Interaction sequence prediction method and feature enhancement model training method

Abstract

The invention discloses an interaction sequence prediction method and a feature enhancement model training method, and relates to the field of large model technology and prediction. The method comprises the steps that a plurality of historical interaction sequences formed by a user in a historical time period are obtained, and different historical interaction sequences are used for representing interaction sequences of the user in different dimensions; performing feature extraction on the plurality of historical interaction sequences by using a feature extraction model to obtain a plurality of historical interaction features; guiding a feature enhancement model to perform feature enhancement on the plurality of historical interaction features based on an association relationship among the plurality of historical interaction sequences to obtain a plurality of enhanced interaction features; and inputting the plurality of enhanced interaction features into a sequence prediction model, and predicting the plurality of historical interaction sequences by using the sequence prediction model to obtain predicted interaction sequences corresponding to the plurality of historical interaction sequences in a preset time period. The technical problem that the prediction accuracy of the interaction sequence is poor in the prior art is solved.

Description

Technical Field

[0001] This application relates to large model technology and data processing, specifically to a method for predicting interactive sequences and a method for training feature enhancement models. Background Technology

[0002] Currently, the widespread adoption of online services has made interaction sequence prediction a core element in improving the quality of recommendation systems and user experience. While related technologies utilize models that enhance prediction accuracy, they suffer from limitations in handling the randomness and noise of the sequences. Improper handling of randomness and noise in interaction sequences affects the robustness and accuracy of the models, resulting in poor prediction accuracy for interaction sequences.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This application provides a method for predicting interaction sequences and a method for training feature enhancement models, so as to at least solve the technical problem of poor prediction accuracy of interaction sequences in related technologies.

[0005] According to one aspect of the embodiments of this application, an interaction sequence prediction method is provided, comprising: acquiring multiple historical interaction sequences formed by a user in a historical time period, wherein different historical interaction sequences are used to represent user interaction sequences in different dimensions; extracting features from the multiple historical interaction sequences using a feature extraction model to obtain multiple historical interaction features; guiding a feature enhancement model to enhance the multiple historical interaction features based on the correlation between the multiple historical interaction sequences to obtain multiple enhanced interaction features; inputting the multiple enhanced interaction features into a sequence prediction model, and using the sequence prediction model to predict the multiple historical interaction sequences to obtain predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period.

[0006] According to one aspect of the embodiments of this application, a training method for a feature enhancement model is provided, comprising: acquiring multiple sample interaction sequences; performing sequence enhancement on the multiple sample interaction sequences to obtain sample enhancement interaction sequences corresponding to the multiple sample interaction sequences; constructing a sample contrast loss based on the multiple sample interaction sequences and the sample enhancement interaction sequences; and training an initial feature extraction model using the sample contrast loss to obtain a feature extraction model.

[0007] According to one aspect of the embodiments of this application, an interaction sequence prediction method is provided, comprising: acquiring a product interaction sequence and a scene interaction sequence formed by a user in a historical time period, wherein the product interaction sequence is used to represent an interaction sequence associated with a product, and the scene interaction sequence is used to represent an interaction sequence associated with a scene; extracting features from the product interaction sequence and the scene interaction sequence using a feature extraction model to obtain product interaction features and scene interaction features; and guiding a feature enhancement model to enhance the product interaction features and scene interaction features based on the correlation between the product interaction features and the scene interaction features to obtain enhanced product interaction features and enhanced scene interaction features.

[0008] According to one aspect of the embodiments of this application, an interaction sequence prediction method is provided, comprising: inputting product enhanced interaction features and scene enhanced interaction features into a sequence prediction model, using the sequence prediction model to predict product interaction sequences and scene interaction sequences, and obtaining predicted product interaction sequences corresponding to product interaction sequences and predicted scene interaction sequences corresponding to scene interaction sequences within a preset time period.

[0009] According to one aspect of the embodiments of this application, an interaction sequence prediction method is provided, comprising: obtaining multiple historical interaction sequences formed by a user in a historical time period by calling a first interface, wherein the first interface includes a first parameter, the parameter value of the first parameter includes multiple historical interaction sequences, and different historical interaction sequences are used to represent user interaction sequences in different dimensions; extracting features from the multiple historical interaction sequences using a feature extraction model to obtain multiple historical interaction features; guiding a feature enhancement model to enhance the multiple historical interaction features based on the correlation between the multiple historical interaction sequences to obtain multiple enhanced interaction features; inputting the multiple enhanced interaction features into a sequence prediction model, and using the sequence prediction model to predict the multiple historical interaction sequences to obtain predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period; and outputting the predicted interaction sequences by calling a second interface, wherein the second interface includes a second parameter, the parameter value of the second parameter includes the predicted interaction sequences.

[0010] According to another aspect of the embodiments of this application, a computing device is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0011] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory storing an executable program; and a processor connected to the memory via a bus for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0012] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0013] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0014] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the methods in various embodiments of this application.

[0015] According to another aspect of the embodiments of this application, a computer program is also provided, which, when executed by a processor, implements the methods of the various embodiments of this application.

[0016] In this embodiment, multiple historical interaction sequences formed by the user over a historical time period are obtained, where different historical interaction sequences represent user interaction sequences in different dimensions. A feature extraction model is used to extract features from the multiple historical interaction sequences to obtain multiple historical interaction features. Based on the correlation between the multiple historical interaction sequences, a feature enhancement model is guided to enhance the multiple historical interaction features, resulting in multiple enhanced interaction features. These enhanced interaction features are then input into a sequence prediction model, which predicts the multiple historical interaction sequences to obtain predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period, thereby improving the accuracy of interaction sequence prediction. It is noteworthy that through sequence enhancement and correlation-based feature enhancement, the model can more comprehensively capture the dynamic changes in user interaction and the interactions between different dimensions. Sequence enhancement can enhance the model's robustness to the randomness and noise of user interaction, while feature enhancement utilizes the correlation between historical interaction sequences, enabling the model to not only understand user interaction in a single dimension but also capture the dynamic interactions of scene sequences and product sequences. This allows the sequence prediction model to more accurately predict interaction sequences within a preset time period, significantly improving prediction accuracy and thus solving the technical problem of poor prediction accuracy of interaction sequences in related technologies.

[0017] It is worth noting that the general description above and the detailed description that follow are merely for illustrative purposes and do not constitute a limitation on this application. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a schematic diagram illustrating an application scenario of an interactive sequence prediction method according to an embodiment of this application;

[0020] Figure 2 This is a flowchart of an interaction sequence prediction method according to an embodiment of this application;

[0021] Figure 3 This is a flowchart of an interactive sequence prediction structure according to an embodiment of this application;

[0022] Figure 4 This is a flowchart of a training method for a feature enhancement model according to an embodiment of this application;

[0023] Figure 5 This is a flowchart of an interactive sequence prediction method according to an embodiment of this application;

[0024] Figure 6 This is a flowchart of an interactive sequence prediction method according to an embodiment of this application;

[0025] Figure 7 This is a schematic diagram of an interactive sequence prediction device according to an embodiment of this application;

[0026] Figure 8 This is a schematic diagram of a training device for a feature enhancement module according to an embodiment of this application;

[0027] Figure 9 This is a schematic diagram of an interactive sequence prediction device according to an embodiment of this application;

[0028] Figure 10 This is a schematic diagram of an interactive sequence prediction device according to an embodiment of this application;

[0029] Figure 11 This is a structural block diagram of a computing device according to an embodiment of this application;

[0030] Figure 12 This is a structural block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0032] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] The technical solution provided in this application is mainly implemented using large-scale model technology. Here, "large-scale model" refers to a deep learning model with a massive number of parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of parameters. Large-scale models can also be called foundation models. They are pre-trained using large-scale unlabeled corpora to produce pre-trained models with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include Large Language Models (LLMs) and multi-modal pre-training models.

[0034] It's important to note that in practical applications, large models can be fine-tuned using a small number of samples after pre-training, allowing them to be applied to various tasks. For example, large models can be widely used in Natural Language Processing (NLP), computer vision, and speech processing. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and Image Generation, as well as NLP tasks such as text-based sentiment classification, text summarization, and machine translation. Therefore, the main application scenarios for large models include, but are not limited to, digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0035] First, some nouns or terms that appear in the description of the embodiments of this application shall be interpreted as follows:

[0036] In search and recommendation systems, full-domain data refers to interactive data that integrates multiple different functional domains, such as recommendation systems, text-to-product search (T2PS), and image-to-product search (I2PS), forming a data set that comprehensively reflects the user's interaction patterns across functional domains.

[0037] A scenario refers to a specific environment or module in e-commerce applications where user interaction occurs, including but not limited to text search modules, image search modules, homepage recommendation modules, live streaming modules, etc. These modules not only reflect the diverse shopping needs of users, but also reflect the level of user participation and interaction patterns in different environments.

[0038] Sequence models are a type of neural network model that takes a series of ordered information, such as an interactive sequence, as input and generates embedding vectors to capture dependencies and temporal dynamics within the sequence.

[0039] A multilayer perceptron is a multilayer feedforward neural network consisting of an input layer, one or more hidden layers, and an output layer. It can handle complex nonlinear relationships and is the foundation for building deep learning models.

[0040] An interaction sequence is a series of interactions performed by a user in chronological order on a specific application or website, such as browsing products, adding items to the shopping cart, and making a purchase. These sequences can reveal the user's interaction patterns and changes in interests.

[0041] User representation is a high-dimensional vector extracted from interaction sequences. It contains rich information about user interests and can help the system understand users' personalized needs, thereby achieving more accurate recommendations and search results.

[0042] Sequence augmentation is a data augmentation technique that uses methods such as random masking or random permutation to construct sequences that are similar to but slightly different from the original sequence, thereby enhancing the robustness and generalization ability of the model. Random masking involves randomly replacing a portion of the data in the sequence with mask characters, while random permutation shuffles the order of elements in the sequence.

[0043] Contrastive learning is a self-supervised learning method that constructs positive samples similar to the original sequence and naturally dissimilar negative samples. This makes similar samples more closely packed in the feature space and dissimilar samples more distant, thereby improving the model's representation learning ability.

[0044] Conditional sampling is the process of collecting or selecting data under specific conditions or constraints. This method can select relevant samples according to needs to train models, thereby improving learning efficiency and prediction accuracy.

[0045] This application can effectively capture the interdependencies between scene sequences and item sequences to accurately predict future user interactions. During the learning process, this application can effectively model the relationship between scenes and items, providing strong support for interaction prediction. The network provided in this application can better model user-wide data and learn a broader range of user representations.

[0046] According to an embodiment of this application, an interactive sequence prediction method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0047] Considering the large number of model parameters in large models and the limited computing resources of mobile terminals, the method provided in this application embodiment can be applied to, for example, Figure 1 The application scenarios shown are not limited to these. Figure 1 This is a schematic diagram illustrating an application scenario of an interactive sequence prediction method according to an embodiment of this application. Figure 1In the application scenario shown, the large model is deployed on server 10. Server 10 can connect to one or more client devices 20 via a local area network (LAN), wide area network (WAN), internet connection, or other types of data network. These client devices 20 may include, but are not limited to, smartphones, tablets, laptops, PDAs, personal computers, smart home devices, and in-vehicle devices. Client devices 20 can interact with users through a graphical user interface to access the large model, thereby implementing the method provided in this embodiment.

[0048] In this embodiment, the system consisting of a client device and a server can perform the following steps: The client device stores multiple historical interaction sequences formed by the user within a historical time period. The server acquires the multiple historical interaction sequences formed by the user within a historical time period; it uses a feature extraction model to extract features from the multiple historical interaction sequences to obtain multiple historical interaction features; based on the correlation between the multiple historical interaction sequences, it guides a feature enhancement model to enhance the features of the multiple historical interaction features to obtain multiple enhanced interaction features; it inputs the multiple enhanced interaction features into a sequence prediction model, and uses the sequence prediction model to predict the multiple historical interaction sequences to obtain predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period.

[0049] It should be noted that with the rapid development of high-performance computing units, the methods provided in this application embodiment can also be applied to model-in-machine systems in other application scenarios. In one optional embodiment, the model-in-machine system has multiple built-in models, and users can select one model to adjust as needed to obtain their own model. The high-performance computing unit built into the model-in-machine system can then directly call the adjusted model to execute the methods provided in this application embodiment. In another optional embodiment, the large model-in-machine system has a pre-trained model built-in, and the high-performance computing unit built into the model-in-machine system can then directly call that model to execute the methods provided in this application embodiment.

[0050] Furthermore, when users need to train their own models, they can upload their own datasets via the client. These datasets are then sent to the server, allowing the server to adjust the pre-trained model using the dataset to obtain the user's customized model, which can then be deployed to the production environment. To facilitate users' model adjustment needs, the server provides complete adjustment tools, development frameworks, and processes, supporting multiple adjustment strategies. This allows the adjusted model to better adapt to different application domains and achieve a high degree of customization.

[0051] Under the aforementioned operating environment, this application provides the following: Figure 2 The interactive sequence prediction method shown. Figure 2This is a flowchart of an interaction sequence prediction method according to an embodiment of this application. For example... Figure 2 As shown, the method may include the following steps:

[0052] Step S202: Obtain multiple historical interaction sequences formed by the user within a historical time period;

[0053] Among them, different historical interaction sequences are used to represent the user's interaction sequences in different dimensions.

[0054] The aforementioned historical interaction sequences can be a series of user interactions within a specific historical time window, including but not limited to scene interactions and product interactions, constituting time-series data of user interactions. These sequences can be records of user operations such as searching, browsing, and purchasing on e-commerce platforms. The interaction sequence itself can be an interaction sequence in which the user interacts with a product.

[0055] The aforementioned interaction sequences across different dimensions can be categorized as scenario-based or product-based. Scenario-based interaction sequences can be understood as user interaction sequences in different scenarios, such as search scenarios, recommendation scenarios, and live streaming scenarios. Product-based interaction sequences, on the other hand, focus on user interaction sequences with specific products, such as browsing history for beverages, shoes, and electronic products.

[0056] The above steps can be performed on user interaction data from e-commerce platforms, aiming to extract the user's interaction history in different scenarios and for different products from the overall user data, providing basic data for subsequent interaction prediction and model training.

[0057] In one alternative embodiment, historical time periods can be acquired. First, a suitable historical time window is defined, such as the past 30 days, a week, or a day, depending on the accuracy requirements of interaction prediction and data availability; this is not limited here. The length of the window should be long enough to capture long-term patterns of user interaction, but also avoid being too long and causing the interaction patterns to become outdated. Within the defined time window, all user interaction data is collected, including but not limited to searches, browsing, purchases, favorites, and likes. For these interactions, multiple historical interaction sequences are formed based on the scenario in which the interaction occurs (e.g., text search scenario, image search scenario, homepage recommendation scenario, live streaming scenario, etc.) and the product categories involved (e.g., beverages, cups, headphones, pants, clothes, shoes, etc.). For example, from the user's interaction data, a scenario sequence and a product sequence can be constructed separately. The scenario sequence records the order of user interactions in different scenarios, while the product sequence records the order of user interactions with different products. For example, the interactive sequence can be a scene sequence (S): Search → Recommendation → Live Stream → Search → Homepage Recommendation; or a product sequence (T): Sneakers → Sports Watch → Sneakers → Sports Apparel → Sneakers. This is just an example and is not limited to these examples.

[0058] In one alternative embodiment, the interaction sequences of different dimensions can be "scenario sequences" and "product sequences." Scenario sequences reflect user interaction patterns in different scenarios, while product sequences focus on the user's interaction history with specific products. By constructing these two sequences separately, a more comprehensive understanding of user interaction habits and interests can be achieved, thereby improving the accuracy of predictions.

[0059] In e-commerce applications, understanding users' past interaction patterns is crucial for predicting future shopping interests and improving the accuracy and personalization of recommendation systems. By extracting historical interaction sequences from comprehensive data, it is possible to uncover changes in user preferences across different times and scenarios. This directly contributes to building smarter recommendation algorithms and enhancing user experience.

[0060] Step S204: Use the feature extraction model to extract features from multiple historical interaction sequences to obtain multiple historical interaction features;

[0061] The feature extraction model described above is a machine learning model used to extract user interaction features from historical interaction sequences. The feature extraction model can be designed as a sequence model, but is not limited to this; it can be used to capture dependencies and contextual information within the sequence.

[0062] Feature vectors extracted from historical interaction sequences are high-dimensional and contain information such as user interaction patterns, interests, and scene context. The purpose of extracting these features is for subsequent feature enhancement and interaction prediction.

[0063] The steps described above are used to convert multiple user interaction sequences over different historical time periods into feature vectors that machine learning algorithms can understand and manipulate, enabling subsequent deep learning models to learn and predict more efficiently based on these features. This process involves the application of feature extraction models, which can capture and transform the inherent patterns and dynamic changes of user interactions from the interaction sequences, forming a preliminary construction of a representation of user interactions.

[0064] Optionally, the acquired historical interaction sequences can be cleaned and formatted to ensure the integrity and consistency of each sequence data. At the same time, the interactions, scenarios, and products in the sequences can be encoded so that machine learning models can recognize and process them.

[0065] Multiple historical interaction sequences can be taken as input, and a feature extraction model can be used to calculate and transform them, outputting corresponding historical interaction features. These features are typically represented in vector form, comprehensively reflecting the characteristics of each element in the sequence and the dependencies between them. Optionally, the extracted historical interaction features can be dimensionalized by adding additional information such as timestamps and interaction types to increase their expressive power. Furthermore, considering the correlation between multiple sequences, feature aggregation or cross-referencing may be necessary to enhance the comprehensiveness and accuracy of the final features.

[0066] In one optional embodiment, the feature extraction model extracts features from the scene sequence and the product sequence separately, obtaining scene interaction features and product interaction features. The model encodes each element (scene or product) in each historical interaction sequence. The encoded result for each element includes an embedding vector that comprehensively reflects the element's position in the sequence, its association with other elements, and the sequence's contextual information. Through this encoding, the original interaction sequence is transformed into a set of historical interaction features. These features not only include the element's own attributes but also implicitly contain the dynamic changes in user interaction and the contextual influence of the scene.

[0067] For example, for the scenario sequence "search → recommendation → live streaming", the feature extraction model might generate a set of vectors, each vector corresponding to a scenario in the sequence, reflecting the user's activity level, preference changes, and transition patterns between scenarios within that scenario. For the product sequence "sports shoes → sports watch → sportswear", the model similarly generates a set of vectors, each vector corresponding to a product, reflecting the user's interaction history with that product, depth of interest, and correlation between product categories.

[0068] The above steps ensure that feature enhancement and sequence prediction are based on structured and information-rich historical interaction features, thereby improving the efficiency and accuracy of the entire framework in learning user dynamic interests and predicting interactions. By using an efficient feature extraction model, which captures high-order dependencies in the interaction sequence, we can understand user interaction patterns and product preferences in different scenarios, thus achieving accurate interaction prediction and personalized recommendations.

[0069] Step S206: Based on the correlation between multiple historical interaction sequences, guide the feature enhancement model to enhance the features of multiple historical interaction features to obtain multiple enhanced interaction features;

[0070] The aforementioned feature enhancement refers to further processing historical interaction features to improve their expressive power, enabling them to more comprehensively and accurately reflect users' dynamic interests and interaction patterns. Feature enhancement can be achieved by capturing the interdependencies between scene and product sequences through dual-sequence coupling modeling. The feature enhancement model aims to enhance the features extracted from historical interaction sequences by capturing the correlations between these sequences, thereby generating a more comprehensive and accurate representation of user interactions.

[0071] Feature enhancement is achieved through dual-sequence coupling modeling, primarily used to capture the interdependence between scene and product sequences. A single feature extraction model independently encodes both the scene and product sequences. The goal of this step is to extract preliminary features from the scene and product sequences, namely scene interaction features and product interaction features. These features reflect the dynamics and information of their respective sequences. Next, feature enhancement is performed using a feature enhancement model. Specifically, feature enhancement can be achieved through bidirectional input and an attention mechanism.

[0072] In this approach, bidirectional input can involve incorporating scene interaction features into the encoding network of the product sequence, while simultaneously incorporating product interaction features into the encoding network of the scene sequence. This method allows the model to consider scene information while encoding the product sequence, and vice versa, thereby capturing the interaction between the scene and the product.

[0073] Attention mechanisms allow models to directly focus on the relationships between scene sequences and product sequences during the encoding process, which helps the model learn the mutually reinforcing relationships between scenes and products.

[0074] After encoding and feature fusion, the enhanced features of both the scene sequence and the product sequence are used to train the sequence prediction model. This means that when predicting future interactions of a product sequence, the model also considers the dynamics of the scene sequence, and vice versa. Through this joint training approach, the model can more comprehensively understand the complexity of user interactions and improve prediction accuracy.

[0075] Dual-sequence coupling modeling can independently encode scene sequences and product sequences, and can also capture the interdependence between them through cross-feature fusion and joint training, thereby achieving feature enhancement. These enhanced features can more accurately reflect user interaction patterns and product preferences in different scenarios, providing a richer and more accurate information foundation for subsequent interaction prediction and recommendation systems.

[0076] In one alternative embodiment, the feature enhancement model enhances the features of each element by analyzing the relationships between sequences across different dimensions. For example, after a user purchases a pair of sneakers in a live-streaming scenario, they might become interested in sportswear in a recommended scenario. The feature enhancement model leverages this connection between the scenario and the product, fusing features from the live-streaming purchase interaction into the features of the recommended scenario, thereby enhancing its ability to predict potential future user interactions.

[0077] Furthermore, Conditional Contrastive Regularization Loss (CCR) can be used to enhance the feature enhancement process. Through CCR loss, the model can learn how to more effectively compare different enhanced sequences with the original sequence, thereby guiding the feature enhancement model to more accurately capture the dependencies between sequences and weaken the potential association between sequences and negative samples, thus improving representation quality. Under the influence of the feature enhancement model, historical interaction features are integrated and strengthened, forming multiple enhanced interaction features. These features are no longer limited to a single sequence view but incorporate the interactions of multi-dimensional sequences, including the symbiotic relationship between scene and product sequences, as well as the randomness and noise handling of user interactions.

[0078] For example, when a user frequently searches for "sneakers" in a text search scenario but interacts less in an image search scenario, it may indicate a high reliance on text descriptions. The feature enhancement model analyzes these interactions and discovers that a user's browsing and purchase of sneakers in a homepage recommendation scenario may be related to previous text search interactions. Therefore, the model enhances the text search interaction features with the recommendation scenario interaction features, generating an enhanced feature that integrates the user's interest in sneakers and their reliance on text descriptions. This enhanced feature will provide richer information and more accurate predictions when forecasting future user interactions in image search scenarios.

[0079] Furthermore, the feature enhancement model also considers the randomness and noise of user interactions, using CCR loss to improve the self-supervised learning process, making the enhanced features more robust and accurate in representing user interests. For example, even if a user browses sneakers in a homepage recommendation scenario but doesn't purchase them, and then makes a purchase later in a live stream scenario, the feature enhancement model can accurately identify and strengthen the impact of this scenario transition on the product purchase interaction through conditional sampling, thereby generating a more accurate predicted interaction sequence.

[0080] Guided by the feature enhancement model, multiple enhanced interaction features were generated based on the correlation between historical interaction sequences. These features are more comprehensive and accurate in capturing users' dynamic interests and predicting future interactions, effectively improving the model's predictive performance, especially in dealing with multi-dimensional interaction data and randomness and noise issues.

[0081] Step S208: Input multiple enhanced interaction features into the sequence prediction model, and use the sequence prediction model to predict multiple historical interaction sequences to obtain the predicted interaction sequences corresponding to multiple historical interaction sequences within a preset time period.

[0082] The sequence prediction model described above can be used to predict future user interaction sequences based on enhanced interaction features. This model effectively handles the randomness and noise issues of sequence data, improving prediction accuracy.

[0083] The aforementioned sequence prediction model predicts user interaction sequences within a preset time period based on enhanced interaction features of the input. This includes predicting scenarios the user might visit and products they might interact with. The model considers the randomness and noise of user interactions, as well as the complex relationships between scenarios and products, to improve the robustness of the prediction. After completing the prediction, the model outputs the predicted interaction sequences within the preset time period. These sequences can be used by recommendation systems or other downstream applications to provide more personalized and accurate services.

[0084] In an optional embodiment, two sequence prediction networks—a scene sequence prediction network and a product sequence prediction network—can be used in dual-sequence coupling modeling to predict the scene sequence and product sequence, respectively. Enhanced interaction features, processed by a feature enhancement model (i.e., feature vectors incorporating scene and product information), are input into the sequence prediction model. These feature vectors are further optimized through conditional sampling contrastive regularization (CCR loss) to enhance the model's representation invariance and robustness. The model performs predictions and outputs predicted interaction sequences for a preset future time period. This model, in its prediction, not only considers the randomness and noise of user interactions but also improves prediction robustness by predicting the interaction distribution over a period of time.

[0085] During training, methods such as multi-label binary classification and cross-entropy loss can be used. In the prediction phase, metrics can be used to compare the model's predictions with actual user interactions to assess prediction accuracy. Simultaneously, the contrastive learning optimization objective included in the loss function further optimizes the model's predictive ability by adjusting the sensitivity of similarity calculation. The model will output possible scene sequences and product sequences for users within a preset time period. These predictions can be used in recommendation systems to achieve personalized recommendations and improve user experience.

[0086] By combining dual-sequence coupling modeling and conditional sampling contrastive regularization (CCR loss), this application not only predicts user interaction sequences based on enhanced features but also handles noise and randomness in interaction data, improving the accuracy and reliability of predictions. This enhanced predictive capability is of significant value for optimizing recommendation algorithms on e-commerce platforms, and can significantly improve user satisfaction and platform conversion rates.

[0087] The above steps describe a deep learning-based user interaction prediction framework. Its core lies in extracting features from historical user interaction data, enhancing these features, and then predicting potential future user interactions. This framework comprises four main steps, which will be explained in detail below.

[0088] In one optional embodiment, multiple historical interaction sequences can be collected and organized from user interaction records across different historical time periods. Each sequence represents the user's interaction trajectory in a specific dimension or scenario, such as a search interaction sequence, a shopping cart interaction sequence, and a purchase interaction sequence. These sequences contain the user's past interaction records in various scenarios, such as products viewed, keywords searched, items added to the shopping cart, and items purchased. Interaction sequences across different dimensions can comprehensively reflect the user's interaction patterns in different scenarios, providing a foundation for subsequent feature extraction and interaction prediction.

[0089] The feature extraction model is used to perform deep learning on the multiple historical interaction sequences collected in the previous step to extract representative features for each sequence. These features, namely historical interaction features, contain the inherent representation of user interaction patterns in various scenarios and are an important quantitative reflection of user interaction. For each historical interaction sequence, the feature extraction model generates a high-dimensional vector. This vector integrates the temporal order of each interaction in the sequence, product information, scenario information, etc., thus forming an abstract representation of the user interaction pattern.

[0090] By analyzing the correlations and interdependencies between interaction sequences across different dimensions, these features are integrated and enhanced to capture deeper interaction patterns and dynamic changes. This application allows the encoded features of one sequence (such as scene sequence features) to be used as input to another sequence (such as a product sequence), thereby considering the dynamic changes of both the scene and the product when predicting future interactions. This step enhances the expressive power of the features through the interaction between sequences, further improving the predictive performance of the model.

[0091] Finally, the enhanced historical interaction features are input into the sequence prediction model to predict future interactions. The prediction model uses these features to predict the user's interaction sequence within a preset future time period, including possible scenarios and products. The predicted output sequence provides important information for the recommendation system, helping it understand user interest trends and shopping habits, thereby providing more personalized and accurate product recommendations. Through these steps, the entire framework can not only independently model user interactions in different scenarios, but also capture dynamic changes in user interactions and predict possible future interactions through feature enhancement and sequence prediction, providing strong support for recommendation systems and information retrieval services.

[0092] By leveraging deep learning models to capture the complexity and dynamism of interaction sequences, and further enhancing the model's predictive capabilities through feature augmentation, we can provide users with a more personalized and accurate experience in recommendation systems and information retrieval services. The combination of dual-sequence coupling modeling and conditional sampling contrastive regularization learning enables more accurate predictions of future interactions, demonstrating superior performance even in environments with uneven data distribution and significant long-tail effects.

[0093] For example, multiple historical interaction sequences formed by a user over a historical period can be obtained. For instance, all search and purchase interaction records of a user within a week can be retrieved from the platform; these records constitute the user's historical interaction sequences in terms of scenario and product dimensions. Feature extraction models are used to extract features from the scenario sequences and product sequences respectively, obtaining scenario interaction features and product interaction features. For example, the model might extract user interest features for outdoor activities from search and live streaming scenarios, and user preference information for outdoor shoes from the product sequences. Based on the correlation between the scenario sequences and product sequences, feature enhancement models are used to enhance features. For example, the model might consider a user purchasing outdoor shoes in a live streaming scenario to enhance the interaction features of searching for outdoor shoes in the scenario sequence. Simultaneously, based on the user's purchase interaction, the features of outdoor shoes in the product sequence are enhanced to better reflect the user's actual needs. The enhanced scenario interaction features and product interaction features are then input into a sequence prediction model, which is used to predict the user's future interactions. For example, predicting which products a user might be interested in in the next few days, and the possible interaction sequences.

[0094] Overall, this application enhances the model's ability to handle randomness and noise in interaction sequences by deeply exploring the interdependence between scene sequences and product sequences during feature extraction and enhancement stages. This improves the accuracy of prediction sequences during the prediction stage and solves the problem of low prediction accuracy in related technologies. For example, on e-commerce platforms, the system can more accurately predict users' potential future scene transitions (such as switching from search to live streaming) and product interactions (such as possible purchases of other outdoor equipment), providing users with more personalized and accurate recommendation services.

[0095] Through the above steps, multiple historical interaction sequences formed by users over a historical time period are obtained. These different historical interaction sequences represent user interaction sequences across different dimensions. A feature extraction model is used to extract features from each of the multiple historical interaction sequences, resulting in multiple historical interaction features. Based on the correlation between these sequences, a feature enhancement model is guided to enhance these historical interaction features, resulting in multiple enhanced interaction features. These enhanced features are then input into a sequence prediction model, which predicts the historical interaction sequences within a preset time period, thus improving the accuracy of interaction sequence prediction. It is noteworthy that through sequence enhancement and correlation-based feature enhancement, the model can more comprehensively capture the dynamic changes in user interactions and the interactions between different dimensions. Sequence enhancement enhances the model's robustness to the randomness and noise of user interactions, while feature enhancement utilizes the correlation between historical interaction sequences, enabling the model to understand not only user interactions in a single dimension but also the dynamic interactions of scene sequences and product sequences. This allows the sequence prediction model to more accurately predict interaction sequences within a preset time period, significantly improving prediction accuracy and thus solving the technical problem of poor prediction accuracy for interaction sequences in related technologies.

[0096] In the above embodiments of this application, a feature enhancement model is guided to enhance multiple historical interaction features based on the correlation between multiple historical interaction sequences to obtain multiple enhanced interaction features. This includes: determining a second historical interaction sequence corresponding to a first historical interaction sequence based on the correlation between multiple historical interaction sequences, wherein the second historical interaction sequence is at least one historical interaction sequence among the multiple historical interaction sequences that has a correlation with the first historical interaction sequence; and enhancing multiple historical interaction features based on the first historical interaction sequence and the second historical interaction sequence to obtain multiple enhanced interaction features.

[0097] The first historical interaction sequence mentioned above can be a product sequence, and the second historical interaction sequence can be a scenario sequence. It should be noted that products interacted with through search, recommendations, or live streaming are all product sequences, while scenario sequences represent scenarios such as search, recommendation, and live streaming.

[0098] The aforementioned relationships are used to represent the interdependence or influence between different historical interaction sequences. These relationships could be changes in a user's interest in a product under different scenarios, or correlations in user interaction patterns over different time periods.

[0099] The aforementioned feature enhancement model can be used to enhance historical interaction features, making them more comprehensive and accurate in reflecting users' dynamic interests and interaction patterns. The feature enhancement model generates enhanced interaction features by analyzing the correlations between sequences and integrating scene sequence features and product sequence features.

[0100] For example, in the field of e-commerce search and recommendation, a user's interactions within a historical time period might include browsing "sports shoes" in a "search" scenario, then watching a sports equipment live stream and purchasing a pair of "sports shoes" in a "live stream" scenario. A user might also browse "sports apparel" in a "recommendation" scenario. These interactions between scenarios and products constitute a sequence of user interactions.

[0101] In one optional embodiment, a "second historical interaction sequence" corresponding to the "first historical interaction sequence" can be determined based on the association relationship. For example, the sequence of products interacted by the user in various scenarios can be determined as the first historical interaction sequence, and the sequence of scenarios (such as live streaming, search, recommendation, etc.) in which the associated interactions occur can be determined as the second historical interaction sequence. Then, based on the first and second historical interaction sequences, the scenario sequence features and product sequence features are integrated and enhanced. For example, the feature enhancement model will analyze how the user's interest in "sneakers" in the search scenario affects their purchase decision in the live streaming scenario, and how the user's interaction in the live streaming scenario, in turn, affects their interaction in the search and recommendation scenarios.

[0102] Through feature enhancement based on association relationships, the model not only independently learns the features of each historical interaction sequence but also captures the interdependencies and influences between different sequences, achieving feature integration and enhancement. This mechanism enables the model to more comprehensively understand user interaction patterns. When processing multi-dimensional interaction sequences, the model can learn the dynamic interaction relationships between scenarios and products, improving the accuracy of predicting future interaction sequences. For example, in e-commerce recommendation scenarios, the model can more accurately predict users' possible future scenario switching and product preferences, thereby providing more personalized and accurate recommendation services.

[0103] For example, user A had the following interactions within a week: on Monday, they browsed sneakers in a search context; on Wednesday, they watched a sports equipment livestream and purchased a pair of sneakers; and on Friday, they browsed several sportswear items in a recommendation context. Based on these interactions, the model can construct historical interaction sequences for the user across both context and product dimensions. Then, based on the relationships between these sequences—such as the dependence of sneaker purchases on livestreams and recommendation contexts—the model guides the feature enhancement model to integrate and enhance the context sequence features and product sequence features. Finally, the enhanced interaction features are fed into a sequence prediction model to predict user A's possible interaction sequences in the following days, such as potentially re-entering a livestream to view sports equipment or searching for other sneaker-related products in a search context. Through this feature enhancement and prediction based on relationships, the model can more accurately understand and predict user interactions, improving the accuracy and personalization of recommendations.

[0104] In the above embodiments of this application, based on a first historical interaction sequence and a second historical interaction sequence, feature enhancement is performed on multiple historical interaction features to obtain multiple enhanced interaction features, including: determining a first historical interaction feature corresponding to the first historical interaction sequence from multiple historical interaction features, and determining a second historical interaction feature corresponding to the second historical interaction sequence; performing feature enhancement on the first historical interaction feature based on the second historical interaction feature to obtain an enhanced interaction feature corresponding to the first historical interaction feature; performing feature enhancement on the second historical interaction feature based on the first historical interaction feature to obtain an enhanced interaction feature corresponding to the second historical interaction feature; and obtaining multiple enhanced interaction features based on the enhanced interaction feature corresponding to the first historical interaction feature and the enhanced interaction feature corresponding to the second historical interaction feature.

[0105] The aforementioned historical interaction features refer to feature representations extracted from different historical interaction sequences of a user. Each historical interaction sequence, after being processed by the feature extraction model, generates a high-dimensional vector, i.e., a historical interaction feature, which contains comprehensive information about the user's interactions in the sequence.

[0106] The aforementioned feature enhancement is a process that improves the representational power of original features through information exchange and interaction between sequences. The feature enhancement mechanism in this application allows the encoded features of one sequence to be used as input to another sequence, thereby capturing the interdependencies between sequences and enhancing the model's ability to predict future interactions.

[0107] In one optional embodiment, from multiple historical interaction features, a first historical interaction feature corresponding to a first historical interaction sequence and a second historical interaction feature corresponding to a second historical interaction sequence are explicitly distinguished and extracted. The second historical interaction feature is used to enrich and strengthen the first historical interaction feature, generating enhanced interaction features; simultaneously, the first historical interaction feature is used to enhance the second historical interaction feature, obtaining corresponding enhanced interaction features. The enhanced interaction features of the first and second historical interaction features obtained in the above two steps are merged to form a comprehensive enhanced interaction feature set containing enhanced information of scene sequences and product sequences. This process, through bidirectional feature enhancement and comprehensive feature fusion, effectively improves the model's predictive ability for user interactions, especially when dealing with complex dependencies and randomness between sequences, generating more robust and comprehensive user interaction representations, thereby improving prediction accuracy and personalized recommendation effects.

[0108] The feature enhancement process in this application can be achieved through dual-sequence coupling modeling. First, first historical interaction features and second historical interaction features are extracted from multiple historical interaction sequences, respectively. Then, the first historical interaction features are enhanced based on the second historical interaction features, and vice versa. This bidirectional feature enhancement allows the model to simultaneously consider the dynamic changes of the scenario and the product, thus making the prediction of future user interactions more accurate and comprehensive.

[0109] Through feature enhancement, this application can more accurately capture the interdependencies between scenarios and products in the user's overall data. Specifically, feature enhancement not only considers the user's interaction patterns in a single scenario but also the mutual influence between different scenarios and the complex relationship between product information and scenario information. For example, in e-commerce applications, if a user searches for "sports shoes" in a search scenario and views recommendations for sports equipment in the homepage recommendation scenario on the same day, feature enhancement allows the model to understand the potential correlation between these two interactions. That is, the user's interest in sports equipment may also reflect a potential demand for sports shoes. Therefore, when predicting future user interactions, the model will more accurately consider the user's potential interest in sports shoes or related sports equipment in multiple scenarios, thereby improving the personalization and accuracy of recommendations.

[0110] In the above embodiments of this application, the method further includes: performing sequence enhancement on multiple historical interaction sequences to obtain enhanced interaction sequences corresponding to multiple historical interaction sequences; constructing a target contrast loss based on multiple historical interaction sequences and enhanced interaction sequences; and updating the model parameters of the feature extraction model using the target contrast loss.

[0111] The sequence enhancement described above can be achieved through random masking and random rearrangement. Random masking involves randomly selecting some elements in the sequence and replacing them with specific markers, while random rearrangement involves randomly rearranging a portion of a continuous subsequence while leaving the rest unchanged.

[0112] The sequences generated by sequence augmentation techniques are similar to the user's original historical interaction sequences in some aspects, but differ in details, and are used to train models to learn more robust representations.

[0113] In contrastive learning, the model is guided to learn by calculating the similarity between the original sequence and the enhanced sequence, as well as the difference between the original sequence and other sequences. This enables the model to distinguish between similar and dissimilar sequences, thereby learning more discriminative feature representations.

[0114] Sequence augmentation generates enhanced interaction sequences that are similar to but contain minor differences from the user's historical interaction sequences. These enhanced sequences, along with multiple original historical interaction sequences, are then used to construct a target contrastive loss to update the parameters of the feature extraction model. This allows the model to learn more robust and discriminative feature representations. Specifically, sequence augmentation constructs positive samples (enhanced sequences and original sequences) and negative samples (original sequences and other sequences or perturbed sequences) through random masking and random rearrangement. Then, using the target contrastive loss in the contrastive learning framework, the model focuses on learning the similarity of positive samples and the differences of negative samples. Through this process, the model's representation learning ability is enhanced.

[0115] Through sequence augmentation and contrastive learning, the model can learn the intrinsic structure and dynamic changes of interaction sequences, while exhibiting stronger robustness to noise and randomness in the data. Sequence augmentation creates diverse samples, allowing the model to cover a wider range of situations during the learning process, while targeted contrastive loss ensures that the model can distinguish between similar and dissimilar sequences, thereby learning more stable and richer feature representations during training. This is crucial for accurately predicting future user interactions, especially when faced with sparse and noisy interaction data, effectively improving the model's generalization ability and prediction accuracy.

[0116] For example, in an e-commerce scenario, a user might have multiple historical interaction sequences: searching for "sneakers" in a text search scenario, viewing images of sports equipment in an image search scenario, browsing sportswear in a homepage recommendation scenario, and watching a live stream of sports products in a live stream scenario. First, sequence augmentation techniques enhance these historical interaction sequences. For example, the sequences are randomly masked, perhaps replacing "sneakers" with a mask marker, or the order of some items in the sequence is rearranged. The resulting enhanced interaction sequences are similar to the original sequences in content, but with differences in specific details. Next, a target contrastive loss is constructed to calculate the similarity between the original and enhanced sequences, as well as the differences between the original sequence and other sequences, guiding the feature extraction model to learn more robust feature representations. The feature extraction model updates its parameters by optimizing the target contrastive loss, enabling it to more accurately capture users' interest in sneakers and sports equipment, even with randomness in the order and noise in the data.

[0117] In the above embodiments of this application, constructing a target contrast loss based on multiple historical interaction sequences and enhanced interaction sequences includes: constructing a first contrast loss based on the similarity between multiple historical interaction sequences and enhanced interaction sequences; constructing a second contrast loss based on the similarity between multiple historical interaction sequences; and determining the target contrast loss based on the first contrast loss and the second contrast loss.

[0118] The aforementioned targeted contrastive loss, within the contrastive learning framework, enhances the model's learning ability by quantifying and boosting the similarity between positive and negative samples. The targeted contrastive loss combines the first and second contrastive losses, aiming to learn robust representations from user historical interaction sequences and their enhanced variants.

[0119] The first contrastive loss mentioned above is a loss calculated based on the similarity between the original historical interaction sequence and the enhanced interaction sequence, which is used to help the model learn feature representations that are invariant to the enhancement operation.

[0120] The second contrastive loss mentioned above is a loss calculated by comparing the similarity between different historical interaction sequences. It is used to enhance the model's ability to distinguish between different user interactions and improve the accuracy of personalized recommendations.

[0121] In e-commerce recommendation systems, a user's historical interaction sequences may include diverse interaction records such as searching, browsing, and favorites. Enhanced interaction sequences are generated through random masking or rearrangement and used to train the model. The first contrastive loss ensures that even after random changes to the sequence, the model can still identify that it represents the same user's interaction pattern, thus improving the robustness of the representation. The second contrastive loss distinguishes users by calculating the similarity between different user sequences, ensuring personalized recommendation results.

[0122] In one alternative embodiment, starting with multiple original historical interaction sequences, the model generates enhanced interaction sequences through sequence augmentation. Next, the model constructs a first contrastive loss using the similarity between the enhanced interaction sequences and the original sequences to enhance feature invariance; simultaneously, a second contrastive loss is constructed by calculating the similarity between different historical interaction sequences, aiming to improve the discriminative power of the features. Finally, these two losses are combined to determine a target contrastive loss, guiding the model to learn a more complete and stable representation of user interactions.

[0123] By constructing a target contrastive loss, the system can significantly improve the performance of the recommendation system, especially when dealing with the randomness and noise of interaction sequences. The first contrastive loss enhances the similarity between the sequence and the original sequence, enabling the model to learn feature representations that are resistant to data transformations, thereby more accurately predicting user interactions. The second contrastive loss enhances the model's ability to personalize recommendations by distinguishing the historical interaction sequences of different users, improving user satisfaction and platform conversion rates. Combining the two losses, the system can not only handle complex interaction sequences but also provide more personalized and accurate recommendations.

[0124] For example, in an e-commerce platform scenario, a user's historical interaction sequence includes records of interactions with products in different scenarios (such as search, recommendation, and live streaming). The system generates enhanced sequences through random masking and rearrangement to construct the first contrastive loss, ensuring the model maintains predictive stability even with minor changes in the sequence. Simultaneously, a second contrastive loss calculates the similarity of interaction sequences among different users in the same scenario, enabling the model to distinguish users' personalized interaction patterns and provide more accurate recommendations for each user. Combining these two losses, the model trained by the system can more accurately understand users' overall interaction patterns, predict possible future scenario transitions and product interactions, thereby achieving high-quality personalized recommendations and improving user experience.

[0125] In the above embodiments of this application, sequence enhancement is performed on multiple historical interaction sequences to obtain enhanced interaction sequences corresponding to multiple historical interaction sequences, including: performing masking processing on multiple historical interaction sequences respectively to obtain masked interaction sequences and unmasked interaction sequences corresponding to multiple historical interaction sequences; rearranging the elements of the masked interaction sequences to obtain rearranged interaction sequences corresponding to multiple historical interaction sequences; and obtaining enhanced interaction sequences of multiple historical interaction sequences based on the unmasked interaction sequences and the rearranged interaction sequences.

[0126] The masking process described above is a technique in sequence augmentation. By randomly replacing or masking some elements in a sequence, a masked interaction sequence is created to simulate the incomplete observability of user interactions, helping the model learn stability when faced with missing data.

[0127] The element rearrangement mentioned above is another technique in sequence augmentation. By randomly changing the order of elements in a sequence, rearranged interactive sequences are created to simulate the randomness and uncertainty of user interactions and enhance the model's adaptability to changes in order.

[0128] In the context of e-commerce recommendation systems, a user's historical interaction sequence includes a series of purchase, browsing, and click interactions. Sequence augmentation allows the system to generate masked interaction sequences (randomly obscuring some product interaction information) and rearranged interaction sequences (changing the temporal order of user interactions with products). These augmented sequences, used in conjunction with the original sequences, help the model learn a more stable and comprehensive representation of user interests, maintaining predictive accuracy even with missing information or altered order.

[0129] In one optional embodiment, multiple historical interaction sequences are masked to generate masked and unmasked interaction sequences. Subsequently, the masked interaction sequences are rearranged element by element to obtain rearranged interaction sequences. Finally, using the unmasked and rearranged interaction sequences, the system constructs an enhanced interaction sequence. This sequence contains information from the original sequences and enhances the model's predictive ability through random perturbations. Through this series of sequence enhancement steps, the system can significantly improve the robustness and generalization ability of the recommendation model. Masking enables the model to make accurate predictions even with incomplete information, while element rearrangement helps the model understand the randomness of interactions, enhancing its adaptability to changes in time sequence. The resulting enhanced interaction sequence is used for model training, enabling the model to learn feature representations invariant to data transformations and perturbations. Therefore, when predicting future user interactions, even with incomplete or random data, it can maintain high prediction accuracy and personalized recommendation capabilities.

[0130] For example, in an e-commerce platform, user A's historical interaction sequence might include records of clicking, browsing, favoriting, and purchasing different products. The system first performs masking, randomly removing some interaction records to generate a masked interaction sequence. Then, it rearranges the elements of the masked interaction sequence to create a rearranged interaction sequence with a different order than the original sequence. Next, it combines the unmasked interaction sequence with the rearranged interaction sequence to form an enhanced interaction sequence. This sequence is used for model training, helping the model learn to accurately predict user interests even with missing information and changes in the interaction order. Ultimately, the trained model demonstrates higher accuracy and stability in user interaction prediction, providing user A with more personalized and accurate recommendation services, improving user satisfaction and platform conversion rates.

[0131] The aforementioned sequence augmentation method for multiple historical interaction sequences not only improves the model's ability to predict user interactions, but also enhances its ability to handle incomplete and random data.

[0132] In the above embodiments of this application, the method further includes: obtaining the actual interaction sequences corresponding to the plurality of historical interaction sequences; constructing a third contrastive loss based on the actual interaction sequences and the predicted interaction sequences; and updating the model parameters of the feature enhancement model based on the third contrastive loss.

[0133] The above-mentioned actual interaction sequence is the sequence of product or scenario interactions that actually occur to the user within a preset time period.

[0134] The predicted interaction sequence mentioned above is a product or scene interaction sequence predicted by the sequence prediction model based on product enhanced interaction features and scene enhanced interaction features.

[0135] The third contrastive loss mentioned above is a loss function that improves the feature enhancement model by comparing the similarity between the actual interaction sequence and the predicted interaction sequence, thereby ensuring the accuracy and reliability of the model's predictions.

[0136] In one alternative embodiment, it is necessary not only to collect historical interaction sequences of users in past time periods, but also to continuously track users' actual interactions in the current time period, including searching, browsing, and purchasing. These actual interaction sequences are compared with predicted interaction sequences to evaluate the model's predictive performance and update the model parameters. For example, if the model predicts that a user will switch to a live streaming scenario and browse a specific category of products, and the actual interaction sequence confirms this prediction, then the model's parameters will receive positive feedback; otherwise, they will be adjusted using a third contrastive loss to improve prediction accuracy.

[0137] After acquiring historical interaction sequences and enhanced interaction features obtained through the feature enhancement model, the system uses a sequence prediction model to generate predicted interaction sequences. Next, actual interaction sequences are collected from real user interactions, and a third contrastive loss is constructed by comparing the similarity between the predicted and actual sequences. This loss function serves as a feedback signal to update the parameters of the feature enhancement model, ensuring that the model can learn from actual interactions and continuously optimize its predictive capabilities.

[0138] By introducing a third contrastive loss and updating the parameters of the feature enhancement model, the system can continuously iterate and optimize the model's predictive accuracy. This process helps reduce the deviation between model predictions and actual user interactions, ensuring that the recommendation system can provide recommendations that better match the user's current interests and interaction patterns, thereby improving user satisfaction and platform conversion rates. In a global data context, this mechanism can also help the model better understand the complex relationship between scenarios and products, and capture dynamic changes in user interests.

[0139] For example, suppose the system collects user A's historical interaction sequences of goods and scenarios over a period of time, and then generates product-enhanced interaction features and scenario-enhanced interaction features using feature extraction and enhancement methods. Next, the system uses a sequence prediction model to predict the possible product and scenario interaction sequences of user A in the following preset time period. These predictions are compared with user A's actual interaction sequences during that time period to construct a third contrastive loss. If the predicted interaction sequences differ significantly from the actual sequences, the system will adjust the parameters of the feature enhancement model based on the third contrastive loss to improve the subsequent prediction quality. This process not only improves the model's prediction accuracy but also enhances the model's adaptability and robustness in the face of dynamic changes in user interests.

[0140] Through the above process, when processing user interaction prediction in the context of global data, this application can not only enhance the robustness of the model through sequence augmentation and contrastive learning, but also continuously optimize the feature augmentation model through feedback from actual interaction sequences, ensuring that its prediction results are highly consistent with the actual user interaction and improving the user experience.

[0141] This application designs a modeling method based on dual-sequence coupling modeling for scene information and product sequences, and proposes a conditional sampling contrastive learning method to address the randomness of interaction sequences. Figure 3 This is a flowchart of an interactive sequence prediction structure according to an embodiment of this application, such as... Figure 3 As shown, the system includes a scene history sequence and a product history sequence. A dual-sequence coupling network is used to predict the scene prediction sequence and the product prediction sequence. Conditional sampling contrastive regularization learning can be used to construct a conditional sampling first contrastive loss between the enhanced scene features and the original scene features output by the sequence model, and similarly, a conditional sampling first contrastive loss can be constructed between the enhanced product features and the original product features output by the sequence model. This improves the learning efficiency and accuracy of the sequence model in representing dynamically changing interests. A sequence feature enhancement module enhances the product and scene features, resulting in enhanced product and scene features. Prediction is then performed on these enhanced product and scene features to obtain the scene prediction sequence and the product prediction sequence. A prediction loss method can be used to improve the accuracy of the sequence model's predictions.

[0142] The aforementioned scenario history sequence is used to represent user interaction scenarios at different points in time, including but not limited to liking, favorites, shopping cart, sharing, and searching. The product history sequence is used to represent products that users interact with at different points in time, such as beverages, cups, headphones, pants, clothes, and shoes.

[0143] The aforementioned dual-sequence coupling network comprises a sequence enhancement module, a sequence model, and a sequence feature enhancement module. The sequence enhancement module creates enhanced sequences through random masks or reordering elements to increase the model's robustness, enabling it to handle randomness and noise within the sequences. The sequence model encodes the scene sequence and the product sequence using neural network structures or other sequence models, extracting their feature representations.

[0144] The sequence feature enhancement module described above can use the encoded features of one sequence as input to another sequence, allowing for the simultaneous consideration of dynamic changes in the scenario and the product, and capturing the interdependence between the two.

[0145] The conditional sampling in the above-mentioned conditional sampling contrastive regularization learning can be sampling through specific conditions to provide positive and negative samples for contrastive learning. The conditional contrastive regularization loss uses the learned conditional weights to enhance contrastive learning and improve the learning efficiency and accuracy of sequence models in representing dynamically changing interests.

[0146] The aforementioned prediction loss function is used to quantify the accuracy of the model in predicting the next stage scenario or product sequence. It calculates the cross-entropy loss based on user representation and product representation by comparing it with the set sample labels, in order to train the model.

[0147] Figure 3 The architecture shown processes historical sequence data to obtain feature representations of scenes and products, while enhancing the model's robustness through a sequence augmentation module. The sequence feature augmentation module combines information from scene and product sequences, then a prediction model predicts the next stage of interaction. Finally, prediction loss and conditional first contrastive loss (CCRLoss) are used for model training and optimization. This application aims to leverage dual-sequence coupling modeling to simultaneously learn the dynamic changes of scene and product sequences, and to handle randomness and noise in sequence data through conditional sampling contrastive regularization, thereby improving the model's accuracy and robustness in predicting future user interactions.

[0148] The dual-sequence coupling network proposed in this application focuses on capturing the interaction between scenes and items in the user's historical interaction sequence. This application demonstrates that the method can model the joint distribution of scenes and items by maximizing the lower bound of the joint log-likelihood. This analysis not only provides theoretical support for the model design, but also reveals that dual-sequence learning can effectively model the dynamic changes of user interactions and the complex relationship between scenes and items.

[0149] The aforementioned dual-sequence coupled network consists of two parallel networks that predict scene sequences and item sequences for a future time period, respectively. Each network uses a sequence model to extract features; this application can use a neural network structure as the sequence model for feature extraction. The two networks are interconnected through their respective feature enhancement modules to capture the interdependencies between sequences. The randomness and noise of user interactions are considered during the prediction process, and the robustness of the prediction is improved by predicting the interaction distribution over a period of time. Specifically, taking the item sequence as an example, this application sets the labels of items interacted with by the user in the next stage as positive samples and the rest as negative samples. This application determines the cross-computation score of items using the user representation sequence obtained from the model and the representation sequences of different items, and applies a multi-label binary classification cross-entropy loss to the score and the set sample labels. Similarly, this application sets the labels of scenes interacted with by the user in the next stage as positive samples and the rest as negative samples. This application determines the cross-computation score of scenes using the user representation sequence obtained from the model and the representation sequences of different scenes, and applies a multi-label binary classification cross-entropy loss to the score and the set sample labels.

[0150] The aforementioned sequence feature enhancement module can improve the predictive power of the sequence model by fusing the features of the scene and the product. The sequence model not only learns the historical sequences of the scene and the product independently, but also captures the mutual influence between the two through a feature enhancement mechanism. This enhancement is achieved by using the encoded features of one sequence as the input of another sequence, thereby allowing the sequence model to take into account the dynamic changes of the scene and the product when predicting the user's future interaction sequence.

[0151] Considering the randomness and noise typically present in online user interactions, dual-sequence coupled networks introduce Conditional Contrastive Regularization Loss (CCR) to enhance representation invariance, thereby improving the efficiency of learning dynamic interests. CCR loss uses learned conditional weights to more effectively promote similarity between sequences that have undergone different enhancement operations. Conditional sampling contrastive regularization learning specifically includes sequence enhancement and conditional weights.

[0152] Sequence augmentation creates augmented sequences by randomizing a portion of the input sequence and reordering the elements. The masking involves randomly masking a portion of the input sequence, while the reordering involves randomly selecting a size, uniformly selecting a continuous subsequence within that size and shuffling its elements, while the elements outside the subsequence remain in their original order.

[0153] Conditional weights, designed using a binary classification function (soft max function), focus on samples that are difficult to optimize, thus enabling more efficient representation learning. These weights reflect the different contributions of various samples in the optimization process. Conditional weights characterize the dependency between negative sample sampling and the original sequence, and innovatively investigate the dependency among multiple positive samples.

[0154] The loss function of conditional contrastive regularization learning improves the model by calculating the similarity between the original sequence and the augmented sequence, as well as the similarity between the original sequence and different sampled sequences. The loss function uses similarity to calculate conditional weights and construct the contrastive learning optimization objective. Simultaneously, the loss function includes two temperature hyperparameters to adjust the sensitivity of the similarity calculation.

[0155] The current modeling example in this application is based on the identification features of scene and item sequences, but the input features of the model are flexible and scalable, and can be compatible with using richer attribute information of users or goods as input to the encoding network. At the same time, the choice of sequence encoder is also very flexible.

[0156] This application employs dual-sequence coupling modeling to characterize the symbiotic relationship between scenarios and products. Current work largely fails to adequately explore the symbiotic characteristics of scenario information, resulting in underutilization of user-wide interaction trajectory data. This application proposes dual-sequence coupling modeling to model the dependency between scenarios and products. This modeling approach can more comprehensively capture user interaction patterns in different scenarios, thereby improving the accuracy of interaction prediction. Compared to traditional methods, this scheme can better utilize user-wide interaction trajectory data to uncover the symbiotic characteristics of scenario information, which is crucial for improving the performance of recommendation systems and information retrieval services.

[0157] This application employs conditional sampling to improve the Info NCE loss, making it more suitable for recommendation scenarios. Addressing the issues of high randomness and noise in interaction sequences within the recommendation domain, this application proposes an enhanced loss training method for contrastive learning in self-supervised learning. This enhancement method can more effectively handle noise and randomness in user interaction data, improving the model's robustness. Compared to traditional contrastive learning loss, this enhancement method better utilizes the relationship between positive and negative samples, improving the model's learning efficiency and prediction accuracy. Furthermore, this application also enhances the model's generalization ability when facing sparse and noisy data.

[0158] This application proposes a dual-sequence coupling model that integrates scene and product sequences, enabling more accurate prediction of future user interactions. The dual-sequence coupling model is proposed to model the dependency between scenes and products. This modeling approach can more comprehensively capture user interaction patterns in different scenarios, thereby improving the accuracy of interaction prediction. This application also proposes an improved loss training method for contrastive learning in self-supervised learning. This improved method can more effectively handle noise and randomness in user interaction data, enhancing the model's robustness.

[0159] In a global data context, this application observes significant differences in user interactions across different scenarios. For example, in e-commerce applications, user interaction patterns in search scenarios differ drastically from those in recommendation scenarios. These differences are not only reflected in user interactions with items but also in user conversion rates and engagement across different scenarios. Therefore, by simultaneously considering the sequential dynamics of scenarios and items, this application can truly understand users' complex interaction patterns and accurately predict their future interactions.

[0160] To achieve this goal, this application constructs two parallel networks for predicting scene sequences and item sequences, respectively, and introduces a sequence feature enhancement module to capture the interdependence between them. Specifically, the scene sequence prediction network and the item sequence prediction network encode their respective dynamic features from historical interactions, while the sequence feature enhancement module allows the encoded features of one network to be used as input to the other network, thus enabling the two prediction networks to fully utilize their interdependence during the learning process.

[0161] This application not only learns the individual dynamics of scenes and item sequences but also captures the complex relationships between them. This modeling of joint distributions provides a powerful method for understanding and predicting users' dynamic interests and interactions. Conditional contrastive regularization loss introduces conditional weights into the contrastive learning framework. These weights are learned and adaptively adjust based on the similarity and differences between samples. When calculating the first contrastive loss, these weights are used to strengthen the model's similarity representation of positive samples (enhanced sequences versus original sequences) while simultaneously improving its ability to distinguish negative samples (different sequences or perturbed sequences).

[0162] This application considers that the mutual information between the interaction sequence and negative samples may not be zero, and reduces this potential correlation through conditional weights. This allows the model to focus more on samples that are truly similar to the target sequence during learning, thereby improving learning efficiency and representation quality. Furthermore, when facing a long-tailed distribution of user interaction data, the CCR loss, through conditional weights, gives more attention to items with fewer interactions, thus mitigating the impact of uneven data distribution. This method helps improve the model's learning performance on long-tailed items, enhancing the model's fairness and diversity.

[0163] It should be noted that 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, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0164] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0165] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0166] According to the embodiments of this application, a training method for a feature enhancement model is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here. Figure 4 This is a flowchart illustrating a training method for a feature enhancement model according to an embodiment of this application. Figure 4 As shown, the method includes:

[0167] Step S402: Obtain multiple sample interaction sequences;

[0168] In one optional embodiment, the system extracts multiple sample interaction sequences from user historical interaction data. These sequences contain user interaction sequences at different points in time, such as product browsing, purchasing, and searching, and form the basis for building the model. In e-commerce applications, this may mean extracting representative sequences from a large amount of user data, with each sequence representing the interaction pattern of a user or a group of users.

[0169] Step S404: Perform sequence enhancement on multiple sample interaction sequences to obtain sample-enhanced interaction sequences corresponding to the multiple sample interaction sequences;

[0170] After acquiring sample interaction sequences, the system applies techniques such as random masking or element rearrangement to enhance these sequences, generating enhanced interaction sequences. The purpose of sequence augmentation is to increase the robustness and generalization ability of the model, enabling it to maintain good predictive performance even when dealing with incomplete or sequentially varying data. For example, random masking can simulate missing data in user interaction data; element rearrangement can simulate the randomness of user interactions.

[0171] Step S406: Construct a sample contrast loss based on multiple sample interaction sequences and sample enhancement interaction sequences;

[0172] A sample contrastive loss can be constructed using the original sample interaction sequences and the generated sample-enhanced interaction sequences. This loss, through contrastive learning, quantifies and enhances the similarity difference between positive samples (original and enhanced sequences) and negative samples (other unrelated sequences), guiding the model to learn feature representations invariant to masking and rearrangement operations. In e-commerce recommendation scenarios, positive samples might be obtained through conditional sampling of the same user's interaction sequences at different time points, as well as sequence variants obtained through sequence enhancement.

[0173] Step S408: Train the initial feature extraction model using sample contrast loss to obtain the feature extraction model.

[0174] The system uses a pre-contrast loss to train the initial feature extraction model. Through this contrastive loss-guided training process, the model learns more discriminative and invariant feature representations that more accurately reflect user interests and interaction patterns, even in the presence of noise or random variations in the data. Once trained, the resulting feature extraction model efficiently extracts high-quality user representations from comprehensive user interaction data for subsequent interaction prediction or recommendation tasks.

[0175] Through the above steps, multiple sample interaction sequences are obtained; sequence augmentation is performed on these multiple sample interaction sequences to obtain sample-augmented interaction sequences corresponding to the multiple sample interaction sequences; a sample contrast loss is constructed based on the multiple sample interaction sequences and the sample-augmented interaction sequences; the initial feature extraction model is trained using the sample contrast loss to obtain the feature extraction model, thereby improving the accuracy of interaction sequence prediction. It is worth noting that through feature augmentation based on correlation, the model can more comprehensively capture the dynamic changes of user interaction and the interaction between different dimensions. Feature augmentation utilizes the correlation between historical interaction sequences, enabling the model to not only understand user interaction in a single dimension, but also capture the dynamic interaction of scene sequences and product sequences. This allows the subsequent sequence prediction model to more accurately predict interaction sequences within a preset time period, significantly improving prediction accuracy and thus solving the technical problem of poor prediction accuracy of interaction sequences in related technologies.

[0176] In the above embodiments of this application, a sample comparison loss is constructed based on multiple sample interaction sequences and sample enhancement interaction sequences, including: constructing a first sample comparison loss based on the similarity between multiple sample interaction sequences and sample enhancement interaction sequences; constructing a second sample comparison loss based on the similarity between multiple sample interaction sequences; and determining the sample comparison loss based on the first sample comparison loss and the second sample comparison loss.

[0177] According to the embodiments of this application, an interactive sequence prediction method is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here. Figure 5 This is a flowchart of an interactive sequence prediction method according to an embodiment of this application.

[0178] like Figure 5 As shown, the method includes:

[0179] Step S502: Obtain the product interaction sequence and scene interaction sequence formed by the user in the historical time period;

[0180] Among them, the product interaction sequence is used to represent the interaction sequence associated with the product, and the scene interaction sequence is used to represent the interaction sequence associated with the scene.

[0181] Two specific types of sequences can be extracted from users' historical interaction data: product interaction sequences and scenario interaction sequences. Product interaction sequences record the user's interaction with products, such as browsing, purchasing, and adding to favorites, while scenario interaction sequences describe the user's interactions in different scenarios, such as search scenarios, recommendation scenarios, and live streaming scenarios. These sequences are fundamental to understanding user interests and interaction patterns, providing a rich data source for subsequent feature extraction and prediction.

[0182] Step S504: Use the feature extraction model to extract features from the product interaction sequence and the scene interaction sequence respectively to obtain product interaction features and scene interaction features.

[0183] Pre-trained feature extraction models can be used to process product interaction sequences and scene interaction sequences separately, extracting product interaction features and scene interaction features. These feature extraction models can capture the dependencies between elements in the sequence, generating low-dimensional feature vectors that contain temporal sequence information and contextual relationships. These feature vectors condense key information about user interactions, providing structured input for subsequent feature enhancement and prediction.

[0184] Step S506: Based on the correlation between product interaction features and scene interaction features, guide the feature enhancement model to enhance the product interaction features and scene interaction features to obtain product enhanced interaction features and scene enhanced interaction features.

[0185] By analyzing the correlation between product interaction features and scene interaction features, feature enhancement models can be used to further optimize these two feature sets. Feature enhancement involves adding additional information or transformations to the feature space to improve the model's understanding of user interaction patterns. For example, a feature enhancement model might use scene interaction features as conditional information to enhance product interaction features, and vice versa. This enhancement process can capture the symbiotic relationship between scenes and products, enabling predictive models to more accurately understand users' interaction tendencies with products in specific scenarios.

[0186] Step S508: Input the product enhanced interaction features and scene enhanced interaction features into the sequence prediction model, and use the sequence prediction model to predict the product interaction sequence and scene interaction sequence to obtain the predicted product interaction sequence and the predicted scene interaction sequence corresponding to the product interaction sequence within the preset time period.

[0187] Enhanced product and scene interaction features can be input into a sequence prediction model, which can then predict user product and scene interaction sequences within a preset future time period based on these enhanced features. The sequence prediction model utilizes the learned feature dynamics to predict possible user interaction sequences and scene transitions, providing predictive information for recommendation systems or information retrieval services. Through this prediction process, the model can not only understand past user dynamics but also predict future scene and product preferences, thereby providing more accurate personalized recommendations.

[0188] The above process, through feature extraction, enhancement, and prediction, achieves a deep understanding of users' historical interactions and accurate prediction of future interactions, significantly improving the performance of recommendation systems and information retrieval services, especially when dealing with complex and dynamic user interaction data. The key to this process lies in its ability to effectively integrate user interactions across scenario and product dimensions, providing a comprehensive representation of user interactions and thus delivering higher-quality prediction results for subsequent applications.

[0189] Through the above steps, product interaction sequences and scene interaction sequences formed by users over historical time periods are obtained. Product interaction sequences represent interaction sequences associated with products, and scene interaction sequences represent interaction sequences associated with scenes. Feature extraction models are used to extract features from both product and scene interaction sequences, resulting in product interaction features and scene interaction features. Based on the correlation between these features, a feature enhancement model is used to enhance both product and scene interaction features, resulting in enhanced product interaction features and enhanced scene interaction features. These enhanced features are then input into a sequence prediction model to predict the product and scene interaction sequences, yielding product interaction features within a preset time period. The model predicts product interaction sequences corresponding to product interaction sequences and predicts scene interaction sequences corresponding to scene interaction sequences, thereby improving the accuracy of interaction sequence prediction. Notably, through sequence augmentation and feature augmentation based on correlation, the model can more comprehensively capture the dynamic changes in user interactions and the interactions between different dimensions. Sequence augmentation can enhance the model's robustness to the randomness and noise of user interactions, while feature augmentation utilizes the correlation between historical interaction sequences, enabling the model to not only understand user interactions in a single dimension but also capture the dynamic interactions between scene sequences and product sequences. This allows the sequence prediction model to more accurately predict interaction sequences within a preset time period, significantly improving prediction accuracy and thus solving the technical problem of poor prediction accuracy of interaction sequences in related technologies.

[0190] According to the embodiments of this application, an interactive sequence prediction method is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here. Figure 6 This is a flowchart of an interactive sequence prediction method according to an embodiment of this application.

[0191] like Figure 6 As shown, the method includes:

[0192] Step S602: Obtain multiple historical interaction sequences formed by the user within a historical time period by calling the first interface;

[0193] The first interface includes a first parameter, the value of which includes multiple historical interaction sequences. Different historical interaction sequences are used to represent the user's interaction sequences in different dimensions.

[0194] The aforementioned first interface can be an interface for data interaction between the cloud server and the client. Multiple historical interaction sequences can be passed into the interface function as the first parameter of the interface function to achieve the purpose of uploading multiple historical interaction sequences to the cloud server.

[0195] Step S604: Use the feature extraction model to extract features from multiple historical interaction sequences to obtain multiple historical interaction features;

[0196] Step S606: Based on the correlation between multiple historical interaction sequences, guide the feature enhancement model to enhance the features of multiple historical interaction features to obtain multiple enhanced interaction features;

[0197] Step S608: Input multiple enhanced interaction features into the sequence prediction model, use the sequence prediction model to predict multiple historical interaction sequences, and obtain the predicted interaction sequences corresponding to multiple historical interaction sequences within a preset time period.

[0198] Step S610: Output the predicted interaction sequence by calling the second interface.

[0199] The second interface includes a second parameter, the value of which includes the predicted interaction sequence.

[0200] The aforementioned second interface can be an interface for data interaction between the cloud server and the client. The cloud server can pass the predicted interaction sequence into the interface function as the second parameter of the interface function, thereby achieving the purpose of sending the predicted interaction sequence to the client.

[0201] Through the above steps, multiple historical interaction sequences formed by the user within a historical time period are obtained by calling the first interface. The first interface includes a first parameter, the value of which includes multiple historical interaction sequences. Different historical interaction sequences represent user interaction sequences at different dimensions. Features are extracted from the multiple historical interaction sequences using a feature extraction model to obtain multiple historical interaction features. Based on the correlation between the multiple historical interaction sequences, a feature enhancement model is guided to enhance the features of the multiple historical interaction features, resulting in multiple enhanced interaction features. These enhanced interaction features are then input into a sequence prediction model, which predicts the multiple historical interaction sequences to obtain predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period. Finally, the predicted interaction sequences are output by calling the second interface. The interaction sequence is predicted, and the second interface includes a second parameter whose value includes the predicted interaction sequence, thereby improving the accuracy of interaction sequence prediction. It is noteworthy that through sequence augmentation and feature augmentation based on association, the model can more comprehensively capture the dynamic changes of user interaction and the interaction between different dimensions. Sequence augmentation can enhance the model's robustness to the randomness and noise of user interaction, while feature augmentation utilizes the association between historical interaction sequences, enabling the model to not only understand user interaction in a single dimension, but also capture the dynamic interaction of scene sequences and product sequences. This allows the sequence prediction model to more accurately predict the interaction sequence within a preset time period, significantly improving prediction accuracy and thus solving the technical problem of poor prediction accuracy of interaction sequences in related technologies.

[0202] According to embodiments of this application, an interactive sequence prediction apparatus for implementing the above-described interactive sequence prediction method is also provided. Figure 7 This is a schematic diagram of an interactive sequence prediction device according to an embodiment of this application, such as... Figure 7 As shown, the device 700 includes: an acquisition module 702, an extraction module 704, an enhancement module 706, and a prediction module 708.

[0203] The system comprises the following modules: an acquisition module, which acquires multiple historical interaction sequences formed by the user over a historical time period, wherein different historical interaction sequences represent user interaction sequences at different dimensions; an extraction module, which uses a feature extraction model to extract features from the multiple historical interaction sequences to obtain multiple historical interaction features; an enhancement module, which guides a feature enhancement model to enhance the features of the multiple historical interaction sequences based on the correlation between them, to obtain multiple enhanced interaction features; and a prediction module, which inputs the multiple enhanced interaction features into a sequence prediction model and uses the sequence prediction model to predict the multiple historical interaction sequences to obtain predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period.

[0204] It should be noted that the acquisition module 702, extraction module 704, enhancement module 706, and prediction module 708 mentioned above correspond to steps S202 to S208 in the above embodiments. The four modules and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware or software components stored in memory and processed by one or more processors. The above modules can also run as part of the device in the server 10 provided in the above embodiments.

[0205] In the above embodiments of this application, the enhancement module is used to determine the second historical interaction sequence corresponding to the first historical interaction sequence based on the association relationship between multiple historical interaction sequences, wherein the second historical interaction sequence is at least one historical interaction sequence that has an association relationship with the first historical interaction sequence; based on the first historical interaction sequence and the second historical interaction sequence, multiple historical interaction features are enhanced to obtain multiple enhanced interaction features.

[0206] In the above embodiments of this application, the enhancement module is used to determine a first historical interaction feature corresponding to a first historical interaction sequence from multiple historical interaction features, and to determine a second historical interaction feature corresponding to a second historical interaction sequence; to enhance the first historical interaction feature based on the second historical interaction feature to obtain an enhanced interaction feature corresponding to the first historical interaction feature; to enhance the second historical interaction feature based on the first historical interaction feature to obtain an enhanced interaction feature corresponding to the second historical interaction feature; and to obtain multiple enhanced interaction features based on the enhanced interaction feature corresponding to the first historical interaction feature and the enhanced interaction feature corresponding to the second historical interaction feature.

[0207] In the above embodiments of this application, the device further includes: a construction module and an update module.

[0208] The enhancement module is used to enhance multiple historical interaction sequences to obtain enhanced interaction sequences corresponding to the multiple historical interaction sequences; the construction module is used to construct a target contrast loss based on multiple historical interaction sequences and enhanced interaction sequences; and the update module is used to update the model parameters of the feature extraction model using the target contrast loss.

[0209] In the above embodiments of this application, the construction module is used to construct a first contrast loss based on the similarity between multiple historical interaction sequences and enhanced interaction sequences; construct a second contrast loss based on the similarity between multiple historical interaction sequences; and determine a target contrast loss based on the first contrast loss and the second contrast loss.

[0210] In the above embodiments of this application, the enhancement module is used to perform masking processing on multiple historical interaction sequences to obtain masked interaction sequences and unmasked interaction sequences corresponding to multiple historical interaction sequences; to rearrange the elements of the masked interaction sequences to obtain rearranged interaction sequences corresponding to multiple historical interaction sequences; and to obtain enhanced interaction sequences of multiple historical interaction sequences based on the unmasked interaction sequences and the rearranged interaction sequences.

[0211] In the above embodiments of this application, the device further includes: an acquisition module for acquiring actual interaction sequences corresponding to multiple historical interaction sequences; a construction module for constructing a third contrastive loss based on the actual interaction sequences and the predicted interaction sequences; and an update module for updating the model parameters of the feature enhancement model based on the third contrastive loss.

[0212] It should be noted that the preferred embodiments involved in the above embodiments of this application are the same as the solutions, application scenarios and implementation processes provided in the above embodiments, but are not limited to the solutions provided in the above embodiments.

[0213] According to an embodiment of this application, a training apparatus for a feature enhancement module used to implement the training method of the above-described feature enhancement model is also provided. Figure 8 This is a schematic diagram of a training device for a feature enhancement module according to an embodiment of this application, as shown below. Figure 8 As shown, the device 800 includes: an acquisition module 802, an enhancement module 804, a construction module 806, and a training module 808.

[0214] The module includes: an acquisition module for acquiring multiple sample interaction sequences; an enhancement module for performing sequence enhancement on the multiple sample interaction sequences to obtain sample-enhanced interaction sequences corresponding to the multiple sample interaction sequences; a construction module for constructing a sample contrast loss based on the multiple sample interaction sequences and the sample-enhanced interaction sequences; and a training module for training the initial feature extraction model using the sample contrast loss to obtain the feature extraction model.

[0215] It should be noted that the acquisition module 802, enhancement module 804, construction module 806, and training module 808 mentioned above correspond to steps S402 to S408 in the above embodiments. The four modules and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware or software components stored in memory and processed by one or more processors. The above modules can also run as part of the device in the server 10 provided in the above embodiments.

[0216] In the above embodiments of this application, the construction module is further configured to construct a first sample comparison loss based on the similarity between multiple sample interaction sequences and sample enhancement interaction sequences; construct a second sample comparison loss based on the similarity between multiple sample interaction sequences; and determine the sample comparison loss based on the first sample comparison loss and the second sample comparison loss.

[0217] According to embodiments of this application, an interactive sequence prediction apparatus for implementing the above-described interactive sequence prediction method is also provided. Figure 9 This is a schematic diagram of an interactive sequence prediction device according to an embodiment of this application, such as... Figure 9 As shown, the device 900 includes: an acquisition module 902, an extraction module 904, an enhancement module 906, and a prediction module 908.

[0218] The system comprises the following modules: an acquisition module for acquiring product interaction sequences and scene interaction sequences formed by users over a historical time period, where product interaction sequences represent interaction sequences associated with products and scene interaction sequences represent interaction sequences associated with scenes; an extraction module for extracting features from product interaction sequences and scene interaction sequences using a feature extraction model to obtain product interaction features and scene interaction features; an enhancement module for guiding a feature enhancement model to enhance product interaction features and scene interaction features based on the correlation between product interaction features and scene interaction features to obtain enhanced product interaction features and enhanced scene interaction features; and a prediction module for inputting the enhanced product interaction features and enhanced scene interaction features into a sequence prediction model to predict the product interaction sequences and scene interaction sequences within a preset time period.

[0219] It should be noted that the acquisition module 902, extraction module 904, enhancement module 906, and prediction module 908 mentioned above correspond to steps S502 to S508 in the above embodiments. The four modules and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware or software components stored in memory and processed by one or more processors. The above modules can also run as part of the device in the server 10 provided in the above embodiments.

[0220] According to embodiments of this application, an interactive sequence prediction apparatus for implementing the above-described interactive sequence prediction method is also provided. Figure 10 This is a schematic diagram of an interactive sequence prediction device according to an embodiment of this application, such as... Figure 10As shown, the device 1000 includes: an acquisition module 1002, an extraction module 1004, an enhancement module 1006, a prediction module 1008, and an output module 1010.

[0221] The system comprises the following modules: an acquisition module, which acquires multiple historical interaction sequences formed by a user over a historical time period by calling a first interface, wherein the first interface includes a first parameter whose value includes multiple historical interaction sequences, and different historical interaction sequences represent user interaction sequences in different dimensions; an extraction module, which extracts features from the multiple historical interaction sequences using a feature extraction model to obtain multiple historical interaction features; an enhancement module, which guides a feature enhancement model to enhance the multiple historical interaction features based on the correlation between the multiple historical interaction sequences to obtain multiple enhanced interaction features; a prediction module, which inputs the multiple enhanced interaction features into a sequence prediction model and uses the sequence prediction model to predict the multiple historical interaction sequences to obtain predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period; and an output module, which outputs the predicted interaction sequences by calling a second interface, wherein the second interface includes a second parameter whose value includes the predicted interaction sequences.

[0222] It should be noted that the acquisition module 1002, extraction module 1004, enhancement module 1006, prediction module 1008, and output module 1010 mentioned above correspond to steps S602 to S610 in the above embodiments. The five modules and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware or software components stored in memory and processed by one or more processors. The above modules can also run as part of the device in the server 10 provided in the above embodiments.

[0223] Embodiments of this application may provide a computing device. Figure 11 This is a structural block diagram of a computing device according to an embodiment of this application. Figure 11 As shown, the computing device 100 may include one or more (only one is shown in the figure) processors 102, memory 104, memory controller, and peripheral interfaces.

[0224] The aforementioned computing device can be understood as an integrated smart terminal, including but not limited to servers, desktop computers, PCs (Personal Computers), all-in-one model machines, etc., and the computing device may have the model in the above embodiments of this application pre-installed.

[0225] Specifically, this computing device can pre-install various types of models, including but not limited to models in natural language processing, visual processing, speech processing, code processing, and multimodal task processing, thus providing diverse model selection. In different product forms, this computing device can support one or more model usage methods, including but not limited to model training, model invocation, model fine-tuning, model deployment, model inference, and application. In some product forms, this computing device also supports model management, including but not limited to multi-type model management (supporting the management of discriminative, generative, and other types of models), model version control (supporting the control of different model versions), and model evaluation (evaluating model performance and effectiveness based on model evaluation tools). In other product forms, this computing device can also create applications based on models, providing API calling capabilities, allowing models to be called into created applications through API interfaces, and providing application management tools for application management and monitoring.

[0226] Furthermore, the computing device may also include data management (supporting the creation and management of model tuning datasets), a training center (providing abundant training resources to help users learn and master AI technology), and basic control capabilities (providing enterprise-level basic control capabilities to ensure the security and efficient operation of the system). Through the above functions, it provides a comprehensive and integrated device for AI development, training, deployment, and application.

[0227] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the methods and apparatus in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the methods in the above embodiments. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to terminal A via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0228] The processor can invoke an executable program stored in memory via a transmission device to execute the method described in any of the above embodiments.

[0229] Embodiments of this application may provide an electronic device. Figure 12 This is a structural block diagram of an electronic device according to an embodiment of this application. Figure 2 As shown, the electronic device may include: an input / output device 112; a memory 114; and a processor 116, wherein the processor 116 is connected to the input / output device 112 and the memory 114 via a bus 118.

[0230] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the methods and apparatus in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the methods in the above embodiments. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to terminal A via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0231] The processor can invoke an executable program stored in memory via a transmission device to execute the method described in any of the above embodiments.

[0232] Those skilled in the art will understand that, Figure 11 The structure shown is for illustrative purposes only. The computing device may also be a smartphone (such as an Android phone, iOS phone, etc.), tablet computer, PDA, mobile Internet device (MID), PAD, and other terminal devices. This figure does not limit the structure of the computing device described above. For example, the computing device 100 may include more or fewer components (such as network interfaces, display devices, etc.) than shown in this figure, or may have a different configuration than shown in this figure.

[0233] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0234] Embodiments of this application also provide a computer-readable storage medium. Optionally, in this embodiment, the computer-readable storage medium can be used to store program code executed by the method provided in the above embodiments.

[0235] Optionally, in this embodiment, the storage medium may be located in a computing device.

[0236] Optionally, in this embodiment, the computer-readable storage medium is configured to store an executable program, which, when the executable program is running, controls the device where the computer-readable storage medium is located to execute the method described in any of the above embodiments.

[0237] Embodiments of this application also provide a computer program product. Optionally, in this embodiment, the computer program product may include a computer program that, when executed by a processor, implements the methods provided in the embodiments described above.

[0238] Embodiments of this application also provide a computer program product. Optionally, the computer program product may include a non-volatile computer-readable storage medium, which can be used to store a computer program that, when executed by a processor, implements the method provided in the above embodiments.

[0239] Embodiments of this application also provide a computer program. Optionally, in this embodiment, when the computer program is executed by a processor, it implements the method provided in the above embodiments.

[0240] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0241] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0242] The units described as separate components may or may not be physically separate. The 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 the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0243] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0244] 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 storage medium. Based on this understanding, the technical solution of this application, in essence, 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 storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0245] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An interactive sequence prediction method, characterized in that: include: Acquire multiple historical interaction sequences formed by the user in a historical time period, wherein different historical interaction sequences are used to represent the interaction sequences of the user in different dimensions; Using a feature extraction model to extract features from the multiple historical interaction sequences respectively, to obtain multiple historical interaction features; Based on the association relationship between the multiple historical interaction sequences, guiding the feature enhancement model to perform feature enhancement on the multiple historical interaction features to obtain multiple enhanced interaction features; The multiple enhanced interaction features are input into a sequence prediction model, and the multiple historical interaction sequences are predicted using the sequence prediction model to obtain predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period.

2. The interactive sequence prediction method according to claim 1, characterized in that: Based on the association relationship between the multiple historical interaction sequences, a feature enhancement model is guided to enhance the multiple historical interaction features to obtain multiple enhanced interaction features, including: Based on the association relationship between the multiple historical interaction sequences, determine a second historical interaction sequence corresponding to the first historical interaction sequence, wherein the second historical interaction sequence is at least one historical interaction sequence in the multiple historical interaction sequences that has the association relationship with the first historical interaction sequence; Based on the first historical interaction sequence and the second historical interaction sequence, feature enhancement is performed on the multiple historical interaction features to obtain the multiple enhanced interaction features.

3. The interactive sequence prediction method according to claim 2, characterized in that: Based on the first historical interaction sequence and the second historical interaction sequence, feature enhancement is performed on the multiple historical interaction features to obtain the multiple enhanced interaction features, including: Determine a first historical interaction feature corresponding to the first historical interaction sequence from the multiple historical interaction features, and determine a second historical interaction feature corresponding to the second historical interaction sequence; Performing feature enhancement on the first historical interaction feature based on the second historical interaction feature to obtain an enhanced interaction feature corresponding to the first historical interaction feature; Performing feature enhancement on the second historical interaction feature based on the first historical interaction feature to obtain an enhanced interaction feature corresponding to the second historical interaction feature; The multiple enhanced interaction features are obtained based on the enhanced interaction feature corresponding to the first historical interaction feature and the enhanced interaction feature corresponding to the second historical interaction feature.

4. The interactive sequence prediction method according to claim 2, characterized in that: The method further comprises: Performing sequence enhancement on the multiple historical interaction sequences to obtain enhanced interaction sequences corresponding to the multiple historical interaction sequences; constructing a target contrast loss based on the multiple historical interaction sequences and the enhanced interaction sequence; The target contrast loss is used to update the model parameters of the feature extraction model.

5. The interactive sequence prediction method according to claim 4, characterized in that: Constructing a target contrast loss based on the multiple historical interaction sequences and the enhanced interaction sequence, including: constructing a first contrast loss based on the similarity between the plurality of historical interaction sequences and the enhanced interaction sequence; constructing a second contrast loss based on the similarities between the plurality of historical interaction sequences; The target contrast loss is determined based on the first contrast loss and the second contrast loss.

6. The interactive sequence prediction method according to claim 4, characterized in that: Performing sequence enhancement on the multiple historical interaction sequences to obtain enhanced interaction sequences corresponding to the multiple historical interaction sequences includes: Performing masking processing on the multiple historical interaction sequences respectively to obtain masked interaction sequences and unmasked interaction sequences corresponding to the multiple historical interaction sequences; Rearranging elements of the masked interaction sequence to obtain a re-arranged interaction sequence corresponding to the multiple historical interaction sequences; Based on the unmasked interaction sequence and the rearranged interaction sequence, an enhanced interaction sequence of the multiple historical interaction sequences is obtained.

7. The interactive sequence prediction method according to claim 1, characterized in that: The method further comprises: Obtaining actual interaction sequences corresponding to the multiple historical interaction sequences; constructing a third contrast loss based on the actual interaction sequence and the predicted interaction sequence; Model parameters of the feature enhancement model are updated based on the third contrast loss.

8. A method for training a feature enhancement model, characterized in that: include: Obtain multiple sample interaction sequences; Performing sequence enhancement on the multiple sample interaction sequences to obtain sample enhanced interaction sequences corresponding to the multiple sample interaction sequences; constructing a sample contrast loss based on the multiple sample interaction sequences and the sample enhancement interaction sequence; The initial feature extraction model is trained using the sample contrast loss to obtain a feature extraction model.

9. The method for training a feature enhancement model according to claim 8, characterized in that: Constructing a sample contrast loss based on the multiple sample interaction sequences and the sample enhancement interaction sequence, including: constructing a first sample contrast loss based on the similarity between the plurality of sample interaction sequences and the sample enhancement interaction sequence; constructing a second sample contrast loss based on the similarities between the plurality of sample interaction sequences; The sample contrast loss is determined based on the first sample contrast loss and the second sample contrast loss.

10. An interactive sequence prediction method, characterized in that: include: Acquire a product interaction sequence and a scene interaction sequence formed by the user in a historical time period, wherein the product interaction sequence is used to represent an interaction sequence associated with a product, and the scene interaction sequence is used to represent an interaction sequence associated with a scene; Using a feature extraction model to extract features from the product interaction sequence and the scene interaction sequence respectively, to obtain product interaction features and scene interaction features; Based on the correlation between the product interaction feature and the scene interaction feature, guiding the feature enhancement model to perform feature enhancement on the product interaction feature and the scene interaction feature to obtain a product enhanced interaction feature and a scene enhanced interaction feature; The product enhancement interaction features and the scene enhancement interaction features are input into a sequence prediction model, and the sequence prediction model is used to predict the product interaction sequence and the scene interaction sequence to obtain a predicted product interaction sequence corresponding to the product interaction sequence and a predicted scene interaction sequence corresponding to the scene interaction sequence within a preset time period.

11. An interactive sequence prediction method, characterized in that: include: Acquire multiple historical interaction sequences formed by the user in a historical time period by calling a first interface, wherein the first interface includes a first parameter, a parameter value of the first parameter includes multiple historical interaction sequences, and different historical interaction sequences are used to represent the interaction sequences of the user in different dimensions; Using a feature extraction model to extract features from the multiple historical interaction sequences respectively, to obtain multiple historical interaction features; Based on the association relationship between the multiple historical interaction sequences, guiding the feature enhancement model to perform feature enhancement on the multiple historical interaction features to obtain multiple enhanced interaction features; Inputting the multiple enhanced interaction features into a sequence prediction model, using the sequence prediction model to predict the multiple historical interaction sequences, and obtaining predicted interaction sequences corresponding to the multiple historical interaction sequences within a preset time period; The predicted interaction sequence is output by calling a second interface, wherein the second interface includes a second parameter, and a parameter value of the second parameter includes the predicted interaction sequence.

12. An electronic device, characterized in that: include: A memory storing an executable program; A processor, connected to the memory via a bus, and configured to run the program, wherein the program executes the method described in any one of claims 1 to 11 when running.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored executable program, wherein when the executable program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 11.

14. A computer program product, characterized in that The invention comprises a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 11.

Images