Gender identification method and device for online car-hailing driver and electronic equipment

Abstract

The invention relates to the technical field of data processing, in particular to an online car-hailing driver gender recognition method and device and electronic equipment, and is used for accurately and efficiently carrying out gender recognition on online car-hailing drivers. The operation data of the multiple modal types comprises the operation data of the multiple modal types. Performing feature recognition on the operation data of each modal type to obtain an initial gender prediction score of each modal type and a corresponding initial confidence coefficient; then, based on the data acquisition rate of each modal type, correcting the initial confidence coefficient of each modal type to obtain a first target confidence coefficient corresponding to each modal type; and finally, based on the initial gender prediction score of each modal type and the first target confidence, determining a gender prediction result of the target driver.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, device and electronic device for gender identification of ride-hailing drivers. Background Technology

[0002] Currently, with the rapid development of ride-hailing platforms and the increasing demand for refined operations, platforms are placing higher demands on the accuracy of driver profiles. Among these, driver gender, as a key attribute in driver profiles, not only relates to the realization of personalized services but also directly impacts core business processes such as intelligent dispatching, resource matching, and user experience optimization.

[0003] However, existing technologies generally rely on the gender information that drivers fill in during registration, which is often incomplete, incorrect, or unverified. Although some ride-hailing platforms use regular random facial recognition checks to verify drivers' gender, facial verification consumes significant computing and storage resources, making it difficult to achieve high-frequency, full-coverage real-time verification on a scale of millions of drivers.

[0004] Meanwhile, due to limitations in sampling strategies and triggering conditions, a large number of drivers remain unverified for extended periods, resulting in issues such as high false negative rates, delayed updates, and unreliable identification results, which fail to meet the platform's business requirements for real-time and accurate profiles. Summary of the Invention

[0005] This application provides a method, device, and electronic device for accurately and efficiently identifying the gender of ride-hailing drivers. The technical solution of this application is as follows: In a first aspect, embodiments of this application provide a method for gender identification of ride-hailing drivers, the method comprising: Acquire operational data of multiple modalities generated by the target driver within a specified time period; Feature identification is performed on the operational data of each modality type to obtain an initial gender prediction score and a corresponding initial confidence level for each modality type; Based on the data acquisition rate of each modality type, the initial confidence level of each modality type is corrected to obtain the first target confidence level corresponding to each modality type; wherein, the data acquisition rate of any modality type is determined based on the amount of operational data of that modality type; Based on the initial gender prediction score for each modality and the first target confidence level, the gender prediction result for the target driver is determined.

[0006] Secondly, embodiments of this application provide a gender recognition device for ride-hailing drivers, the device comprising: The data acquisition unit is configured to acquire operational data of multiple modalities generated by the target driver within a specified time period; The initial prediction unit is configured to: perform feature recognition on the operational data of each modality type to obtain an initial gender prediction score for each modality type, and a corresponding initial confidence level; The data correction unit is configured to: correct the initial confidence level of each modality based on the data acquisition rate of each modality to obtain a first target confidence level corresponding to each modality; the data acquisition rate of any modality is determined based on the amount of operational data of that modality; The gender prediction unit is configured to: determine the gender prediction result of the target driver based on the initial gender prediction score for each modality type and the first target confidence level. Thirdly, embodiments of this application provide an electronic device, including: Memory, used to store program instructions; A processor is configured to invoke program instructions stored in the memory and execute the steps of the method described in any one of the first aspects according to the obtained program instructions.

[0007] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the method described in any one of the first aspects.

[0008] Fifthly, embodiments of this application provide a computer program product comprising: computer program code, which, when executed on a computer, causes the computer to perform the method described in any of the first aspects.

[0009] In this embodiment, feature identification is performed on the operational data of the target driver in multiple modalities within a specified time period to obtain an initial gender prediction score and corresponding initial confidence level for each modality. Based on this, the initial confidence level is corrected by combining the data collection rate of each modality to obtain a first target confidence level that is closer to the actual data quality.

[0010] This effectively suppresses the problem of inflated confidence levels caused by sparse, missing, or noisy data in certain modalities, making the final fusion result more realistically reflect the driver's actual gender attributes and improving the accuracy of gender identification for the target driver.

[0011] Finally, the initial gender prediction scores for each modality are fused with the corrected first target confidence score to determine the gender prediction result for the target driver. This process avoids reliance on a single information source (such as registration information) or high-cost verification methods (such as facial recognition), enabling efficient and accurate gender identification for a massive driver population without increasing system load. This enhances the ride-hailing platform's decision-making capabilities in core business scenarios such as safe service matching and dynamic capacity scheduling.

[0012] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0013] 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: Figure 1 An application scenario diagram of the gender recognition method for ride-hailing drivers provided in the embodiments of this application; Figure 2 A flowchart illustrating the overall process of a ride-hailing driver gender identification method provided in this application embodiment; Figure 3 A schematic diagram of the inference process of the initial prediction model provided in the embodiments of this application; Figure 4 A schematic diagram of the inference process of the adaptive weight model provided in the embodiments of this application; Figure 5 Another flowchart of the method for identifying the gender of ride-hailing drivers provided in the embodiments of this application; Figure 6 A schematic diagram of the structure of the gender recognition device for ride-hailing drivers provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0016] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0017] The design concept of the embodiments of this application is briefly introduced below: As mentioned earlier, in the intelligent dispatch and personalized services of ride-hailing platforms, driver gender is a key attribute in driver profiles. It not only relates to the realization of personalized services, but also directly affects core business aspects such as intelligent dispatch, resource matching, and user experience optimization.

[0018] For example, some female passengers prefer female drivers (especially at night or for longer trips) to feel safer and more comfortable. Based on reliable gender profiles, the platform proactively prioritizes matching these passengers with female drivers, significantly improving user experience and enhancing platform trust.

[0019] For example, in certain scenarios that require a high level of physical strength or vehicle control (such as assisting with luggage, transporting large equipment, long-distance orders on off-road terrain or in inclement weather), some passengers may prefer services from male drivers. Based on reliable gender profiles, the platform intelligently recommends drivers with a higher match to passengers with such needs, thereby improving service efficiency and user satisfaction.

[0020] However, existing technologies generally rely on the gender information that drivers fill in during registration, which is often incomplete, incorrect, or unverified. Although some ride-hailing platforms use regular random facial recognition checks to verify drivers' gender, facial verification consumes significant computing and storage resources, making it difficult to achieve high-frequency, full-coverage real-time verification on a scale of millions of drivers.

[0021] In view of this, this application provides a method for gender identification of ride-hailing drivers. In this method, feature identification is performed on the operational data of the target driver in multiple modalities within a specified time period to obtain an initial gender prediction score and a corresponding initial confidence level for each modality.

[0022] Based on this, the initial confidence level is corrected by combining the data collection rate of each modality (i.e., the amount of operational data actually collected under that modality) to obtain a first target confidence level that is closer to the actual data quality. In this way, the problem of inflated confidence levels caused by data sparsity, missing data, or noise interference in certain modalities can be effectively suppressed, making the final fusion result more realistically reflect the actual gender attributes of the driver and improving the accuracy of gender identification of the target driver.

[0023] Finally, the initial gender prediction scores for each modality are fused with the corrected first target confidence score to determine the gender prediction result for the target driver. This process avoids reliance on a single information source (such as registration information) or high-cost verification methods (such as facial recognition), enabling efficient and accurate gender identification for a massive number of drivers without increasing system burden. This enhances the ride-hailing platform's decision-making capabilities in core business scenarios such as safe service matching and dynamic capacity scheduling.

[0024] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0025] like Figure 1 As shown, it is a schematic diagram of an application scenario in an embodiment of this application, which includes: driver terminal device 110 and ride-hailing platform server 120.

[0026] In this application embodiment, the driver-end device 110 includes, but is not limited to, mobile or embedded devices such as smartphones, in-vehicle smart terminals, and tablets that can access the ride-hailing platform and perform functions such as order acceptance, navigation, voice calls, and message interaction.

[0027] Server 120 is the backend system supporting the core business of the ride-hailing platform. Server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing, big data processing, and artificial intelligence model services. This application does not make any specific limitations in this regard.

[0028] It should be noted that the ride-hailing driver gender recognition method in the various embodiments of this application can be executed by an electronic device, which can be the driver-side device 110 or the server 120. That is, the method can be executed by the server 120 alone, or it can be executed jointly by the driver-side device 110 and the server 120.

[0029] For example, in a scenario where driver-side device 110 and server 120 collaborate, driver-side device 110 continuously collects and locally caches multimodal operational data during the driver's daily operations. This data includes, but is not limited to: driver order-taking behavior logs (such as online time periods, order completion rates, and cancellation rates), text chat records with passengers, audio clips from voice calls, and basic device information. Driver-side device 110 can periodically upload the aforementioned multimodal data to the ride-hailing platform's server 120.

[0030] When there is a business need to identify the gender of a target driver, server 120 can acquire operational data of multiple modalities generated by the target driver within a specified time period, and obtain an initial gender prediction score and corresponding initial confidence level for each modality by performing feature recognition on the operational data of each modality. Subsequently, server 120 corrects the initial confidence level of each modality based on the data collection rate of each modality to obtain a first target confidence level for each modality. Then, based on the initial gender prediction score and the first target confidence level of each modality, the gender prediction result of the target driver is determined.

[0031] In one alternative implementation, the server 120 may use the gender prediction result to construct a dynamic driver profile and combine it with passenger preferences (such as female passengers tending to be served by female drivers in nighttime orders) for accurate matching during the order scheduling stage, thereby improving service safety and user experience.

[0032] In one alternative implementation, the driver-side device 110 and the server 120 can communicate via a communication network. The communication network can be a wired network or a wireless network, such as 4G, 5G, Wi-Fi, Bluetooth, etc.

[0033] It should be noted that, Figure 1 The examples shown are merely illustrative; in reality, the number of driver-side devices and servers is unlimited and not specifically limited in this application embodiment. For instance, when there are multiple servers, these servers can form a blockchain, with each server acting as a node on the blockchain. Key data such as multimodal data, gender recognition results, and confidence correction records disclosed in this application embodiment can be stored on the blockchain to enhance data traceability and tamper-proof capabilities.

[0034] The specific implementation of this application involves data collection and use, such as operational data of various modalities and driver registration information. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the relevant parties is required, and the collection, storage, processing and analysis of all data shall strictly comply with the privacy protection laws and regulations and data security standards of the relevant countries and regions.

[0035] See Figure 2 The diagram shown is an overall flowchart of a method for gender identification of ride-hailing drivers provided in this application embodiment, specifically including the following steps: S201, Obtain operational data of multiple modalities generated by the target driver within a specified time period; In this application embodiment, the operational data of multiple modalities includes: order data, text data generated from communication with passengers, and voice data.

[0036] The order data in this application embodiment includes, but is not limited to: the number of orders received within a specified time period, order completion rate and cancellation rate, average order duration, average order revenue, and percentage of working hours. This type of data is used to describe the driver's behavioral patterns and time preference characteristics during operation. Since drivers of different genders often have identifiable differences in the distribution of driving hours, service continuity, and order-taking behavior patterns, this type of data can serve as strong support for determining the driver's gender.

[0037] The text data in this application includes, but is not limited to: average text length and text quantity within a specified time period, frequency of words reflecting male and female preferences, and data on the use of emoticons (e.g., the frequency of using cute or formal emoticons). This type of data is used to describe the language style and emotional expression tendencies exhibited by drivers in their communication with passengers. Since drivers of different genders often have identifiable differences in word choice, tone of voice, and emotional expression, this type of data can serve as strong support for determining the gender of drivers.

[0038] The speech data in this application includes, but is not limited to: speech rate characteristics within a specified time period, mean value of Mel-frequency cepstral coefficients, variance of Mel-frequency cepstral coefficients, and fundamental frequency related data (such as fundamental frequency range, fundamental frequency variance, and fundamental frequency mean). This type of data is obtained by extracting acoustic features from recordings of conversations between drivers and passengers during order services, and is used to characterize the driver's voiceprint characteristics. Since drivers of different genders often have identifiable differences in their vocal physiological structure and speech expression habits, this type of data can serve as strong support for determining the driver's gender.

[0039] S202, perform feature recognition on the operational data of each modality type to obtain the initial gender prediction score and the corresponding initial confidence level for each modality type; In this embodiment of the application, an initial prediction model is pre-constructed for gender prediction of operational data for each modality type. The initial prediction model has multiple classification heads, each corresponding to a modality type, for gender prediction of operational data for that modality type.

[0040] During the sample construction phase, training samples are pre-built based on massive historical driver data within the ride-hailing platform. In specific implementation, historical operational data of various modalities generated by drivers from each platform within historical time periods can be extracted. The modalities of the historical operational data must be consistent with the modalities of the operational data in the aforementioned step S201; that is, the historical operational data of various modalities must include historical order data, historical text data, and historical voice data.

[0041] Based on this, for each modality type, the historical operational data of that modality type is used as the sample content, and the real gender verified by the platform is used as the sample label, thereby constructing the training sample for that modality type.

[0042] During the model training phase, training samples corresponding to each modality type are input into the corresponding classifiers in the initial prediction model for joint or independent training. Each classifier uses historical operational data of that modality type as input, performs feature encoding and discrimination learning through a neural network structure, and finally outputs two continuous values: the initial gender prediction score and the initial confidence level.

[0043] The initial gender prediction score represents the probability that the model judges the driver to be a certain gender, with a value range of [0,1]. The initial confidence score reflects the reliability of the model's current prediction result, and is also a value in the range of [0,1], rather than a hard category label "male" or "female".

[0044] During training, verified real gender values ​​from the platform are used as a supervision signal. Combined with a specific loss function (such as confidence-weighted binary cross-entropy loss or basis composite loss), the model is guided to simultaneously optimize prediction accuracy and confidence reasonableness. This ensures that each classifier not only provides a quantitative assessment of gender preference (i.e., the initial gender prediction score) but also adaptively assesses its own credibility level under the current data conditions (i.e., the initial confidence level).

[0045] In an alternative implementation, scenario description data can be introduced into the training samples during the training phase to enhance the adaptability of the initial prediction model to complex business environments and improve gender discrimination accuracy.

[0046] Specifically, drivers' behavioral patterns, language styles, and speech expressions may differ significantly across different service scenarios. For example, they may speak more cautiously in complaint handling scenarios, use more casual language in casual conversations, and speak faster when confirming orders at night. If these contextual factors are ignored, the model may misclassify scenario-specific features as gender-related features, leading to decreased generalization performance. However, by introducing scenario description data, the model can learn gender discrimination rules based on scenario awareness, thereby improving the accuracy and robustness of predictions.

[0047] The scenario description data in this application embodiment is used to describe the business categories produced by the data source (i.e., text and voice), such as "order confirmation", "route consultation", "complaint handling", "fee dispute", "casual chat" and so on.

[0048] The aforementioned scenario description data can be obtained through automatic semantic analysis of historical text and speech data. In practice, keyword matching rules and pre-trained language models (such as BERT) can be used to perform intent recognition and sentiment classification on historical chat text and speech-to-text. This is then combined with auxiliary information such as the time of data source output (e.g., whether it is nighttime), order status (e.g., whether it is close to cancellation), and passenger feedback to comprehensively determine the business scenario to which the current interaction belongs. Finally, the generated scenario description data is added as supplementary contextual features to the training samples to guide the initial prediction model to more accurately capture gender-related discriminative patterns in different business contexts, thereby improving the model's generalization ability and prediction stability in real-world complex scenarios.

[0049] Next, as follows Figure 3 As shown, when performing step S202 above, operational data for each modality can be input into the trained initial prediction model. The model will pre-process the text and speech data for semantic recognition to obtain scene description data.

[0050] The scene description data is then fused with the operational data of each modality to obtain the first fused feature (order data + scene description data), the second fused feature (text data + scene description data), and the third fused feature (voice data + scene description data).

[0051] Finally, the three sets of fused features are processed by the classification heads corresponding to each modality type in the model: each classification head performs forward computation independently, and outputs the initial gender prediction score (P1 / P2 / P3) and the corresponding initial confidence (C1 / C2 / C3) for each modality type.

[0052] S203, based on the data acquisition rate of each modality type, the initial confidence level of each modality type is corrected to obtain the first target confidence level corresponding to each modality type; wherein, the data acquisition rate of any modality type is determined based on the amount of operational data of the modality type; As mentioned earlier, this application sets corresponding data collection dimensions for each modality type. For example, the data collection dimensions for order data may include: the number of orders received by the target driver within a specified time period, order completion rate and cancellation rate, average order duration, average order revenue, and percentage of working hours.

[0053] The data collection rate for each modality type represents the proportion of valid data actually collected under that modality type. Taking order data as an example, assuming its data collection dimension is 10, and 8 order data are actually collected, then the data collection rate is 8 / 10 = 0.8.

[0054] Because missing or sparse data can lead to unreliable confidence assessments of model outputs, this embodiment uses the data collection rate to correct the initial confidence level for each modality type, thereby obtaining a first target confidence level that more closely reflects the actual data quality.

[0055] In one alternative implementation, the weighted features corresponding to the target driver can be obtained by concatenating the data acquisition rate and initial confidence level of each modality type. Then, feature recognition is performed on these weighted features to obtain the first target confidence level for each modality type.

[0056] In practice, an adaptive weight model can be trained to implement the correction process for the initial confidence level.

[0057] In the sample construction phase, training samples for training the adaptive weight model can be constructed using the already trained initial prediction model. First, the training samples of the already trained initial prediction model are reused, and the historical operating data of each historical driver for multiple modalities are input into the already trained initial prediction model to obtain the initial prediction score and initial confidence level for each modality.

[0058] Then, the data collection rate and initial confidence level for each modality are concatenated into a long vector, i.e., the weight vector, according to a preset order. This weight vector is used as the sample content for training the adaptive weight model, and the real gender of the historical driver is used as the sample label, thereby constructing the training samples for the adaptive weight model.

[0059] During the model training phase, the aforementioned weight vectors are input into the adaptive weight model. Multi-layer nonlinear transformations are used to perform feature recognition on the weight vectors, outputting the corrected confidence score for each modality type (order / text / voice). During training, the corrected confidence score for each modality type is weighted and fused with the corresponding initial gender prediction score. Then, the binary cross-entropy between the weighted fusion result and the sample label representing the true gender is used as the loss function to perform end-to-end optimization of the adaptive weight model. The resulting adaptive weight model learns how to jointly determine the reliability of each modality type based on the initial confidence score and the data collection rate.

[0060] During the model inference phase, this adaptive weight model can be directly used to dynamically correct the initial confidence level for each modality type. Specifically, as follows... Figure 4As shown, when performing step S203 above, the data acquisition rate (r1 / r2 / r3) corresponding to each modality type and the initial confidence level (C1 / C2 / C3) of each modality type can be concatenated into a weight vector Z{C1, C2, C3, r1, r2, r3}.

[0061] Then, the weight vector Z is input into the trained adaptive weight model, and the first target confidence (W) corresponding to each modality type is obtained by performing feature recognition on the weight vector Z. A1 / W A2 / W A3 In this way, without relying on manual rules, the influence of unreliable modalities caused by missing, sparse, or noisy data can be automatically suppressed, making the confidence assessment more in line with the actual data quality, thereby significantly improving the accuracy, stability, and generalization ability of multimodal gender recognition results.

[0062] S204, based on the initial gender prediction score of each modality type and the first target confidence level, determine the gender prediction result of the target driver.

[0063] In one alternative implementation, for each modality, a gender prediction score can be determined based on the product of the initial gender prediction score for that modality and the first target confidence level. Then, the gender prediction result for the target driver is determined based on the sum of the gender prediction scores for all modality types (i.e., the total gender prediction score for the target driver).

[0064] In practice, a reasonable gender judgment threshold can be set, and the total gender prediction score for each modality type can be compared with this threshold. For example, a score greater than the threshold is considered male, and a score less than the threshold is considered female; or, a score greater than the threshold is considered female, and a score less than the threshold is considered male.

[0065] The gender determination threshold can be set through joint optimization of models or through offline statistical analysis of historical driver data; this application does not limit this method.

[0066] Taking joint model optimization as an example, after training the initial prediction model and the adaptive weight model, the total gender prediction score for each driver can be calculated on an independent validation set. Then, different threshold points are iterated through, and the score that maximizes gender classification accuracy is selected as the gender judgment threshold. This process ensures that the selected threshold matches the data distribution of the total gender prediction score through collaborative optimization of the model output. After the system goes live, the gender recognition accuracy under the current threshold can be reassessed periodically using driver samples whose gender has been confirmed through real-name authentication or manual verification. The gender judgment threshold can be periodically reviewed and adjusted to ensure the accuracy and adaptability of gender judgment in different scenarios.

[0067] For ease of understanding, the above will still be used. Figure 4 Let's illustrate with an example. Suppose the gender judgment threshold is 0.7; anything above this threshold is male, and anything below is female.

[0068] After obtaining the initial gender prediction scores (P1 / P2 / P3) and the first target confidence level (W) for each modality type through steps S201~S203, A1 / W A2 / W A3 After that, the product of the initial gender score for each modality type and the first target confidence level can be used as the gender prediction score for that modality type. That is, the gender prediction score of the order data. 1 = P1 × W A1 Gender prediction score for text data 2 = P2 × W A2 Gender prediction score of voice data 3 = P3 × W A3 .

[0069] Furthermore, the gender prediction scores corresponding to each modality type are summed to obtain the total gender prediction score P for the target driver. 总 = 1+ 2+ 3. When P 总 If the value exceeds the gender determination threshold, the target driver is confirmed to be male; otherwise, they are female.

[0070] This method has a simple computational structure, consumes little computing resources, and is easy to deploy on a large scale on driver-end devices or server-side devices. At the same time, when calculating the total score for gender prediction, it introduces a first target confidence level based on data collection rate correction, which allows the contribution weight of each modality to be dynamically adjusted according to its actual data quality. This mechanism can effectively improve the accuracy and stability of gender judgment without increasing additional computational complexity, enabling the system to maintain low resource consumption and high inference efficiency while also having high gender recognition accuracy.

[0071] Furthermore, considering the high mobility of ride-hailing drivers in practical applications, newly registered drivers or those with short operating histories have relatively little accumulated operational data on the current ride-hailing platform. This results in severely sparsity in the collected operational data. Under these circumstances, the data collection rate for each modality is relatively low, directly rendering the first target confidence score corrected by the data collection rate unreliable. Consequently, the total gender prediction score calculated from the first target confidence score and the initial prediction scores for each modality will have accuracy issues.

[0072] Based on this, embodiments of this application define data completeness conditions for determining whether a target driver possesses a stable and usable foundation of multimodal behavioral data, taking into account practical application scenarios. These data completeness conditions include at least one of the following: Condition 1: The registration duration on the ride-hailing platform must reach the registration duration threshold; for example, registration must be no less than 14 days.

[0073] Condition 2: The number of orders received in the past is not less than the order quantity threshold; for example, not less than 5 orders. Condition 3: The data acquisition rate for at least one modality type is higher than the acquisition rate threshold; for example, the data acquisition rate for each modality type is not less than 0.6, or the data acquisition rate for at least two modality types is not less than 0.7.

[0074] By setting the aforementioned data completeness conditions, it is possible to effectively distinguish between drivers who have accumulated sufficient operational data and new drivers who are still in the cold start phase and have sparse operational data. This avoids relying on unreliable first-target confidence levels for fusion decisions when multimodal operational data is severely insufficient.

[0075] This mechanism no longer relies solely on the localized indicator of data collection rate for each modality. Instead, it combines objective indicators reflecting a driver's actual operational activity, such as registration duration and historical order volume, to comprehensively assess whether a driver possesses a reliable data foundation for multimodal gender prediction. This prevents gender misjudgments caused by sparse operational data while ensuring full utilization of information from each modality when multimodal operational data is abundant, thereby improving the overall accuracy, robustness, and coverage of gender identification.

[0076] Specifically, when the target driver meets the above data completeness conditions, it means that he has accumulated a sufficiently dense amount of multi-modal operational data. At this time, the gender of the target driver can be identified through the fusion weighting method described in step S204 above.

[0077] Correspondingly, when the target driver does not meet the above data completeness conditions, it indicates that they are in the cold start phase, and the operational data of multimodal types is sparse. In this case, prior information needs to be introduced to assist in prediction. During implementation, a second target confidence level suitable for the cold start scenario can be dynamically generated based on the overall distribution of data collection rates for each modality type to replace the unreliable first target confidence level. At the same time, based on the proportion of reference drivers similar to the target driver in the platform in the specified registration information, a prior score reflecting the gender tendency of the group is calculated. By fusing this prior score with the initial gender prediction scores of each modality and the corrected second target confidence level, the total gender prediction score of the target driver is calculated, i.e., the gender prediction result.

[0078] This approach effectively avoids gender prediction bias caused by distortion of the first target confidence level in driver scenarios with severely sparse data. A second target confidence level is then generated based on the data collection rate, ensuring that the weights of each modality in the fusion process better reflect its actual usability during the cold start phase. Simultaneously, a prior score calculated based on a reference driver group matched with specified registration information is introduced to supplement the gender distribution prior when individual multimodal operational data is missing. The synergistic effect of these two approaches significantly improves the reliability and discriminative power of the overall gender prediction score for new drivers, thereby effectively improving gender recognition accuracy in cold start scenarios while ensuring the consistency of the overall system process.

[0079] For ease of understanding, Figure 5 This illustrates the specific implementation process for gender identification of the target driver when the target driver does not meet the data completeness requirement, as follows: Figure 5 As shown, the following steps may be included: S205, when the target driver does not meet the completeness condition several times, for each modality type, based on the proportion of the data collection rate of the modality type in the data collection rate of each modality type, determine the prior confidence of each modality type; In practice, the prior confidence level W for each modality type can be calculated using the following formula (1). p : (1) in, Let r represent the data acquisition rate of modality m, and K represent the number of each modality. This represents the sum of the data acquisition rates for each modality.

[0080] S206, Based on the prior confidence and data acquisition rate of each modality type, the first target confidence of each modality type is corrected to obtain the second target confidence corresponding to each modality type; In one alternative implementation, a prior influence factor for the target driver can be determined based on the data acquisition rate for each modality. For each modality, this prior influence factor is used to interpolate the first target confidence level and the prior confidence level for that modality to obtain a second target confidence level.

[0081] Specifically, the prior influence factor is used to represent the degree of influence of the prior score on the gender prediction result. That is, the sparser the acquired operational data, the more the gender identification of the target driver needs to rely on the prior information of the driver group of the ride-hailing platform.

[0082] In practice, a reasonable linear function (such as monotonic nonlinear compression algorithms like Sigmoid and Tanh) can be set. Under the premise that the prior influence factor is negatively correlated with the data acquisition rate, the prior influence factor can be calculated based on the data acquisition rate of each modality.

[0083] Taking the Sigmoid algorithm as an example, the prior influence factor can be calculated using the following formula (2). : (2) in, For the prior influence coefficient, These are control coefficients used to avoid invalid calculations; This represents the average data acquisition rate for each modality. This is to avoid Take the empirical value of 0, for example =0.3.

[0084] Using the above formulas (1) and (2), the prior confidence W for each modality type is obtained respectively. p and the prior influence factors of the target driver. Then, prior influence factors can be used. The first target confidence level W for each modality type A and prior confidence level W p Interpolation is performed to obtain the second target confidence level W for each modality type. B : (3) in, The first target confidence level W for each modality type A The corresponding interpolation results, This represents the sum of the interpolation results for each modality type.

[0085] In the above formula (3), W B The calculation formula represents the normalization of the interpolation results for each modality type, and the normalization result for each modality type is used as the second target confidence level W for that modality type. B This ensures that the sum of the second objective confidence scores for each modality is 1, thereby guaranteeing the reasonable allocation of weights to different modality types in subsequent processing.

[0086] S207, Based on the ratio of the number of reference drivers in the ride-hailing platform to the total number of drivers in the ride-hailing platform, determine the prior score of the target driver; wherein, any reference driver is a driver whose registration information is the same as that of the target driver, and the specified registration information includes: operating area, operating vehicle type, and part or all of the operating time period; The prior score (also known as prior probability) in this application represents the statistical tendency of gender distribution in a group of drivers with similar registration characteristics to the target driver, and serves as an important supplementary basis for gender prediction in the cold start phase.

[0087] The designated registration information in this application embodiment includes some or all of the following: operating area, operating vehicle type, and operating time period. The operating area refers to the target driver's daily activity range, which can be divided into granularities such as street blocks, administrative districts, and cities. The operating vehicle type refers to the service category or vehicle class selected by the target driver, such as special offer vehicles, premium vehicles, and business vehicles. The operating time period refers to the main time periods during which the target driver typically picks up passengers, such as daytime hours (e.g., 6:00–18:00) and nighttime hours (e.g., 18:00–6:00 the next day).

[0088] When performing step S207, the prior score of the target driver can be calculated by combining the specified registration information that is discriminative or representative from the target driver's registration information and using conditional frequency estimation or smoothed group proportion calculation methods. .

[0089] In practice, the designated registration information of the target driver can be obtained in advance, and then reference drivers with the same designated registration information as the target driver can be selected from the ride-hailing platform.

[0090] For example, if the target driver is located in City A and uses the premium vehicle option during nighttime hours, then other drivers on the ride-hailing platform who are also located in City A and use the premium vehicle option during nighttime hours can be used as reference drivers.

[0091] Then, select any gender (e.g., male) from the reference drivers and use the proportion of male drivers among the reference drivers in the ride-hailing platform as the prior score of the target driver.

[0092] For example, if there are 600,000 male drivers among the reference drivers, and a total of 3 million male drivers on ride-hailing platforms, then the prior score... =60 / 300=0.2.

[0093] It should be understood that choosing male or female as the reference gender is merely a difference in computational perspective and does not affect the validity of the prior score in gender prediction. This is because in the binary gender classification scenario involved in this application, the driver's gender has only two mutually exclusive and exhaustive categories, "male" and "female," and the sum of their corresponding probabilities is always 1.

[0094] If we use males as a reference, the resulting prior score reflects the probability that drivers on the ride-hailing platform whose registered information matches the target driver's is male. Conversely, if we use females as a reference, the resulting prior score reflects the probability that such drivers are female. The two can be converted into each other through a complementary relationship. Therefore, by determining the prior score for one category (e.g., males), the prior score for the other category (females) can be directly derived from "1 - male prior score".

[0095] Therefore, it can be seen that regardless of which gender is selected as the calculation basis, it can effectively represent the gender distribution tendency of the same driver group, and this application does not limit it.

[0096] S208, based on the prior score, combined with the initial gender prediction score and the second target confidence score for each modality type, the gender prediction result of the target driver is determined.

[0097] In one alternative implementation, the initial gender prediction score P for each modality type can be compared with a second target confidence level. The product of these factors determines the gender prediction score for each modality type. Then, the prior influence factor is used. Prior scores and gender prediction scores for each modality Interpolation calculations were performed to obtain the total predicted score P for the target driver's gender. 总 Finally, based on the total gender prediction score P... 总 The driver's gender prediction result is determined by comparing it with the aforementioned gender judgment threshold.

[0098] In practice, the following formula (4) can be used to obtain the total gender prediction score P of the target driver. 总 : (4) In this way, when the target driver does not meet the aforementioned data completeness condition, a prior score obtained by matching specified registration information is introduced, and combined with a second target confidence score corrected for data collection rate to perform weighted fusion of the initial gender prediction score. This effectively compensates for the information gaps caused by the sparsity of multimodal operational data. Thus, during the cold start phase, the system no longer relies on the unreliable first target confidence score, but dynamically fuses group statistical priors and finite modal signals, making the total gender prediction score closer to the true gender distribution and ensuring the overall gender judgment accuracy of the system under full driver coverage.

[0099] Based on the same inventive concept, this application also provides a gender recognition device for ride-hailing drivers, which specifically can be as follows: Figure 6 As shown, it includes: The data acquisition unit 601 is configured to acquire operational data of multiple modalities generated by the target driver within a specified time period; The initial prediction unit 602 is configured to: perform feature recognition on the operational data of each modality type to obtain an initial gender prediction score for each modality type and a corresponding initial confidence level; The data correction unit 603 is configured to: correct the initial confidence level of each modality based on the data acquisition rate of each modality to obtain a first target confidence level corresponding to each modality; the data acquisition rate of any modality is determined based on the amount of operational data of the modality. The gender prediction unit 604 is configured to determine the gender prediction result of the target driver based on the initial gender prediction score of each modality type and the first target confidence level.

[0100] In an optional implementation, before performing the initial gender prediction score based on each of the modality types and the corresponding first target confidence level to determine the gender prediction result of the target driver, the data correction unit 603 is further configured to: Determine that the target driver meets the preset data completeness conditions; The data completeness condition includes at least one of the following: The registration duration on the ride-hailing platform has reached the registration duration threshold; The historical order volume is no less than the order volume threshold; The data acquisition rate for at least one modality type is higher than the acquisition rate threshold.

[0101] In an optional implementation, the data correction unit 603 is further configured to: If the target driver does not meet the data completeness condition, then for each modality type, the prior confidence level of each modality type is determined based on the proportion of the data collection rate of that modality type in the data collection rate of each modality type. Based on the prior confidence and data acquisition rate of each modality type, the first target confidence of each modality type is corrected to obtain the second target confidence corresponding to each modality type; The prior score of the target driver is determined based on the ratio of the number of reference drivers in the ride-hailing platform to the total number of drivers on the ride-hailing platform; wherein, any reference driver is a driver whose registration information is the same as that of the target driver, and the specified registration information includes: operating area, operating vehicle type, and part or all of the operating time period; Based on the prior score, combined with the initial gender prediction score and the second target confidence score for each modality type, the gender prediction result of the target driver is determined.

[0102] In one optional implementation, the first target confidence level for each modality type is corrected based on the prior confidence level and data acquisition rate for each modality type to obtain a second target confidence level corresponding to each modality type. Specifically, the data correction unit 603 is configured as follows: Based on the data collection rate of each modality type, the prior influence factor of the target driver is determined; wherein, the prior influence factor represents the degree of influence of the prior score on the gender prediction result; the prior influence factor is negatively correlated with the data collection rate; For each modality type, the prior influence factor is used to interpolate the first target confidence level and the prior confidence level of the modality type to obtain the second target confidence level.

[0103] In one optional implementation, the process of determining the gender prediction result of the target driver based on the prior score, combined with the initial gender prediction score and the second target confidence score for each modality type, is performed. Specifically, the data correction unit 603 is configured to: For each modality, the gender prediction score for that modality is determined based on the product of the initial gender prediction score for that modality and the second target confidence level. Using the aforementioned prior influence factor, interpolation is performed on the prior score and the gender prediction score for each modality type to determine the gender prediction result for the target driver.

[0104] In one optional implementation, the multi-modal operational data includes: order data, text data generated from communication with passengers, and voice data; the feature recognition of the operational data for each modality is performed to obtain an initial gender prediction score and a corresponding initial confidence level for each modality, wherein the initial prediction unit 602 is specifically configured as follows: Semantic recognition is performed on the text data and the voice data to obtain scenario description data of the multiple modal types of operational data; Feature recognition is performed on the first fusion feature of the order data and the scene description data to obtain the initial prediction score and initial confidence level corresponding to the order data; The second fusion feature of the text data and the scene description data is used to perform feature recognition to obtain the initial prediction score and initial confidence level corresponding to the text data; Feature recognition is performed on the third fusion feature of the speech data and the scene description data to obtain the initial prediction score and initial confidence level corresponding to the speech data.

[0105] In one optional implementation, the data acquisition rate based on each modality type is executed to correct the initial confidence level of each modality type, thereby obtaining a first target confidence level corresponding to each modality type. Specifically, the data correction unit 603 is configured as follows: The data acquisition rate and initial confidence of each modality type are concatenated to obtain the weighted features corresponding to the target driver; Feature recognition is performed on the weighted features to obtain the first target confidence level for each modality type.

[0106] In one optional implementation, the initial gender prediction score based on each of the modality types and the first target confidence level are executed to determine the gender prediction result of the target driver. Specifically, the gender prediction unit 604 is configured as follows: For each modality type, the gender prediction score for that modality type is determined based on the product of the initial gender prediction score for that modality type and the first target confidence score. The gender prediction result of the target driver is determined based on the sum of the gender prediction scores for each modality.

[0107] Based on the same inventive concept, this application also provides an electronic device, specifically as follows: Figure 7 As shown, the electronic device 130 is presented in the form of a general-purpose electronic device. The components of the electronic device 130 may include, but are not limited to: at least one processor 131, at least one memory 132, and a bus 133 connecting different system components (including memory 132 and processor 131).

[0108] Bus 133 represents one or more of several bus architectures, including a memory bus or memory controller, peripheral bus, processor, or a local bus using any of the various bus architectures.

[0109] The memory 132 may include a readable medium in the form of volatile memory, such as random access memory (RAM) 1321 and / or cache memory 1322, and may further include read-only memory (ROM) 1323.

[0110] The memory 132 may also include a program / utility 1325 having a set (at least one) of program modules 1324, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0111] Electronic device 130 may also communicate with one or more external devices 134 (e.g., keyboard, pointing device, etc.), and may also communicate with one or more devices that enable a user to interact with electronic device 130, and / or with any device that enables electronic device 130 to communicate with one or more other electronic devices (e.g., router, modem, etc.).

[0112] This communication can be performed via input / output (I / O) interface 135. Furthermore, electronic device 130 can also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapter 136. As shown, network adapter 136 communicates with other modules used in electronic device 130 via bus 133. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 130, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0113] In an optional implementation, this application also provides a computer-readable storage medium including instructions, such as a memory 132 including instructions, which can be executed by the processor 131 of the aforementioned device to perform the aforementioned method. Optionally, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0114] This application also provides a computer program product, including a computer program / instruction, which, when executed by a processor 831, implements any of the methods in the online ride-hailing driver gender identification method provided in this application.

[0115] In an alternative implementation, various aspects of the gender identification method for ride-hailing drivers provided in this application can also be implemented in the form of a program product, which includes program code. When the program product is run on a computer device, the program code is used to cause the computer device to perform the steps in the gender identification method for ride-hailing drivers according to the various exemplary embodiments of this application described above.

[0116] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable 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 thereof.

[0117] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0120] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for gender identification of ride-hailing drivers, characterized in that, The method comprises: acquiring a plurality of modal types of operation data generated by a target driver within a specified period of time; performing feature recognition on the operation data of each modal type to obtain an initial gender prediction score of each modal type and a corresponding initial confidence; based on the data collection rate of each modal type, correcting the initial confidence of each modal type to obtain a first target confidence corresponding to each modal type; wherein the data collection rate of any modal type is determined based on the number of operation data of the modal type; based on the initial gender prediction score of each modal type and the first target confidence, determining the gender prediction result of the target driver.

2. The method of claim 1, wherein, Before the method based on the initial gender prediction score of each modal type and the corresponding first target confidence determines the gender prediction result of the target driver, the method further comprises: determining that the target driver meets a preset data completeness condition; wherein the data completeness condition comprises at least one of the following: the registration time length of the target driver on the online car-hailing platform reaches a registration time length threshold; the historical order quantity is not less than an order quantity threshold; the data collection rate corresponding to at least one modal type is higher than a collection rate threshold.

3. The method of claim 2, wherein, The method further comprises: if the target driver does not meet the data completeness condition, for each modal type, based on the data collection rate of the modal type, the proportion in the data collection rate of each modal type is determined to obtain the prior confidence of each modal type; based on the prior confidence and the data collection rate of each modal type, the first target confidence of each modal type is corrected to obtain a second target confidence corresponding to each modal type; based on the ratio of the number of reference drivers in the online car-hailing platform to the total number of drivers in the online car-hailing platform, the prior score of the target driver is determined; wherein any reference driver is a driver with the same specified registration information as the target driver, and the specified registration information includes part or all of the operation area, the operation vehicle type and the operation period; based on the prior score, combined with the initial gender prediction score and the second target confidence of each modal type, the gender prediction result of the target driver is determined.

4. The method of claim 3, wherein, The method based on the prior confidence and the data collection rate of each modal type, correcting the first target confidence of each modal type to obtain a second target confidence corresponding to each modal type, comprises: based on the data collection rate of each modal type, determining the prior influence factor of the target driver; wherein the prior influence factor represents the influence degree of the prior score on the gender prediction result; the prior influence factor is negatively correlated with the data collection rate; for each modal type, using the prior influence factor to perform interpolation calculation on the first target confidence and the prior confidence of the modal type to obtain the second target confidence.

5. The method of claim 4, wherein, The method based on the prior score, combined with the initial gender prediction score and the second target confidence of each modal type, to determine the gender prediction result of the target driver, comprises: determining, for each modality type, a gender prediction score of the modality type based on a product of the initial gender prediction score of the modality type and the second target confidence; interpolating the prior score and the gender prediction score of each modality type by using the prior influence factor to determine the gender prediction result of the target driver.

6. The method according to any one of claims 1 to 5, characterized in that, The operation data of the plurality of modality types includes order data, text data and voice data generated in communication with passengers; The feature recognition of the first fusion feature of the order data and the scene description data obtains the initial prediction score and the initial confidence corresponding to the order data; The feature recognition of the second fusion feature of the text data and the scene description data obtains the initial prediction score and the initial confidence corresponding to the text data; The feature recognition of the third fusion feature of the voice data and the scene description data obtains the initial prediction score and the initial confidence corresponding to the voice data. The correction of the initial confidence of each modality type based on the data collection rate of each modality type obtains the first target confidence corresponding to each modality type, and the data collection rate of any modality type is determined based on the operation data quantity of the modality type. The feature splicing of the data collection rate and the initial confidence of each modality type obtains the weight feature corresponding to the target driver; 7. The method according to any one of claims 1 to 5, characterized in that, The feature recognition of the weight feature obtains the first target confidence of each modality type. The determination of the gender prediction result of the target driver based on the initial gender prediction score of each modality type and the first target confidence includes: determining, for each modality type, a gender prediction score of the modality type based on a product of the initial gender prediction score of the modality type and the first target confidence; 8. The method of any one of claims 1-5, wherein, determining the gender prediction result of the target driver based on a sum value of the gender prediction scores of each modality type. The device includes: a data collection unit configured to acquire operation data of a plurality of modality types generated by a target driver within a specified period; 9.A device for identifying a gender of a driver of a network car, the device comprising: an initial prediction unit configured to perform feature recognition on the operation data of each modality type to obtain an initial gender prediction score of each modality type and a corresponding initial confidence; a data correction unit configured to correct the initial confidence of each modality type based on a data collection rate of each modality type to obtain a first target confidence corresponding to each modality type, and the data collection rate of any modality type is determined based on the operation data quantity of the modality type; a gender prediction unit configured to determine the gender prediction result of the target driver based on the initial gender prediction score of each modality type and the first target confidence. including: a memory for storing program instructions; 10. An electronic device, comprising: ​ ​ a processor for invoking program instructions stored in the memory to perform the steps of the method of any of claims 1-8, in accordance with the obtained program instructions.

Images