User representation construction method and device, equipment, storage medium and program product

Abstract

This disclosure relates to the field of artificial intelligence technology, and in particular provides a method, apparatus, device, storage medium, and program product for constructing user representations. The method includes: acquiring user attribute data and historical trajectory sequence data of a target user, and base station attribute data of multiple base stations; the historical trajectory sequence data includes identification information of multiple base stations accessed by the target user in chronological order within a preset time period; using a large language model to perform semantic understanding on the user attribute data, historical trajectory sequence data, and base station attribute data of multiple base stations to generate a user semantic representation; using a sequence recommendation model to perform collaborative filtering processing on the historical trajectory sequence data to obtain a user collaborative filtering representation; and performing multi-layer interactive fusion of the user semantic representation and the user collaborative filtering representation to construct a user representation of the target user. This disclosure can construct user representations based on the joint use of a large language model and a sequence recommendation model, improving the generalization ability and robustness of user representations.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, storage medium, and program product for constructing user representations. Background Technology

[0002] In the digital age, massive amounts of user behavior data enable various industries to gain a deeper understanding of user characteristics and intentions. Especially in scenarios such as culture and tourism, emergency response, finance, and anti-fraud, the accurate depiction of user trajectories and profiles is of significant value.

[0003] In related technologies, sequence recommendation models and time-series prediction models are commonly used to analyze user trajectory sequence data to predict future user trajectories. However, these methods are limited by challenges such as sparse data, processing long or periodic sequences, and semantic understanding, resulting in low accuracy in constructing user trajectory behavior and profiles. Summary of the Invention

[0004] This disclosure is made in view of the above-mentioned problems. This disclosure provides a method, apparatus, device, storage medium, and program product for constructing user representations.

[0005] According to one aspect of this disclosure, a method for constructing user representations is provided, comprising: The system acquires user attribute data and historical trajectory sequence data of the target user, as well as base station attribute data of multiple base stations; wherein, the historical trajectory sequence data includes the identification information of the multiple base stations accessed by the target user in chronological order within a preset time period; Using a large language model, semantic understanding is performed on the user attribute data, the historical trajectory sequence data, and the base station attribute data of the multiple base stations to generate a user semantic representation of the target user; The historical trajectory sequence data is processed by a sequence recommendation model to obtain a user collaborative filtering representation of the target user. The user semantic representation and the user collaborative filtering representation are fused together through multi-layer interaction to construct the user representation of the target user.

[0006] According to another aspect of this disclosure, a user representation construction apparatus is provided, comprising: The acquisition module is used to acquire user attribute data and historical trajectory sequence data of the target user, as well as base station attribute data of multiple base stations; wherein, the historical trajectory sequence data includes the identification information of the multiple base stations accessed by the target user in chronological order within a preset time period; The semantic understanding module is used to perform semantic understanding on the user attribute data, the historical trajectory sequence data and the base station attribute data of the multiple base stations using a large language model, and generate a user semantic representation of the target user. The collaborative filtering module is used to perform collaborative filtering processing on the historical trajectory sequence data using a sequence recommendation model to obtain the user collaborative filtering representation of the target user. The multi-layer interaction fusion module is used to perform multi-layer interaction fusion of the user semantic representation and the user collaborative filtering representation to construct the user representation of the target user.

[0007] In another aspect of exemplary embodiments of this disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the methods described in exemplary embodiments of this disclosure.

[0008] In another aspect of exemplary embodiments of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the methods described in exemplary embodiments of the present disclosure.

[0009] In another aspect of the exemplary embodiments of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the methods described in the exemplary embodiments of this disclosure.

[0010] As will be described in detail below, the user representation construction method according to embodiments of this disclosure involves acquiring user attribute data and historical trajectory sequence data of a target user, as well as base station attribute data of multiple base stations. The historical trajectory sequence data includes identification information of multiple base stations accessed by the target user in chronological order within a preset time period. A large language model is used to perform semantic understanding on the user attribute data, historical trajectory sequence data, and base station attribute data of multiple base stations to generate a semantic representation of the target user. A sequence recommendation model is used to perform collaborative filtering processing on the historical trajectory sequence data to obtain a collaborative filtering representation of the target user. The semantic representation and the collaborative filtering representation are then fused through multi-layer interaction to construct the user representation of the target user. This method enables the joint construction of user representations based on a large language model and a sequence recommendation model, achieving complementary advantages between semantic features and collaborative filtering features. This results in a user representation with high accuracy and comprehensiveness, and improves the generalization ability and robustness of the user representation.

[0011] It should be understood that both the foregoing general description and the following detailed description are exemplary and intended to provide further illustration of the claimed technology. Attached Figure Description

[0012] The above and other objects, features, and advantages of this disclosure will become more apparent from the more detailed description of the embodiments thereof in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the disclosure and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0013] Figure 1 A flowchart illustrating a method for constructing a user representation provided in an exemplary embodiment of this disclosure is shown. Figure 2 A schematic diagram of the structure of the user representation construction apparatus provided in an exemplary embodiment of this disclosure is shown; Figure 3 A schematic diagram of the structure of an electronic device provided in an exemplary embodiment of this disclosure is shown; Figure 4 A schematic diagram of the structure of a computer system provided in an exemplary embodiment of this disclosure is shown. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this disclosure more apparent, exemplary embodiments according to this disclosure will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments of this disclosure. It should be understood that this disclosure is not limited to the exemplary embodiments described herein.

[0015] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0016] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below. It should be noted that the concepts of "first", "second", etc., used in this disclosure are only used to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0017] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0018] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0019] In numerous fields such as culture and tourism, emergency response, finance, and anti-fraud, accurate descriptions of user behavior and profiles are particularly important. For example, in the culture and tourism industry, by accurately analyzing tourists' interests and preferences and combining this with historical data to predict their movement trends, scenic spots can optimize opening hours, ensure the orderly opening of popular attractions during peak tourist seasons, and allocate tour guides appropriately, thereby significantly improving tourist satisfaction. Similarly, in the financial sector, understanding users' work and living environments and income and consumption levels allows for more accurate assessment of credit risk, rational planning of loan amounts, increased success rates in lending, reduced non-performing loan rates, and efficient allocation of financial resources.

[0020] With the help of deep learning, sequence recommendation models and time-series prediction models have demonstrated unique advantages in processing user trajectory sequence data. They can capture subtle information in user trajectory sequence data, and by analyzing the chronological order of users' historical trajectories, they can outline the dynamic evolution of users' trajectory behavior, thereby making relatively accurate predictions of users' future trajectory behavior. For example, in the cultural tourism scenario, the model can predict the next attraction a user might visit based on the order of attractions they have previously visited, or recommend attractions that the user might want to visit next.

[0021] However, these models related to user behavior sequences also reveal significant shortcomings in real-world applications. On one hand, user time-series behavior data is generally sparse, making it difficult for models to accurately capture user behavior patterns and interest preferences, thus severely impacting the accuracy of recommendation and prediction results. On the other hand, when dealing with long or periodic data sequences, the models struggle to build long-term dependencies. For example, certain key behaviors of users from months or even years ago may still influence their current behavior patterns and interests.

[0022] Furthermore, relying solely on sequence data for recommending or predicting user behavior trajectories lacks semantic understanding, resulting in limited and simplistic learned representations that fail to reveal complex user motivations. When analyzing user trajectories, the lack of the ability to uncover long-term dependencies makes it difficult to effectively capture differences in travel patterns across different time periods, leading to significant limitations in understanding user behavior.

[0023] In other words, most existing technical solutions only process and analyze user trajectory sequence data from a single dimension, failing to integrate diverse data for semantic-assisted understanding. This makes it difficult to extract rich information from massive amounts of data, resulting in low accuracy in user trajectory behavior prediction and profile construction, and failing to meet the refined needs of multiple fields for accurate construction of user trajectory behavior and profiles.

[0024] Therefore, in order to solve the above problems, this disclosure provides a method for constructing user representations, which can jointly construct user representations based on large language models and sequence recommendation models, realize the complementary advantages of the models, improve the accuracy and comprehensiveness of user representations, and thus facilitate more accurate user trajectory prediction, user profile construction, location service recommendation or precision service marketing for multiple fields.

[0025] The user representation construction method provided in this disclosure can be executed by a terminal or by a chip applied to the terminal.

[0026] For example, the terminal may include one or more of the following: mobile phone, tablet computer, wearable device, in-vehicle device, laptop computer, ultra-mobile personal computer (UMPC), netbook, PDA, and wearable device based on augmented reality (AR) and / or virtual reality (VR) technology. The exemplary embodiments disclosed herein do not impose specific limitations on these.

[0027] Figure 1 A flowchart illustrating a method for constructing user representations provided in an exemplary embodiment of this disclosure is shown. Figure 1 As shown, the method for constructing this user representation includes: S101, acquire user attribute data and historical trajectory sequence data of the target user, as well as base station attribute data of multiple base stations; wherein, the historical trajectory sequence data includes the identification information of multiple base stations accessed by the target user in chronological order within a preset time period; S102, using a large language model to perform semantic understanding on user attribute data, historical trajectory sequence data and base station attribute data of multiple base stations, to generate a user semantic representation of the target user; S103, use a sequence recommendation model to perform collaborative filtering on historical trajectory sequence data to obtain the user collaborative filtering representation of the target user; S104 integrates user semantic representation with user collaborative filtering representation through multi-layer interaction to construct user representation of the target user.

[0028] Specifically, the aforementioned user attribute data, also known as basic user tags, may include, but is not limited to, information such as user age, gender, place of residence, area of ​​residence, service package type, and terminal device model.

[0029] The aforementioned historical trajectory sequence data can be the identification information of multiple base stations accessed by the target user in chronological order within a preset time period. For example, the identification information can be at least one of base station number, base station location code, etc., but is not limited to these. Here, the preset time period can be the period most recent to the current time, and the length of the preset time period can be set according to actual needs; this embodiment does not specifically limit it.

[0030] This embodiment of the disclosure can pre-associate and store the identification information of base stations with base station attribute data. After obtaining the identification information of each base station, the corresponding base station attribute data can be obtained based on the identification information. The base station attribute data includes, but is not limited to, base station location information (such as latitude and longitude), point of interest (POI) category information within the base station coverage area, and information such as the functional attributes of the area to which the base station belongs.

[0031] After obtaining the aforementioned user attribute data, historical trajectory sequence data, and base station attribute data from multiple base stations, this raw data can be cleaned first. Through operations such as data deduplication, outlier detection and handling, and missing value imputation, the messy data can be transformed into standardized and clean data, providing high-quality data input for subsequent processing. For example, missing values ​​in user attribute data or historical trajectory sequence data can be appropriately imputed according to their type.

[0032] The aforementioned large language model can be constructed through pre-training. It involves textually concatenating user attribute data, historical trajectory sequence data, and corresponding base station attribute data to form a unified semantic description text. This semantic description text is then input into the large language model, which, through its semantic encoding and feature extraction capabilities, outputs a high-dimensional user semantic representation rich in semantic information. For example, if a user frequently visits historical and cultural base stations, the large language model will highlight their interest in history and culture in the generated user semantic representation. Based on this, this embodiment of the disclosure, by integrating multi-source heterogeneous data such as user attribute data, historical trajectory sequence data, and base station attribute data, comprehensively portrays the target user from multiple dimensions. The generated user representation reflects the target user's behavioral habits, activity preferences, and location scene characteristics, effectively compensating for the information limitations inherent in user representation relying solely on single trajectory data, resulting in a more complete and accurate user representation.

[0033] For processing the historical trajectory sequence data of the target user, this embodiment of the disclosure will employ a sequence recommendation model for collaborative filtering side representation learning. The historical trajectory sequence data of the target user is used as input to the sequence recommendation model. This historical trajectory sequence data contains identification information of multiple base stations connected to by the target user at different time points, recording not only the target user's movement trajectory but also reflecting the interaction order between the target user and different base stations. For example, the target user connects to base stations sequentially within a day... Base stations Base stations Such sequence information will be received by the model.

[0034] When processing the sequence of target users' visits to base stations, the sequence recommendation model analyzes information such as the order and time interval of the target user's visits to different base stations, thereby uncovering the potential correlation patterns of the target user's visits to base stations at different time points. Combining the common characteristics of base station visits of a massive user group, the collaborative filtering algorithm is used to mine the implicit correlations between the target user and the base station (for example, when a user frequently visits a few adjacent base stations during a specific time period on weekday evenings, the model will learn this behavioral pattern and reflect it in the collaborative filtering side representation), and between the target user and other users. This generates a user collaborative filtering representation that can reflect the target user's mobile behavior pattern and the common characteristics of the group, solving the problem of insufficient utilization of time series information in traditional collaborative filtering and improving the reliability of subsequent user behavior prediction.

[0035] Here, during the training phase of the sequence recommendation model, as training progresses, the model's parameters are optimized, such as adjusting attention weights and updating the connection weights of the neural network, to maximize the accuracy of predicting user visits to base stations. This means that as the model learns, it can more accurately predict which base stations a user might visit in the future based on their previous base station visit sequences. After multiple rounds of training, the sequence recommendation model finally generates a user collaborative filtering representation. Collaborative filtering characterization with base stations .

[0036] Among them, user collaborative filtering representation This comprehensively reflects the target user's access preferences, behavioral patterns, and similarities to other users' base station access behaviors at different historical moments; base station collaborative filtering representation. This reflects the characteristics of how the base station is accessed by different users. These characteristics will provide important data support for subsequent user trajectory prediction and profile construction, and help to gain a deeper understanding of user behavior and the role of the base station in user behavior.

[0037] Based on this, the embodiments of this disclosure can perform multi-layer interactive fusion of user semantic representation and user collaborative filtering representation. This multi-layer interactive fusion can include feature alignment, weighted fusion, and deep interaction, fully combining personalized user semantic features from the semantic side and user collaborative filtering representation from the collaborative filtering side to generate a user representation of the target user. This achieves complementary advantages between semantic features and collaborative filtering features, making the user representation highly accurate and comprehensive, improving its generalization ability and robustness, and enabling it to more accurately reflect users' true behavioral preferences. This user representation has strong practicality and engineering application value when applied to scenarios such as user behavior prediction, user profile construction, location service recommendation, or precision marketing.

[0038] As can be seen, this disclosure effectively addresses the shortcomings of existing technologies in terms of data utilization, semantic understanding, and application scenarios by integrating multi-source data and employing innovative model application methods. This provides more accurate user trajectory prediction and profile construction services for multiple fields. This disclosure uses signaling data as the core data source for user trajectories. By integrating diverse information such as user attribute data, historical trajectory sequence data spanning multiple days at different granularities, and base station attribute data, a comprehensive and dynamic dataset is constructed. Leveraging the powerful semantic understanding capabilities of large language models, it deeply mines the behavioral logic and semantic features behind user trajectories. Simultaneously, it fully utilizes the learning advantages of sequence recommendation models on behavioral sequences, mining effective information from sparse data, capturing the temporal dependence and continuity of user behavior, and constructing long-term dependency relationships. This solves the problems of sparse user temporal behavior data and the difficulty of processing long sequences in related technologies. The feature representations generated by large language models and sequence recommendation models are fused through multi-layer interaction to achieve complementary advantages, thereby generating more accurate and comprehensive user representations. For example, when analyzing the impact of key user behaviors from months or even years ago on the present, it outperforms existing technologies, improving the accuracy of recommendation and prediction results.

[0039] The user representations generated by this disclosure can be applied to multiple fields. For example, in the cultural tourism sector, scenic spots can use user representations to understand tourists' interests, preferences, and behavioral patterns, optimize opening hours and resource allocation, improve tourist satisfaction, attract more tourists, and increase revenue. In the financial sector, financial institutions can use user representations to assess credit risk, plan lending limits, reduce non-performing loans, improve the efficiency of financial resource allocation, and enhance market competitiveness. In the emergency response sector, relevant departments can use user representations to track the movement and distribution of people in disaster-stricken areas, achieve precise and efficient rescue, protect people's lives and property, and enhance the government's emergency management capabilities and social image. In the anti-fraud sector, the application of user representations helps to promptly detect and prevent fraudulent activities, protect users' property, and maintain financial order. In trajectory prediction scenarios, the system combines users' historical travel trajectories, time patterns, and semantic information to predict users' future travel paths and plan the optimal route.

[0040] According to the technical solution of the exemplary embodiments of this disclosure, user attribute data and historical trajectory sequence data of the target user, as well as base station attribute data of multiple base stations are obtained. The historical trajectory sequence data includes the identification information of multiple base stations accessed by the target user in chronological order within a preset time period. A large language model is used to perform semantic understanding on the user attribute data, historical trajectory sequence data, and base station attribute data of multiple base stations to generate a semantic representation of the target user. A sequence recommendation model is used to perform collaborative filtering processing on the historical trajectory sequence data to obtain a collaborative filtering representation of the target user. The semantic representation and the collaborative filtering representation are then fused through multi-layer interaction to construct the target user's user representation. This method enables the joint construction of user representation based on a large language model and a sequence recommendation model, achieving complementary advantages between semantic features and collaborative filtering features. This results in a user representation with high accuracy and comprehensiveness, and improves the generalization ability and robustness of the user representation.

[0041] In some embodiments, a large language model is used to perform semantic understanding on user attribute data, historical trajectory sequence data, and base station attribute data from multiple base stations to generate a user semantic representation of the target user, including: Using a large language model, semantic understanding of the base station attribute data of each base station is performed based on base station semantic prompts, and a base station semantic representation of each base station is generated. By leveraging a large language model to generate prompts based on user profiles, a comprehensive analysis is performed on user attribute data, historical trajectory sequence data, and semantic representations of multiple base stations to generate a semantic representation of the target user.

[0042] Specifically, firstly, by using a large language model combined with pre-defined semantic prompts for base stations, semantic understanding is performed on base station attribute data such as base station location information, point of interest category information within the base station coverage area, and functional attributes of the area to which the base station belongs. This can transform structured and discrete base station attribute data into semantically related base station semantic representations, achieving a unified semantic expression of base station information and improving the expressive power and interpretability of base station features.

[0043] Here, the semantic prompts for base stations can be the instruction text that drives the large language model to perform unified semantic parsing of base station attribute data. This instruction text can standardize the understanding dimension and output format of the large language model for base station attribute data, ensure that the representations of different base stations are aligned in the same semantic space, and provide standardized and fusionable base station semantic features for subsequent comprehensive analysis at the user level.

[0044] Secondly, by using a large language model combined with pre-defined user profiles to generate prompt words, a comprehensive analysis is performed on the semantic representations of multiple base stations arranged in chronological order in user attribute data, historical trajectory sequence data, and historical trajectory sequence data. This fully leverages the advantages of the large language model in multi-source heterogeneous data fusion and semantic understanding, and can automatically uncover the deep correlation between user behavior and location scenarios, avoiding the limitations of manual feature design and improving the richness and accuracy of user semantic representation.

[0045] Here, the user profile generation prompts can be used to drive the large language model to comprehensively understand user attribute data, historical trajectory sequence data, and base station semantic representations of multiple base stations, and generate instruction text for user semantic representation. This instruction text clarifies the analysis objectives, processing objects, and output requirements of the large language model, enabling the large language model to automatically mine the deep correlation between user behavior and location scenarios, forming a unified and highly expressive user semantic representation, and providing support for subsequent user representation construction.

[0046] Based on this, the embodiments of this disclosure first perform semantic representation on the base station separately, and then combine it with user attribute data to generate user semantic representation, thereby realizing hierarchical modeling from scene semantics to user semantics. This ensures the full extraction of base station features and achieves overall semantic characterization at the user level, making the user semantic representation more in line with the user's real behavior and preferences, which is conducive to improving the accuracy of subsequent user representation construction.

[0047] In some embodiments, a large language model is used to perform semantic understanding of the base station attribute data of each base station based on base station semantic cue words, generating a base station semantic representation for each base station, including: Convert the base station attribute data of each base station into a structured text description of the base station; Obtain semantic prompts for base stations, and input the semantic prompts and structured text descriptions of base stations into a large language model to generate semantic representations of each base station.

[0048] Specifically, when constructing the semantic representation of a base station, the core information of the base station is sorted out, and key information such as the base station location information, the interest point category information within the base station coverage area, and the functional attributes of the area to which the base station belongs are organized into a clear text description, forming a structured text description of the base station.

[0049] The structured text description of the base station is concatenated with pre-defined semantic prompts to form the input text for base station semantic parsing. This input text is then fed into a trained large language model, which performs unified semantic encoding and feature extraction on the base station's location information, scene features, and functional attributes. This generates a semantic representation of the base station that accurately reflects its environmental characteristics, its relationship with the surrounding environment, and its role in user behavior. In this way, the structured base station attribute data is transformed into a semantic representation of the base station with semantic relationships.

[0050] The processing of semantic prompts and structured text descriptions for base stations by the large language model can be represented as follows:

[0051] in, Represents the semantic characterization of a base station; These are semantic prompts for base stations, specifically pre-designed system prompts tailored to base station attribute data. This represents the structured text description of the base station, i.e., the base station's original semantic data; Represents a large language model.

[0052] In some embodiments, a large language model is used to generate prompts based on user profiles, comprehensively analyzing user attribute data, historical trajectory sequence data, and base station semantic representations from multiple base stations to generate a user semantic representation of the target user, including: User attribute data, historical trajectory sequence data, and base station semantic representations from multiple base stations are converted into a comprehensive structured text description. The system obtains user profiles to generate prompts, and inputs these prompts and a comprehensive structured text description into a large language model to generate a semantic representation of the target user.

[0053] Specifically, user attribute data, historical trajectory sequence data, and the semantic representations of multiple base stations corresponding to each base station arranged chronologically within the historical trajectory sequence data are structurally integrated and transformed into a standardized and orderly text input format to obtain a comprehensive structured text description. Here, for the historical trajectory sequence data, the trajectories for each day are serialized chronologically. Taking a day as an example, from 0:00 to 24:00, with a 1-hour time interval, the base station identifiers corresponding to the user's location at each time point are recorded sequentially, forming a trajectory sequence text similar to [0:00 location - base station A, 1:00 location - base station B... 24:00 location - base station N]. At the same time, the generated base station semantic representations are added to the associated base stations.

[0054] The comprehensive structured text description is concatenated with prompts generated from user profiles to form the input text for the large language model. The large language model performs joint semantic understanding and deep feature mining on this input text to extract high-level semantic information such as user activity patterns, scene preferences, and behavioral patterns. Finally, it generates a user semantic representation that integrates user attributes, trajectory temporal features, and base station scene semantics. This user semantic representation can accurately reflect user interests, preferences, behavioral patterns, and other characteristics.

[0055] The process by which a large language model generates prompts and comprehensive structured text descriptions for user profiles can be represented as follows:

[0056] in, Represents the user's semantic representation; This refers to user profile generation prompts, specifically pre-designed prompts tailored to the user. This represents a comprehensive, structured text description. Represents a large language model.

[0057] In some embodiments, user semantic representation and user collaborative filtering representation are fused through multi-layer interaction to construct a user representation of the target user, including: The user semantic representation and the user collaborative filtering representation are concatenated to obtain the fused input features; A pre-trained multi-layer interactive fusion network is obtained, and the fused input features are input into the multi-layer interactive fusion network for multi-layer feature extraction and nonlinear transformation to construct a user representation of the target user.

[0058] Specifically, firstly, the user semantic representation and the user collaborative filtering representation are concatenated to obtain fused input features. The user semantic representation is a high-dimensional semantic feature vector obtained through semantic understanding using a large language model, while the user collaborative filtering representation is a temporal behavioral feature vector obtained through sequence recommendation and collaborative filtering processing. These two representations are concatenated and fused along their feature dimensions to form fused input features that simultaneously contain semantic and collaborative filtering behavioral information. This initial fusion of the two heterogeneous representations preserves the effective information of each representation, avoids the information loss problem inherent in single representations, and provides a complete feature foundation for subsequent deep fusion.

[0059] This disclosure allows for the pre-training of a multi-layer interactive fusion network to promote deep interaction between collaborative filtering-side representations and semantic representations, generating more in-depth and accurate user and base station representations. The fused input features are then fed into the multi-layer interactive fusion network for multi-layer feature extraction and nonlinear transformation to construct the user representation of the target user.

[0060] Here, the multi-layer interactive fusion network can be constructed using structures such as multi-layer perceptron, attention layer, or cross-interaction layer. Through multi-layer neural networks, the spliced ​​fusion input features are subjected to layer-by-layer feature interaction, weight learning, and nonlinear transformation. The network adaptively learns the correlation and complementary information between user semantic representation and user collaborative filtering representation, and outputs a user representation of the target user that integrates high-level semantic features and temporal behavioral features, thereby improving the expressive power and discriminative power of the user representation.

[0061] In some embodiments, the multi-layer interactive fusion network is a multi-layer perceptron, which includes multiple fusion layers. During the training phase, the multi-layer perceptron constructs a loss function based on the mean square error and mean absolute error of user trajectory prediction, and evaluates the performance of the outputs of multiple fusion layers on a validation set. The fusion layer with the optimal loss function value is selected as the output layer of the multi-layer perceptron.

[0062] Specifically, in the user representation fusion stage, embodiments of this disclosure may employ a multilayer perceptron. As a multi-layer interactive fusion network, this multi-layer perceptron It can be composed of multiple sequentially connected fusion layers. Each fusion layer uses a non-linear activation function to perform linear and non-linear transformations on the fused input features, realizing feature interaction and feature purification at different granularities, and can automatically mine high-order correlation information between two heterogeneous representations, thereby improving the expressive and generalization capabilities of user representations.

[0063] In the process of the above linear and nonlinear transformations, the user semantic representation in the input features is fused. Collaborative filtering representation with users The information carried is intertwined, multilayer sensor It can learn complex relationships between information from different sources. The specific calculation formula is as follows:

[0064] ...

[0066] in, The first step in the user representation fusion process The output of the hidden layer, The first step in the user representation fusion process The weight matrix of the hidden layer, The first step in the user representation fusion process The bias vector of the hidden layer, The first step in the user representation fusion process Activation function of hidden layer.

[0067] To determine the optimal fusion layer, the model performance of different fusion layers is evaluated on a validation set during the training phase of the multilayer perceptron. This embodiment of the present disclosure can construct a loss function based on the user trajectory prediction task. The loss function uses the mean squared error (MSE) and mean absolute error (MAE) between the predicted trajectory and the true trajectory as optimization objectives, enabling the multilayer perceptron to balance the overall level of prediction error and the suppression of abnormal biases during training, thereby improving the stability and reliability of user representation learning.

[0068] After each round of training, the validation set is input into the multilayer perceptron, and the user trajectory prediction performance of the outputs of multiple fusion layers is evaluated, and the corresponding loss function values ​​are calculated. The fusion layer with the optimal loss function value and the best trajectory prediction effect (i.e., the smallest sum of MSE and MAE) is selected as the final output layer of the multilayer perceptron, and the features corresponding to the output layer are the user representation of the target user.

[0069] Assume the first If the loss function value of the fusion layer is minimized, then through the fusion layer... The computation of each fusion layer yields the user representation. ,Right now .

[0070] In some embodiments, the base station representation fusion stage adopts the same approach as the user representation fusion stage, combining the base station semantic representation... Collaborative filtering characterization with base stations Input to multilayer perceptron In the middle, calculate according to the following formula:

[0071] ...

[0073] in, The first step in the base station characterization fusion process. The output of the hidden layer, The first step in the base station characterization fusion process. The weight matrix of the hidden layer, The first step in the base station characterization fusion process. The bias vector of the hidden layer, The first step in the base station characterization fusion process. The activation function of the hidden layer. This ultimately yields the base station fusion representation. : .

[0074] Multi-layer interactive fusion, leveraging the powerful feature learning capabilities of multi-layer perceptrons, can deeply explore the potential connections between collaborative filtering and semantic representation information, enabling the model to learn more complex patterns. This provides more valuable support for user trajectory prediction and profile construction in fields such as culture and tourism, emergency response, finance, and anti-fraud. For example, when faced with user trajectory data containing rich semantic information and large-scale collaborative filtering data, it can not only consider the sequence information of user visits to base stations but also combine information about user interests, preferences, and behavioral patterns in semantic representations for comprehensive analysis at different levels. This approach can better adapt to different types and scales of data, improving the model's adaptability and generalization ability under various data conditions.

[0075] In practical applications, precise user and base station representations generated through multi-layered interaction fusion can be applied to various fields. In the cultural tourism sector, based on precise user representations, scenic spots can gain a deeper understanding of tourists' interests, preferences, and behavioral patterns. For example, by analyzing user representations, scenic spots can discover that some tourists have a strong preference for historical and cultural sites and tend to visit popular attractions in the morning and less-known but distinctive locations in the afternoon. Based on this information, scenic spots can then prioritize opening popular historical and cultural sites in the morning and allocate tour guide resources accordingly; in the afternoon, they can selectively open less-known sites and launch unique cultural experience activities to attract tourists who prefer these types of attractions, thereby improving the overall tourist experience and increasing the scenic spot's attractiveness and reputation.

[0076] In the financial sector, financial institutions can more accurately assess user credit risk by utilizing user profiles. By combining basic user characteristics with information such as consumption patterns and job stability reflected in behavioral patterns, financial institutions can construct more precise credit assessment models. For example, when a user's behavioral patterns indicate that they reside in upscale communities, have a stable workplace, and frequently visit high-end consumption venues on weekends, these behavioral characteristics can serve as important evidence reflecting their strong spending power and stable income in a credit assessment model. Based on this information, financial institutions can rationally plan loan amounts to ensure that business matches users' actual debt repayment ability, significantly improving the success rate of lending, while effectively reducing the probability of non-performing loans. This optimizes the allocation of financial resources and promotes the sustainable development of financial business within a sound risk management framework.

[0077] In the field of emergency response, relevant departments can leverage user representations to track the movement and distribution of people in disaster-stricken areas. By analyzing base station information connected by users at different times and semantic data reflecting the user's environment and needs, the location and number of affected people can be quickly determined, enabling timely allocation of relief supplies and personnel. For example, after disasters such as earthquakes or floods, user representations can be used to determine which areas have concentrated populations and urgently need relief supplies, and which areas have people with mobility impairments or special needs, thereby achieving precise rescue and improving emergency response efficiency.

[0078] In trajectory prediction scenarios, for users who commute daily, the system can predict their travel routes over a future period by combining historical travel patterns, time regularities, and other semantic information included in their user profile, such as workplace and home address. For example, an office worker who regularly travels from home to work every weekday morning, passing through several specific locations, can have their travel trajectory accurately predicted on future weekday mornings by analyzing their user profile, allowing the system to plan the optimal route in advance and avoid congested areas.

[0079] The foregoing mainly describes the solutions provided by the embodiments of this disclosure. It is understood that, in order to achieve the above functions, the electronic device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this disclosure can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0080] This disclosure embodiment can divide the electronic device into functional units according to the above method example. For example, each function can be divided into a separate functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in this disclosure embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0081] In the case of dividing each functional module according to its corresponding functions, an exemplary embodiment of this disclosure provides a user representation construction apparatus, which may be a terminal or a chip applied to the terminal. Figure 2 A schematic diagram of the structure of a user representation construction apparatus provided in an exemplary embodiment of this disclosure is shown. Figure 2 As shown, the device 200 includes: The acquisition module 201 is used to acquire user attribute data and historical trajectory sequence data of the target user, as well as base station attribute data of multiple base stations; wherein, the historical trajectory sequence data includes the identification information of the multiple base stations accessed by the target user in chronological order within a preset time period; The semantic understanding module 202 is used to perform semantic understanding on the user attribute data, the historical trajectory sequence data and the base station attribute data of the multiple base stations using a large language model, and generate a user semantic representation of the target user. Collaborative filtering module 203 is used to perform collaborative filtering processing on the historical trajectory sequence data using a sequence recommendation model to obtain the user collaborative filtering representation of the target user; The multi-layer interaction fusion module 204 is used to perform multi-layer interaction fusion of the user semantic representation and the user collaborative filtering representation to construct the user representation of the target user.

[0082] In some embodiments, the semantic understanding module 202 is further configured to use a large language model to perform semantic understanding on the base station attribute data of each base station based on base station semantic prompts, and generate base station semantic representations of each base station; and use the large language model to generate prompts based on user profiles to perform comprehensive analysis on the user attribute data, the historical trajectory sequence data and the base station semantic representations of the multiple base stations, and generate user semantic representations of the target user.

[0083] In some embodiments, the semantic understanding module 202 is further configured to convert the base station attribute data of each base station into a base station structured text description; obtain base station semantic prompts, and input the base station semantic prompts and the base station structured text descriptions into a large language model to generate a base station semantic representation for each base station.

[0084] In some embodiments, the semantic understanding module 202 is further configured to convert the user attribute data, the historical trajectory sequence data, and the base station semantic representations of the multiple base stations into a comprehensive structured text description; obtain user profile generation prompts, and input the user profile generation prompts and the comprehensive structured text descriptions into the large language model to generate the user semantic representation of the target user.

[0085] In some embodiments, the multi-layer interaction fusion module 204 is further configured to concatenate the user semantic representation with the user collaborative filtering representation to obtain fused input features; obtain a pre-trained multi-layer interaction fusion network, and input the fused input features into the multi-layer interaction fusion network for multi-layer feature extraction and nonlinear transformation to construct the user representation of the target user.

[0086] In some embodiments, the multi-layer interactive fusion network is a multi-layer perceptron, which includes multiple fusion layers. During the training phase, the multi-layer perceptron constructs a loss function based on the mean square error and mean absolute error of user trajectory prediction, and evaluates the performance of the outputs of the multiple fusion layers on a validation set, selecting the fusion layer with the optimal loss function value as the output layer of the multi-layer perceptron.

[0087] This disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the methods disclosed in this disclosure.

[0088] Figure 3 A schematic diagram of the structure of an electronic device provided in an exemplary embodiment of this disclosure is shown. For example... Figure 3 As shown, the electronic device 300 includes at least one processor 301 and a memory 302 coupled to the processor 301, which can perform the corresponding steps in the methods disclosed in the embodiments of this disclosure.

[0089] The processor 301 described above can also be called a Central Processing Unit (CPU), which can be an integrated circuit chip with signal processing capabilities. Each step in the method disclosed in this embodiment can be implemented by the integrated logic circuitry in the processor 301 or by software instructions. The processor 301 can be a general-purpose processor, a digital signal processor (DSP), an ASIC, a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this embodiment can be directly implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can be located in the memory 302, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor 301 reads information from the memory 302 and, in conjunction with its hardware, completes the steps of the method described above.

[0090] Furthermore, various operations / processes according to this disclosure, implemented via software and / or firmware, can be transmitted from a storage medium or network to a computer system with a dedicated hardware architecture, for example, Figure 4 The computer system 400 shown is equipped with the programs that constitute the software. When various programs are installed, the computer system is able to perform various functions, including functions such as those described above. Figure 4 A schematic diagram of the structure of a computer system provided in an exemplary embodiment of this disclosure is shown.

[0091] Computer system 400 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0092] like Figure 4As shown, the computer system 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the computer system 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0093] Multiple components in the computer system 400 are connected to the I / O interface 405, including: an input unit 406, an output unit 407, a storage unit 408, and a communication unit 409. The input unit 406 can be any type of device capable of inputting information into the computer system 400. The input unit 406 can receive input numerical or character information and generate key signal inputs related to user settings and / or function control of the electronic device. The output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a monitor, speaker, video / audio output terminal, vibrator, and / or printer. The storage unit 408 may include, but is not limited to, a hard disk and an optical disk. The communication unit 409 allows the computer system 400 to exchange information / data with other devices via a network such as the Internet, and may include, but is not limited to, a modem, network card, infrared communication device, wireless communication transceiver, and / or chipset, such as Bluetooth™ device, WiFi device, WiMax device, cellular communication device, and / or the like.

[0094] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above. For example, in some embodiments, the methods disclosed in this disclosure can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 can be configured to perform the methods disclosed in this disclosure by any other suitable means (e.g., by means of firmware).

[0095] This disclosure also provides a computer-readable storage medium, wherein when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is able to perform the methods disclosed in this disclosure.

[0096] The computer-readable storage medium in this disclosure can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. The aforementioned computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specifically, the aforementioned computer-readable storage medium may include electrical connections based on one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0097] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0098] This disclosure also provides a computer program product, including a computer program, wherein when the computer program is executed by a processor, it implements the methods disclosed in the embodiments of this disclosure.

[0099] In embodiments of this disclosure, computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network (including a local area network (LAN) or a wide area network (WAN)), or it can be connected to an external computer.

[0100] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0101] The modules, components, or units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the modules, components, or units do not necessarily constitute a limitation on the module, component, or unit itself.

[0102] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary hardware logic components that can be used include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0103] The above description is merely an illustration of some embodiments of this disclosure and the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0104] While specific embodiments of this disclosure have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of this disclosure. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of this disclosure. The scope of this disclosure is defined by the appended claims.

Claims

1. A method for constructing user representations, characterized in that, include: The system acquires user attribute data and historical trajectory sequence data of the target user, as well as base station attribute data of multiple base stations; wherein, the historical trajectory sequence data includes the identification information of the multiple base stations accessed by the target user in chronological order within a preset time period; Using a large language model, semantic understanding is performed on the user attribute data, the historical trajectory sequence data, and the base station attribute data of the multiple base stations to generate a user semantic representation of the target user; The historical trajectory sequence data is processed by a sequence recommendation model to obtain a user collaborative filtering representation of the target user. The user semantic representation and the user collaborative filtering representation are fused together through multi-layer interaction to construct the user representation of the target user.

2. The method as described in claim 1, characterized in that, The step of using a large language model to perform semantic understanding on the user attribute data, the historical trajectory sequence data, and the base station attribute data of the multiple base stations to generate a user semantic representation of the target user includes: Using a large language model, semantic understanding of the base station attribute data of each base station is performed based on base station semantic prompts, and base station semantic representations of each base station are generated. The large language model is used to generate prompts based on user profiles and comprehensively analyze the user attribute data, the historical trajectory sequence data, and the semantic representations of the base stations of the multiple base stations to generate the user semantic representation of the target user.

3. The method as described in claim 2, characterized in that, The step of using a large language model to perform semantic understanding of the base station attribute data of each base station based on base station semantic cue words, and generating base station semantic representations for each base station, includes: The base station attribute data of each of the base stations is converted into a structured text description of the base station; Obtain semantic prompts for base stations, and input the semantic prompts and the structured text descriptions of the base stations into a large language model to generate semantic representations of each base station.

4. The method as described in claim 2, characterized in that, The process of generating a user semantic representation for the target user by comprehensively analyzing the user attribute data, the historical trajectory sequence data, and the base station semantic representations of multiple base stations using the large language model based on user profiles includes: The user attribute data, the historical trajectory sequence data, and the base station semantic representations of the multiple base stations are converted into a comprehensive structured text description; The system obtains user profile-generated prompt words and inputs the user profile-generated prompt words and the comprehensive structured text description into the large language model to generate the user semantic representation of the target user.

5. The method according to any one of claims 1 to 4, characterized in that, The step of fusing the user semantic representation with the user collaborative filtering representation through multi-layer interaction to construct the user representation of the target user includes: The user semantic representation and the user collaborative filtering representation are concatenated to obtain the fused input features; A pre-trained multi-layer interactive fusion network is obtained, and the fusion input features are input into the multi-layer interactive fusion network for multi-layer feature extraction and nonlinear transformation to construct the user representation of the target user.

6. The method as described in claim 5, characterized in that, The multi-layer interactive fusion network is a multi-layer perceptron, which includes multiple fusion layers. During the training phase, the multi-layer perceptron constructs a loss function based on the mean square error and mean absolute error of user trajectory prediction, and evaluates the performance of the outputs of the multiple fusion layers on the validation set. The fusion layer with the optimal loss function value is selected as the output layer of the multi-layer perceptron.

7. A user representation construction apparatus, characterized in that, include: The acquisition module is used to acquire user attribute data and historical trajectory sequence data of the target user, as well as base station attribute data of multiple base stations; wherein, the historical trajectory sequence data includes the identification information of the multiple base stations accessed by the target user in chronological order within a preset time period; The semantic understanding module is used to perform semantic understanding on the user attribute data, the historical trajectory sequence data and the base station attribute data of the multiple base stations using a large language model, and generate a user semantic representation of the target user. The collaborative filtering module is used to perform collaborative filtering processing on the historical trajectory sequence data using a sequence recommendation model to obtain the user collaborative filtering representation of the target user. The multi-layer interaction fusion module is used to perform multi-layer interaction fusion of the user semantic representation and the user collaborative filtering representation to construct the user representation of the target user.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1 to 6.

Images