiency of the algorithm;

[0144] S3-4: Based on the real-time fitness function, use the iterative optimization module of the menu category optimization model to iteratively optimize several initial solutions to obtain the optimal solution, including the following steps:

[0145] S3-4-1: Set the MOCPO population parameters and the maximum number of iterations, introduce a cyclic population reduction mechanism, limit the number of individuals in the MOCPO population parameters, and obtain the updated MOCPO population parameters for the next iteration;

[0146] The formula is:

[0147]

[0148] Where S t+1 is the number of individuals in the MOCPO population parameter of the t+1th iteration; S tis the number of individuals in the MOCPO population parameter of the tth iteration; S min is the minimum number of individuals in the MOCPO population parameter; a' is the function evaluation parameter; V is the function evaluation loop parameter; V max is the maximum function evaluation loop parameter; t is the number of iterations indicator;

[0149] S3-4-2: Calculate the initial fitness value of the initial MOCPO individual in the initial MOCPO population according to the real-time fitness function;

[0150] S3-4-3: Based on the initial fitness value and the updated MOCPO population parameters, the first defense strategy, the second defense strategy, the third defense strategy, and the fourth defense strategy are used to update the initial MOCPO population to obtain an updated MOCPO population;

[0151] The formula for the first defense strategy is:

[0152]

[0153] Where, For the updated MOCPO individuals within the first defense range; is the initial MOCPO individual within the first defense range; τ1 is a random number based on normal distribution; τ2 is a random value in the interval [0,1]; It is the optimal solution within the first defense range; is the vector generated between the true optimal solution within the first defense range and the optimal solution randomly selected from the MOCPO population; i' is the MOCPO individual indicator; t is the iteration indicator;

[0154] The formula for the second defense strategy is:

[0155]

[0156] Where, For the updated MOCPO individuals within the second defense range; The initial MOCPO individual within the second defense range; is the search upper limit vector of the second defense range; τ3 is a random value in the interval [0,1]; are the r1th and r2th initial MOCPO individuals respectively; r1 and r2 are two random integers between [1, S]; is the vector generated between the true optimal solution within the second defense range and the optimal solution randomly selected from the MOCPO population;

[0157] The formula for the third defense strategy is:

[0158]

[0159] Where, For the updated MOCPO individuals within the third defense range; The initial MOCPO individual within the third defense range; is the search upper limit vector of the third defense range; are the r2th and r3th initial MOCPO individuals respectively; r3 is a random integer between [1, S]; The odor diffusion factor defined for the fitness function; λ t It is a defense factor; Control parameters for search direction;

[0160] The formula for the fourth defense strategy is:

[0161]

[0162] Where, For the updated MOCPO individuals within the fourth defense range; For the initial MOCPO individual within the fourth defense range; is the optimal solution within the fourth defense range; τ4 and τ5 are both random values ​​in the interval [0,1]; λ t It is a defense factor; Control parameters for search direction; is the average force affecting the search direction; a' is the convergence speed factor;

[0163] S3-4-4: Use the dynamic reverse learning algorithm to perform dynamic reverse learning on the updated MOCPO population to generate a dynamic reverse MOCPO population;

[0164] The formula is:

[0165]

[0166] Where, is the dynamically reversed MOCPO individual; γ' is the decreasing inertia coefficient; L max 、L min are the maximum and minimum values ​​of the vector space respectively; For the newer MOCPO individuals;

[0167] S3-4-5: Calculate the fitness values ​​of all MOCPO individuals in the updated MOCPO population and the dynamically reversed MOCPO population according to the real-time fitness function, take the MOCPO individual with the minimum fitness value as the optimal individual, and retain the optimal individual;

[0168] S3-4-6: If the number of iterations of iterative optimization reaches the maximum number of iterations or the fitness value of the optimal individual meets the requirements, the optimal solution corresponding to the optimal individual is output;

[0169] S3-5: Use the vector decoding module of the menu category optimization model to decode the individual vectors of the optimal solution to obtain the optimal real-time menu category optimization solution;

[0170] S4: According to the real-time menu category optimization plan, the preset menu category data of the ordering platform is optimized to obtain optimized menu category data.

[0171] Example 2:

[0172] like Figure 2 As shown, this embodiment provides a menu category optimization system based on historical meal ordering data, which is used to implement a menu category optimization method, including a multi-dimensional combination feature extraction unit, a menu category demand analysis unit, a menu category optimization unit, and an optimization solution execution unit connected in sequence;

[0173] A multi-dimensional combination feature extraction unit is used to collect historical ordering data from the ordering platform and extract real-time multi-dimensional combination features of the historical ordering data at the current moment based on a preset multi-dimensional feature engineering space;

[0174] A menu category demand analysis unit is used to analyze the user menu category demand based on real-time multi-dimensional combination features using a pre-trained user menu category demand analysis model to obtain real-time user menu category demand analysis results;

[0175] The menu category optimization unit is used to generate a solution based on the real-time user menu demand analysis results using a pre-trained menu category optimization model to obtain a real-time menu category optimization solution;

[0176] The optimization plan execution unit is used to optimize the preset menu category data of the ordering platform according to the real-time menu category optimization plan to obtain the optimized menu category data.

[0177] The present invention provides a menu category optimization method and system based on historical meal ordering data, which constructs a three-dimensional feature engineering space including time, space, and user behavior, and realizes the comprehensive capture of complex user needs. This multi-dimensional feature fusion method effectively overcomes the defect of poor comprehensiveness of the existing technology, making the optimization result closer to actual needs; uses a user menu category demand analysis model for analysis, explores the deep relationship between user needs and menu categories in historical meal ordering data, uses a menu category optimization model to dynamically generate menu category optimization solutions, and dynamically adjusts and continuously optimizes according to real-time user needs, which can respond to market changes in a timely manner and effectively improve the market adaptability of menu categories and the accuracy of optimization solutions; adopts automated feature extraction, intelligent analysis and dynamic optimization to realize the construction of an automated and intelligent system for menu category optimization. Through this automated decision-making mechanism, human intervention is reduced and the degree of intelligence is improved.

[0178] The present invention is not limited to the above optional embodiments. Anyone can derive various other forms of products based on the teachings of the present invention. The above specific embodiments should not be construed as limiting the scope of protection of the present invention. The scope of protection of the present invention shall be based on the scope defined in the claims, and the description can be used to interpret the claims.

Claims

1. A menu category optimization method based on historical meal ordering data, characterized by: The steps include: Collect historical ordering data from the ordering platform and extract real-time multi-dimensional combined features of the historical ordering data at the current moment based on the preset multi-dimensional feature engineering space; Based on the real-time multi-dimensional combination features, the pre-trained user menu category demand analysis model is used to perform analysis and obtain the real-time user menu category demand analysis results; Based on the real-time user dish demand analysis results, a pre-trained menu category optimization model is used to generate a real-time menu category optimization solution. According to the real-time menu category optimization plan, the preset menu category data of the ordering platform is optimized to obtain the optimized menu category data.

2. The menu category optimization method based on historical ordering data according to claim 1, characterized in that: The multidimensional feature engineering space includes a time feature engineering architecture, a space feature engineering architecture, a user behavior feature engineering architecture and a multidimensional fusion feature engineering architecture. The multidimensional fusion feature engineering architecture is connected to the time feature engineering architecture, the space feature engineering architecture and the user behavior feature engineering architecture respectively.

3. The menu category optimization method based on historical ordering data according to claim 2, characterized in that: The time feature engineering is provided with a periodic consumption feature extraction model, the space feature engineering is provided with a consumption circle feature extraction model, the user behavior feature engineering architecture is provided with a user portrait feature extraction model, and the multi-dimensional combination feature engineering architecture is provided with a tensor combination model.

4. The menu category optimization method based on historical ordering data according to claim 3, characterized in that: Collect historical ordering data from the ordering platform and extract real-time multi-dimensional combined features of the historical ordering data at the current moment based on the preset multi-dimensional feature engineering space. This includes the following steps: Collect historical ordering data from the ordering platform and pre-process the historical ordering data to obtain pre-processed historical ordering data; Use the periodic consumption feature extraction model of time feature engineering in the preset multi-dimensional feature engineering space to extract the real-time time features of historical ordering data at the current moment; Use the consumer circle feature extraction model of spatial feature engineering in the multidimensional feature engineering space to extract the real-time spatial features of historical ordering data at the current moment; Use the user profile feature extraction model of user behavior feature engineering in the multi-dimensional feature engineering space to extract real-time user behavior features of historical ordering data at the current moment; The tensor combination model of the multidimensional combination feature engineering architecture in the multidimensional feature engineering space combines the real-time time features, the real-time spatial features and the real-time user behavior features to obtain real-time multidimensional combination features.

5. The menu category optimization method based on historical ordering data according to claim 4, characterized in that: The user menu category demand analysis model is constructed based on the LSTM-DAGN algorithm, and the user menu category demand analysis model includes a time feature channel, a spatial feature channel, a user behavior feature channel and a dynamic gated attention module. The time feature channel, the spatial feature channel and the user behavior feature channel are set in parallel, and the time feature channel, the spatial feature channel and the user behavior feature channel are all connected to the dynamic gated attention module.

6. The menu category optimization method based on historical meal ordering data according to claim 5, characterized in that: Based on the real-time multi-dimensional combination features, a pre-trained user menu category demand analysis model is used to perform analysis to obtain the real-time user menu category demand analysis results, including the following steps: Use the time feature channel in the pre-trained user menu category demand analysis model to perform additional feature enhancement on the real-time time feature in the real-time multi-dimensional combination feature to obtain the enhanced real-time time feature; Use the spatial feature channel in the user menu category demand analysis model to embed the real-time spatial features in the real-time multi-dimensional combination features into graph features to obtain the embedded real-time spatial features; The user behavior feature channel in the user menu category demand analysis model is used to perform preference evolution on the real-time user behavior features in the real-time multi-dimensional combination features, and the real-time user behavior features after preference evolution are obtained; According to the dynamic gated attention weight, the dynamic gated attention module in the user menu category demand analysis model is used to fuse the enhanced real-time temporal features, the embedded real-time spatial features, and the real-time user behavior features after preference evolution to obtain real-time fused features. Based on the real-time fusion features, user menu category demand analysis is performed to obtain real-time user menu category demand analysis results.

7. The menu category optimization method based on historical ordering data according to claim 6, characterized in that: The holiday vector is fused with the real-time time feature in the real-time multi-dimensional combined feature to obtain the enhanced real-time time feature; Embed the merchant distribution map feature into the real-time spatial feature in the real-time multi-dimensional combination feature to obtain the embedded real-time spatial feature; According to historical user behavior characteristics, a contrastive learning algorithm is used to perform preference evolution on the real-time user behavior characteristics in the real-time multi-dimensional combination characteristics, and the real-time user behavior characteristics after preference evolution are obtained.

8. The menu category optimization method based on historical ordering data according to claim 7, characterized in that: The menu category optimization model is constructed based on the MOCPO algorithm, and the menu category optimization model includes an influence factor generation module, an optimization target update module, an initialization module, an iterative optimization module and a vector decoding module which are connected in sequence.

9. The menu category optimization method based on historical ordering data according to claim 8, characterized in that: Based on the real-time user menu demand analysis results, a pre-trained menu category optimization model is used to generate a real-time menu category optimization solution, which includes the following steps: Based on the real-time user menu demand analysis results, the influence factor generation module of the pre-trained menu category optimization model is used to generate the corresponding real-time influence factors; Based on the real-time influencing factors, the optimization target update module of the menu category optimization model is used to update the predicted optimization target, obtain the real-time optimization target, and set the real-time fitness function based on the real-time optimization target; Encode the initial real-time menu category optimization plan into individual vectors of the initialization module, and use the initialization module of the menu category optimization model to generate several initial solutions based on the individual vectors; Based on the real-time fitness function, the iterative optimization module of the menu category optimization model is used to iteratively optimize several initial solutions to obtain the optimal solution; Use the vector decoding module of the menu category optimization model to decode the individual vectors of the optimal solution and obtain the optimal real-time menu category optimization plan.

10. A menu category optimization system based on historical ordering data, used to implement the menu category optimization method according to any one of claims 1 to 9, characterized in that: It includes a multi-dimensional combination feature extraction unit, a menu category demand analysis unit, a menu category optimization unit and an optimization plan execution unit which are connected in sequence.

Images