E-commerce recommendation method and computing device

Abstract

The invention provides a method for e-commerce recommendation and computing equipment. The method comprises the following steps: performing joint training of a global multi-scene recommendation model through sample mixing and weight mixing; performing global behavior recognition in combination with behavior sequence features, and perfecting the global multi-scene recommendation model; identifying an identification feature of the scene; and performing scene customization screening according to the weight of the identification feature, so that the global multi-scene recommendation model realizes e-commerce recommendation under multiple scenes. According to the technical scheme, the user data utilization rate can be improved, different scene features can be processed in a standardized and unified mode, the prediction precision, performance and efficiency of the model are improved, the model can better adapt to the requirements of different scenes, and more accurate and personalized recommendation services are provided for users.

Description

Technical Field

[0001] The present invention relates to the technical field of software development, and in particular to a method and computing device for e-commerce recommendation. Background Art

[0002] With the development of deep learning, breakthroughs have been made in large-scale models, and more and more large-scale models are emerging. Recommendation algorithms are now ubiquitous in current e-commerce software. Private marketing scenarios use them to recommend special prices or quality products to users within the context of the site, and promotional event pages also rely on recommendation algorithms. The "Guess You'll Also Like" section on the homepage of e-commerce software is a key channel for attracting users within the site. These scenarios, including during and after the purchase process, influence users' purchasing decisions.

[0003] To increase click-through and conversion rates, recommendation systems tend to recommend products users might like. However, this can lead to overly narrow recommendations and fail to meet users' need for novelty. Therefore, ensuring accurate recommendations while providing a diverse selection of products is a major challenge. Given the diverse range of recommendation scenarios and complex application requirements, developing universal, customized algorithms and engineering services to meet global recommendation algorithm requirements is a key issue.

[0004] To this end, a technical solution is needed to enable the model to better adapt to the needs of different scenarios. Summary of the Invention

[0005] The present invention aims to provide a method and computing device for e-commerce recommendations, which can improve the utilization rate of user data, standardize and unify the characteristics of different scenarios, improve the prediction accuracy, performance and efficiency of the model, enable the model to better adapt to the needs of different scenarios, and provide users with more accurate and personalized recommendation services.

[0006] According to one aspect of the present invention, a method for e-commerce recommendation is provided, the method comprising:

[0007] Joint training of global multi-scenario recommendation models through sample mixing and weight mixing;

[0008] Combine behavior sequence features to perform global behavior recognition and improve the global multi-scenario recommendation model;

[0009] Identify the scene's signature features;

[0010] Scenario-customized screening is performed according to the weights of the identification features, so that the global multi-scenario recommendation model can realize e-commerce recommendations in multiple scenarios.

[0011] According to some embodiments, joint training of a global multi-scenario recommendation model through sample mixing and weight mixing includes:

[0012] The homepage recommendation samples and the global recommendation samples are coordinated and trained to achieve the sample mixing, wherein the homepage recommendation samples and the global recommendation samples share an embedding vector.

[0013] According to some embodiments, the joint training of the global multi-scenario recommendation model by sample mixing and weight mixing further includes:

[0014] Perform proportional fusion and training on samples from all scenes to generate a unified embedding vector, combined with the embedding vector of scene-specific features;

[0015] The scene-specific features are added to the global multi-scene recommendation model through a scene-specific embedding layer and fused with the general features to improve the adaptability of the model to different scenes.

[0016] According to some embodiments, the joint training of the global multi-scenario recommendation model by sample mixing and weight mixing further includes:

[0017] The embedding vector and weight of the homepage recommendation model are imported into the global recommendation model through a learning rate warm-up operation to achieve the weight mixing.

[0018] According to some embodiments, global behavior recognition is performed in combination with behavior sequence features to improve the global multi-scenario recommendation model:

[0019] Extracting user behavior data in different scenarios in the behavior sequence;

[0020] Combining the user behavior data with relevant pages and context information of the different scenarios to construct the behavior sequence features at the scenario level;

[0021] The user's global behavior in the different scenarios is subjected to page scenario-level attribution modeling, so that the global multi-scenario recommendation model can perceive the data of the global and all behaviors.

[0022] According to some embodiments, the global multi-scenario recommendation model upgrades the behavior sequence features from a single product granularity to a scenario-level granularity.

[0023] According to some embodiments, identifying identification features of a scene includes:

[0024] Use a gated unit structure or attention mechanism network to identify the scene's features and assign weights.

[0025] According to some embodiments, the recommendation performance of the global multi-scenario recommendation model in different scenarios has different characteristics.

[0026] According to some embodiments, for the difference features in different scenarios, the generalization of the model is adjusted by setting an auxiliary task loss function.

[0027] According to another aspect of the present invention, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the method as described above is implemented.

[0028] According to another aspect of the present invention, there is provided a computing device comprising:

[0029] processor;

[0030] A memory stores a computer program, and when the computer program is executed by the processor, implements any of the methods described above.

[0031] According to an embodiment of the present invention, the global multi-scenario recommendation model is jointly trained through sample mixing and weight mixing to improve the utilization rate of user data and realize the standardization and unified processing of different scenario features. Global behavior recognition is performed in combination with behavior sequence features to improve the recommendation capability of the global multi-scenario recommendation model, so that the model can better adapt to the needs of different scenarios. By identifying the identification features of the scene and performing customized scene screening according to the weight of the identification features, the global multi-scenario recommendation model can realize e-commerce recommendations in multiple scenarios, provide users with more accurate and personalized recommendation services, and improve the prediction accuracy, performance and efficiency of the model.

[0032] According to some embodiments, joint training through sample mixing and weight mixing allows the model to learn the commonalities and characteristics of different scenarios, thereby improving the model's generalization and adaptability in multiple scenarios. This helps to overcome the limitations that single-scenario models may face and enhance the model's generalization capabilities.

[0033] According to some embodiments, global behavior recognition, combined with behavioral sequence features, can more accurately capture user interest changes and preference trends. This can provide users with more personalized and accurate product recommendations based on user behavior sequences, thereby improving user satisfaction and platform conversion rates.

[0034] According to some embodiments, scenario-customized screening is performed based on the weights of identification features, so that the global multi-scenario recommendation model can provide targeted recommendation strategies based on the characteristics of different scenarios, effectively taking into account the differences between different scenarios, and can flexibly respond to diverse user needs and market environments, supporting scenario-customized screening.

[0035] According to some embodiments, resource utilization is optimized: Compared to developing and maintaining a recommendation model for each scenario, a global, multi-scenario recommendation model can reduce maintenance costs and more efficiently utilize computing resources. Compared to a single-scenario model, a global, multi-scenario recommendation model is more stable and robust, and is less susceptible to individual outliers or noise.

[0036] It should be understood that the foregoing general description and the following detailed description are exemplary only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments.

[0038] Figure 1 A flow chart of a method for e-commerce recommendation according to an example embodiment is shown.

[0039] Figure 2 A schematic diagram illustrating the structure and training process of an e-commerce multi-scenario recommendation ranking model according to an example embodiment.

[0040] Figure 3 A schematic diagram illustrating a global behavior sequence according to an example embodiment.

[0041] Figure 4 A block diagram of a computing device is shown according to an exemplary embodiment. DETAILED DESCRIPTION

[0042] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. Like reference numerals in the drawings represent like or similar parts, and thus repeated description thereof will be omitted.

[0043] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, it will be appreciated by those skilled in the art that the technical solutions of the present invention can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present invention.

[0044] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0045] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, while others may be combined or partially combined. Therefore, the actual execution order may vary depending on the actual situation.

[0046] It should be understood that although the terms first, second, third, etc. may be used herein to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Thus, the first component discussed below could be referred to as the second component without departing from the teachings of the present inventive concept. As used herein, the term "and / or" includes any one and all combinations of one or more of the associated listed items.

[0047] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0048] Those skilled in the art will understand that the drawings are merely schematic diagrams of example embodiments, and the modules or processes in the drawings are not necessarily necessary for implementing the present invention, and therefore cannot be used to limit the scope of protection of the present invention.

[0049] Although e-commerce platforms boast a vast number of users and products, data on each user's purchasing behavior or ratings is often very limited, making it difficult to build accurate personalized recommendation models. For newly registered users, the system lacks sufficient historical behavioral data to make personalized recommendations. When new products are added to the shelves, the lack of user interaction data makes it difficult to assess their quality and determine which users might be interested in them. User preferences and needs constantly evolve over time. For example, seasonal changes, fashion trends, and personal growth all influence user shopping choices. Capturing and responding to these changes in a timely manner is also a challenge.

[0050] The "Guess You'll Like" section on e-commerce software homepages is a key channel for attracting users within the site. Scenarios throughout the shopping process, including during and after purchase, influence users' purchasing decisions. To improve click-through and conversion rates, recommendation systems tend to recommend products users are likely to like. However, this can lead to overly narrow recommendations and fail to meet users' need for novelty. Therefore, ensuring accurate recommendations while providing a diverse product selection is a major challenge. Given the diverse range of recommendation scenarios and complex application requirements, developing universal, customized algorithms and engineering services to meet the needs of a comprehensive recommendation algorithm becomes a key issue.

[0051] To this end, the present invention proposes a method for e-commerce recommendation, which can improve the utilization rate of user data, standardize and unify the characteristics of different scenarios, improve the prediction accuracy, performance and efficiency of the model, enable the model to better adapt to the needs of different scenarios, and provide users with more accurate and personalized recommendation services.

[0052] According to some embodiments, the present invention adopts a method that integrates Embedding, multi-scenario recognition structure and scenario-specific network structure, and identifies important scene identification features through Gate structure or Attention network, performs scene-customized screening of neurons, networks or vectors according to recognition weights, and adjusts the generalization of the model by setting auxiliary task loss functions for significantly differentiated scenes and indicators. The present invention uses a unified global multi-scenario recommendation model to solve the recommendation needs of multiple scenes. On the algorithm side, the scene features are modeled through the multi-scenario recommendation model algorithm, so that the model can intelligently adapt to the accurate recommendation of all scenes; on the engineering side, the rapid development of multiple scene application needs is completed through the configuration and plug-in abstraction of the required application logic.

[0053] According to some embodiments, the present invention proposes a global, multi-scenario recommendation model that can address recommendation needs across multiple e-commerce scenarios with a single model. This model intelligently adapts to multiple scenarios by identifying and addressing the differences between different application scenarios and leveraging prior knowledge to solve specific problems.

[0054] According to some embodiments, during the cold start phase, transfer learning techniques are used to transfer the capabilities of the homepage recommendation model to the global recommendation model. This invention employs weight blending (importing the embeddings and weights of the homepage recommendation model into the global recommendation model) and sample blending (jointly training data samples for homepage and global recommendations), significantly reducing the model's cold start time and accelerating the model's convergence in the target domain.

[0055] Exemplary embodiments of the present invention are described below with reference to the accompanying drawings.

[0056] Figure 1A flow chart of a method for e-commerce recommendation according to an example embodiment is shown.

[0057] See also Figure 1 ,In S101, the global multi-scenario recommendation model is jointly trained through sample mixing and weight mixing.

[0058] According to some embodiments, the global, multi-scenario recommendation model paradigm is a key feature of global recommendation, aiming to solve problems in multiple scenarios with a single model. Because different application scenarios have varying data distributions, pages, categories, and demographics, this paradigm needs to recognize and address these differences. By incorporating prior knowledge into the model, a universal approach can be used to solve specific problems. This is the essence of the global, multi-scenario recommendation model.

[0059] See also Figure 2 ,According to some embodiments, the global multi-scenario recommendation model adopts a two-stage ,training approach.

[0060] According to some embodiments, in the first stage, a global multi-scenario recommendation model is trained through sample mixing. Homepage recommendation samples and global recommendation samples are trained in a coordinated manner to achieve the sample mixing, wherein the homepage recommendation samples and the global recommendation samples share an embedding vector.

[0061] Sample mixing refers to the joint training of data samples for home page recommendations and global recommendations to improve the overall performance of the model. Home page recommendation samples are added to the global recommendation model in the form of co-training, and are jointly trained together with the global recommendation data samples. The embedding vectors (Embedding) are only shared with the home page recommendation model, retaining their independent model structures. Through joint training, the features and weights of the home page recommendation model can assist the training of the global recommendation model and improve the model's adaptability to global data. The sample mixing method enhances the quality of the embedding without adding any additional complex structure and without significantly increasing resource overhead online.

[0062] According to some embodiments, in the second phase, the global multi-scenario recommendation model trained in the first phase is further trained through weight blending. The capabilities of the global multi-scenario recommendation model completed in the first phase are further trained in specific scenario domains to achieve customization and blending of capabilities for multiple scenarios. The embedding vectors and weights of the homepage recommendation model are imported into the global recommendation model through a learning rate warm-up operation to achieve the weight blending.

[0063] Weight blending involves importing the embeddings and weights of the homepage recommendation model into the global recommendation model through a learning rate warmup (Warmup). Warmup import initializes the embeddings and weights of the homepage recommendation model into the global recommendation model, helping the global recommendation model converge quickly during the cold start phase. These initial parameters are adjusted through a small number of training steps (i.e., the Warmup phase) to adapt them to the global recommendation task.

[0064] In S103, global behavior recognition is performed in combination with behavior sequence features to improve the global multi-scenario recommendation model.

[0065] According to some embodiments, user behavior data from different scenarios in a behavior sequence is extracted and combined with relevant page and contextual information from the different scenarios to construct scenario-level behavior sequence features. Attribution modeling is performed on the user's global behavior in the different scenarios at the page-scenario level, allowing the global multi-scenario recommendation model to perceive data from all domains and all behaviors. The global multi-scenario recommendation model then upgrades the behavior sequence features from the granularity of a single product to the granularity of the scenario.

[0066] See also Figure 3 Global behavior recognition is added to sequence features. Within a user's behavior sequence, user behavior patterns in different scenarios are identified and upgraded from the item level to the scene level. This implementation involves extracting user behavior data in different scenarios, such as clicks, purchases, and browsing; combining this behavior data with scene-related page and contextual information to construct scene-level behavior features.

[0067] According to some embodiments, global recommendations utilize a method that integrates embeddings, multi-scene recognition structures, and scene-specific network structures. First, samples from all scenes are proportionally integrated and trained to generate a unified embedding, which is then combined with scene-specific feature embeddings.

[0068] According to some embodiments, unique features refer to features that are unique to a particular scenario and are different from the general features of other scenarios. For example, in a short video scenario, features such as the video's length and number of views are unique features. In the e-commerce shopping scenario of the present invention, features such as product sales and reviews are unique features. These unique features are added to the model through a scenario-specific embedding layer and fused with the general features to improve the model's adaptability to different scenarios.

[0069] According to some embodiments, by learning these features, the global multi-scenario recommendation model and recommendation system can better understand user behavior patterns in different scenarios. By combining the existing feature system of the homepage recommendation model, including clicks and long click sequences, purchases, and multiple behavior sequences, page-scenario-level attribution modeling is performed on the global user behavior, enabling the model to perceive the global and all-behavior footprint.

[0070] At S105 , identification features of the scene are identified.

[0071] According to some embodiments, a gated unit structure or an attention mechanism network is used to identify the signature features of the scene and assign weights.

[0072] Use a gate (gated unit) structure or an attention network to identify important scene identification features. Identification features are used to distinguish characteristics of different scenes or user behaviors, including scene-specific identification features and user behavior identification features. Scene-specific identification features include page type (homepage, mid-purchase, post-purchase), event type (big sale, daily recommendations), etc.; user behavior identification features include click-through rate and purchase conversion rate of users in a specific scene. These identification features are identified and weighted using the gate or attention mechanism.

[0073] In S107, scenario-customized screening is performed according to the weights of the identification features, so that the global multi-scenario recommendation model can realize e-commerce recommendations in multiple scenarios.

[0074] In some embodiments, neurons, networks, or vectors are screened for specific scenarios based on gate and attention recognition weights. These screening criteria are used to select important features or neurons, including criteria such as a feature's importance score exceeding a certain threshold, a feature's high contribution in a specific scenario, or a feature's strong relevance to the target scenario. These criteria are dynamically adjusted through gating units or attention mechanisms to optimize model performance.

[0075] According to some embodiments, the global multi-scenario recommendation model has different recommendation performance characteristics in different scenarios. For the different characteristics and the different indicators in different scenarios, the generalization of the model is adjusted by setting an auxiliary task loss function.

[0076] For significantly differentiated scenarios and metrics, the model's generalization is adjusted by setting an auxiliary task loss function. Significantly differentiated scenarios and metrics refer to features or metrics whose performance varies significantly across different scenarios. For example, significantly differentiated scenarios include homepage recommendations and post-purchase recommendations in e-commerce software. Homepage recommendations focus more on the breadth of user interests, while post-purchase recommendations focus more on repeat purchase behavior. Significantly differentiated metrics include click-through rate and conversion rate. Homepage recommendations have a higher click-through rate but a lower conversion rate, while post-purchase recommendations have a higher conversion rate but a lower click-through rate.

[0077] According to some embodiments, an auxiliary task loss function (Auxiliary Task Loss) refers to an additional loss function introduced in addition to the main task (such as recommendation accuracy) to optimize model generalization. In significantly differentiated scenarios, an auxiliary task loss function is introduced to optimize the indicators of a specific scenario. Through multi-task learning, the loss functions of the main task and the auxiliary task are weighted and summed to obtain the final optimization target. The auxiliary task loss function can help the model better adapt to different scenarios and improve generalization ability.

[0078] In a multi-scenario recommendation system, the objectives and optimization metrics for different scenarios may differ significantly. For example, homepage recommendations may focus more on click-through rate (CTR), while post-purchase recommendations may focus more on conversion rate (CVR). To optimize these different objectives, a specific loss function can be designed for each scenario and combined into an overall loss function through a weighted summation.

[0079] The following examples illustrate two significantly different scenarios.

[0080] Scenario A: Click-through rate optimization. Use binary cross entropy loss (BCE Loss) to optimize click-through rate.

[0081] LCTR=-N1 i =1∑N[y i log(pi)+(1-y i )log(1-p i )]

[0082] Among them, y i is the actual click label of the sample (0 or 1), p i is the click probability predicted by the model.

[0083] Scenario B: Conversion rate optimization. Binary cross entropy loss is also used to optimize conversion rate.

[0084] LCVR=-M1 j =1∑M[z j log(qj)+(1-z j )log(1-q j )]

[0085] Among them, z j is the actual conversion label of the sample (0 or 1), q j is the conversion probability predicted by the model.

[0086] The final total loss function can be expressed as:

[0087] Ltotal=αLCTR+βLCVR

[0088] Among them, α and β are weight coefficients used to balance the losses in different scenarios.

[0089] According to some embodiments, if auxiliary tasks are added, it is necessary to include an auxiliary task loss function for multi-task learning to optimize the model's generalization ability. For example, suppose there are two tasks: the primary task is click-through rate optimization, and the auxiliary task is conversion rate optimization. The loss function of the auxiliary task can be weighted into the total loss.

[0090] Assuming that the main task loss is Lmain and the auxiliary task loss is Laux, the total loss function can be expressed as:

[0091] Ltotal=Lmain+λLaux

[0092] Among them, λ is the weight coefficient of the auxiliary task loss, which is used to adjust the contribution of the auxiliary task to the total loss. If the main task is click-through rate optimization (CTR) and the auxiliary task is conversion rate optimization (CVR), then:

[0093] Ltotal=LCTR+λLCVR

[0094] According to some embodiments, global recommendation sorting not only improves data utilization, but also solves the problems of feature alignment and resource consumption, and avoids target field prediction deviations caused by the existence of global data. In addition, by introducing the global interest model, the prediction accuracy of the fine ranking model is further improved, and a seamless connection from search to recommendation is achieved. The global recommendation capability can feed back the technical iteration of homepage recommendation. This case uses global behavioral data to bury data (i.e., data collection) wherever user behavior occurs in all different scenarios. In this way, user behavior data in different scenarios can be comprehensively collected for model training and optimization. After the global behavioral data is fully buried, it can be used for training the homepage recommendation model. The global general model can be configured with the output value of the homepage recommendation model through weights to serve all recommendation scenarios in the e-commerce software.

[0095] According to some embodiments, feature alignment refers to standardizing and unifying features across different scenarios in multi-scenario recommendations so that the model can better process them. Feature weights are dynamically adjusted through model learning to ensure that features from different scenarios have equal contributions to the model. An attention mechanism is used to assign weights to features from different scenarios.

[0096] In order to better utilize global data while avoiding the difficulties of feature alignment and resource consumption, the solution of the present invention trains a global model on the global data by extracting a small number of common features, producing a series of features representing the user's global interests, and then sending these features to the refined ranking model for training and prediction.

[0097] According to some embodiments, the solution of the present invention effectively solves the problem of feature alignment and resource consumption, and does not cause prediction deviation in the target field due to the existence of global data. The global model features do not affect the training data distribution of the field model.

[0098] According to some embodiments, the global multi-scenario recommendation model effectively improves the performance and efficiency of the global recommendation system by integrating multiple technologies and methods, enabling the model to better adapt to the needs of different scenarios and provide more accurate and personalized recommendation services.

[0099] According to some embodiments, the present invention addresses the cold start problem with a multi-scenario e-commerce recommendation model based on transfer learning and LHUC (Learning Hidden Unit Contributions). Using transfer learning, knowledge can be extracted from pre-trained models in one or more source domains and applied to the target domain. This can significantly reduce the cold start time of the target domain model because the initial parameters have been pre-trained and have a certain degree of generalization capability. Since the initial parameters have been optimized, the model converges faster on the target domain, thereby reducing training time.

[0100] According to some embodiments, the solution of the present invention integrates scene-specific features through the LHUC mechanism. LHUC is a method for adaptively adjusting the weights of network layers, which dynamically adjusts the output of each layer by introducing a learnable scaling factor. This method allows the model to automatically adjust its behavior in different scenarios to better capture scene-specific features. By dynamically adjusting the network layer weights through LHUC, the model can adaptively capture the specific features of different scenarios, thereby improving the accuracy of recommendations. In addition to using general scene features, specific embedding layers can also be designed for each scene. These embedding layers can capture feature information unique to the scene and fuse it with general features.

[0101] According to some embodiments, the present invention implements cross-optimization of product and user characteristics in e-commerce. Product and user characteristics are crucial in e-commerce recommendation systems. By introducing cross-features, the relationship between user preferences and product characteristics can be more accurately captured. Combining features from multiple modalities, such as text, images, and videos, further enriches product and user representations, improving recommendation effectiveness.

[0102] According to some embodiments, the present invention optimizes multi-scenario sequence modeling. Using sequence modeling techniques such as Target Attention, it models user behavior sequences across different scenarios, capturing both long-term and short-term changes in user interests. The present invention not only focuses on sequence patterns within a single scenario but also considers user switching behavior between different scenarios, building a cross-scenario sequence model. For example, a user might first see a product in a short video scenario and then purchase it in a shopping scenario. This cross-scenario behavior pattern can be captured through joint modeling.

[0103] According to some embodiments, the present invention upgrades user behavior from the item level to the scene level within sequence features, incorporating scene-specific page and contextual information. Furthermore, using sequence modeling techniques such as Target Attention, it models user behavior sequences across different scenarios, capturing both long-term and short-term changes in user interests. This constructs a cross-scenario sequence model, optimizing the modeling capabilities of multi-scenario sequences.

[0104] According to some embodiments, the present invention extracts a small number of common features from global data to train features representing users' global interests and feeds them into a refined ranking model. This approach effectively solves the problem of feature alignment and resource consumption, and avoids target field prediction bias caused by global data.

[0105] According to some embodiments, global recommendation capabilities can feed back into the technical iteration of homepage recommendations. Global behavioral data, once collected across all points, can be used to train the homepage recommendation model. This universal model, combined with the output of the homepage recommendation model through weighting, can serve all recommendation scenarios within e-commerce software, achieving a seamless transition from search to recommendation.

[0106] This solution effectively solves the cold start problem, feature alignment problem, and resource consumption problem in multi-scenario recommendation by integrating technologies such as transfer learning, LHUC mechanism, global behavior recognition, multi-scenario sequence modeling, and global interest model. It improves the model's adaptability to different scenarios and recommendation accuracy, and realizes the efficient operation and precise service of the global recommendation system.

[0107] Figure 4 A block diagram of a computing device is shown according to an exemplary embodiment.

[0108] like Figure 4 As shown, computing device 30 includes processor 12 and memory 14. Computing device 30 may also include bus 22, network interface 16, and I / O interface 18. Processor 12, memory 14, network interface 16, and I / O interface 18 may communicate with each other via bus 22.

[0109] The processor 12 may include one or more general-purpose CPUs (Central Processing Units), microprocessors, or application-specific integrated circuits, etc., for executing relevant program instructions. According to some embodiments, the computing device 30 may also include a high-performance graphics card (GPU) 20 for accelerating the processor 12.

[0110] The memory 14 may include machine-readable media in the form of volatile memory, such as random access memory (RAM), read-only memory (ROM), and / or cache memory. The memory 14 is used to store one or more programs including instructions and data. The processor 12 may read the instructions stored in the memory 14 to execute the method according to the embodiment of the present invention described above.

[0111] The computing device 30 may also communicate with one or more networks via the network interface 16. The network interface 16 may be a wireless network interface.

[0112] The bus 22 may include an address bus, a data bus, a control bus, etc. The bus 22 provides a path for exchanging information between various components.

[0113] It should be noted that, in the specific implementation process, the computing device 30 may also include other components necessary for normal operation. In addition, those skilled in the art will understand that the above-mentioned device may also only include components necessary to implement the embodiments of this specification, and does not necessarily include all components shown in the figure.

[0114] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above method. The computer-readable storage medium may include, but is not limited to, any type of disk, including a floppy disk, an optical disk, a DVD, a CD-ROM, a microdrive and a magneto-optical disk, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory device, a magnetic or optical card, a nanosystem (including a molecular memory IC), a network storage device, a cloud storage device, or any type of medium or device suitable for storing instructions and / or data.

[0115] An embodiment of the present invention further provides a computer program product, which includes a computer program. The computer program is operable to enable a computer to execute part or all of the steps of any one of the methods described in the above method embodiments.

[0116] Those skilled in the art will readily appreciate that the technical solutions of the present invention can be implemented using software and / or hardware. "Unit" and "module" in this specification refer to software and / or hardware that can independently perform specific functions or work in conjunction with other components. Examples of hardware include field programmable gate arrays and integrated circuits.

[0117] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the present invention is not limited by the order of the actions described, because according to the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0118] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0119] In the several embodiments provided herein, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is merely a logical functional division. In actual implementation, other division methods may be used. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not implemented. In addition, the coupling or direct coupling or communication connection shown or discussed may be through some service interface. The indirect coupling or communication connection of devices or units may be electrical or other forms.

[0120] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0121] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0122] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. This computer software product is stored in a memory and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of the present invention.

[0123] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0124] The exemplary embodiments of the present invention are specifically shown and described above. It should be understood that the present invention is not limited to the detailed structure, configuration or implementation described herein; on the contrary, the present invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended clauses.

Claims

1. A method for e-commerce recommendation, characterized in that: The method comprises: Joint training of global multi-scenario recommendation models through sample mixing and weight mixing; Combine behavior sequence features to perform global behavior recognition and improve the global multi-scenario recommendation model; Identify the scene's signature features; Scenario-customized screening is performed according to the weights of the identification features, so that the global multi-scenario recommendation model can realize e-commerce recommendations in multiple scenarios.

2. The method according to claim 1, characterized in that Joint training of global multi-scenario recommendation models through sample mixing and weight mixing, including: The homepage recommendation samples and the global recommendation samples are coordinated and trained to achieve the sample mixing, wherein the homepage recommendation samples and the global recommendation samples share an embedding vector.

3. The method according to claim 2, characterized in that Joint training of global multi-scenario recommendation models through sample mixing and weight mixing also includes: Perform proportional fusion and training on samples from all scenes to generate a unified embedding vector, combined with the embedding vector of scene-specific features; The scene-specific features are added to the global multi-scene recommendation model through a scene-specific embedding layer and fused with the general features to improve the adaptability of the model to different scenes.

4. The method according to claim 3, characterized in that Joint training of global multi-scenario recommendation models through sample mixing and weight mixing also includes: The embedding vector and weight of the homepage recommendation model are imported into the global recommendation model through a learning rate warm-up operation to achieve the weight mixing.

5. The method according to claim 1, wherein Combined with behavioral sequence features, global behavior recognition is performed to improve the global multi-scenario recommendation model: Extracting user behavior data in different scenarios in the behavior sequence; Combining the user behavior data with relevant pages and context information of the different scenarios to construct the behavior sequence features at the scenario level; The user's global behavior in the different scenarios is subjected to page scenario-level attribution modeling, so that the global multi-scenario recommendation model can perceive the data of the global and all behaviors.

6. The method according to claim 5, characterized in that The global multi-scenario recommendation model upgrades the behavior sequence features from the granularity of a single product to the granularity of a scenario level.

7. The method according to claim 1, characterized in that Identify the scene's signature features, including: Use a gated unit structure or attention mechanism network to identify the scene's features and assign weights.

8. The method according to claim 1, characterized in that Also includes: For the different features in different scenarios, the model generalization is adjusted by setting the auxiliary task loss function.

9. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 8 when the computer program is executed by a processor.

10. A computing device, characterized in that include: processor; A memory storing a computer program, wherein when the computer program is executed by the processor, the method according to any one of claims 1 to 8 is implemented.

Images