User cold start recommendation method based on meta learning

Abstract

The invention belongs to the technical field of recommendation systems, and discloses a user cold start recommendation method based on meta-learning, and the recommendation technology comprises the detailed steps: step 1, user behavior log collection: collecting behavior data of a user on each platform, including behaviors of preference, resource browsing, clicking, searching and sharing; step 2, user feature preprocessing: key user features are selected based on feature distribution and significance indexes; and step 3, constructing a double-tower multi-target recall model: constructing the double-tower multi-target recall model based on the characteristics of the users and the articles. Through deep learning of consumption interest distribution of old users and correlation of articles, a more detailed and accurate new user portrait can be constructed, so that interests and preferences of the users can be matched more accurately; on the basis of accurate construction of the new user portrait, the recommendation system can formulate a personalized recommendation strategy, a one-by-one recommendation mode is avoided, and recommendation content better meeting user requirements is provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation systems, and in particular relates to a user cold start recommendation method based on meta-learning. Background Art

[0002] With the rapid development of the Internet and e-commerce, the problem of information overload has become more and more serious. The cost of users filtering information to meet their own needs has continued to rise. Intelligent recommendation technology has emerged, aiming to help users quickly discover valuable content by connecting massive information with user interests, and achieve a win-win situation for information consumers and producers. The intelligent recommendation process is usually divided into two stages: recall and sorting. In the recall stage, the system selects a relatively small candidate set from the huge resource library based on the user's interests and historical behaviors. After entering the sorting stage, the system calculates the click probability for each resource in the candidate set, so as to accurately select the content that the user is most interested in. Modern recommendation systems are no longer satisfied with traditional collaborative filtering and content filtering methods, but instead use generative models of deep learning technology to simulate users' consumption interests and provide users with more personalized recommendations. However, the cold start problem of new users is still a major challenge facing current recommendation systems. Due to the lack of historical behavior data of new users, it is difficult for recommendation models to accurately capture their consumption interests and preferences. In order to solve the cold start problem of new users, although the existing recommendation methods have alleviated it to a certain extent, they are limited by the accuracy of item features and the limitations of user personalization mining, and the effect is not ideal. Although the recommendation method based on collaborative filtering recommends by analyzing user behavior or item associations, it also faces problems such as data sparsity, subjectivity in similarity calculation, and insufficient new user behavior data, resulting in the accuracy and personalization of recommendation results need to be improved. Therefore, for intelligent recommendation systems, how to better solve the cold start problem of new users and improve the accuracy and personalization of recommendations is a hot topic and difficulty in current research, which requires continuous improvement and optimization. Summary of the invention

[0003] The purpose of the present invention is to provide a user cold start recommendation method based on meta-learning to solve the problems raised in the above background technology.

[0004] In order to achieve the above object, the present invention provides the following technical solution: a user cold start recommendation method based on meta-learning, wherein the detailed steps of the recommendation technology include:

[0005] Step 1: Collect user behavior logs: Collect user behavior data on various platforms, including preferences, resource browsing, clicks, searches, and sharing behaviors;

[0006] Step 2: User feature preprocessing: Select key user features based on feature distribution and significance indicators;

[0007] Step 3: Building a dual-tower multi-target recall model: Building a dual-tower multi-target recall model based on the characteristics of users and items;

[0008] Step 4: Meta-learning user cold start sub-network: Introduce the concept of meta-learning and build a cold start sub-network;

[0009] Step 5: Gated network function: Add a gated network to distinguish between new and old users;

[0010] Step 6: Data synchronization and caching: synchronize the trained user embedding and item embedding to the system cache;

[0011] Step 7: Online user incremental feature update: collect new user behavior data in real time and merge it into historical behavior data;

[0012] Step 8, online user feature preprocessing: preprocess the user features updated in real time to keep consistent with the offline module;

[0013] Step 9: Inferring online new user behavior profiles: based on real-time updated user features;

[0014] Step 10: Recall results based on the overall model: Combine the user's current behavior profile and basic profile to obtain recommended resources from the resource library.

[0015] Preferably, the specific steps of collecting user behavior logs in step 1 include:

[0016] A1, Preference collection: Track the different models, brands and price ranges browsed by users, analyze users' specific preferences, and record users' stay time, scroll depth and interactive behavior on the browsing page;

[0017] A2, resource browsing collection: monitor the resource types of articles, pictures, and videos browsed by users, and record the browsing time and browsing order;

[0018] A3, click behavior collection: Track the user's click behavior on the page, including the links, buttons, and image elements clicked;

[0019] A4, Search behavior collection: monitor the keywords entered by users in the search box, record the search frequency and clicks on search results;

[0020] A5, Sharing behavior collection: Tracking the content shared by users, sharing channels (including social media, email) and sharing objects;

[0021] A6, context feature collection: real-time monitoring of user network status;

[0022] A7, User feature collection: Collect user's geographical information through user registration information or IP address positioning;

[0023] A8, distinguish between new and old users: distinguish between new and old users based on user registration time and behavior data;

[0024] A9. Collect characteristics of new and old users separately: For new users, focus on collecting initial data such as registration information, first search and browsing behavior to build an initial user portrait. For old users, in addition to continuously collecting behavioral data, you also need to pay attention to the long-term characteristics of behavioral changes, interest shifts, and loyalty changes.

[0025] Preferably, the specific steps of user feature preprocessing in step 2 include:

[0026] B1, Feature selection: Analyze the distribution of each feature, including its value range and frequency distribution, to determine the contribution of the feature to the user profile and evaluate the significance of each feature in distinguishing user groups or predicting user behavior;

[0027] B2, feature compression coding: using simhash technology and hashing to generate fixed-length feature codes, compressing the selected user features and mapping the original high-dimensional feature space to a low-dimensional coding space;

[0028] B3, bucket-able data processing: For some features, including age group and region, which have clear value ranges or categories, bucket-based processing can be performed;

[0029] B4, improve processing efficiency: through feature compression coding and bucket-dividable data processing, the amount of data that needs to be processed in subsequent calculations can be greatly reduced, thereby improving processing efficiency;

[0030] B5, Note: When performing feature compression encoding, maintain the representativeness and discrimination of the original data to avoid excessive compression leading to information loss.

[0031] Preferably, the specific steps of constructing the dual-tower multi-target recall model in step three include:

[0032] C1, model construction:

[0033] C1-1, feature selection: including user features and item features;

[0034] C1-2, feature encoding: Use deep neural networks to encode user and item features to generate user vectors and item vectors;

[0035] C1-3, dual-tower structure: construct two independent neural network structures, one for processing user features (user tower) and the other for processing item features (item tower);

[0036] C1-4, similarity calculation: After obtaining the user vector and item vector, the similarity is calculated using the cosine similarity and dot product methods;

[0037] C2, loss function design:

[0038] C2-1, multi-objective loss: design multiple loss functions according to specific recommendation goals and business needs, such as click-through rate prediction loss and purchase conversion rate prediction loss;

[0039] C2-2, optimization algorithm: select a suitable optimization algorithm to minimize the total loss function;

[0040] C3, model training

[0041] C3-1, data preparation: preprocess and extract features from the collected user behavior logs to generate a data set for training;

[0042] C3-2, model initialization: Initialize the parameters of the twin-tower model, including the weights and biases of the neural network;

[0043] C3-3, iterative training: iteratively train the model using the training dataset, and update the model parameters by calculating the loss function and back-propagating the gradient;

[0044] C3-4, Model evaluation: Use the test set to evaluate the trained model, calculate various performance indicators to evaluate the performance of the model, and tune and improve the model based on the evaluation results.

[0045] Preferably, the specific operations of meta-learning the user cold start sub-network in step 4 include:

[0046] D1, build a cold start sub-network: design a sub-network specifically for dealing with the cold start problem, which learns to assist in the recommendation or prediction of new users;

[0047] D2, set up the memory matrix: the memory matrix is ​​used to store the basic portraits of old users;

[0048] D3, store and learn old user portraits: Basic portraits: collect and store basic information of old users, use machine learning algorithms to learn preferences and interests from old users' historical behaviors, and store the learned information in a memory matrix;

[0049] D4, adopts the scaling and transfer method; the scaling method: according to the similarity between the new user and the old user, the old user portrait is scaled or weighted; the transfer method: the scaled old user portrait information is transferred to the new user as the initial portrait or prediction basis of the new user.

[0050] D5, ensure that new users benefit from old user portraits: Ensure that new users can benefit from old user portraits through verification and testing. Based on the verification and testing results, continuously optimize and improve the meta-learning cold start sub-network to improve its performance and adaptability.

[0051] Preferably, the specific steps of data synchronization and caching in step six include:

[0052] E1, Data synchronization: Select a cache system that provides high-performance key-value storage and fast data access capabilities; export the trained user embedding and item embedding data from the training environment, and import the exported data into the cache system;

[0053] E2, design cache update strategy: for applications that require high real-time performance, update the cache immediately when user and item information changes. For applications that do not require high real-time performance, set a fixed time interval to update the cache regularly: when only some user and item data changes, use incremental update;

[0054] E3, ensure data real-time and accuracy: implement a monitoring and alarm mechanism for the cache system to promptly detect and handle problems that arise during data synchronization and updating, set a reasonable expiration time (TTL) for the cache, and verify the data during data synchronization and updating to ensure its accuracy and completeness. If problems arise during data synchronization and updating that lead to data inconsistency or corruption, a rollback mechanism needs to be implemented to restore the data to its previous state.

[0055] Preferably, the specific steps of updating the online user incremental features in step seven include:

[0056] F1, real-time collection of new user behavior data:

[0057] Data source: Determine the source of the data source to collect user behavior data in real time, including records of user clicks, browsing, purchases, comments, and sharing behaviors;

[0058] Data format: Ensure that the collected data is in a uniform format and is easy to process;

[0059] Real-time: Use real-time data stream processing technology to ensure real-time data collection and processing;

[0060] F2, merge the new data into the historical behavior data:

[0061] Data cleaning: Clean the collected new data to remove duplicate, invalid and abnormal data;

[0062] Data merging: Merge the cleaned new data into the user's historical behavior data;

[0063] Data persistence: store the merged data persistently;

[0064] F3, update user characteristics:

[0065] Feature extraction: Extract features that reflect user interests and preferences based on the merged user historical behavior data;

[0066] Feature update: replace or add the extracted new features or updated feature values ​​to the user's feature vector;

[0067] F4, ensure the real-time and accuracy of model recommendations:

[0068] Online learning: Use online learning algorithms to update model parameters in real time to reflect changes in user behavior data;

[0069] Model evaluation: Regularly evaluate the accuracy of model recommendations, use offline and online evaluation indicators, and adjust model parameters or optimize algorithms based on the evaluation results;

[0070] Feedback loop: Collect user feedback as new behavior data and merge it into historical behavior data to form a closed feedback loop.

[0071] Preferably, the specific steps of online user feature preprocessing in step eight include:

[0072] G1, feature selection and alignment:

[0073] Feature consistency: Ensure that the features used online are exactly the same as those used in offline training;

[0074] Feature missing processing: For feature missing problems that occur in online environments, formulate processing strategies in advance;

[0075] G2, feature scaling and normalization:

[0076] Normalization: For numerical features, normalization techniques are used to scale them to the same scale;

[0077] Standardization: transform the features into a distribution with a mean of 0 and a standard deviation of 1 through zero-mean normalization;

[0078] G3, discrete feature processing:

[0079] One-hot Encoding: For categorical features, one-hot encoding is used to convert them into binary vectors.

[0080] Feature embedding: For high-dimensional or sparse discrete features, feature embedding technology can be used to convert them into low-dimensional dense vectors;

[0081] G4, Feature Update and Synchronization:

[0082] Real-time update: When the user's features change, the user's feature vector is updated in real time;

[0083] Synchronization mechanism: ensures synchronization between online feature database and offline feature database;

[0084] G5, Monitoring and Evaluation:

[0085] Feature quality monitoring: Regularly evaluate the quality of online features;

[0086] Model performance evaluation: Monitor the performance of the online model to ensure that when using the preprocessed features for inference, the model's performance is comparable to or better than that during offline training.

[0087] Preferably, the specific steps of inferring the online new user behavior portrait in step nine include:

[0088] H1, inference of new user behavior profile:

[0089] H1-1, real-time feature collection: collect new users’ behavior data in real time, including clicks, browsing, and searches;

[0090] H1-2, feature preprocessing: preprocess the collected new user behavior data, including data cleaning, formatting, and normalization steps to ensure that the processed features are consistent with the feature format and range used in offline training;

[0091] H1-3, Application of meta-learning sub-network: Using the trained meta-learning sub-network, the characteristics of new users are inferred. The meta-learning sub-network can quickly adapt to and predict users' potential preferences and behavior patterns based on limited user behavior data;

[0092] H1-4, New user portrait construction: Based on the output of the meta-learning sub-network and combined with the basic information of the new user, a new user behavior portrait is constructed. The portrait will serve as an important reference for new users in future behavior analysis and recommendation systems;

[0093] H2, update of old user portraits:

[0094] H2-1, historical data collection: For old users, collect their historical behavior data;

[0095] H2-2, feature extraction and updating: extract the features of old users from historical data, including static features and dynamic features;

[0096] H2-3, Application of meta-learning sub-network: Similar to the inference of new user portraits, the meta-learning sub-network is used to infer the characteristics of old users;

[0097] H2-4, portrait update: merge and update the output of the meta-learning sub-network with the original portrait of the old user.

[0098] Preferably, the specific steps of recalling the results based on the overall model in step ten include:

[0099] S1, integrated user portrait:

[0100] Current behavior profile: Based on the user's recent behavior data, the user's current behavior profile is obtained through online feature preprocessing and meta-learning sub-network inference;

[0101] Basic profile: stable characteristics and preferences accumulated by users over a long period of time;

[0102] Integration: Integrate the current behavior profile and the basic profile to form a comprehensive and multi-dimensional understanding of the user;

[0103] S2, obtain recommended resources from the resource library:

[0104] Resource library: contains a collection of all recommended resources, such as products, articles, and videos;

[0105] Resource characteristics: Each resource has a set of data describing its characteristics, such as the category, price, and sales volume of the product; the topic and keywords of the article;

[0106] Matching: Based on the integrated user profile and resource features, an algorithm is used to obtain recommended resources that match the user's interests from the resource library.

[0107] S3, design recommendation strategy:

[0108] Diversified recommendations: Introduce resources of different types, styles, and fields into the recommendation results;

[0109] Relevance recommendation: Based on the user's historical behavior and interests, highly relevant resources are recommended;

[0110] Novelty recommendation: recommend new resources that users have not yet come into contact with but may be of interest;

[0111] Personalized recommendation: Provide customized recommendation services based on users’ unique preferences and needs;

[0112] S4, sorting and filtering recall results:

[0113] Sorting: Use a suitable sorting algorithm to sort the recall results.

[0114] Screening: Screen the recall results according to certain rules or standards to remove low-quality or unsuitable resources for recommendation.

[0115] S5, Monitoring and Optimization:

[0116] Real-time monitoring: Real-time monitoring of the quality of recommendation results and user feedback.

[0117] A / B testing: Compare the effects of different recommendation strategies through A / B testing and select the best strategy for deployment.

[0118] Continuous optimization: Based on user feedback and data analysis results, we continuously optimize recommendation algorithms and strategies to improve the quality of recommendation results and user experience.

[0119] The beneficial effects of the present invention are as follows:

[0120] 1. The present invention can construct a more detailed and accurate new user portrait by deeply studying the consumption interest distribution of old users and the correlation of items, thereby more accurately matching the user's interests and preferences; based on the accurate construction of the new user portrait, the recommendation system can formulate personalized recommendation strategies, avoid a one-size-fits-all recommendation method, and provide recommendation content that better meets user needs.

[0121] 2. By optimizing the initialization distribution of new users, the present invention can ensure that new users are not affected by existing users or popular items, thereby reducing the risk of cold start problem transmission. It can also reduce the bias generated by the system based on historical data, so that new users can obtain more fair and objective recommendation results.

[0122] 3. The present invention can improve user experience and engagement. New users can immediately obtain recommended content that meets their interests and needs without going through a long "trial and error" process. Accurate and personalized recommendations can enhance users' reliance on and trust in the recommendation system, thereby improving user engagement and loyalty. By recommending content that users are truly interested in, it can increase user interaction with recommended content, thereby improving conversion rates.

[0123] 4. The present invention is not only applicable to platforms such as Autohome through a meta-learning-based method, but can also be expanded to other fields, such as e-commerce, news, video, etc.; the system can adaptively capture the dynamic changes of users and items, such as changes in user interests, the introduction of new items, etc., and adjust the recommendation strategy accordingly; for new users or new items, the system can use meta-learning methods to effectively deal with data sparsity problems and ensure the accuracy and effectiveness of recommendations; the system can continuously learn new user data and item data, and continuously optimize the recommendation model to adapt to the ever-changing environment and user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0124] Figure 1 It is a flow chart of the technical solution of the present invention. DETAILED DESCRIPTION

[0125] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0126] like Figure 1 As shown, the embodiment of the present invention provides a user cold start recommendation method based on meta-learning, and the detailed steps of the recommendation technology include:

[0127] Step 1: Collect user behavior logs: Collect user behavior data on various platforms, including preferences, resource browsing, clicks, searches, and sharing behaviors. At the same time, collect context features (including network status and phone model) and user features (including region and age group), distinguish between new and old users, and collect their features separately.

[0128] Step 2: User feature preprocessing: Based on feature distribution and significance indicators, select key user features, and use simhash and other technologies to compress and encode the features to accelerate subsequent similarity calculations. For bucketable data, perform bucket processing based on compression coding to improve processing efficiency.

[0129] Step 3: Construction of dual-tower multi-target recall model: Based on the characteristics of users and items, a dual-tower multi-target recall model is constructed, a reasonable loss function and optimization algorithm are designed, and model training is performed;

[0130] Step 4: Meta-learning user cold start sub-network: Introduce the concept of meta-learning, build a cold start sub-network, set up a memory matrix to store and learn the basic profiles and learning profiles of old users, and use scaling and transfer methods to ensure that new users can benefit from the profiles of old users;

[0131] Step 5: Gated network function: Add a gating network to distinguish new and old users. For new users, reduce the learning rate and mainly rely on the model prediction results; for old users, increase the learning rate and participate in the parameter update of the meta-learning sub-network;

[0132] Step 6: Data synchronization and caching: synchronize the trained user embedding and item embedding to the system cache, and design a reasonable cache update strategy to ensure data real-time and accuracy;

[0133] Step 7: Online user incremental feature update: collect new user behavior data in real time, merge it into historical behavior data, update user features, and ensure the real-time and accuracy of model recommendations;

[0134] Step 8: Online user feature preprocessing: Preprocess the user features updated in real time to keep them consistent with the offline module, ensuring that the quality of the features used in online reasoning is consistent with that used in offline training.

[0135] Step 9: Inferring online new user behavior profiles: Based on real-time updated user features, use the meta-learning sub-network and the gating network to infer new user behavior profiles. For old users, use historical data and the meta-learning sub-network to update the profiles.

[0136] Step 10, recall results based on the overall model: Combine the user's current behavior profile and basic profile, obtain recommendable resources from the resource library, design reasonable recommendation strategies, such as diversified recommendations, relevance recommendations, etc., to enrich the recall results, sort and filter the recall results, and ensure the quality of the recommendation results and user experience.

[0137] The specific steps of collecting user behavior logs in step 1 include:

[0138] A1, Preference collection: Track the different models, brands and price ranges browsed by users, analyze users' specific preferences, record users' stay time, scroll depth and interactive behaviors (including likes and comments) on the browsing page, so as to gain a deeper understanding of users' attention to car series;

[0139] A2, resource browsing collection: monitor the types of resources of articles, pictures, and videos browsed by users, record browsing time and browsing order, analyze user preferences for different types of resources, and the frequency and pattern of user browsing resources;

[0140] A3, click behavior collection: Track the user's click behavior on the page, including the links, buttons, and image elements clicked, analyze the user's click path and click depth to understand the user's interest points and interaction habits on the page content;

[0141] A4, Search behavior collection: monitor the keywords entered by users in the search box, record the search frequency and clicks on search results, and analyze users' search intentions and search habits in order to provide more accurate search recommendations and result optimization;

[0142] A5, Sharing behavior collection: Track the content shared by users, sharing channels (including social media, email) and sharing objects, analyze users' sharing intentions and the characteristics of shared content, so as to understand users' recognition of specific content and willingness to spread it;

[0143] A6, context feature collection: real-time monitoring of the user's network status, recording network speed, stability and connection type, collecting information such as the user's mobile phone model, operating system version and device performance, in order to understand the compatibility and usage of the user's device;

[0144] A7, User feature collection: Collect user's geographical information (including country, province, city, etc.) through user registration information or IP address positioning, analyze user's behavior data and historical records, and infer user's age group, gender and other demographic characteristics;

[0145] A8, distinguish new and old users: distinguish new and old users based on user registration time and behavior data.

[0146] For new users, focus on their first visit behavior, registration process, and initial interests. For old users, analyze their long-term interests and preference changes based on their historical behavior data.

[0147] A9. Collect characteristics of new and old users separately: For new users, focus on collecting initial data such as registration information, first search and browsing behavior to build an initial user portrait. For old users, in addition to continuously collecting behavioral data, you also need to pay attention to the long-term characteristics of behavioral changes, interest shifts, and loyalty changes.

[0148] The specific steps of user feature preprocessing in step 2 include:

[0149] B1, Feature selection: Analyze the distribution of each feature, including its value range and frequency distribution, to determine the contribution of the feature to the user profile, evaluate the significance of each feature in distinguishing user groups or predicting user behavior, select features that have a significant impact on the results, and select key user features based on the results of feature distribution and significance analysis. The features will be used in subsequent user profile construction and recommendation algorithms;

[0150] B2, feature compression coding: using simhash technology and hashing to generate fixed-length feature codes, compressing the selected user features, and mapping the original high-dimensional feature space to a low-dimensional coding space to reduce storage space requirements and computational complexity;

[0151] B3, bucket-able data processing: For certain features, including age group and region, which have clear value ranges or categories, bucket processing can be performed. Based on the compression coding results, users are bucketed according to the value ranges or categories of the features, and similar users are divided into the same bucket for subsequent batch processing and analysis;

[0152] B4, improve processing efficiency: through feature compression coding and bucket-divided data processing, the amount of data that needs to be processed in subsequent calculations can be greatly reduced, thereby improving processing efficiency. Since similar users are divided into the same bucket, batch calculations can be performed within the bucket, further reducing the amount of calculations and improving processing speed;

[0153] B5. Notes: When performing feature compression encoding, it is necessary to maintain the representativeness and discrimination of the original data to avoid information loss due to excessive compression. When performing bucketing, it is necessary to select an appropriate bucketing strategy based on actual conditions to ensure that users in the bucket have similar features and behavior patterns.

[0154] Among them, the specific steps of building the dual-tower multi-target recall model in step three include:

[0155] C1, model construction:

[0156] C1-1, feature selection: including user features and item features; user features include basic user information (such as region, age group), behavior logs (such as browsing, clicking, searching, sharing, etc.), context features (such as network status, mobile phone model), etc.; item features: including item metadata (such as model, brand, price, etc.), user-item interaction history (such as clicking, purchasing, etc.), etc.;

[0157] C1-2, feature encoding: Use deep neural networks (such as DNN, DCN, etc.) to encode the features of users and items to generate user vectors and item vectors. In the feature encoding process, various technologies (such as embedding, attention mechanism, etc.) can be used to improve the accuracy and effectiveness of feature representation;

[0158] C1-3, dual-tower structure: build two independent neural network structures, one for processing user features (user tower) and the other for processing item features (item tower). The structures of the two towers can be the same or different, depending on the actual situation and performance requirements;

[0159] C1-4, similarity calculation: After obtaining the user vector and item vector, the cosine similarity and dot product methods are used to calculate the similarity. The higher the similarity value, the greater the user's interest in the item, which can be used as a basis for recall;

[0160] C2, loss function design:

[0161] C2-1, multi-objective loss: design multiple loss functions according to specific recommendation goals and business requirements, such as click-through rate prediction loss and purchase conversion rate prediction loss. These loss functions can be combined into a total loss function using a weighted sum method to optimize multiple goals simultaneously during training.

[0162] C2-2, Optimization algorithm: Select a suitable optimization algorithm (such as Adam, SGD, etc.) to minimize the total loss function. During the training process, you can use optimization methods such as gradient descent to update the model parameters;

[0163] C3, model training

[0164] C3-1, Data preparation: Preprocess and extract features from the collected user behavior logs to generate a training data set. The data set should contain positive samples (i.e., items that users are interested in) and negative samples (i.e., items that users are not interested in) to train the model's ability to distinguish between positive and negative samples.

[0165] C3-2, model initialization: Initialize the parameters of the dual-tower model, including the weights and biases of the neural network. You can use random initialization or pre-trained models to initialize the model.

[0166] C3-3, iterative training: Use the training data set to iteratively train the model, and update the model parameters by calculating the loss function and back-propagating the gradient. During the training process, the validation set can be used to monitor the performance of the model so that the hyperparameters can be adjusted in time or the training can be stopped;

[0167] C3-4, model evaluation: Use the test set to evaluate the trained model, calculate various performance indicators (such as accuracy, recall, F1 value, etc.) to evaluate the performance of the model, and tune and improve the model based on the evaluation results.

[0168] The specific operations of meta-learning the user cold start sub-network in step 4 include:

[0169] D1, build a cold start sub-network: design a sub-network specifically for handling the cold start problem. This sub-network can be an independent model or a component of the main model. The sub-network learns to assist in the recommendation or prediction of new users.

[0170] D2, set up the memory matrix: The memory matrix is ​​used to store the basic profiles of old users (such as demographic information, historical behavior, etc.) and learning profiles (such as user preferences learned by the model). Various data structures can be used to implement the memory matrix, such as matrices, graphs, key-value pairs, etc., to ensure that the data structure supports efficient query and update operations;

[0171] D3, store and learn old user portraits: Basic portrait: collect and store basic information of old users, such as age, gender, geographic location, etc. Learning portrait: use machine learning algorithms to learn their preferences and interests from their historical behaviors, and store the learned information in the memory matrix;

[0172] D4, adopts scaling and transfer methods; scaling method: scaling or weighting the old user portrait according to the similarity between the new user and the old user, which can be achieved by calculating the feature similarity or behavior similarity between the new user and the old user; transfer method: transferring the scaled old user portrait information to the new user as the initial portrait or prediction basis of the new user, which can be achieved by direct copying, weighted averaging or other advanced methods.

[0173] D5, ensure that new users benefit from the old user portraits: Ensure that new users can benefit from the old user portraits through verification and testing. This can be achieved by comparing indicators such as new user satisfaction and conversion rate before and after using the meta-learning cold start sub-network. Based on the verification and testing results, the meta-learning cold start sub-network is continuously optimized and improved to improve its performance and adaptability.

[0174] The specific steps of data synchronization and caching in step six include:

[0175] E1, Data synchronization: Select a cache system: First, you need to select a suitable cache system, such as Redis, Memcached, etc. These systems usually provide high-performance key-value storage and fast data access capabilities; export the trained user embedding and item embedding data from the training environment, which usually involves storing the model weights or embedding vectors in a file or database in an appropriate format (such as binary, JSON, etc.), and importing the exported data into the cache system by writing custom scripts or using existing tools to ensure the integrity and accuracy of the data during the import process;

[0176] E2, design cache update strategy: For applications that require high real-time performance, update the cache immediately when user and item information changes. For applications that do not require high real-time performance, set a fixed time interval to regularly update the cache: When only some user and item data changes, use incremental updates to update only the changed data to reduce the pressure on the cache system; set version numbers for cached data to facilitate version comparison and conflict resolution during the update process, which can ensure consistency and accuracy when synchronizing data between multiple services or processes;

[0177] E3, ensure data real-time and accuracy: implement a monitoring and alarm mechanism for the cache system to promptly detect and handle problems that occur during data synchronization and updating, set a reasonable expiration time (TTL) for the cache to ensure that the data in the cache does not become obsolete when the data changes, and manage the use of cache space through appropriate expiration strategies (such as LRU, LFU, etc.) During data synchronization and updating, verify the data to ensure its accuracy and integrity. This can be achieved by calculating hash values, comparing data sizes, etc. If problems occur during data synchronization and updating that lead to data inconsistency or damage, a rollback mechanism needs to be implemented to restore the data to its previous state.

[0178] Among them, the specific steps of updating the incremental features of online users in step seven include:

[0179] F1, real-time collection of new user behavior data:

[0180] Data source: Determine the source of the data source to collect user behavior data in real time, including records of user clicks, browsing, purchases, comments, and sharing behaviors;

[0181] Data format: Ensure that the collected data is in a uniform format and easy to process, usually including fields such as user ID, behavior type, behavior time, and behavior object (such as item ID);

[0182] Real-time: Use real-time data stream processing technology (such as Kafka, Flink, etc.) to ensure real-time data collection and processing;

[0183] F2, merge the new data into the historical behavior data:

[0184] Data cleaning: Clean the collected new data to remove duplicate, invalid and abnormal data;

[0185] Data merging: Merge the cleaned new data into the user's historical behavior data. This can be achieved through database query, update operations, or using a dedicated data processing framework (such as Spark Streaming).

[0186] Data persistence: The merged data is persistently stored for subsequent feature extraction and model training;

[0187] F3, update user characteristics:

[0188] Feature extraction: Based on the merged user historical behavior data, extract features that can reflect user interests and preferences. These features may include user activity, user preferred item categories, user behavior time distribution, etc.

[0189] Feature update: Replace or add the extracted new features or updated feature values ​​to the user's feature vector. This can be achieved by updating the user portrait database or online feature storage system.

[0190] F4, ensure the real-time and accuracy of model recommendations:

[0191] Online learning: Use online learning algorithms (such as incremental learning, streaming learning, etc.) to update model parameters in real time to reflect changes in user behavior data, thus ensuring the real-time nature of model recommendations;

[0192] Model evaluation: Regularly evaluate the accuracy of model recommendations, using offline evaluation indicators (such as accuracy, recall, F1 value, etc.) and online evaluation indicators (such as click-through rate, conversion rate, etc.), and adjust model parameters or optimize algorithms based on the evaluation results;

[0193] Feedback loop: Collect user feedback (such as ratings, comments, etc.) as new behavior data and merge it into historical behavior data to form a closed feedback loop, so that the accuracy of model recommendations can be continuously optimized through user feedback.

[0194] Among them, the specific steps of online user feature preprocessing in step eight include:

[0195] G1, feature selection and alignment:

[0196] Feature consistency: Ensure that the features used online are exactly the same as those used in offline training, including the number, type, name, and order of features;

[0197] Missing feature processing: For missing features in online environments, formulate processing strategies in advance, such as using default values, median filling, or model-based methods (such as predictive filling);

[0198] G2, feature scaling and normalization:

[0199] Normalization: For numerical features, normalization techniques are used to scale them to the same scale, which helps the model better handle features of different dimensions;

[0200] Standardization: The features are transformed into a distribution with a mean of 0 and a standard deviation of 1 through zero-mean normalization. Standardization is particularly effective when dealing with features with different scales and distributions;

[0201] G3, discrete feature processing:

[0202] One-hot Encoding: For categorical features (such as gender, region, etc.), one-hot encoding is used to convert them into binary vectors to ensure that the online encoding method is consistent with the offline encoding method.

[0203] Feature embedding: For high-dimensional or sparse discrete features, feature embedding technology can be used to convert them into low-dimensional dense vectors. The online embedding matrix should be consistent with the embedding matrix used in offline training.

[0204] G4, Feature Update and Synchronization:

[0205] Real-time update: When the user's characteristics change (such as new purchase behavior, ratings, etc.), the user's feature vector is updated in real time;

[0206] Synchronization mechanism: ensure the synchronization between the online feature library and the offline feature library, which can be achieved through regular data synchronization tasks or using a real-time data stream processing framework;

[0207] G5, Monitoring and Evaluation:

[0208] Feature quality monitoring: Regularly evaluate the quality of online features, including feature completeness, accuracy, consistency, and effectiveness;

[0209] Model performance evaluation: Monitor the performance of the online model to ensure that when using the preprocessed features for inference, the model's performance is comparable to or better than that during offline training.

[0210] Among them, the specific steps of inferring the online new user behavior portrait in step nine include:

[0211] H1, inference of new user behavior profile:

[0212] H1-1, real-time feature collection: collect new users’ behavior data in real time, including clicks, browsing, and searches, and use website statistics tools (such as Baidu Statistics) or customized data collection systems to capture these behavior data;

[0213] H1-2, feature preprocessing: preprocess the collected new user behavior data, including data cleaning, formatting, and normalization steps to ensure that the processed features are consistent with the feature format and range used in offline training;

[0214] H1-3, Application of meta-learning sub-network: Using the trained meta-learning sub-network, the characteristics of new users are inferred. The meta-learning sub-network can quickly adapt to and predict users' potential preferences and behavior patterns based on limited user behavior data;

[0215] H1-4, new user portrait construction: Based on the output of the meta-learning sub-network and combined with the basic information of the new user (such as the information filled in during registration), a new user behavior portrait is constructed. The portrait will serve as an important reference for new users in future behavior analysis and recommendation systems.

[0216] H2, update of old user portraits:

[0217] H2-1, historical data collection: For old users, collect their historical behavior data, including past clicks, purchases, ratings, etc.

[0218] H2-2, feature extraction and updating: extract the features of old users from historical data, including static features and dynamic features. Static features include basic information of users, and dynamic features include the latest behavior patterns of users.

[0219] H2-3, Application of meta-learning sub-network: Similar to the inference of new user portraits, the meta-learning sub-network is used to infer the characteristics of old users;

[0220] H2-4, portrait update: The output of the meta-learning sub-network is integrated and updated with the original portrait of the old user. When the behavior of the old user changes significantly, the label and weight of his portrait can be adjusted accordingly.

[0221] Among them, the specific steps of recall results based on the overall model in step ten include:

[0222] S1, integrated user portrait:

[0223] Current behavior profile: Based on the user's recent behavior data, the user's current behavior profile is obtained through online feature preprocessing and meta-learning sub-network inference;

[0224] Basic profile: The stable characteristics and preferences accumulated by users over a long period of time, usually based on the user's historical data and registration information;

[0225] Integration: Integrate the current behavior profile and the basic profile to form a comprehensive and multi-dimensional understanding of the user;

[0226] S2, obtain recommended resources from the resource library:

[0227] Resource library: contains a collection of all recommended resources, such as products, articles, and videos;

[0228] Resource characteristics: Each resource has a set of data describing its characteristics, such as the category, price, and sales volume of the product; the topic and keywords of the article;

[0229] Matching: Based on the integrated user profile and resource features, use appropriate algorithms (such as content-based recommendation, collaborative filtering, etc.) to obtain recommended resources that match the user's interests from the resource library.

[0230] S3, design recommendation strategy:

[0231] Diversified recommendations: Introduce resources of different types, styles, and fields into the recommendation results to meet the diverse needs of users;

[0232] Relevance recommendation: Based on the user's historical behavior and interests, highly relevant resources are recommended to improve the accuracy of recommendations and user satisfaction;

[0233] Novelty recommendation: recommend new resources that users have not yet come into contact with but may be interested in, so as to stimulate users' desire to explore;

[0234] Personalized recommendation: Provide customized recommendation services based on users’ unique preferences and needs;

[0235] S4, sorting and filtering recall results:

[0236] Sorting: Use a suitable sorting algorithm (such as a learning-to-rank algorithm) to sort the recall results to ensure that the resources that best meet the user's interests and needs are ranked first.

[0237] Screening: Screen the recall results according to certain rules or standards to remove low-quality or unsuitable resources for recommendation.

[0238] S5, Monitoring and Optimization:

[0239] Real-time monitoring: Real-time monitoring of the quality of recommendation results and user feedback to identify and resolve problems in a timely manner.

[0240] A / B testing: Compare the effects of different recommendation strategies through A / B testing and select the best strategy for deployment.

[0241] Continuous optimization: Based on user feedback and data analysis results, we continuously optimize recommendation algorithms and strategies to improve the quality of recommendation results and user experience.

[0242] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0243] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A user cold start recommendation method based on meta-learning, characterized by: The detailed steps of the recommended technique include: Step 1: Collect user behavior logs: Collect user behavior data on various platforms, including preferences, resource browsing, clicks, searches, and sharing behaviors; Step 2: User feature preprocessing: Select key user features based on feature distribution and significance indicators; Step 3: Building a dual-tower multi-target recall model: Building a dual-tower multi-target recall model based on the characteristics of users and items; Step 4: Meta-learning user cold start sub-network: Introduce the concept of meta-learning and build a cold start sub-network; Step 5: Gated network function: Add a gated network to distinguish between new and old users; Step 6: Data synchronization and caching: synchronize the trained user embedding and item embedding to the system cache; Step 7: Online user incremental feature update: collect new user behavior data in real time and merge it into historical behavior data; Step 8, online user feature preprocessing: preprocess the user features updated in real time to keep consistent with the offline module; Step 9: Inferring online new user behavior profiles: based on real-time updated user features; Step 10: Recall results based on the overall model: Combine the user's current behavior profile and basic profile to obtain recommended resources from the resource library.

2. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific steps of collecting user behavior logs in step 1 include: A1, Preference collection: Track the different models, brands and price ranges browsed by users, analyze users' specific preferences, and record users' stay time, scroll depth and interactive behavior on the browsing page; A2, resource browsing collection: monitor the resource types of articles, pictures, and videos browsed by users, and record the browsing time and browsing order; A3, click behavior collection: Track the user's click behavior on the page, including the links, buttons, and image elements clicked; A4, Search behavior collection: monitor the keywords entered by users in the search box, record the search frequency and clicks on search results; A5, Sharing behavior collection: Tracking the content shared by users, sharing channels (including social media, email) and sharing objects; A6, context feature collection: real-time monitoring of user network status; A7, User feature collection: Collect user's geographical information through user registration information or IP address positioning; A8, distinguish between new and old users: distinguish between new and old users based on user registration time and behavior data; A9. Collect characteristics of new and old users separately: For new users, focus on collecting initial data such as registration information, first search and browsing behavior to build an initial user portrait. For old users, in addition to continuously collecting behavioral data, you also need to pay attention to the long-term characteristics of behavioral changes, interest shifts, and loyalty changes.

3. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific steps of user feature preprocessing in step 2 include: B1, Feature selection: Analyze the distribution of each feature, including its value range and frequency distribution, to determine the contribution of the feature to the user profile and evaluate the significance of each feature in distinguishing user groups or predicting user behavior; B2, feature compression coding: using simhash technology and hashing to generate fixed-length feature codes, compressing the selected user features and mapping the original high-dimensional feature space to a low-dimensional coding space; B3, bucketable data processing: For certain features, including age group and region, which have clear value ranges or categories, bucket processing is performed; B4, improve processing efficiency: through feature compression coding and bucket-divided data processing, the amount of data to be processed in subsequent calculations is reduced, thereby improving processing efficiency; B5, Note: When performing feature compression encoding, maintain the representativeness and discrimination of the original data to avoid excessive compression leading to information loss.

4. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific steps of constructing the dual-tower multi-target recall model in step 3 include: C1, model construction: C1-1, feature selection: including user features and item features; C1-2, feature encoding: Use deep neural networks to encode user and item features to generate user vectors and item vectors; C1-3, dual-tower structure: construct two independent neural network structures, one for processing user features (user tower) and the other for processing item features (item tower); C1-4, similarity calculation: After obtaining the user vector and item vector, the similarity is calculated using the cosine similarity and dot product methods; C2, loss function design: C2-1, multi-objective loss: design multiple loss functions according to specific recommendation goals and business needs, such as click-through rate prediction loss and purchase conversion rate prediction loss; C2-2, optimization algorithm: select a suitable optimization algorithm to minimize the total loss function; C3, model training C3-1, data preparation: preprocess and extract features from the collected user behavior logs to generate a data set for training; C3-2, model initialization: Initialize the parameters of the twin-tower model, including the weights and biases of the neural network; C3-3, iterative training: iteratively train the model using the training dataset, and update the model parameters by calculating the loss function and back-propagating the gradient; C3-4, Model evaluation: Use the test set to evaluate the trained model, calculate various performance indicators to evaluate the performance of the model, and tune and improve the model based on the evaluation results.

5. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific operations of meta-learning the user cold start sub-network in step 4 include: D1, build a cold start sub-network: design a sub-network specifically for dealing with the cold start problem, which learns to assist in the recommendation or prediction of new users; D2, set up the memory matrix: the memory matrix is ​​used to store the basic portraits of old users; D3, store and learn old user portraits: Basic portraits: collect and store basic information of old users, use machine learning algorithms to learn preferences and interests from old users' historical behaviors, and store the learned information in a memory matrix; D4, using the scale-and-shift approach; Scaling method: scale or weight the old user portrait according to the similarity between the new user and the old user. Transfer method: transfer the scaled old user portrait information to the new user as the initial portrait or prediction basis for the new user. D5, ensure that new users benefit from old user portraits: Ensure that new users can benefit from old user portraits through verification and testing. Based on the verification and testing results, continuously optimize and improve the meta-learning cold start sub-network to improve its performance and adaptability.

6. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific steps of data synchronization and caching in step six include: E1, Data synchronization: Select a cache system that provides high-performance key-value storage and fast data access capabilities; export the trained user embedding and item embedding data from the training environment, and import the exported data into the cache system; E2, design cache update strategy: for applications that require high real-time performance, update the cache immediately when user and item information changes. For applications that do not require high real-time performance, set a fixed time interval to update the cache regularly: when only some user and item data changes, use incremental update; E3, ensure data real-time and accuracy: implement a monitoring and alarm mechanism for the cache system to promptly detect and handle problems that arise during data synchronization and updating, set a reasonable expiration time (TTL) for the cache, and verify the data during data synchronization and updating to ensure its accuracy and completeness. If problems arise during data synchronization and updating that lead to data inconsistency or corruption, a rollback mechanism needs to be implemented to restore the data to its previous state.

7. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific steps of updating the online user incremental features in step seven include: F1, real-time collection of new user behavior data: Data source: Determine the source of the data source to collect user behavior data in real time, including records of user clicks, browsing, purchases, comments, and sharing behaviors; Data format: Ensure that the collected data is in a uniform format and is easy to process; Real-time: Use real-time data stream processing technology to ensure real-time data collection and processing; F2, merge the new data into the historical behavior data: Data cleaning: Clean the collected new data to remove duplicate, invalid and abnormal data; Data merging: Merge the cleaned new data into the user's historical behavior data; Data persistence: store the merged data persistently; F3, update user characteristics: Feature extraction: Extract features that reflect user interests and preferences based on the merged user historical behavior data; Feature update: replace or add the extracted new features or updated feature values ​​to the user's feature vector; F4, ensure the real-time and accuracy of model recommendations: Online learning: Use online learning algorithms to update model parameters in real time to reflect changes in user behavior data; Model evaluation: Regularly evaluate the accuracy of model recommendations, use offline and online evaluation indicators, and adjust model parameters or optimize algorithms based on the evaluation results; Feedback loop: Collect user feedback as new behavior data and merge it into historical behavior data to form a closed feedback loop.

8. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific steps of online user feature preprocessing in step eight include: G1, feature selection and alignment: Feature consistency: Ensure that the features used online are exactly the same as those used in offline training; Feature missing processing: For feature missing problems that occur in online environments, formulate processing strategies in advance; G2, feature scaling and normalization: Normalization: For numerical features, normalization techniques are used to scale them to the same scale; Standardization: transform the features into a distribution with a mean of 0 and a standard deviation of 1 through zero-mean normalization; G3, discrete feature processing: One-hot encoding: For categorical features, one-hot encoding is used to convert them into binary vectors; Feature Embedding: For high-dimensional or sparse discrete features, feature embedding technology is used to convert them into low-dimensional dense vectors; G4, Feature Update and Synchronization: Real-time update: When the user's features change, the user's feature vector is updated in real time; Synchronization mechanism: ensures synchronization between online feature database and offline feature database; G5, Monitoring and Evaluation: Feature quality monitoring: Regularly evaluate the quality of online features; Model performance evaluation: Monitor the performance of the online model to ensure that when using the preprocessed features for inference, the model's performance is comparable to or better than that during offline training.

9. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific steps of inferring the online new user behavior profile in step nine include: H1, inference of new user behavior profile: H1-1, real-time feature collection: collect new users’ behavior data in real time, including clicks, browsing, and searches; H1-2, feature preprocessing: preprocess the collected new user behavior data, including data cleaning, formatting, and normalization steps to ensure that the processed features are consistent with the feature format and range used in offline training; H1-3, Application of meta-learning sub-network: Using the trained meta-learning sub-network, the characteristics of new users are inferred. The meta-learning sub-network can quickly adapt to and predict users' potential preferences and behavior patterns based on limited user behavior data; H1-4, New user portrait construction: Based on the output of the meta-learning sub-network and combined with the basic information of the new user, a new user behavior portrait is constructed. The portrait will serve as an important reference for new users in future behavior analysis and recommendation systems; H2, update of old user portraits: H2-1, historical data collection: For old users, collect their historical behavior data; H2-2, feature extraction and updating: extract the features of old users from historical data, including static features and dynamic features; H2-3, Application of meta-learning sub-network: Similar to the inference of new user portraits, the meta-learning sub-network is used to infer the characteristics of old users; H2-4, portrait update: merge and update the output of the meta-learning sub-network with the original portrait of the old user.

10. The user cold start recommendation method based on meta-learning according to claim 1, characterized in that: The specific steps of the recall result based on the overall model in step 10 include: S1, integrated user portrait: Current behavior profile: Based on the user's recent behavior data, the user's current behavior profile is obtained through online feature preprocessing and meta-learning sub-network inference; Basic profile: stable characteristics and preferences accumulated by users over a long period of time; Integration: Integrate the current behavior profile and the basic profile to form a comprehensive and multi-dimensional understanding of the user; S2, obtain recommended resources from the resource library: Resource library: contains a collection of all recommended resources, such as products, articles, and videos; Resource characteristics: Each resource has a set of data describing its characteristics, such as the category, price, and sales volume of the product; the topic and keywords of the article; Matching: Based on the integrated user profile and resource features, an algorithm is used to obtain recommended resources that match the user's interests from the resource library. S3, design recommendation strategy: Diversified recommendations: Introduce resources of different types, styles, and fields into the recommendation results; Relevance recommendation: Based on the user's historical behavior and interests, highly relevant resources are recommended; Novelty recommendation: recommend new resources that users have not yet come into contact with but may be of interest; Personalized recommendation: Provide customized recommendation services based on users’ unique preferences and needs; S4, sorting and filtering recall results: Sorting: Use a suitable sorting algorithm to sort the recall results. Screening: Screen the recall results according to certain rules or standards to remove low-quality or unsuitable resources for recommendation. S5, Monitoring and Optimization: Real-time monitoring: Real-time monitoring of the quality of recommendation results and user feedback. A / B testing: Compare the effects of different recommendation strategies through A / B testing and select the best strategy for deployment. Continuous optimization: Based on user feedback and data analysis results, we continuously optimize recommendation algorithms and strategies to improve the quality of recommendation results and user experience.

Images