User online ordering preference analysis method based on deep learning

Abstract

The invention relates to the technical field of online ordering, in particular to a user online ordering preference analysis method based on deep learning, and the method comprises the steps: collecting ordering data of a user in a target application, and carrying out the preprocessing of the ordering data; constructing a multi-model coordination analysis framework; based on the preprocessed ordering data, analyzing the ordering preference of the user through a multi-model coordination analysis framework; the ordering data comprises user behavior data and dish feature data. Through coordination and cooperation of a plurality of models in the multi-model coordination analysis framework, user preferences are analyzed from multiple perspectives, so that comprehensive and diversified preference analysis is provided.

Description

Technical Field

[0001] This invention relates to the field of online food ordering technology, and in particular to a method for analyzing user online food ordering preferences based on deep learning. Background Technology

[0002] Online ordering refers to the process where customers browse menus, select items, and complete ordering and payment through online ordering applications using internet-connected devices (such as mobile phones, tablets, and computers), eliminating the need for traditional paper menus or direct communication with waiters, thus improving ordering efficiency. Behind this efficiency improvement is the analysis of users' online ordering preferences through algorithmic models, an intelligent process encompassing user data collection, data analysis and modeling, and menu recommendation.

[0003] However, while pursuing efficiency, algorithmic models also introduce problems of rigidity and narrowing. Specifically, in pursuit of click-through rates or conversion rates, algorithmic models continuously reinforce existing user preferences, limiting the possibility of users exploring new flavors and diverse choices. Therefore, we propose a deep learning-based method for analyzing online food ordering preferences to address these issues. Summary of the Invention

[0004] This invention analyzes user preferences from multiple perspectives by coordinating multiple models within a multi-model coordination analysis framework, thereby providing comprehensive and diverse preference analysis and alleviating the problems of rigidity and narrowing.

[0005] The technical solution proposed in this invention is: a method for analyzing users' online food ordering preferences based on deep learning, the method comprising: collecting users' food ordering data in a target application and preprocessing the food ordering data; Construct a multi-model coordination analysis framework; Based on the preprocessed order data, user order preferences are analyzed using a multi-model coordination analysis framework. The ordering data includes user behavior data and dish feature data.

[0006] Preferred options also include: Dynamically set the calling frequency, calling status, and calling order of models within the multi-model coordinated analysis framework; The model selection decision framework determines the models included within the multi-model coordinated analysis framework during deployment.

[0007] Preferably, the construction of the multi-model coordinated analysis framework includes: Obtain the collaborative filtering model, analyze the user-food interaction behavior through the collaborative filtering model, and discover the relationship between users and food. Obtain a content-based model and analyze the relationship between the attributes of dishes and the user's historical interactions with the content of dishes through the content-based model; Obtain the time series model and the exploration model, and analyze the user's real-time dynamic needs through the time series model and the exploration model; Obtain a context-aware model and analyze the impact of contextual information on user preferences using the context-aware model; Obtain the fusion ranking model, and generate a recommendation list by fusing the analysis results of the collaborative filtering model, content-based model, time-series model, knowledge graph model, and context-aware model.

[0008] Preferably, the step of analyzing user-dish interaction behavior through a collaborative filtering model to discover the association between users and dishes includes: A collaborative filtering model is constructed using the Item-CF algorithm. Calculate the similarity between different dishes ;in, Indicates user With dishes Interaction behavior judgment function between them Indicates user With dishes Interaction behavior judgment function between them Indicates an indicator function; This represents the collection of all users in the target application; Obtain user behavior data, extract the menu items that the user has recently interacted with, and construct the user profile. Historical interactive food collection The user behavior data includes query, click, and order data.

[0009] Then the user For dishes The predicted interest score is ;in, Indicates user For dishes The intensity of interaction; The analysis of the relationship between dish attributes and user's historical interaction with dish content through a content-based model includes: Obtain dish feature data and construct dish feature vectors. ;in, and Embedding codes that represent the cuisine, main ingredients, flavor, price range, cooking method, rating, and availability of a dish; users The user interest vector is obtained by averaging the feature vectors of all dishes that have been interacted with in the past. ,in, This indicates the number of dishes that have been interacted with in the past; Calculate the similarity score between the user vector and the feature vector of the candidate dishes. ;in, This represents the feature vector of the candidate dish.

[0010] Preferably, the step of analyzing the user's real-time dynamic needs through time-series models and exploratory models includes: The popularity score of dishes within a rolling calculation time window is calculated using a time-series model, including: Define the time window; Acquire user behavior data and dish feature data within multiple consecutive time windows within the target application; Click-through rate and order volume of dishes are calculated based on user behavior data, and rating and negative review rate of dishes are calculated based on dish feature data; Calculate the popularity score of the dish. ;in, and Each represents a dish Order volume, click-through rate, and negative review rate; This represents the weight hyperparameter; Based on the dish's listing time, an exploration model is used to randomly explore from the set of explored dishes to obtain the popularity score of the explored dishes, including: The Bandit algorithm, an exploratory algorithm, was used to explore within the menu to obtain the available time ranges. And the click-through rate is lower than the click-through rate threshold. The dishes; Calculate the popularity score of the dishes discovered. ; Each represents a dish Order volume, click-through rate, rating, and negative review rate; The explored menu is based on popularity score ranges. A collection of dishes from outside the region; among them, Indicates the popularity score threshold; The recommended menu will consist of dishes within the popularity score range and dishes discovered through exploration, in order to reflect the dynamic needs of users. The analysis of the impact of contextual information on user preferences using a context-aware model includes: Predicting in a given scenario using a context-aware model The probability of a user interacting with a dish is calculated using a binary classification model to predict the probability score. ; in, This represents the Sigmoid activation function. Indicates the safety bias term; and They represent the feature weights, Represents the scenario feature vector; ;in, These respectively represent the dining time period, day of the week, delivery address type code, weather condition code, and special events; Preferably, the step of generating a recommendation list by fusing the analysis results of a collaborative filtering model, a content-based model, a time-series model, a knowledge graph model, and a context-aware model through a fusion ranking model includes: Get , constitute the model feature vector ; Will As the input vector, it is fed into a pre-trained initial ranking model, which outputs a click-through rate score. The preliminary sorting model is the GBDT (Boosted Tree Data) model. Based on the dishes Perform preliminary sorting, generate a preliminary preference list, and save it in the online log; Reordering is performed using a reordering model, specifically as follows: Calculate relevant scores for the dishes in the initial preference list. ;in, Indicates dishes Click-through rate score; Indicates dishes Dishes Similarity; This indicates the selected set of recommended dishes. This represents a set of candidate dishes; Indicates the trade-off coefficient; according to The dishes in the initial preference list are rearranged to generate the final preference list, which is then saved in the online log. The reordering model is the maximum edge correlation algorithm model.

[0011] Preferably, the dynamic setting of the calling frequency, calling status, and calling order of models within the multi-model coordination analysis framework includes: The multi-model coordination analysis framework is divided into three layers: offline, near-line, and online, to balance computational cost and practicality. Specifically: Collaborative filtering, content-based analysis, time-series, and fusion ranking models are assigned to the offline layer; exploration models are assigned to the near-line layer; and context-aware models are assigned to the online layer. The offline layer model is called on an hourly or daily basis; the near-line layer model is called on a minute-by-minute basis; and the online layer model is called on a millisecond-by-millisecond basis. During the analysis phase, the online layer model enters a parallel invocation state upon receiving a user request, and is processed in parallel with the offline and near-line layer models. During the sorting phase, after the collaborative filtering model, content-based analysis model, and time-series model output results, the fusion sorting model serially calls the state and sorts the dishes based on the outputs of the collaborative filtering model, content-based analysis model, and time-series model to generate a preliminary preference list and a final preference list. Based on the efficiency constraints of the target application, the calling frequency and update frequency of the fusion ranking model are dynamically adjusted, including: If the efficiency constraint of the target application is to efficiently analyze user preferences, then the call frequency of the fusion ranking model will be adjusted from daily to minute-level, and the GBDT model will be replaced by an online regression model or a neural network model to adapt to online learning; and the online training frequency of the initial ranking model will be increased to once per hour. Otherwise, maintain the original call frequency and training frequency of the fusion ranking model.

[0012] Preferably, the step of generating a recommendation list by fusing the analysis results of a collaborative filtering model, a content-based model, a time-series model, a knowledge graph model, and a context-aware model through a fusion ranking model further includes: updating the GBDT training sample set during the GBDT training phase, including: Identify dishes whose rankings have been boosted by the re-ranking model. If users have interacted with these dishes, obtain their feature vectors and increase the weights of training samples containing these feature vectors. ;in, This indicates the promotion bonus coefficient. This indicates the percentage increase in ranking.

[0013] Identify dishes whose ranking has been lowered by the re-ranking model. If there is no user interaction with the dish, obtain the dish's feature vector and reduce the weight of training samples containing that dish's feature vector. .

[0014] Preferably, the model selection decision framework determines the models included within the multi-model coordinated analysis framework during deployment, including: Add user features to the user interest vector to create a multi-dimensional user profile. ;in This indicates the user's age, activity frequency encoding, and activity time period encoding; User-preference segmentation is performed using user segmentation algorithms, including: use Cluster analysis is performed using a pre-trained K-means clustering algorithm to classify users by value, dividing them into high-value users, medium-value users, and low-value users. Based on value classification, for each type of user, an optimal model combination is set, including: For high-value users and considering the application efficiency constraints of the target application, models within the multi-model coordination analysis framework are selected, including: If the efficiency constraint is to efficiently analyze user preferences, then the collaborative filtering model, content-based model, time series model, exploratory model, context-aware model, and fusion ranking model are selected, and the fusion ranking model is selected as the initial ranking model, either an online regression model or a neural network model. Otherwise, choose the collaborative filtering model, content-based model, time-series model, exploratory model, context-aware model, or fusion ranking model, where the initial ranking model in the fusion ranking model is the GBDT model. For mid-value users, the model, combined with the application efficiency constraints of the target application, selects models within a multi-model coordinated analysis framework, including: If the efficiency constraint is to efficiently analyze user preferences, then the collaborative filtering model, the content-based model, and the fusion ranking model are selected, and the fusion ranking model is selected as the initial ranking model, either an online regression model or a neural network model. Otherwise, choose the collaborative filtering model, the content-based model, or the fusion ranking model, where the initial ranking model in the fusion ranking model is the GBDT model. For low-value users, select models within the multi-model coordinated analysis framework, including: The selection criteria include collaborative filtering, time series, exploratory, and fusion ranking models; the initial ranking model within the fusion ranking model is the GBDT model.

[0015] A computer-readable storage medium storing a computer program that is executed by a processor to implement the deep learning-based method for analyzing online user ordering preferences.

[0016] The beneficial effects of this invention are: This invention analyzes users' ordering preferences through a multi-model coordination analysis framework. The collaborative work of multiple machine learning models improves the comprehensiveness of the analysis and the diversity of recommendations. Furthermore, by dynamically setting the call frequency, call status, and call order within the multi-model coordination framework, the invention ensures the implementation of preference analysis functions while also meeting the latency and resource utilization requirements of online applications. Attached Figure Description

[0017] Figure 1 This is a flowchart of the deep learning-based method for analyzing online food ordering preferences according to the present invention. Detailed Implementation

[0018] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious modifications will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0019] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0020] Example: Reference Figure 1 The technical solution provided by this invention is: a method for analyzing online food ordering preferences based on deep learning, the method comprising: Step 1: Collect user ordering data in the target application and preprocess the ordering data; the ordering data includes user behavior data and dish feature data.

[0021] Step 2: Construct a multi-model coordination analysis framework; Step 3: Based on the preprocessed order data, analyze user order preferences using a multi-model coordination analysis framework; Specifically, the following steps are included: Obtain the collaborative filtering model, analyze the user-food interaction behavior through the collaborative filtering model, and discover the association between users and foods. Specifically: A collaborative filtering model was constructed using the Item-CF algorithm; the similarity between different dishes was calculated. ;in, Indicates user With dishes Interaction behavior judgment function between them Indicates user With dishes Interaction behavior judgment function between them Indicates an indicator function; This represents the collection of all users in the target application.

[0022] If user With dishes If there is interaction between them, then Otherwise, it is 0; If user With dishes If there is interactive behavior between them, then Otherwise, it is 0; if ,but ;if ,but .

[0023] Obtain user behavior data, extract the menu items that the user has recently interacted with, and construct the user profile. Historical interactive food collection The user behavior data includes query, click, and order data.

[0024] Then, the user For dishes The predicted interest score is ;in, Indicates user For dishes The intensity of the interaction; if it's a click, then... If it is an order, then If it's a query, then .

[0025] Obtain a content-based model, and analyze the relationship between dish attributes and user's historical interaction with dish content using this model. Specifically: Obtain dish feature data and construct dish feature vectors. The feature vector of this dish can be used as a dish profile. Among them, and Embedding codes that represent the cuisine, main ingredients, flavor, price range, cooking method, rating, and availability of a dish.

[0026] users The user interest vector is obtained by averaging the feature vectors of all dishes that have been interacted with in the past. This user interest vector can serve as a preliminary user profile. This indicates the number of dishes that have been interacted with in the past.

[0027] Calculate the similarity score between the user vector and the feature vector of the candidate dishes. ;in, This represents the feature vector of the candidate dish.

[0028] Obtain the time-series model and the exploratory model, and analyze the user's real-time dynamic needs through the time-series model and the exploratory model, specifically including: The popularity score of dishes within a rolling calculation time window is calculated using a time-series model, including: Define the time window; Acquire user behavior data and dish feature data within multiple consecutive time windows within the target application; Click-through rate and order volume of dishes are calculated based on user behavior data, and rating and negative review rate of dishes are calculated based on dish feature data; Calculate the popularity score of the dish. ;in, and Each represents a dish Order volume, click-through rate, and negative review rate; This represents the weight hyperparameter, which can be tuned through A / B testing. In this embodiment, values ​​of 0.5, 0.3, 0.15, and 0.05 can be selected.

[0029] Based on the dish's listing time, an exploration model is used to randomly explore from the set of explored dishes to obtain the popularity score of the explored dishes, including: The Bandit algorithm, an exploratory algorithm, was used to explore within the menu to obtain the available time ranges. And the click-through rate is lower than the click-through rate threshold. The dishes.

[0030] Calculate the popularity score of the dishes discovered. ; Each represents a dish Order volume, click-through rate, rating, and negative review rate.

[0031] The explored menu is based on popularity score ranges. A collection of dishes from outside the region; among them, This represents the popularity score threshold.

[0032] The recommended menu will consist of dishes within the popularity score range and dishes discovered through exploration, in order to reflect the dynamic needs of users.

[0033] Because user preferences are not static, but rather the result of the combined effects of user attributes, dish characteristics, and context, the following steps address the issue of the same user having different needs in different scenarios.

[0034] Obtain a context-aware model and analyze its impact on user preferences. Specifically: Predicting in a given scenario using a context-aware model The probability of a user interacting with a dish (the occurrence of user behavior) is predicted using a binary classification model, specifically by calculating the probability score. .

[0035] in, This represents the Sigmoid activation function. Indicates the safety bias term; and They represent the feature weights, Represents the scenario feature vector; ;in, These represent the meal time period (0 for breakfast, 1 for lunch, 0 for dinner, and 0 for late-night snack), day of the week, delivery address type code (e.g., 1 for home, 0 for office, and 0 for school), weather condition code (e.g., 0 for sunny and 1 for rainy), and special events (e.g., 1 for holidays and 0 for no holidays).

[0036] For example, in the scenario If the scenario is {weekday, lunch, location: company}, the model will tend to give high scores to fast food, light meals, and business set meals, meaning it predicts that the user's preference is for fast food, light meals, and business set meals, and the probability of the user interacting with these dishes will increase.

[0037] The system acquires a fusion ranking model, which integrates the analysis results from collaborative filtering, content-based, time-series, knowledge graph, and context-aware models to generate a recommendation list. Specifically: Get , constitute the model feature vector ; Will As the input vector, it is fed into a pre-trained initial ranking model, which outputs a click-through rate score. The preliminary sorting model is the GBDT (Boosted Tree Data) model. Based on the dishes Perform preliminary sorting, generate a preliminary preference list, and save it in the online log; Reordering is performed using a reordering model, specifically as follows: Calculate relevant scores for the dishes in the initial preference list. ;in, Indicates dishes Click-through rate score; Indicates dishes Dishes Similarity; This indicates the selected set of recommended dishes. This represents a set of candidate dishes; This represents the tradeoff coefficient, a parameter used to weigh correlation and diversity.

[0038] according to The dishes in the initial preference list are rearranged to generate the final preference list, which is then saved in the online log. The reordering model is the maximum edge correlation algorithm model.

[0039] Example: Due to the high requirements for latency and throughput in online services, in order to both utilize multiple models to mine user preferences and meet the requirements of online services, we propose the following technical solution based on Example 1. The specific solution is as follows: The calling frequency, calling status, and calling order of models within the multi-model coordinated analysis framework can be dynamically set. Specifically: The multi-model coordination analysis framework is divided into three layers: offline, near-line, and online, to balance computational cost and practicality. Specifically: Collaborative filtering, content-based analysis, time-series, and fusion ranking models are assigned to the offline layer; exploration models are assigned to the near-line layer; and context-aware models are assigned to the online layer. The offline layer model is called on an hourly or daily basis; the near-line layer model is called on a minute-by-minute basis; and the online layer model is called on a millisecond-by-millisecond basis. During the preference analysis phase, the online layer model enters a parallel invocation state upon receiving a user request, processing tasks in parallel with the offline and nearline layer models. During the sorting phase, after the collaborative filtering model, content-based analysis model, and time-series model output results, the fusion sorting model serially calls the state and sorts the dishes based on the outputs of the collaborative filtering model, content-based analysis model, and time-series model to generate a preliminary preference list and a final preference list. Based on the efficiency constraints of the target application, the calling frequency and update frequency of the fusion ranking model are dynamically adjusted, including: If the efficiency constraint of the target application is to efficiently analyze user preferences, then the call frequency of the fusion ranking model should be adjusted from daily to minute-level, and the GBDT model should be replaced by an online regression model or a neural network model to adapt to online learning; and the online training frequency of the initial ranking model should be increased to once per hour; otherwise, the original call frequency and training frequency of the fusion ranking model should be maintained.

[0040] The model selection decision framework determines the models included within the multi-model coordinated analysis framework during deployment. Specifically: Add user features to the user interest vector to create a multi-dimensional user profile. ;in This indicates the user's age, activity frequency code (e.g., 1 for weekly activity, 2 for daily activity, and 0 for monthly activity), and activity time code (e.g., 1 for breakfast, 2 for lunch, and 3 for dinner).

[0041] User-preference segmentation is performed using user segmentation algorithms, including: use The system performs cluster analysis using a pre-trained K-means clustering algorithm to classify users by value, dividing them into high-value users (high activity), medium-value users (stable activity), and low-value users (low activity). Based on value classification, for each type of user, an optimal model combination is set, including: For high-value users and considering the application efficiency constraints of the target application, models within the multi-model coordination analysis framework are selected, including: If the efficiency constraint is to efficiently analyze user preferences, then the collaborative filtering model, content-based model, time series model, exploratory model, context-aware model, and fusion ranking model are selected, and the fusion ranking model is selected as the initial ranking model, either an online regression model or a neural network model. Otherwise, choose the collaborative filtering model, content-based model, time-series model, exploratory model, context-aware model, or fusion ranking model, where the initial ranking model in the fusion ranking model is the GBDT model. For mid-value users, the model, combined with the application efficiency constraints of the target application, selects models within a multi-model coordinated analysis framework, including: If the efficiency constraint is to efficiently analyze user preferences, then the collaborative filtering model, the content-based model, and the fusion ranking model are selected, and the fusion ranking model is selected as the initial ranking model, either an online regression model or a neural network model. Otherwise, choose the collaborative filtering model, the content-based model, or the fusion ranking model, where the initial ranking model in the fusion ranking model is the GBDT model. For low-value users, select models within the multi-model coordinated analysis framework, including: The selection criteria include collaborative filtering, time series, exploratory, and fusion ranking models; the initial ranking model within the fusion ranking model is the GBDT model.

[0042] Example: Since the goal of the GBDT model is to maximize CTR, and the goal of the Maximum Marginal Relevance (MMR) algorithm model is to optimize relevance and diversity, the GBDT model considers a perfect list to be destroyed by the MMR model in pursuit of diversity. However, the GBDT model does not learn from this destruction, resulting in its next output still being a list that will be reordered by the MMR model, thus causing internal friction. This not only reduces efficiency but also wastes computational resources. The optimization approach is to form a complete closed loop from GBDT generating an initial recommendation list to MMR reordering to optimize diversity, from user clicks on diverse content to GBDT learning this optimization, to GBDT generating an initial preference list that considers diversity, and finally to reducing MMR reordering. This reduces computational load and improves efficiency. Therefore, to enable the GBDT model to understand the intent of the MMR model in a timely manner, we propose the following technical solution based on Example 2: During the GBDT training phase, the GBDT training sample set is updated, including: Identify dishes whose rankings have been boosted by the re-ranking model. If users have interacted with these dishes, obtain their feature vectors and increase the weights of training samples containing these feature vectors. ;in, This indicates the promotion bonus coefficient. Indicates the percentage increase in ranking. ,in, Indicates dishes Rank in the initial preference list Indicates dishes Rank in the final preference list; Identify dishes whose ranking has been lowered by the re-ranking model. If there is no user interaction with the dish, obtain the dish's feature vector and reduce the weight of training samples containing that dish's feature vector. .

[0043] The present invention also provides a computer-readable storage medium storing a computer program that is executed by a processor to implement the deep learning-based method for analyzing user online ordering preferences.

[0044] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0045] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0046] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the principles described, the implementation of the present invention may have any changes or modifications.

Claims

1. A method for analyzing online food ordering preferences based on deep learning, characterized in that, The method includes: Collect user order data in the target application and preprocess the order data; Construct a multi-model coordination analysis framework; Based on the preprocessed order data, user order preferences are analyzed using a multi-model coordination analysis framework. The ordering data includes user behavior data and dish feature data.

2. The method for analyzing online food ordering preferences based on deep learning according to claim 1, characterized in that, Also includes: Dynamically set the calling frequency, calling status, and calling order of models within the multi-model coordinated analysis framework; The model selection decision framework determines the models included within the multi-model coordinated analysis framework during deployment.

3. The method for analyzing online food ordering preferences based on deep learning according to claim 2, characterized in that, The construction of the multi-model coordinated analysis framework includes: Obtain the collaborative filtering model, analyze the user-food interaction behavior through the collaborative filtering model, and discover the relationship between users and food. Obtain a content-based model and analyze the relationship between the attributes of dishes and the user's historical interactions with the content of dishes through the content-based model; Obtain the time series model and the exploration model, and analyze the user's real-time dynamic needs through the time series model and the exploration model; Obtain a context-aware model and analyze the impact of contextual information on user preferences using the context-aware model; Obtain the fusion ranking model, and generate a recommendation list by fusing the analysis results of the collaborative filtering model, content-based model, time-series model, knowledge graph model, and context-aware model.

4. The method for analyzing online food ordering preferences based on deep learning according to claim 3, characterized in that, The method of analyzing user-dish interaction behavior using a collaborative filtering model to discover the association between users and dishes includes: A collaborative filtering model is constructed using the Item-CF algorithm. Calculate the similarity between different dishes ;in, Indicates user With dishes Interaction behavior judgment function between them Indicates user With dishes Interaction behavior judgment function between them Indicates an indicator function; This represents the collection of all users in the target application; Obtain user behavior data, extract the menu items that the user has recently interacted with, and construct the user profile. Historical interactive food collection The user behavior data includes query, click, and order data. Then the user For dishes The predicted interest score is ;in, Indicates user With dishes The intensity of interaction; The analysis of the relationship between dish attributes and user's historical interaction with dish content through a content-based model includes: Obtain dish feature data and construct dish feature vectors. ;in, and Embedding codes that represent the cuisine, main ingredients, flavor, price range, cooking method, rating, and availability of a dish; users The user interest vector is obtained by averaging the feature vectors of all dishes that have been interacted with in the past. ,in, This indicates the number of dishes that have been interacted with in the past; Calculate the similarity score between the user vector and the feature vector of the candidate dishes. in, This represents the feature vector of the candidate dish.

5. The method for analyzing online food ordering preferences based on deep learning according to claim 4, characterized in that, The analysis of users' real-time dynamic needs through time-series and exploratory models includes: The popularity score of dishes within a rolling calculation time window is calculated using a time-series model, including: Define the time window; Acquire user behavior data and dish feature data within multiple consecutive time windows within the target application; Click-through rate and order volume of dishes are calculated based on user behavior data, and rating and negative review rate of dishes are calculated based on dish feature data; Calculate the popularity score of the dish. ;in, and Each represents a dish Order volume, click-through rate, and negative review rate; This represents the weight hyperparameter; Based on the menu item's listing time, an exploration model is used to randomly explore from the set of explored menu items to obtain the popularity score of the explored menu items, including: The Bandit algorithm, an exploratory algorithm, was used to explore within the menu to obtain the available time ranges. And the click-through rate is lower than the click-through rate threshold. The dishes; Calculate the popularity score of the dishes discovered. Each represents a dish Order volume, click-through rate, ratings, and negative review rate; The explored menu is based on popularity score ranges. A collection of dishes from outside the region; among them, Indicates the popularity score threshold; The recommended menu will consist of dishes within the popularity score range and dishes discovered through exploration, in order to reflect the dynamic needs of users. The analysis of the impact of contextual information on user preferences using a context-aware model includes: predicting the impact of contextual information on user preferences in a given scenario using a context-aware model. The probability of a user interacting with a dish is predicted using a binary classification model, i.e., by calculating the probability score. ; in, This represents the Sigmoid activation function. Indicates the safety bias term; and They represent the feature weights, Represents the scenario feature vector; ;in, These represent the dining time period, day of the week, delivery address type code, weather condition code, and special event, respectively.

6. The method for analyzing online food ordering preferences based on deep learning according to claim 5, characterized in that, The analysis results from the fusion ranking model, collaborative filtering model, content-based model, time-series model, knowledge graph model, and context-aware model are used to generate a recommendation list, including: Get , constitute the model feature vector ; Will As the input vector, it is fed into a pre-trained initial ranking model, which outputs a click-through rate score. The preliminary sorting model is the GBDT (Boosted Tree Data) model. Based on the dishes Perform preliminary sorting, generate a preliminary preference list, and save it in the online log; Reordering is performed using a reordering model, specifically as follows: Calculate relevant scores for the dishes in the initial preference list. ;in, Indicates dishes Click-through rate score; Indicates dishes Dishes Similarity; This indicates the selected set of recommended dishes. This represents a set of candidate dishes; Indicates the trade-off coefficient; according to The dishes in the initial preference list are rearranged to generate the final preference list, which is then saved in the online log. The reordering model is the maximum edge correlation algorithm model.

7. The method for analyzing online food ordering preferences based on deep learning according to claim 6, characterized in that, The dynamic setting of the calling frequency, calling status, and calling order of models within the multi-model coordination analysis framework includes: The multi-model coordination analysis framework is divided into three layers: offline, near-line, and online, to balance computational cost and practicality. Specifically: Collaborative filtering, content-based analysis, time-series, and fusion ranking models are assigned to the offline layer; exploration models are assigned to the near-line layer; and context-aware models are assigned to the online layer. The offline layer model is called on an hourly or daily basis; the near-line layer model is called on a minute-by-minute basis; and the online layer model is called on a millisecond-by-millisecond basis. During the analysis phase, the online layer model enters a parallel invocation state upon receiving a user request, and is processed in parallel with the offline and near-line layer models. During the sorting phase, after the collaborative filtering model, content-based analysis model, and time-series model output results, the fusion sorting model serially calls the state and sorts the dishes based on the outputs of the collaborative filtering model, content-based analysis model, and time-series model to generate a preliminary preference list and a final preference list. Based on the efficiency constraints of the target application, the calling frequency and update frequency of the fusion ranking model are dynamically adjusted, including: If the efficiency constraint of the target application is to efficiently analyze user preferences, then the call frequency of the fusion ranking model will be adjusted from daily to minute-level, and the GBDT model will be replaced by an online regression model or a neural network model to adapt to online learning; and the online training frequency of the initial ranking model will be increased to once per hour. Otherwise, maintain the original call frequency and training frequency of the fusion ranking model.

8. The method for analyzing online food ordering preferences based on deep learning according to claim 7, characterized in that, The process of generating a recommendation list by fusing the analysis results of a collaborative filtering model, a content-based model, a time-series model, a knowledge graph model, and a context-aware model through a fusion ranking model also includes: updating the GBDT training sample set during the GBDT training phase, including: Identify dishes whose rankings have been boosted by the re-ranking model. If users have interacted with these dishes, obtain their feature vectors and increase the weights of training samples containing these feature vectors. ;in, This indicates the promotion bonus coefficient. Indicates the percentage increase in ranking; Identify dishes whose ranking has been lowered by the re-ranking model. If there is no user interaction with the dish, obtain the dish's feature vector and reduce the weight of training samples containing that dish's feature vector. .

9. The method for analyzing online food ordering preferences based on deep learning according to claim 8, characterized in that, The model selection decision framework determines the models included within the multi-model coordinated analysis framework during deployment, including: Add user features to the user interest vector to create a multi-dimensional user profile. ;in This indicates the user's age, activity frequency encoding, and activity time period encoding; User-preference segmentation is performed using user segmentation algorithms, including: use Cluster analysis is performed using a pre-trained K-means clustering algorithm to classify users by value, dividing them into high-value users, medium-value users, and low-value users. Based on value classification, for each type of user, an optimal model combination is set, including: For high-value users and considering application efficiency constraints related to target applications, models within a multi-model coordination analysis framework are selected, including: If the efficiency constraint is to efficiently analyze user preferences, then the collaborative filtering model, content-based model, time series model, exploratory model, context-aware model, and fusion ranking model are selected, and the fusion ranking model is selected as the initial ranking model, either an online regression model or a neural network model. Otherwise, choose the collaborative filtering model, content-based model, time-series model, exploratory model, context-aware model, or fusion ranking model, where the initial ranking model in the fusion ranking model is the GBDT model. For mid-value users, the model, combined with the application efficiency constraints of the target application, selects models within a multi-model coordinated analysis framework, including: If the efficiency constraint is to efficiently analyze user preferences, then the collaborative filtering model, the content-based model, and the fusion ranking model are selected, and the fusion ranking model is selected as the initial ranking model, either an online regression model or a neural network model. Otherwise, choose the collaborative filtering model, the content-based model, or the fusion ranking model, where the initial ranking model in the fusion ranking model is the GBDT model; For low-value users, select models within the multi-model coordinated analysis framework, including: The selection criteria include collaborative filtering, time series, exploratory, and fusion ranking models; the initial ranking model within the fusion ranking model is the GBDT model.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement the deep learning-based user online ordering preference analysis method according to any one of claims 1-9.

Images