Method and system for customizing personalized home meal delivery matching based on user behavior

Abstract

The application relates to the technical field of user behavior analysis, and provides a user behavior-based customized personalized home meal delivery matching method and system. Current behavior features are obtained by performing time sequence feature extraction on user behavior event data according to a time window, user preference distribution data is obtained by performing preference regression prediction on the current behavior features according to a user behavior feature database, and a personalized matching list is obtained by performing fusion matching degree score calculation and sorting on a candidate merchant and a dish set according to the user preference distribution data. The application collects user behavior and extracts time sequence features, completes preference prediction in combination with historical preferences, and then performs sorting in combination with delivery, taboo and price constraints, so that the matching accuracy and practicability of home meal delivery recommendation are improved.

Description

Technical Field

[0001] This application relates to the field of user behavior analysis technology, and in particular to a method and system for matching personalized home-cooked meals and takeout based on user behavior. Background Technology

[0002] The primary application scenario for home-cooked meal delivery is to precisely connect home cooks with users who have personalized dietary needs. Users typically consider not only the taste of the dishes but also factors such as dietary restrictions, allergies, price tolerance, delivery range, and recent changes in preferences. As the home-cooked meal delivery industry develops, users expect fresh, healthy, and home-style meals that better suit their habits, while platforms aim to improve the efficiency of reaching high-quality chefs and increasing order conversion rates. Therefore, a technological solution is needed that can continuously identify user preferences based on their behavior and perform intelligent matching.

[0003] Existing food delivery matching methods mainly include manual annotation of user preferences, keyword-based search, direct recommendations based on historical order records, and fixed categorization filtering by cuisine or price. These methods share the common characteristic of relying on a single information source or static rules, often focusing only on explicit needs, and rarely processing behavioral data such as browsing history, reviews, and favorites lists. They also struggle to link real-time user behavior changes with dish and chef attributes, thus exhibiting weak matching capabilities in scenarios like home-cooked meals that emphasize personalization and timeliness. Summary of the Invention

[0004] In view of this, this application provides a method and system for matching personalized home-cooked meals based on user behavior to solve the problem of low accuracy in recommending user preferences.

[0005] The first aspect of this application provides a method for matching personalized home-cooked meals and takeout based on user behavior, the method comprising: Based on a preset time window, the acquired user behavior event data is subjected to temporal feature extraction to obtain the current behavior features; Based on a preset user behavior feature database, preference regression prediction is performed on the current behavior features to obtain user preference distribution data; Based on the user preference distribution data, a fusion matching score is calculated and sorted for the preset set of candidate merchants and dishes to obtain a personalized matching list.

[0006] In an optional implementation, the method further includes: The collected raw user behavior data is deduplicated and outliers are filtered to obtain noise-reduced behavior data. Detect missing values ​​in the noise reduction behavior data, and mark the missing values ​​according to preset labels to obtain optimized behavior data; Based on a preset event data structure, the optimized behavior data is standardized to generate user behavior event data.

[0007] In an optional implementation, the step of extracting time-series features from the acquired user behavior event data according to a preset time window to obtain current behavior features includes: Based on the session identifier in the user behavior event data, the user behavior event data is grouped to obtain session event groups; Based on a preset time window and a preset feature type, feature extraction and aggregation statistics are performed on the session event group to obtain the current aggregated features of the same time window. Based on the current time, the event sequence of the user's most recent session is filtered out from the session event group, and a preset number of event data is extracted from the event sequence to form a sequence event group; The event data in the sequence event group are subjected to feature recognition and encoding processing to obtain the feature encoding vector corresponding to each event data. The feature encoding vector is then spliced ​​in time to generate the current sequence feature.

[0008] In an optional implementation, after obtaining the current behavioral characteristics, the method further includes: Based on the user tag corresponding to the current behavior feature, the historical behavior features corresponding to the user tag in the preset user behavior feature database are compared with the current aggregated features to detect incremental feature samples in the current aggregated features that do not belong to the historical behavior features; When the presence of the incremental feature sample is detected, the incremental interaction event feature group of the user on the dish is extracted from the incremental feature sample based on the product identifier in the incremental feature sample; The corresponding interaction intensity value in the preset collaborative filtering matrix is ​​incrementally accumulated and updated according to the incremental interaction event feature group to update the collaborative filtering matrix. The weight coefficients in the preset logistic regression model are iteratively optimized in small batches based on the incremental interactive event feature group to update the logistic regression model.

[0009] In an optional implementation, the step of performing preference regression prediction on the current behavior features based on a preset user behavior feature database to obtain user preference distribution data includes: According to a preset vector format, the current behavioral features and the historical behavioral features are used to construct a behavioral feature vector corresponding to the user tag; Based on the user tags, the corresponding merchant and dish preference combination vectors are extracted from the collaborative filtering matrix to calculate the similarity between the behavioral feature vector and the preference combination vector, and to obtain the collaborative filtering score for each merchant and dish combination. The logistic regression model is used to perform a weighted summation of the feature values ​​in the behavioral feature vector to obtain the preference fusion feature value, and the preference fusion feature value is mapped to the logistic regression prediction probability through a preset Sigmoid function. The collaborative filtering score and the logistic regression prediction probability are weighted and summed according to the preset fusion weight coefficients to obtain user preference distribution data.

[0010] In an optional implementation, the step of calculating and sorting the fusion matching score of the preset candidate merchant and dish set based on the user preference distribution data to obtain a personalized matching list includes: Based on the address information in the user behavior event data and the preset delivery distance threshold, a list of candidate merchants is selected from the preset set of candidate merchants and dishes, and a first list of candidate dishes corresponding to each candidate merchant is extracted from the set of candidate merchants and dishes based on the list of candidate merchants. Based on the dietary restrictions attribute in the user behavior event data, the first candidate dish list is filtered to obtain the second candidate dish list; The user preference distribution data is fused with the attribute features of each candidate dish in the second candidate dish list to obtain a basic matching score for each candidate dish. Based on the price range constraints in the user behavior event data and the current sequence features, the corresponding adjustment weight coefficients are obtained from the preset adjustment weight coefficient set, and the matching degree base score is adjusted according to the adjustment weight coefficients to obtain the final matching degree score of each candidate dish. Based on the final matching score, all candidate dishes in the second candidate dish list are sorted in descending order to obtain a personalized matching list for each candidate merchant.

[0011] A second aspect of this application provides a personalized home-cooked meal delivery matching device based on user behavior, the device comprising: The feature extraction module is used to extract time-series features from the acquired user behavior event data according to a preset time window to obtain the current behavior features; The preference analysis module is used to perform preference regression prediction on the current behavior features based on a preset user behavior feature database to obtain user preference distribution data; The matching and sorting module is used to calculate and sort the matching degree scores of the preset candidate merchants and dishes based on the user preference distribution data to obtain a personalized matching list.

[0012] A third aspect of this application provides an electronic device, the electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the personalized home-cooked meal delivery matching method based on user behavior as described above.

[0013] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the personalized home-cooked meal delivery matching method based on user behavior as described above.

[0014] In summary, this application includes at least the following beneficial technical effects: 1. By jointly modeling based on users' recent behavior, historical preferences, and conversation sequences, food delivery recommendations can better match users' current tastes, ordering habits, and real-time needs, thereby significantly improving the degree of personalized matching.

[0015] 2. By fusing collaborative filtering and logistic regression for prediction, and continuously updating the model and matrix with incremental samples, the system can both utilize group similarity and quickly absorb new behavioral information, thereby improving the accuracy, timeliness and stability of recommendations.

[0016] 3. Before sorting, constraints such as delivery distance, dietary restrictions, and price range are introduced, and candidate merchants and dishes are screened and re-rated. Therefore, not only preferred content can be recommended, but also feasible content can be recommended, reducing the probability of invalid and unsuitable recommendations. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for matching personalized home-cooked meals based on user behavior, as provided in an embodiment of this application. Figure 2 This is a functional module diagram of a personalized home-cooked meal delivery matching device based on user behavior, provided in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application; Figure 4 This is a schematic diagram of the operating environment of the personalized home-cooked meal delivery matching method based on user behavior provided in the embodiments of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] like Figure 4 The diagram illustrates the workflow of the personalized home-cooked meal delivery matching method based on user behavior provided in this embodiment. The operating environment consists of a client and a service platform, which establish a bidirectional data interaction connection via a communication network to jointly support the full-process implementation of the personalized home-cooked meal delivery matching method based on user behavior provided in this application. The client, as a front-end device facing end users, is primarily responsible for collecting raw user behavior data generated in the home-cooked meal delivery scenario and reporting this data to the service platform. It also receives the personalized matching list returned by the service platform and displays recommended merchants and dishes to end users. The service platform, as the core back-end processing and computing device of this method, is the execution carrier of the entire matching method logic. It can connect to and call a preset user behavior feature database and a candidate merchant and dish collection library. Its core function is to receive the raw user behavior data reported by the client, sequentially perform data preprocessing and feature extraction, user preference analysis and prediction, and merchant and dish matching and ranking, ultimately generating a personalized matching list and returning it to the client.

[0021] like Figure 1 The diagram shown is a flowchart of a method for matching personalized home-cooked meals based on user behavior, provided in an embodiment of this application. The method for matching personalized home-cooked meals based on user behavior, provided in an embodiment of this application, includes the following steps.

[0022] Step S1: Extract time-series features from the acquired user behavior event data according to the preset time window to obtain the current behavior features.

[0023] In the operating environment of the home-cooked meal delivery matching method, the client, as the front-end device facing the end user, is responsible for collecting raw user behavior data generated in the home-cooked meal delivery scenario in real time. This data includes explicit information entered by the user through the client interface and implicit behavior records automatically generated by the user during browsing, selection, and evaluation. The client reports the collected raw user behavior data to the service platform via the communication network. The service platform, as the core back-end processing and computing device of the method, receives the raw user behavior data and first performs data preprocessing operations. The service platform performs deduplication processing on the raw user behavior data, that is, it performs uniqueness verification based on the combination of four dimensions: event timestamp, user identifier, event type, and event content. When multiple records have the same value in the above four dimensions, only the record with the earliest timestamp is retained, and the other duplicate records are removed, thereby eliminating the interference of redundant data caused by network retransmission, client duplicate reporting, etc., on subsequent feature extraction. Subsequently, the service platform filters outliers from the deduplicated behavioral data and verifies the legality of each record according to preset data quality rules. These rules include that the event timestamp should not be later than the current system time, the order amount should not be negative or exceed the preset maximum reasonable amount threshold, and the page dwell time should not exceed the preset maximum reasonable duration threshold. Records that do not meet the above rules are judged as abnormal data and removed, thereby ensuring that the data entering the feature extraction stage all come from real and valid user interaction behavior.

[0024] After deduplication and outlier filtering, the service platform obtains denoised behavior data. This data retains the original, dispersed structure from when it was first collected, with some fields containing missing values. The service platform performs missing value detection on the data, iterating through each field of each record to identify fields with empty values ​​or values ​​marked with a preset empty value identifier. For detected missing values, the platform marks them according to the type of missing field using preset identifiers. For dietary restrictions not entered by the user, they are marked as "Unknown Restriction"; for unfilled content in the review text field, they are marked as "Empty Review"; and for missing content in the brand name field, they are marked as "Empty Brand". This marking process ensures that all fields have a definite value, avoiding subsequent data processing anomalies caused by missing values, while retaining the "user did not provide this information" status, providing distinguishable category information for subsequent modeling. The service platform outputs the data processed with missing value marking as optimized behavior data.

[0025] The service platform then standardizes the optimized behavior data according to a pre-defined event data structure. This pre-defined event data structure is a fixed-format data container containing six fields: user identifier, session identifier, event timestamp, event type, event content body, and event content attributes. The service platform extracts corresponding information from each record of the optimized behavior data and populates these six fields. The user identifier is directly mapped from the user identity field of the original record; the session identifier is generated based on the time interval of consecutive user visits; the event timestamp is converted from the time field of the original record to a unified format; the event type field is mapped to a pre-defined enumerated value based on the behavior action in the original record; the event content body field records the dish identifier or chef identifier; and the event content attributes field records supplementary information such as review text and order amount. After standardization, the service platform obtains user behavior event data. Each record in this data is a standardized six-tuple event, and all records together constitute an event-oriented, structurally unified set of user behavior event data. User behavior event data is presented in a unified structured format, representing every interaction between users and the home-cooked meal delivery platform. This provides a standardized data foundation for extracting meaningful temporal features from massive amounts of interaction records.

[0026] After obtaining user behavior event data, the service platform begins performing time-series feature extraction. The goal of this operation is to extract current behavioral features from the user behavior event data. These current behavioral features consist of two parts: current aggregated features and current sequence features. First, the service platform groups all events belonging to the same session identifier based on the session identifier field in the user behavior event data, obtaining session event groups. Each session event group corresponds to a complete user access process, with its start time being the timestamp of the earliest event in the session and its end time being the timestamp of the latest event in the session. This grouping process aggregates scattered events according to the natural boundaries of user access, enabling meaningful statistical analysis of features at the session level.

[0027] The service platform extracts and aggregates features from session event groups based on preset time windows and preset feature types to obtain the current aggregated features within the same time window. In this embodiment, the preset time window is the past 30 calendar days. The service platform counts back 30 days from the current system time, filters out all session event groups whose event timestamps fall within this time window, and aggregates and statistically analyzes user behavior within this time window. The preset feature types include spending power features, consumption frequency features, taste stability features, interaction depth features, and interaction diversity features. Spending power features are obtained by calculating the ratio of the total amount of all user orders within the time window to the total number of user orders, reflecting the user's average spending level. Consumption frequency features are obtained by calculating the ratio of the total number of user orders to the number of days from the user's first order to the current time, reflecting the user's ordering activity level. Taste stability features are obtained by identifying the category of dishes ordered most frequently by the user within the time window and calculating the ratio of the number of orders for this category to the total number of user orders; a higher ratio indicates a more concentrated taste preference. Interaction depth is obtained by calculating the ratio of the total duration of all user sessions within the time window to the total number of user sessions, reflecting the average dwell time per visit. Interaction diversity is obtained by counting the number of chefs and dishes that the user interacted with within the time window, reflecting the breadth of the user's exploration of different chefs and dishes. The service platform concatenates the calculated feature values ​​into a one-dimensional vector according to a preset feature order, which serves as the current aggregated feature. This current aggregated feature characterizes the user's stable behavioral patterns over a longer period, providing a quantitative representation of the user's long-term profile.

[0028] While acquiring the current aggregated features, the service platform, based on the current time point, filters the event sequence of the user's most recent session from the session event group, and extracts a preset number of event data from this event sequence to form a sequence event group. The current time point is the moment when the service platform performs feature extraction. The service platform locates the last session initiated by the user before the current time point, sorts all events in that session in ascending order of event timestamps, and forms a complete event sequence. In this embodiment, the preset number of events is set to 5 based on experience. The service platform extracts the last 5 events from the sorted event sequence. If the total number of events in the session is less than 5, all events are extracted, and the insufficient parts are padded with zero vectors in subsequent encoding processing. The above extraction operation focuses on the user's latest behavioral trajectory in the current session, capturing a series of micro-behaviors before the user makes a final decision. These behaviors often contain the most direct signals of purchase intent. The service platform arranges the 5 extracted events in chronological order to form a sequence event group.

[0029] The service platform performs feature identification and encoding on the event data in the sequence event group to obtain the feature encoding vector corresponding to each event data. For each event in the sequence event group, the service platform first identifies the event type field of the event and maps the event type to a preset numerical code, where browsing events are mapped to a value of zero, favorite events to a value of one, adding to cart events to a value of two, and placing an order events to a value of three. This mapping makes different event types numerically comparable and sequential. The service platform then calculates the time interval between the event and the previous event. For the first event in the sequence, the time interval is set to zero. This time interval, in seconds, reflects the urgency of the user's decision-making process. The smaller the time interval, the faster and more continuous the user's operation, which usually corresponds to a higher purchase intention. The service platform further extracts three values ​​from the content subject field of the event: the dish price, the historical average rating of the corresponding chef, and the global historical sales volume of the dish, as the content features of the event. The service platform arranges the event type encoding value, time interval value, price value, chef rating value, and dish sales volume value in a fixed order to form a five-dimensional vector, which serves as the feature encoding vector of the event. For events with fewer than 5 events in a sequence, the service platform generates a five-dimensional vector of all zeros to fill the corresponding positions, ensuring that each event has a corresponding vector representation in the sequence.

[0030] The service platform concatenates the feature encoding vectors of all events in the sequence event group according to the chronological order of the events to generate the current sequence feature. Specifically, the service platform concatenates the feature encoding vectors of the first, second, third, fourth, and fifth events sequentially to form a 5×5 vector, which is the current sequence feature. This current sequence feature fully preserves the type, timing, and multi-dimensional information of the user's five most recent behaviors, as well as the price of the products involved, chef's reputation, and dish popularity. Its core function is to characterize the user's real-time behavioral patterns in the current session in a fixed-length vector form, providing input data to capture the user's current purchase intent.

[0031] The service platform outputs the current aggregated features and current sequence features generated by the above operations as the current behavioral features. These current behavioral features include both the user's stable behavioral patterns over a long period and the user's immediate behavioral patterns in recent sessions. These two patterns complement each other in preference prediction and matching scoring, jointly supporting the accurate characterization of users' personalized needs. Step S2: Based on the preset user behavior feature database, perform preference regression prediction on the current behavior features to obtain user preference distribution data.

[0032] In the operating environment of the home-cooked meal delivery matching method, the service platform has completed the time-series feature extraction operation in step S1, obtaining the current behavior features composed of the current aggregated features and the current sequence features. The service platform then executes the preference regression prediction operation in step S2. Using historical data stored in the preset user behavior feature database, combined with the current behavior features, and through a dynamically updated collaborative filtering matrix and logistic regression model, it calculates user preference distribution data that reflects the user's preference for various dishes and chefs.

[0033] The service platform first extracts historical behavioral features corresponding to the user tag corresponding to the current behavioral feature from a pre-set user behavioral feature database. This database is a persistent storage unit that the service platform can access. It stores historical behavioral feature data generated after each round of matching operations, using the user tag as the primary key. This data includes historical aggregated features, historical sequence features, and feature version identifiers used by the user in training the preference prediction model historically. The service platform then compares the extracted historical behavioral features with the current aggregated features in the current behavioral features, field by field, to detect whether there are incremental feature samples in the current aggregated features that are not included in the historical behavioral features. Specifically, the service platform compares the value of each feature field in the current aggregated features with the corresponding field value in the historical behavioral features. If a feature field has a value in the current aggregated features, but its value in the historical behavioral features is empty or a preset initial default value, then the feature data corresponding to that field is determined to be an incremental feature sample. For example, when a user adds items to their cart for the first time within the last 30 days, the "number of times added items" feature field in the current aggregated features changes from zero to one value in the historical behavior features. The service platform identifies this change as an incremental feature sample. By identifying new changes in user behavior through the above operations, these changes may be a direct reflection of upgrades in user tastes, adjustments to dietary restrictions, or shifts in consumption habits. Only by incorporating these changes into the model can we ensure that the preference prediction results are consistent with the user's current actual needs.

[0034] When the service platform detects an incremental feature sample, it immediately triggers the incremental update process of the preference prediction model. Based on the product identifier in the incremental feature sample, the service platform extracts incremental interaction event feature groups of users for dishes. Each incremental interaction event feature group is a quintuple set containing user identifier, dish identifier, interaction type, interaction time, and interaction intensity increment. The interaction type is determined based on the event type field in the user behavior event data, and the interaction intensity increment is calculated using preset weight coefficients: browsing behavior has a weight coefficient of one, favorite behavior has a weight coefficient of three, purchase behavior has a weight coefficient of ten, and rating behavior has a weight coefficient of two. This extraction operation aggregates the behavioral information scattered throughout the incremental feature samples into structured interaction event groups, providing directly usable data units for updating the collaborative filtering matrix.

[0035] The service platform incrementally updates the corresponding interaction strength values ​​in the preset collaborative filtering matrix based on incremental interaction event feature groups to update the collaborative filtering matrix. The collaborative filtering matrix is ​​a preset two-dimensional matrix data structure, where row indices represent user identifiers and column indices represent dish identifiers. Each element in the matrix stores the interaction strength value between the corresponding user and the corresponding dish, which is initially zero and increases as user interactions with the dish accumulate. Specifically, for each incremental interaction event in the incremental interaction event feature group, the service platform locates the corresponding row in the collaborative filtering matrix based on the user identifier in the event, locates the corresponding column based on the dish identifier, and then adds the interaction strength increment corresponding to the event to the current value of the matrix element to obtain the updated matrix element value. For example, when a user completes a purchase of a dish, the interaction strength increment corresponding to this purchase is 10. The service platform increases the element value of the corresponding position of that dish in the collaborative filtering matrix by 10, thereby increasing the association strength between the user and the dish. The aforementioned incremental cumulative update operation enables the collaborative filtering matrix to reflect the latest changes in the interaction between users and dishes in real time, ensuring that the matrix data used in similarity calculation is always synchronized with the user's current behavior state.

[0036] Meanwhile, the service platform performs small-batch iterative optimization of the weight coefficients in the preset logistic regression model based on the incremental interaction event feature groups to update the logistic regression model. The logistic regression model is a preset machine learning classification model whose core parameters are a weight coefficient vector corresponding one-to-one with the behavioral feature vector dimension and a bias term coefficient. Specifically, the service platform uses the current behavioral features corresponding to the incremental interaction event feature groups as training samples, and the interaction types in the incremental interaction event feature groups as labels. It then uses the gradient descent algorithm to iteratively update the weight coefficients of the logistic regression model in one or more rounds. In each iteration, the service platform calculates the prediction error of the current model on that training sample, and then fine-tunes the weight coefficients according to the gradient of the error with respect to the weight coefficients, based on a preset learning rate, so that the prediction error of the model on that sample gradually decreases. This small-batch iterative optimization operation allows the logistic regression model to learn new preference patterns from the latest user behavior data without needing to retrain the model completely each time. This ensures that the model's response speed is maintained while allowing the model parameters to continuously approximate the user's actual behavioral patterns.

[0037] After completing the incremental updates of the collaborative filtering matrix and logistic regression model, the service platform begins calculating the preference distribution data. First, according to a preset vector format, the service platform constructs behavioral feature vectors corresponding to user tags from the current and historical behavioral features. The preset vector format specifies the dimensional order of the feature vectors: the first few dimensions are aggregated feature fields, and the latter few dimensions are sequence feature fields. The service platform then performs a weighted average of the corresponding field values ​​in the current and historical aggregated features, and concatenates the current and historical sequence features according to time sequence to form a unified behavioral feature vector. This vector integrates information from the user's long-term historical behavior and the latest behavior, providing a unified input representation for collaborative filtering and logistic regression calculations.

[0038] The service platform extracts corresponding merchant and dish preference combination vectors from the updated collaborative filtering matrix based on user tags. It then calculates the similarity between the behavioral feature vector and the preference combination vector to obtain a collaborative filtering score for each merchant and dish combination. The preference combination vector is a column vector corresponding to the dish column in the collaborative filtering matrix. Each element of this vector represents the interaction strength between different users and that dish, reflecting the dish's popularity among all users and the strength of its association with a specific user. The service platform uses a cosine similarity algorithm to calculate the similarity between the behavioral feature vector and each preference combination vector. The cosine similarity formula is the dot product of the two vectors divided by the product of their magnitudes. The result is between zero and one; a higher value indicates a higher degree of matching between the dish and the user's current behavioral features. The service platform uses the calculated similarity value as the collaborative filtering score, which reflects the user's potential preferences for dishes mined from historical user interaction data.

[0039] The service platform uses an updated logistic regression model to perform a weighted summation of the feature values ​​in the behavioral feature vector to obtain the preference fusion feature value. This fusion feature value is then mapped to a logistic regression prediction probability using a pre-defined Sigmoid function. The logistic regression model multiplies each feature value in the behavioral feature vector by its corresponding weight coefficient, sums the results, and adds a bias term to obtain the preference fusion feature value. This fusion feature value is a real number and can take either positive or negative values. The service platform then substitutes this preference fusion feature value into the Sigmoid function, whose expression is S. , where x is the preference fusion feature value and e is the natural constant. The Sigmoid function maps any real number of preference fusion feature values ​​to a probability interval between zero and one. The mapped value is the logistic regression prediction probability, reflecting the likelihood of a user purchasing a dish based on user behavior characteristics.

[0040] The service platform obtains user preference distribution data by weighting and summing the collaborative filtering score and the logistic regression prediction probability according to preset fusion weight coefficients. The preset fusion weight coefficients are two values ​​between zero and one, with a sum of 1. These coefficients are determined using a grid search method based on historical validation data. The service platform multiplies the collaborative filtering score by the first fusion weight coefficient and the logistic regression prediction probability by the second fusion weight coefficient, then adds the two products to obtain the final user preference distribution data. This distribution data is a mapping structure indexed by dish identifiers and represented by fusion scores. Each dish corresponds to a fusion score between zero and one; a higher score indicates a higher degree of match between the dish and the user's current preference. This weighted summation operation integrates the advantages of collaborative filtering in uncovering latent user preferences with the advantages of logistic regression in accurately predicting using explicit behavioral features. This ensures that the final user preference distribution data reflects both the implicit preference patterns in users' historical interactions and captures the immediate purchase intent reflected in users' current behavioral characteristics.

[0041] Step S3: Based on the user preference distribution data, calculate and sort the fusion matching score of the preset candidate merchant and dish set to obtain a personalized matching list.

[0042] The service platform first filters a list of candidate merchants from a preset set of candidate merchants and dishes based on the address information in the user behavior event data and a preset delivery distance threshold. The address information in the user behavior event data comes from the user's frequently used delivery address collected by the client in step S1, and this address is stored in the user behavior event data in the form of latitude and longitude coordinates. In this embodiment, the delivery distance threshold is set to 1500 meters based on the delivery capacity in the home-cooked meal delivery scenario. The delivery distance threshold represents the maximum straight-line distance that the chef (it should be understood that in the application scenario of this application, the chef is the merchant, and the chef can be used to refer to the merchant below) is willing to undertake delivery service. The service platform reads the latitude and longitude coordinates of the registered addresses of all chefs from the candidate merchant and dish set library, and calculates the straight-line distance between each chef's address and the user's address one by one. The calculation formula adopts the semi-sine formula, which calculates the spherical distance through the trigonometric function relationship between the Earth's radius and the difference in latitude and longitude between two points. The service platform compares the calculated distance value with the preset delivery distance threshold, and only retains those chefs whose distance value is less than or equal to the delivery distance threshold, and stores the identifiers of these chefs in the candidate merchant list. The above filtering process ensures that the matching results are feasible at the physical delivery level, avoiding the recommendation of merchants that cannot be delivered to users. For example, if a user's delivery address is located at the east gate of a residential community, but a chef's delivery range only covers an area 500m beyond the west gate of the community, the chef's delivery distance threshold cannot cover the user's address, and their dishes will not be included in the subsequent matching process.

[0043] The service platform filters the first candidate dish list based on dietary restrictions attributes in user behavior event data to obtain a second candidate dish list. The first candidate dish list is derived from a summary of dishes offered by each chef in the candidate merchant list. The service platform extracts attributes such as dish identifier, dish name, ingredient composition, and flavor tag for each candidate chef from the database of candidate merchants and dish sets to form the first candidate dish list. The dietary restrictions attributes in the user behavior event data come from information actively entered by users in the client's basic information input module. This information is stored in the form of a tag list, such as "does not eat pork," "seafood allergy," and "avoids spicy food." For each candidate dish in the first candidate dish list, the service platform parses its ingredient composition field and performs a string match between the ingredient names in the ingredient composition field and the restrictions in the dietary restrictions attribute list. If a match is found, the candidate dish is removed from the first candidate dish list. This filtering operation respects users' dietary restrictions and health needs, ensuring that the recommended results do not include dishes that users cannot accept. This is a fundamental guarantee of safety and acceptability in personalized matching. For example, if a user enters "does not eat pork" as a taboo in the basic information entry module, and a candidate dish is "pork and cabbage stew with vermicelli", then that dish will be directly filtered out and will not appear in the subsequent matching and scoring process.

[0044] The service platform integrates user preference distribution data with the attribute features of each candidate dish in the second candidate dish list to obtain a basic matching score for each candidate dish. The user preference distribution data is a mapping structure output from step S2, indexed by dish identifiers and with integration scores as values. The integration score for each dish ranges from zero to one, reflecting the degree of matching between the dish and the user's overall preferences. The attribute features of each candidate dish in the second candidate dish list include its corresponding flavor style attribute, price attribute, and historical sales attribute. These attributes are stored numerically in the database of candidate merchants and dish sets. The flavor style attribute is mapped to a preset numerical code based on the cuisine to which the dish belongs; for example, Sichuan cuisine is mapped to 0.9, Cantonese cuisine to 0.7, and light cuisine to 0.5. For each candidate dish in the second list of candidate dishes, the service platform performs a weighted sum of its user preference distribution data, including the fusion score, flavor style attribute value, price attribute value, and historical sales attribute value. The weighting coefficients for this summation are set based on historical validation data: fusion score weight 0.6, flavor style attribute weight 0.2, price attribute weight 0.1, and historical sales attribute weight 0.1. This fusion calculation combines abstract preference scores with specific dish attributes, ensuring that the basic matching score reflects both the user's personalized preferences and the dish's market performance and user acceptance.

[0045] The service platform, based on the price range constraints and current sequence features in the user behavior event data, obtains the corresponding adjustment weight coefficients from a preset set of adjustment weight coefficients, and adjusts the basic matching score according to the adjustment weight coefficients to obtain the final matching score for each candidate dish. The price range constraint originates from the current aggregated features extracted from the user behavior event data in step S1. The spending power feature in this current aggregated feature reflects the average spending level of the user's historical orders. The service platform sets a price range constraint with a fluctuation of 30% above and below this average spending level as the center. The current sequence features originate from the last 5 events captured from the user's most recent session in step S1. The event type encoding sequence in this current sequence feature records the type order of the user's last 5 behaviors, such as the sequence pattern of "browse, browse, favorite, browse, add to cart". The preset set of adjustment weight coefficients is a mapping table indexed by event type encoding sequences and set as adjustment weight coefficients. This mapping table is constructed based on historical behavior data analysis. For example, if a user's last 5 behaviors include adding items to their cart, the adjustment weight coefficient is set to 0.3; if the user's last 5 behaviors include adding items to favorites but not adding items to cart, the adjustment weight coefficient is set to 0.15; and if the user's last 5 behaviors only include browsing, the adjustment weight coefficient is set to 0.05. The service platform uses the event type encoding sequences in the current sequence features as indexes to retrieve the corresponding adjustment weight coefficients from the set of adjustment weight coefficients. Then, it adds these adjustment weight coefficients to the basic matching score to obtain the final matching score. The above score adjustment operation responds to the user's immediate behavioral patterns in the current session. When a user shows a strong purchase intention (such as adding items to cart) in recent behaviors, the matching score of related dishes is increased, enabling the recommendation results to capture the user's immediate intent. For example, if a user browses five Sichuan dishes in the current session, and the last three events include adding items to their cart, the event type encoding sequence in the current sequence features matches a high adjustment weight coefficient of 0.3. The service platform adds this adjustment weight coefficient to the basic matching score of the Sichuan dishes, giving them a higher priority in the final ranking, even if the user preference distribution data scores for these Sichuan dishes are lower than those for other cuisines.

[0046] The service platform sorts all candidate dishes in the second candidate dish list in descending order based on the final matching score, obtaining a personalized matching list for each candidate merchant. The service platform uses a preset sorting algorithm, with the final matching score as the sorting key, to sort the candidate dishes in the second candidate dish list in descending order, placing the dish with the highest score at the top of the list. After sorting, the service platform extracts the top K candidate dishes from the sorted list according to the client's requirement for the number of dishes to be displayed. In this embodiment, K is set to 20, meaning a maximum of 20 dishes are displayed to the user. The service platform groups the extracted top K candidate dishes by merchant, and the dishes corresponding to each merchant are arranged from highest to lowest according to their final matching score, forming a hierarchical data structure with merchants as the main body and dishes as branches. This data structure is the personalized matching list. The service platform transmits this personalized matching list back to the client via a communication network. After receiving the list, the client renders it as the recommended merchants and dishes display area in the user interface, completing the entire process of home-cooked meal delivery matching. The sorting and extraction operations described above provide users with the most relevant and suitable limited options to avoid information overload, while ensuring that the chef's high-quality dishes can get more priority exposure through high matching scores, thereby improving the overall order conversion rate.

[0047] This application applies to the field of user behavior analysis technology. It obtains current behavioral features by extracting temporal features from user behavior event data based on time windows. Then, it performs preference regression prediction on these current behavioral features using a user behavior feature database to obtain user preference distribution data. Finally, it calculates and ranks candidate merchants and menus based on this user preference distribution data to obtain a personalized matching list. This application collects user behavior and extracts temporal features, combines them with historical preferences to complete preference prediction, and then integrates delivery, taboo, and price constraints for ranking, thus improving the matching accuracy and practicality of home-cooked meal delivery recommendations.

[0048] like Figure 2 The diagram shown is a functional block diagram of a personalized home-cooked meal delivery matching device based on user behavior, provided in an embodiment of this application.

[0049] In some embodiments, the personalized home-cooked meal delivery matching device 2 based on user behavior may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the personalized home-cooked meal delivery matching device 2 based on user behavior may be stored in the server's memory and executed by at least one processor to perform (see details). Figure 1 (Description) A function that customizes a personalized home-cooked meal delivery matching method based on user behavior.

[0050] In this embodiment, the personalized home-cooked meal delivery matching device 2 based on user behavior can be divided into multiple functional modules according to its functions. These functional modules may include: a feature extraction module 21, a preference analysis module 22, a matching and ranking module 23, a data standard module 24, and an incremental update module 25. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be detailed in subsequent embodiments.

[0051] Feature extraction module 21 is used to extract time-series features from the acquired user behavior event data according to a preset time window to obtain the current behavior features; Preference analysis module 22 is used to perform preference regression prediction on the current behavior features based on a preset user behavior feature database to obtain user preference distribution data; The matching and sorting module 23 is used to calculate and sort the fusion matching score of the preset candidate merchants and dishes based on the user preference distribution data to obtain a personalized matching list.

[0052] In an optional implementation, the personalized home-cooked meal delivery matching device 2 based on user behavior further includes a data standard module 24, which is used for: The collected raw user behavior data is deduplicated and outliers are filtered to obtain noise-reduced behavior data. Detect missing values ​​in the noise reduction behavior data, and mark the missing values ​​according to preset labels to obtain optimized behavior data; Based on a preset event data structure, the optimized behavior data is standardized to generate user behavior event data.

[0053] In an optional implementation, the feature extraction module 21 is used to: Based on the session identifier in the user behavior event data, the user behavior event data is grouped to obtain session event groups; Based on a preset time window and a preset feature type, feature extraction and aggregation statistics are performed on the session event group to obtain the current aggregated features of the same time window. Based on the current time, the event sequence of the user's most recent session is filtered out from the session event group, and a preset number of event data is extracted from the event sequence to form a sequence event group; The event data in the sequence event group are subjected to feature recognition and encoding processing to obtain the feature encoding vector corresponding to each event data. The feature encoding vector is then spliced ​​in time to generate the current sequence feature.

[0054] In an optional implementation, the personalized home-cooked meal delivery matching device 2 based on user behavior further includes an incremental update module 25, which is used for: Based on the user tag corresponding to the current behavior feature, the historical behavior features corresponding to the user tag in the preset user behavior feature database are compared with the current aggregated features to detect incremental feature samples in the current aggregated features that do not belong to the historical behavior features; When the presence of the incremental feature sample is detected, the incremental interaction event feature group of the user on the dish is extracted from the incremental feature sample based on the product identifier in the incremental feature sample; The corresponding interaction intensity value in the preset collaborative filtering matrix is ​​incrementally accumulated and updated according to the incremental interaction event feature group to update the collaborative filtering matrix. The weight coefficients in the preset logistic regression model are iteratively optimized in small batches based on the incremental interactive event feature group to update the logistic regression model.

[0055] In an optional implementation, the preference analysis module 22 is used for: According to a preset vector format, the current behavioral features and the historical behavioral features are used to construct a behavioral feature vector corresponding to the user tag; Based on the user tags, the corresponding merchant and dish preference combination vectors are extracted from the collaborative filtering matrix to calculate the similarity between the behavioral feature vector and the preference combination vector, and to obtain the collaborative filtering score for each merchant and dish combination. The logistic regression model is used to perform a weighted summation of the feature values ​​in the behavioral feature vector to obtain the preference fusion feature value, and the preference fusion feature value is mapped to the logistic regression prediction probability through a preset Sigmoid function. The collaborative filtering score and the logistic regression prediction probability are weighted and summed according to the preset fusion weight coefficients to obtain user preference distribution data.

[0056] In an optional implementation, the matching and sorting module 23 is used for: Based on the address information in the user behavior event data and the preset delivery distance threshold, a list of candidate merchants is selected from the preset set of candidate merchants and dishes, and a first list of candidate dishes corresponding to each candidate merchant is extracted from the set of candidate merchants and dishes based on the list of candidate merchants. Based on the dietary restrictions attribute in the user behavior event data, the first candidate dish list is filtered to obtain the second candidate dish list; The user preference distribution data is fused with the attribute features of each candidate dish in the second candidate dish list to obtain a basic matching score for each candidate dish. Based on the price range constraints in the user behavior event data and the current sequence features, the corresponding adjustment weight coefficients are obtained from the preset adjustment weight coefficient set, and the matching degree base score is adjusted according to the adjustment weight coefficients to obtain the final matching degree score of each candidate dish. Based on the final matching score, all candidate dishes in the second candidate dish list are sorted in descending order to obtain a personalized matching list for each candidate merchant.

[0057] It should be understood that the various variations and specific embodiments of the methods provided in the above embodiments are also applicable to the user behavior-based personalized home-cooked meal delivery matching device of this embodiment. Through the foregoing detailed description of the user behavior-based personalized home-cooked meal delivery matching method, those skilled in the art can clearly understand the implementation method of the user behavior-based personalized home-cooked meal delivery matching device of this embodiment. For the sake of brevity, it will not be described in detail here.

[0058] like Figure 3 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this application.

[0059] In a preferred embodiment of the present invention, the electronic device 3 may include, but is not limited to, a memory 31, at least one processor 32, and at least one communication bus 33.

[0060] Those skilled in the art should understand that Figure 3 The structure of the electronic device 3 shown does not constitute a limitation of the embodiments of the present invention. The electronic device 3 may also include more or fewer other hardware or software than shown, or different component arrangements.

[0061] In some embodiments, the electronic device 3 is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits, programmable gate arrays, digital processors, and embedded devices.

[0062] It should be noted that the electronic device 3 is merely an example. Other existing or future electronic products that are suitable for this application should also be included within the scope of protection of this application and are incorporated herein by reference.

[0063] In some embodiments, the memory 31 stores a computer program that, when executed by the at least one processor 32, implements all or part of the steps in the user behavior-based personalized home-cooked meal delivery matching method described above. The memory 31 includes read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data. Further, the computer-readable storage medium may primarily include a program storage area and a data storage area, wherein the program storage area may store an operating system, at least one application program required for a function, etc.

[0064] In some embodiments, the at least one processor 32 is the control unit of the electronic device 3, connecting various components of the electronic device 3 via various interfaces and lines. It executes programs or modules stored in the memory 31 and calls data stored in the memory 31 to perform various functions and process data. For example, when the at least one processor 32 executes the computer program stored in the memory 31, it implements all or part of the steps of the personalized home-cooked meal delivery matching method based on user behavior described in this embodiment; or it implements all or part of the functions of the personalized home-cooked meal delivery matching device based on user behavior. The at least one processor 32 may be composed of integrated circuits, such as a single-packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips.

[0065] In some embodiments, the at least one communication bus 33 is configured to enable communication between the memory 31 and the at least one processor 32, etc. Although not shown, the electronic device 3 may also include a power supply (e.g., a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 32 via a power management device, thereby enabling functions such as charging, discharging, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 3 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0066] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause an electronic device (which may be a personal computer, electronic device, or network device, etc.) or processor to execute portions of the methods described in the various embodiments of this application.

[0067] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0068] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0069] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for matching personalized home-cooked meals to customers based on user behavior, characterized in that, The method includes: Based on a preset time window, the acquired user behavior event data is subjected to temporal feature extraction to obtain the current behavior features; Based on a preset user behavior feature database, preference regression prediction is performed on the current behavior features to obtain user preference distribution data; Based on the user preference distribution data, a fusion matching score is calculated and sorted for the preset set of candidate merchants and dishes to obtain a personalized matching list.

2. The method for matching personalized home-cooked meals based on user behavior according to claim 1, characterized in that, The method further includes: The collected raw user behavior data is deduplicated and outliers are filtered to obtain noise-reduced behavior data. Detect missing values ​​in the noise reduction behavior data, and mark the missing values ​​according to preset labels to obtain optimized behavior data; Based on a preset event data structure, the optimized behavior data is standardized to generate user behavior event data.

3. The method for matching personalized home-cooked meals based on user behavior as described in claim 2, characterized in that, The current behavioral features include current aggregate features and current sequence features. The step of extracting time-series features from the acquired user behavior event data according to a preset time window to obtain the current behavioral features includes: Based on the session identifier in the user behavior event data, the user behavior event data is grouped to obtain session event groups; Based on a preset time window and a preset feature type, feature extraction and aggregation statistics are performed on the session event group to obtain the current aggregated features of the same time window; Based on the current time, the event sequence of the user's most recent session is filtered out from the session event group, and a preset number of event data is extracted from the event sequence to form a sequence event group; The event data in the sequence event group are subjected to feature recognition and encoding processing to obtain the feature encoding vector corresponding to each event data. The feature encoding vector is then spliced ​​in time to generate the current sequence feature.

4. The method for matching personalized home-cooked meals based on user behavior according to claim 3, characterized in that, After obtaining the current behavioral characteristics, the method further includes: Based on the user tag corresponding to the current behavior feature, the historical behavior features corresponding to the user tag in the preset user behavior feature database are compared with the current aggregated features to detect incremental feature samples in the current aggregated features that do not belong to the historical behavior features; When the presence of the incremental feature sample is detected, the incremental interaction event feature group of the user on the dish is extracted from the incremental feature sample based on the product identifier in the incremental feature sample; The corresponding interaction intensity value in the preset collaborative filtering matrix is ​​incrementally accumulated and updated according to the incremental interaction event feature group to update the collaborative filtering matrix. The weight coefficients in the preset logistic regression model are iteratively optimized in small batches based on the incremental interactive event feature group to update the logistic regression model.

5. The method for matching personalized home-cooked meals based on user behavior according to claim 4, characterized in that, The step of performing preference regression prediction on the current behavior features based on a preset user behavior feature database to obtain user preference distribution data includes: According to a preset vector format, the current behavioral features and the historical behavioral features are used to construct a behavioral feature vector corresponding to the user tag; Based on the user tags, the corresponding merchant and dish preference combination vectors are extracted from the collaborative filtering matrix to calculate the similarity between the behavioral feature vector and the preference combination vector, and to obtain the collaborative filtering score for each merchant and dish combination. The logistic regression model is used to perform a weighted summation of the feature values ​​in the behavioral feature vector to obtain the preference fusion feature value, and the preference fusion feature value is mapped to the logistic regression prediction probability through a preset Sigmoid function. The collaborative filtering score and the logistic regression prediction probability are weighted and summed according to the preset fusion weight coefficients to obtain user preference distribution data.

6. The method for matching personalized home-cooked meals based on user behavior according to claim 3, characterized in that, The step of calculating and sorting the preset set of candidate merchants and dishes based on the user preference distribution data to obtain a personalized matching list includes: Based on the address information in the user behavior event data and the preset delivery distance threshold, a list of candidate merchants is selected from the preset set of candidate merchants and dishes, and a first list of candidate dishes corresponding to each candidate merchant is extracted from the set of candidate merchants and dishes based on the list of candidate merchants. Based on the dietary restrictions attribute in the user behavior event data, the first candidate dish list is filtered to obtain the second candidate dish list; The user preference distribution data is fused with the attribute features of each candidate dish in the second candidate dish list to obtain a basic matching score for each candidate dish. Based on the price range constraints in the user behavior event data and the current sequence features, the corresponding adjustment weight coefficients are obtained from the preset adjustment weight coefficient set, and the matching degree base score is adjusted according to the adjustment weight coefficients to obtain the final matching degree score of each candidate dish. Based on the final matching score, all candidate dishes in the second candidate dish list are sorted in descending order to obtain a personalized matching list for each candidate merchant.

7. A personalized home-cooked meal delivery matching device based on user behavior, applied to the personalized home-cooked meal delivery matching method based on user behavior as described in claim 1, characterized in that, The device includes: The feature extraction module is used to extract time-series features from the acquired user behavior event data according to a preset time window to obtain the current behavior features; The preference analysis module is used to perform preference regression prediction on the current behavior features based on a preset user behavior feature database to obtain user preference distribution data; The matching and sorting module is used to calculate and sort the matching degree scores of the preset candidate merchants and dishes based on the user preference distribution data to obtain a personalized matching list.

8. An electronic device, characterized in that, The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the personalized home-cooked meal delivery matching method based on user behavior as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the personalized home-cooked meal delivery matching method based on user behavior as described in any one of claims 1 to 6.

Images