Cross-domain recommendation and model training method and device

Abstract

The invention relates to the technical field of recommendation systems, and discloses a cross-domain recommendation and model training method and device, and the method comprises the steps: obtaining and coding source domain behavior data through a source domain user behavior log collection module, and generating a user interest representation; target domain interaction data are collected, a user mapping relation is constructed, and preference alignment between the source domain and the target domain is achieved; constructing a joint training set, taking the source domain interest vector as auxiliary input, training a cross-domain recommendation model and generating model parameters; and finally, generating a personalized recommendation result based on the interest state of the target domain user, and dynamically adjusting a recommendation strategy according to user feedback. The device comprises a source domain representation construction module, a target domain alignment module, a migration training module and a recommendation feedback module, and interest migration and cross-domain recommendation performance optimization among multi-source data can be realized.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, and in particular to a method and apparatus for cross-domain recommendation and model training, applicable to multi-domain integrated recommendation scenarios such as e-commerce, video, and social media. Background Technology

[0002] In existing recommender systems, system performance often heavily relies on user interaction data accumulated within a single platform or domain. When target domain data is sparse and user cold starts are frequent, traditional recommender algorithms struggle to learn accurate user preferences, leading to a decline in recommender quality.

[0003] To address the problem of data sparsity, researchers have proposed the idea of ​​cross-domain recommendation, which involves importing user behavior data from other platforms (source domain) into the target platform, thereby improving the recommendation performance of the target domain by leveraging information from the source domain. However, existing methods often suffer from issues such as semantic inconsistency between the source and target domains, misalignment of user representations, and poor generalization of interest transfer, limiting their practical application effectiveness.

[0004] Furthermore, in most cross-domain recommendation models, the differences in dynamic behavioral features between the source and target domains are often ignored during the model training phase, and there is a lack of effective feedback mechanisms. This results in the model responding slowly when faced with rapid changes in user interests, leading to delayed recommendation results and seriously affecting the user experience. Summary of the Invention

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution: a method for cross-domain recommendation and model training, comprising the following steps:

[0006] Step S1: Obtain user click, rating, and browsing data from the source domain through the user behavior log collection module to construct the original behavior dataset D. src and the original behavior dataset D src After data cleaning, labeling, and formatting, a feature encoding model is used to uniformly vectorize user, item, and contextual information, forming a source domain representation set E. src ;

[0007] Step S2: Obtain the user-item interaction dataset D in the target domain. tar And construct user mapping relationships E between users with overlapping data. map The initial alignment of the source domain interest vector and the target domain preference representation is achieved through weakly supervised learning.

[0008] Step S3: Based on the construction of the source-target domain joint training set, using the source domain interest vector as auxiliary input and the target domain user behavior as supervision signal, train the cross-domain recommendation model. Improve the recommendation accuracy of the model in the target domain through iterative optimization strategies, and finally generate a stable transfer model parameter set P.final ;

[0009] Step S4: Apply the trained model to the target domain recommendation task, and generate a personalized recommendation list R based on the historical behavior and current interest state of users in the target domain. rec Furthermore, by collecting information on click, bounce, and dwell behaviors, the recommendation strategies of some models are adjusted in real time to achieve dynamic evolution and updates.

[0010] Preferably, step S1 includes:

[0011] S1(1) Collect raw user behavior logs from the source domain platform through the behavior collection interface. The raw user behavior logs include click events, rating records, search keywords, and access duration information.

[0012] S1(2) on the original behavior dataset D src Preprocessing is performed, including missing value imputation, outlier removal, text standardization, and timestamp sorting, to obtain the cleaned behavior sequence;

[0013] S1(3) Construct a unified vector encoding structure, and perform independent embedding operations on user ID, item ID and context features respectively, and map them to a unified feature space to form a structured feature representation vector;

[0014] S1(4) employs normalization and redundancy removal to make the source domain representation set E src It possesses high compactness and domain generalization ability, making it easy to use for subsequent transfer training.

[0015] Preferably, step S2 includes:

[0016] S2(1) Construct the target domain user-item interaction record dataset D tar The dataset D tar This includes user actions such as clicking, browsing, and rating items;

[0017] S2(2) If there are partially overlapping users between the source domain and the target domain, then directly establish the user mapping relationship F. map If there is no overlap, a weakly supervised mapping relationship is established based on the cosine similarity between user behavior vectors;

[0018] S2(3) Let the source domain user interest be represented as V src The initial user preferences for the target domain are V. tarinit Then the final user preference vector V in the target domain tar It is obtained by calculation using the following formula:

[0019] V tar =β·V src +(1-β)·V tarinit ;

[0020] Wherein, β is the fusion weight, which is derived from the score of the consistency of the user's historical behavior;

[0021] S2(4) will obtain V tar Perform batch cluster analysis to screen user groups with high consistency in behavioral characteristics, providing a transfer basis for subsequent models.

[0022] Preferably, step S3 includes:

[0023] S3(1) Construct a joint training set of source and target domains, which consists of source domain user interest vectors, target domain item feature vectors and interaction labels as training samples;

[0024] S3(2) Constructs the main body of the recommendation model using a neural network structure, with the front layer of the model receiving the source domain representation set E. src The subsequent layer fuses user features and item features from the target domain, and finally outputs the user's predicted rating for the item;

[0025] S3(3) uses supervised learning, takes the real interactive labels of the target domain as the training target, designs an error function and introduces an early stopping mechanism to prevent overfitting;

[0026] S3(4) optimizes the parameters using mini-batch gradient descent and finally obtains the transfer model parameter set P through several rounds of iterative training. final Once the model is trained, it can be persistently stored.

[0027] Preferably, step S4 includes:

[0028] S4(1) receives real-time requests from users in the target domain and dynamically generates interest vectors based on their historical interaction records, preference tags, and recent behaviors;

[0029] S4(2) Call the trained cross-domain model, combine the interest vector and the candidate item set, and output the predicted rating between user and item;

[0030] S4(3) Generate a personalized recommendation list R based on the predicted scores from high to low. rec and display it to the target user's terminal;

[0031] S4(4) records users’ feedback behavior on recommended content in real time and adjusts the weights of some strategies in the model in real time based on the feedback behavior to optimize the recommendation effect in the next round.

[0032] The present invention also provides a device for cross-domain recommendation and model training, comprising the following modules:

[0033] The source domain representation building module is used to acquire and process source domain user behavior data and construct a multidimensional interest vector representation;

[0034] The target domain alignment module is used to collect user data in the target domain and complete the source-target user preference alignment operation;

[0035] The transfer learning module is used to build and train cross-domain recommendation models;

[0036] The recommendation feedback module is used to generate recommendation results and receive feedback to dynamically optimize the recommendation effect.

[0037] The source domain representation construction module includes:

[0038] The behavior collection unit is used to collect user behavior log data on the platform from the source domain system, and supports unified collection of multiple behavior types.

[0039] The feature encoding unit is used to perform ID mapping, embedding vector construction, and feature normalization processing on user, item, and context information to generate structured representation vectors.

[0040] The preprocessing unit is used to perform outlier removal, standard format conversion, and behavior-time sorting operations on the raw data to ensure data quality and consistency with model input.

[0041] The target domain alignment module includes:

[0042] The user behavior analysis unit is used to construct the interaction records between users and items in the target domain and extract their behavioral feature vectors.

[0043] The mapping relationship construction unit is used to construct an initial mapping table between source and target domain users based on user ID matching;

[0044] The fusion strategy unit is used to automatically adjust the fusion ratio of source domain and target domain preferences based on the behavior consistency factor, and generate a unified user preference vector for input to the recommendation model.

[0045] The transfer training module includes:

[0046] The training data construction unit is used to build a joint training dataset based on source domain user interest representations, target domain item features, and target domain interaction labels. It supports batch processing format to organize samples for efficient model training.

[0047] The model structure generation unit is used to initialize the neural network structure of the cross-domain recommendation model. This structure supports source domain features as auxiliary input and target domain behavior as the main training objective. It supports dynamic combination configuration of the embedding layer, fusion layer and prediction output layer.

[0048] The model training execution unit is used to perform gradient calculation, error backpropagation, and parameter update operations during training. It also performs multiple rounds of iterative optimization based on the set loss function. After training is completed, the model weight parameters are stored locally for subsequent recommendation tasks.

[0049] The recommendation feedback module includes:

[0050] The interest analysis unit is used to generate the latest interest representation by combining the user's current context and historical behavior.

[0051] The recommendation generation unit is used to generate a ranked recommendation list based on the rating results output by the cross-domain model and push it to the user;

[0052] The feedback update unit is used to collect user feedback behavior data and combine it with the strategy module to update some weights of the model in real time, so as to realize the iterative evolution of personalized recommendations.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] Improving recommendation accuracy and generalization ability: This invention introduces user interest representations from the source domain and constructs user mapping and preference alignment mechanisms in the target domain, enabling the recommendation model to maintain high prediction accuracy in scenarios with sparse data or cold start in the target domain, and significantly improving the generalization performance of cross-platform recommendations.

[0055] Enhanced model adaptability and dynamic update capability: This invention adopts a feedback update mechanism, which can dynamically adjust the model strategy based on the user's real-time clicks, exits, and other behaviors, effectively responding to changes in user interests and improving the system's responsiveness to users' long-term preferences and short-term behaviors.

[0056] Modular design facilitates system integration and deployment: The device of this invention adopts a modular structure, including four major functional modules: source domain representation construction, target domain alignment, transfer training, and recommendation feedback. This facilitates flexible integration and deployment in actual recommendation systems and has good engineering practicality and scalability. Attached Figure Description

[0057] Figure 1 A flowchart illustrating the method steps provided in this application;

[0058] Figure 2 A schematic diagram of the system modules provided in this application. Detailed Implementation

[0059] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0060] refer to Figure 1 This invention provides a method for cross-domain recommendation and model training, comprising the following steps:

[0061] Step S1: Obtain user click, rating, and browsing data from the source domain through the user behavior log collection module to construct the original behavior dataset D. src and the original behavior dataset D src After data cleaning, labeling, and formatting, a feature encoding model is used to uniformly vectorize user, item, and contextual information, forming a source domain representation set E. src .

[0062] First, the system periodically captures user behavior logs through the behavior collection interface deployed on the source domain platform log server. The collected logs include click events, rating records, search keywords, and access duration information.

[0063] Subsequently, the system processed the collected raw behavioral dataset D src A series of preprocessing steps are performed. First, data imputation is performed for entries with missing fields, using historical user averages to complete ratings and interpolation to estimate access duration. Second, data with abnormal behavior times or out-of-bounds fields is removed to ensure data consistency. Third, search keywords and contextual text are standardized, including case sensitivity, stop word removal, and typos correction, and the behavior sequences are organized into a time-linked list based on their occurrence order using timestamps for subsequent modeling.

[0064] After preprocessing, the system constructs a unified vector encoding structure. Specifically, user IDs and item IDs are mapped to dense vectors of fixed dimensions using embedding technology. Contextual information such as time period, device type, and user location information is also transformed into structured features using one-hot or multi-value mapping. Multiple features are concatenated in the embedding space to form a complete structured representation vector.

[0065] Finally, to improve the robustness and transferability of the vector representation, the system performs normalization on the embedding vectors and uses principal component analysis to compress redundant parts of the feature dimensions, resulting in a source domain representation set E with high compactness and strong semantic expressiveness. src This set will serve as the source input for the transfer model, supporting subsequent alignment and joint modeling with the target domain.

[0066] Step S2: Obtain the user-item interaction dataset D in the target domain. tar And construct user mapping relationships F between users with overlapping data. map Weakly supervised learning is used to achieve the initial alignment of the source domain interest vector and the target domain preference representation.

[0067] In this embodiment, the target domain is an e-commerce platform, which includes behavioral data generated by users during browsing, searching, and purchasing goods. The system first constructs a user-item interaction dataset D for the target domain through a data interface. tar The data sources include user click logs on the platform, product browsing records, and rating behaviors extracted from order information. The dataset uses a unified format and includes the following fields: User ID, Item ID, Behavior Type (click, browse, purchase), Behavior Timestamp, and Behavior Intensity Value.

[0068] Next, user alignment is performed. If there are some overlapping users between the source and target domains (such as the same account being linked to both a short video account and an e-commerce account), the system directly establishes an explicit mapping relationship F based on the user ID. map For non-overlapping user groups, the system uses a weakly supervised approach to align behavior vectors. Specifically, it aligns the source domain user interest vectors V... src With the initial preference representation V of the target domain tarinit Cosine similarity is calculated between them, and user groups that are close in the vector space are filtered out based on a threshold to construct implicit mapping pairs.

[0069] Subsequently, to achieve interest transfer and preference fusion, the system uses the following weighted formula to initialize the target domain user interests:

[0070] V tar =β·V src +(1-β)·V tarinit ;

[0071] Here, β represents the fusion weight, which the system scores based on the difference in user activity and behavioral consistency between the source and target domains. For example, when a user is active in the source domain but data is sparse in the target domain, the system increases the weight of β; when there is sufficient behavior in the target domain, β is decreased to preserve the original preferences.

[0072] After fusion, the system analyzes the user preference vectors V of all target domains. tar Batch clustering is performed using the K-means algorithm to group users into multiple groups with consistent behavioral characteristics. This data is then used for label balancing, data augmentation, and transfer learning optimization during subsequent model training. This process provides a data foundation for the model to learn stable interest representations in the target domain.

[0073] Step S3: Based on the construction of the source-target domain joint training set, using the source domain interest vector as auxiliary input and the target domain user behavior as supervision signal, train the cross-domain recommendation model. Improve the recommendation accuracy of the model in the target domain through iterative optimization strategies, and finally generate a stable transfer model parameter set P. final .

[0074] First, the system obtains the source domain representation set E from the aforementioned steps. src Target domain user preference vector V tar The system constructs joint training samples using the feature vectors of items in the target domain. Each sample consists of three parts: a set of source domain representations, a final user preference vector in the target domain, and a feature vector of items in the target domain. The labels are the user's interaction behavior with the target item (such as whether to purchase or the rating value). The system divides the training samples into a training set and a validation set to ensure the separation of model training and generalization ability evaluation.

[0075] Next, the system constructs a deep recommendation model. The front layer of the model has a three-way input structure, which processes the source domain representation, target domain user features, and item features respectively; the middle layer has a feature fusion layer, which uses an attention mechanism to assign weights to the three types of features; the output layer is a regression unit that predicts the user's preference rating for candidate items.

[0076] The model is trained using supervised learning. The loss function is designed as a weighted mean squared error, and an early stopping mechanism is introduced to monitor the performance on the validation set and prevent overfitting. The system uses the Adam optimizer, with the learning rate gradually decaying. Gradient calculation and backpropagation are performed in mini-batch mode (batch size = 128) during training.

[0077] After multiple rounds of training and the validation set performance stabilizes, the system will set the model parameter set P. final The model is persistently stored in the model repository, retaining both the model version and training configuration. This model will be used for subsequent recommendation generation tasks.

[0078] Step S4: Apply the trained model to the target domain recommendation task, and generate a personalized recommendation list R based on the historical behavior and current interest state of users in the target domain. rec Furthermore, by collecting information on click, bounce, and dwell behaviors, the recommendation strategies of some models are adjusted in real time to achieve dynamic evolution and updates.

[0079] First, when a user in the target domain initiates a page access request, the system retrieves their historical interaction records from the user's data center in real time, including recently clicked item IDs, rating records, search keywords, and recent access time. The system combines this information with the user's current context (such as access time, device type, and location information) to generate a current interest vector through the interest modeling module, representing the user's current preference state.

[0080] The system calls the deployed cross-domain recommendation model P. final The system takes an interest vector and a set of candidate items as input, and outputs a predicted rating for each user-item pair through a neural network structure. A higher rating indicates a greater likelihood of user interest in the item. The system sorts the candidate set based on the ratings and selects the top-N items to form a personalized recommendation list R. rec And push it to the user's terminal interface.

[0081] After the recommendation results are displayed, the system activates the feedback monitoring module to record user behavior data related to the recommended content in real time, such as click-through rate, dwell time, and bounce rate. The system summarizes user feedback behavior within a certain time window (e.g., 24 hours) and performs anomaly analysis on samples that significantly deviate from the model's predictions.

[0082] Ultimately, the system dynamically adjusts some model strategy parameters based on the feedback results. For example, it may increase the weight of certain item features, correct the rules for generating user interest vectors, or supplement the incremental training dataset with feedback data samples. This mechanism ensures that the model can continuously learn from changes in user preferences, enabling the adaptive evolution of the cross-domain recommendation system.

[0083] refer to Figure 2 This invention provides a cross-domain recommendation and model training apparatus, comprising the following modules:

[0084] The source domain representation building module is used to acquire and process source domain user behavior data to construct a multidimensional interest vector representation.

[0085] The source domain representation construction module includes:

[0086] The behavior collection unit is used to collect user behavior log data on the platform from the source domain system, and supports unified collection of multiple behavior types.

[0087] The feature encoding unit is used to perform ID mapping, embedding vector construction, and feature normalization processing on user, item, and context information to generate structured representation vectors.

[0088] The preprocessing unit is used to perform outlier removal, standard format conversion, and behavior-time sorting operations on the raw data to ensure data quality and consistency with model input.

[0089] The target domain alignment module is used to collect target domain user data and complete the source-target user preference alignment operation.

[0090] The target domain alignment module includes:

[0091] The user behavior analysis unit is used to construct the interaction records between users and items in the target domain and extract their behavioral feature vectors.

[0092] The mapping relationship construction unit is used to construct an initial mapping table between source and target domain users based on user ID matching;

[0093] The fusion strategy unit is used to automatically adjust the fusion ratio of source domain and target domain preferences based on the behavior consistency factor, and generate a unified user preference vector for input to the recommendation model.

[0094] The transfer learning module is used to build and train cross-domain recommendation models.

[0095] The transfer training module includes:

[0096] The training data construction unit is used to build a joint training dataset based on source domain user interest representations, target domain item features, and target domain interaction labels. It supports batch processing format to organize samples for efficient model training.

[0097] The model structure generation unit is used to initialize the neural network structure of the cross-domain recommendation model. This structure supports source domain features as auxiliary input and target domain behavior as the main training objective. It supports dynamic combination configuration of the embedding layer, fusion layer and prediction output layer.

[0098] The model training execution unit is used to perform gradient calculation, error backpropagation, and parameter update operations during training. It also performs multiple rounds of iterative optimization based on the set loss function. After training is completed, the model weight parameters are stored locally for subsequent recommendation tasks.

[0099] The recommendation feedback module is used to generate recommendation results and receive feedback to dynamically optimize the recommendation effect.

[0100] The recommendation feedback module includes:

[0101] The interest analysis unit is used to generate the latest interest representation by combining the user's current context and historical behavior.

[0102] The recommendation generation unit is used to generate a ranked recommendation list based on the rating results output by the cross-domain model and push it to the user;

[0103] The feedback update unit is used to collect user feedback behavior data and combine it with the strategy module to update some weights of the model in real time, so as to realize the iterative evolution of personalized recommendations.

[0104] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0105] Obviously, the embodiments described above are merely some embodiments of the present invention, not all embodiments. The accompanying drawings show preferred embodiments of the present invention, but do not limit the patent scope of the present invention. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the patent protection scope of this invention.

Claims

1. A method for cross-domain recommendation and model training, characterized in that, Includes the following steps: Step S1: Obtain user click, rating, and browsing data from the source domain through the user behavior log collection module to construct the original behavior dataset D. src and the original behavior dataset D src After data cleaning, labeling, and formatting, a feature encoding model is used to uniformly vectorize user, item, and contextual information, forming a source domain representation set E. src ; Step S2: Obtain the user-item interaction dataset D in the target domain. tar And construct user mapping relationships F between users with overlapping data. map The initial alignment of the source domain interest vector and the target domain preference representation is achieved through weakly supervised learning. Step S3: Based on the construction of the source-target domain joint training set, using the source domain interest vector as auxiliary input and the target domain user behavior as supervision signal, train the cross-domain recommendation model. Improve the recommendation accuracy of the model in the target domain through iterative optimization strategies, and finally generate a stable transfer model parameter set P. final ; Step S4: Apply the trained model to the target domain recommendation task, and generate a personalized recommendation list R based on the historical behavior and current interest state of users in the target domain. rec Furthermore, by collecting information on click, bounce, and dwell behaviors, the recommendation strategies of some models are adjusted in real time to achieve dynamic evolution and updates.

2. The method for cross-domain recommendation and model training according to claim 1, characterized in that, Step S1 includes: S1(1) Collect raw user behavior logs from the source domain platform through the behavior collection interface. The raw user behavior logs include click events, rating records, search keywords, and access duration information. S1(2) on the original behavior dataset D src Preprocessing is performed, including missing value imputation, outlier removal, text standardization, and timestamp sorting, to obtain the cleaned behavior sequence; S1(3) Construct a unified vector encoding structure, and perform independent embedding operations on user ID, item ID and context features respectively, and map them to a unified feature space to form a structured feature representation vector; S1(4) employs normalization and redundancy removal to make the source domain representation set E src It possesses high compactness and domain generalization ability, making it easy to use for subsequent transfer training.

3. The method for cross-domain recommendation and model training according to claim 1, characterized in that, Step S2 includes: S2(1) Construct the target domain user-item interaction record dataset D tar The dataset D tar This includes user actions such as clicking, browsing, and rating items; S2(2) If there are partially overlapping users between the source domain and the target domain, then directly establish the user mapping relationship F. map If there is no overlap, a weakly supervised mapping relationship is established based on the cosine similarity between user behavior vectors; S2(3) Let the source domain user interest be represented as V src The initial user preferences for the target domain are V. tarinit Then the final user preference vector V in the target domain tar It is obtained by calculation using the following formula: V tar =β·V src +(1-β)·V tarinit ; Wherein, β is the fusion weight, which is derived from the score of the consistency of the user's historical behavior; S2(4) will obtain V tar Perform batch cluster analysis to screen user groups with high consistency in behavioral characteristics, providing a transfer basis for subsequent models.

4. The method for cross-domain recommendation and model training according to claim 1, characterized in that, Step S3 includes: S3(1) Construct a joint training set of source and target domains, which consists of source domain user interest vectors, target domain item feature vectors and interaction labels as training samples; S3(2) Constructs the main body of the recommendation model using a neural network structure, with the front layer of the model receiving the source domain representation set E. src The subsequent layer fuses user features and item features from the target domain, and finally outputs the user's predicted rating for the item; S3(3) uses supervised learning, takes the real interactive labels of the target domain as the training target, designs an error function and introduces an early stopping mechanism to prevent overfitting; S3(4) optimizes the parameters using mini-batch gradient descent and finally obtains the transfer model parameter set P through several rounds of iterative training. final Once the model is trained, it can be persistently stored.

5. The method for cross-domain recommendation and model training according to claim 1, characterized in that, Step S4 includes: S4(1) receives real-time requests from users in the target domain and dynamically generates interest vectors based on their historical interaction records, preference tags, and recent behaviors; S4(2) Call the trained cross-domain model, combine the interest vector and the candidate item set, and output the predicted rating between user and item; S4(3) Generate a personalized recommendation list R based on the predicted scores from high to low. rec and display it to the target user's terminal; S4(4) records users’ feedback behavior on recommended content in real time and adjusts the weights of some strategies in the model in real time based on the feedback behavior to optimize the recommendation effect in the next round.

6. An apparatus for cross-domain recommendation and model training, based on the method for cross-domain recommendation and model training as described in claim 1, characterized in that, The device includes: The source domain representation building module is used to acquire and process source domain user behavior data and construct a multidimensional interest vector representation; The target domain alignment module is used to collect user data in the target domain and complete the source-target user preference alignment operation; The transfer learning module is used to build and train cross-domain recommendation models; The recommendation feedback module is used to generate recommendation results and receive feedback to dynamically optimize the recommendation effect.

7. The apparatus for cross-domain recommendation and model training according to claim 6, characterized in that, The source domain representation construction module includes: The behavior collection unit is used to collect user behavior log data on the platform from the source domain system, and supports unified collection of multiple behavior types. The feature encoding unit is used to perform ID mapping, embedding vector construction, and feature normalization processing on user, item, and context information to generate structured representation vectors. The preprocessing unit is used to perform outlier removal, standard format conversion, and behavior-time sorting operations on the raw data to ensure data quality and consistency with model input.

8. The apparatus for cross-domain recommendation and model training according to claim 6, characterized in that, The target domain alignment module includes: The user behavior analysis unit is used to construct the interaction records between users and items in the target domain and extract their behavioral feature vectors. The mapping relationship construction unit is used to construct an initial mapping table between source and target domain users based on user ID matching; The fusion strategy unit is used to automatically adjust the fusion ratio of source domain and target domain preferences based on the behavior consistency factor, and generate a unified user preference vector for input to the recommendation model.

9. The apparatus for cross-domain recommendation and model training according to claim 6, characterized in that, The transfer training module includes: The training data construction unit is used to build a joint training dataset based on source domain user interest representations, target domain item features, and target domain interaction labels. It supports batch processing format to organize samples for efficient model training. The model structure generation unit is used to initialize the neural network structure of the cross-domain recommendation model. This structure supports source domain features as auxiliary input and target domain behavior as the main training objective. It supports dynamic combination configuration of the embedding layer, fusion layer and prediction output layer. The model training execution unit is used to perform gradient calculation, error backpropagation, and parameter update operations during training. It also performs multiple rounds of iterative optimization based on the set loss function. After training is completed, the model weight parameters are stored locally for subsequent recommendation tasks.

10. The apparatus for cross-domain recommendation and model training according to claim 6, characterized in that, The recommendation feedback module includes: The interest analysis unit is used to generate the latest interest representation by combining the user's current context and historical behavior. The recommendation generation unit is used to generate a ranked recommendation list based on the rating results output by the cross-domain model and push it to the user; The feedback update unit is used to collect user feedback behavior data and combine it with the strategy module to update some weights of the model in real time, so as to realize the iterative evolution of personalized recommendations.

Images