Multi-source heterogeneous data fusion-based intelligent responsibility judgment method and system for pick-up behavior

Abstract

The invention belongs to the technical field of computer software, and particularly relates to a multi-source heterogeneous data fusion-based intelligent responsibility judgment method and system for a pick-up behavior, and the method comprises the steps: collecting a driver track, a communication recording text, environment information, driver reporting and order metadata, and generating a structured pick-up event through space-time alignment; extracting travel, environment, semantics and behavior reputation multi-dimensional features based on the event, and training a gradient boosting decision tree model; real-time lightweight risk intervention and off-line high-precision responsibility judgment are executed through the model, real-time intervention rapidly triggers reminding based on high-timeliness characteristics, and off-line judgment outputs violation probability based on full-quantity characteristics; artificial rechecking, appeal and arbitration results are returned to drive incremental learning of the model; and finally, differentially weighting the responsibility judgment result according to the historical reputation of the driver and generating an interpretable natural language basis. Through multi-source data fusion and a double-track responsibility judgment mechanism, the problem of high misjudgment rate caused by single data in a traditional method is solved.

Description

A method and system for intelligent liability determination based on multi-source heterogeneous data fusion for pick-up behavior Technical Field

[0001] This invention belongs to the field of Internet information service technology, specifically relating to an intelligent method and system for determining responsibility for pick-up behavior based on multi-source heterogeneous data fusion. Background Technology

[0002] Currently, ride-hailing platforms rely primarily on static rule engines based on fixed thresholds, inefficient manual review and spot checks, or simple models using only a single data source to determine the liability of drivers for violations. These methods have inherent technical limitations.

[0003] The aforementioned methods cannot understand complex contextual scenarios due to static rules, making it difficult to distinguish between reasonable and malicious detours, and between passenger-initiated cancellations and driver-induced cancellations. This results in a high misjudgment rate, frequently wrongly penalizing compliant drivers, low efficiency of manual review, inability to handle massive orders, and difficulty in unifying the standards for liability determination. More importantly, existing methods generally suffer from systemic problems such as data fragmentation, delayed response, rigid rules, and a lack of self-optimization capabilities, leading to inaccurate liability determinations, high driver appeal rates, and high operating costs, which seriously affect the healthy development of the platform ecosystem. Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the problems of high misjudgment rate, low efficiency and inability to adaptively optimize existing online ride-hailing liability determination methods due to single data and rigid rules.

[0005] To address the aforementioned technical problems, a first aspect of this invention discloses an intelligent accountability method for pick-up behavior based on multi-source heterogeneous data fusion, the method comprising:

[0006] S1. Collect driver trajectory data, transcribed text of communication recordings, environmental information, driver reporting records and order metadata, and perform spatiotemporal alignment according to order identifiers and preset time windows to generate structured pick-up events;

[0007] S2. Based on the pick-up event, extract a multi-dimensional feature vector including basic trip features, environmental context features, speech and semantic features, and driver behavior and reputation features, and train a gradient boosting decision tree model based on labeled samples as a liability judgment model.

[0008] S3. Using the aforementioned liability assessment model, perform real-time lightweight risk intervention and offline high-precision liability determination. The real-time lightweight risk intervention is based on some high-time-sensitivity features to perform rapid reasoning and trigger driver-side reminders. The offline high-precision liability determination is based on the complete multi-dimensional feature vector to output the probability of violation.

[0009] S4. Feedback data generated from manual review, driver complaints, or platform arbitration is fed back to the training sample set to drive the judgment model to perform incremental learning or retraining.

[0010] S5. Differentiate the weighting of liability judgment results based on the driver's historical reputation and generate interpretable natural language liability judgment basis.

[0011] As an optional implementation, in the first aspect of the present invention, step S1 specifically includes:

[0012] The driver's trajectory data is constructed by obtaining GPS coordinate sequences, driving speed, yaw rate and network signal quality in real time through the driver's application interface.

[0013] The communication recordings were obtained through a third-party anonymous call service, and then speech recognition and semantic understanding were performed in sequence to extract semantic text containing information on whether there were leading statements and the roles in the conversation.

[0014] Call the map service and weather service application interfaces to obtain the real-time congestion index of the order route, as well as weather type, precipitation intensity, and visibility environmental data;

[0015] The driver's reporting record is formed by receiving the reporting type and reporting timestamp proactively reported by the driver through the application.

[0016] Query the estimated pick-up distance, actual pick-up time, and order metadata (whether a phone call was made) from the platform's order database;

[0017] Using a stream processing engine, the five types of data mentioned above are associated and aligned within a preset time window, using the order identifier as the key, and encapsulated into a unified pick-up event data object.

[0018] As an optional implementation, in the first aspect of the present invention, the extraction of speech semantic features in step S2 specifically includes:

[0019] Automatic speech recognition is performed on communication recordings to convert audio signals into raw text;

[0020] The original text is semantically analyzed using a large language model to identify whether it contains preset keywords or semantics that induce cancellation.

[0021] When a leading semantic is identified, a confidence score for that semantic is calculated.

[0022] Based on the number of conversation rounds and the content, it is determined whether the party who first made the cancellation request in the call was the driver or the passenger.

[0023] The presence of the induced semantics, the confidence score, and the information of the cancel request initiator are used to construct a speech semantic feature vector.

[0024] As an optional implementation, in the first aspect of the present invention, step S3 specifically includes:

[0025] The complete multidimensional feature vector is input into the trained gradient boosting decision tree model to calculate the probability value of the driver's violation.

[0026] The probability value is compared with preset high and low thresholds;

[0027] If the probability value is greater than or equal to the high threshold, the driver's responsibility will be automatically determined and corresponding penalties will be imposed.

[0028] If the probability value is less than the high threshold but greater than or equal to the low threshold, the current order will be marked as a questionable case and automatically pushed to the manual review queue for further processing.

[0029] If the probability value is less than the low threshold, the driver is automatically determined to be not at fault, and the process ends.

[0030] As an optional implementation, in the first aspect of the present invention, step S4 specifically includes:

[0031] Collect and record the conclusions of manual review, successful driver appeals, and platform arbitration results, and use them as sample data with new labels;

[0032] The new label sample data is fed back to the model training sample pool in real time or near real time via a message queue component.

[0033] Based on a predetermined scheduling strategy, the expanded training sample pool is used periodically to incrementally train the judgment model in order to update the model parameters.

[0034] Continuously monitor the key performance indicators of the accountability model in the production environment, and automatically trigger a complete model retraining process when the performance degradation exceeds the preset tolerance.

[0035] As an optional implementation, in the first aspect of the present invention, the generation of interpretable natural language judgment criteria includes:

[0036] The contribution of each feature in the multidimensional feature vector to the current judgment result is calculated using model interpretability techniques.

[0037] Filter out the top N key features with the highest contribution and their direction of influence;

[0038] Based on a predefined template, the key features and their direction of influence are transformed into a coherent natural language description that clearly identifies the main factors leading to the liability determination.

[0039] As an optional implementation, in the first aspect of the present invention, the differential weighting of the liability determination result based on the driver's historical reputation includes:

[0040] Check the driver's recent service ratings and the number of recent valid complaints;

[0041] A reputation weighting factor is calculated based on the historical service ratings and number of complaints.

[0042] The original violation probability output by the accountability model is calculated with the reputation weighting factor to obtain the weighted final violation probability;

[0043] For drivers with high credit ratings, the weighting factor will be adjusted to lower the final probability of violation, thereby implementing a fault-tolerance mechanism.

[0044] For drivers with low credit ratings, the weighting factor will increase the final probability of violations, thereby implementing stricter supervision.

[0045] As an optional implementation, in the first aspect of the present invention, the real-time lightweight risk intervention includes:

[0046] A subset of low-computational-complexity features are extracted in real time from the pick-up event to form a lightweight feature subset, which includes instantaneous yaw status, call connection status, positioning signal quality, and service level.

[0047] The lightweight feature subset is input into a lightweight inference model deployed on an edge server;

[0048] The lightweight inference model outputs the risk probability of the current pick-up behavior within millisecond latency;

[0049] If the risk probability exceeds the intervention threshold, an alert message will be immediately sent to the driver's application or a server-side outbound call will be triggered.

[0050] The second aspect of this invention discloses an intelligent accountability system for pick-up behavior based on multi-source heterogeneous data fusion, used to implement the intelligent accountability method for pick-up behavior based on multi-source heterogeneous data fusion described in any of the above embodiments, the system comprising:

[0051] The data acquisition and alignment module is used to collect driver trajectory data, transcribed text of communication recordings, environmental information, driver reporting records and order metadata, and perform spatiotemporal alignment according to order identifiers and preset time windows to generate structured pick-up events;

[0052] The feature construction and training module is used to extract multi-dimensional feature vectors based on the pick-up event, including basic trip features, environmental context features, speech and semantic features, and driver behavior and reputation features, and to train a gradient boosting decision tree model as a liability judgment model based on labeled samples.

[0053] The dual-track accountability execution module is used to perform real-time lightweight risk intervention and offline high-precision accountability determination using the accountability model.

[0054] The feedback optimization and update module is used to feed back feedback data generated by manual review, driver complaints or platform arbitration to the training sample set, driving the judgment model to perform incremental learning or retraining.

[0055] The differentiated liability assessment output module is used to differentiate and weight the liability assessment results based on the driver's historical reputation and generate interpretable natural language liability assessment criteria.

[0056] A third aspect of this invention discloses another intelligent accountability system for pick-up behavior based on multi-source heterogeneous data fusion, the system comprising:

[0057] Memory containing executable program code;

[0058] A processor coupled to the memory;

[0059] The processor calls the executable program code stored in the memory to execute the intelligent judgment method for pick-up behavior based on multi-source heterogeneous data fusion disclosed in the first aspect of the present invention.

[0060] The fourth aspect of this invention discloses a computer-readable storage medium storing computer instructions, which, when invoked by a processor, are used to execute a method for intelligent judgment of pick-up behavior based on multi-source heterogeneous data fusion disclosed in the first aspect of this invention.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] By aligning and fusing multi-source heterogeneous data in time and space, a panoramic view of pick-up events is constructed, encompassing multiple dimensions such as trajectory, speech semantics, environment, and behavior. This enables the liability assessment model to make decisions based on complete contextual information, overcoming the problems of misjudgment and omission caused by single data. A dual-track liability assessment mechanism is adopted: the lightweight real-time model can provide early warning and intervention for high-risk behaviors within millisecond latency, proactively preventing violations and reducing order cancellation rates; while the offline high-precision model uses full features for in-depth analysis to achieve accurate post-event liability determination. By introducing a closed-loop feedback optimization process, business feedback data such as manual review and driver appeals are automatically fed back to the training sample set, driving the liability assessment model to continuously perform incremental learning or retraining. This invention can provide fault tolerance for long-term compliant, high-reputation drivers and focus on supervising drivers with high frequency of violations. Attached Figure Description

[0063] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein:

[0064] Figure 1 is a flowchart illustrating an intelligent accountability method for pick-up behavior based on multi-source heterogeneous data fusion disclosed in an embodiment of the present invention.

[0065] Figure 2 is a schematic diagram of the structure of an intelligent accountability system for pick-up behavior based on multi-source heterogeneous data fusion disclosed in an embodiment of the present invention;

[0066] Figure 3 is a schematic diagram of another intelligent accountability system for pick-up behavior based on multi-source heterogeneous data fusion disclosed in an embodiment of the present invention;

[0067] Figure 4 is a flowchart of driver behavior recognition in an intelligent judgment method for pick-up behavior based on multi-source heterogeneous data fusion disclosed in an embodiment of the present invention. Detailed Implementation

[0068] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0070] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0071] This invention discloses an intelligent accountability method and system for pick-up behavior based on multi-source heterogeneous data fusion. By aligning and fusing multi-source heterogeneous data in time and space, a panoramic view of pick-up events including trajectory, voice semantics, environment, and behavior is constructed. This enables the accountability model to make decisions based on complete contextual information, overcoming the problems of misjudgment and omission caused by single data. A dual-track accountability mechanism is adopted. The lightweight real-time model can provide early warning and intervention for high-risk behaviors within millisecond latency, proactively preventing violations and reducing order cancellation rates. The offline high-precision model uses full features for in-depth analysis to achieve accurate post-event accountability.

[0072] Example 1

[0073] Please refer to Figure 1, which is a flowchart illustrating an intelligent accountability method for pick-up behavior based on multi-source heterogeneous data fusion disclosed in an embodiment of the present invention. The intelligent accountability method for pick-up behavior based on multi-source heterogeneous data fusion described in Figure 1 is applied to a data processing chip, processing terminal, or processing server. The processing server can be a local server or a cloud server; this embodiment of the present invention does not impose any limitations. As shown in Figure 1, the intelligent accountability method for pick-up behavior based on multi-source heterogeneous data fusion may include the following operations:

[0074] S1. Collect driver trajectory data, transcribed text of communication recordings, environmental information, driver reporting records, and order metadata, and perform spatiotemporal alignment according to order identifiers and preset time windows to generate structured pick-up events.

[0075] Specifically, the above steps integrate data from different sources and structures, including driver's side, passenger's side, third-party services, and the platform itself, according to a unified order identifier and time benchmark. This solves the problems of data isolation and lack of context in traditional methods, transforming the originally scattered trajectory points, call records, environmental parameters, and other information into a structured event object with complete spatiotemporal correlation.

[0076] It is evident that this structured pick-up event provides a unified and standardized input foundation for subsequent feature engineering and model analysis, encapsulating complex behavioral information into data units that can be directly processed by machines, and providing the necessary data foundation for constructing multi-dimensional feature vectors and achieving accurate identification of violations and determination of responsibility.

[0077] S2. Based on the pick-up event, extract a multi-dimensional feature vector including basic trip features, environmental context features, speech and semantic features, and driver behavior and reputation features, and train a gradient boosting decision tree model based on labeled samples as the liability judgment model.

[0078] Specifically, the above steps systematically extract multi-dimensional features covering itinerary, environment, semantics, and behavioral reputation from structured events, constructing an information profile that can comprehensively describe the context of pick-up behavior. The refined feature engineering in the above steps transforms complex behavioral patterns into quantifiable feature vectors, providing an information foundation for the model to go beyond simple rules and perform deep pattern recognition.

[0079] It is evident that training labeled samples with a gradient boosting decision tree model enables the accountability model to automatically learn the complex nonlinear relationship between different dimensional features and violations. This step ensures the accuracy of model discrimination while also taking into account processing efficiency and interpretability of results, providing core algorithmic support for achieving automated and highly accurate accountability determination.

[0080] S3. Using the aforementioned liability assessment model, perform real-time lightweight risk intervention and offline high-precision liability determination.

[0081] Specifically, the above steps, by deploying a lightweight model to perform real-time reasoning on high-time-sensitivity features, build the ability to intervene in risks during the event, enabling the system to instantly identify potential violations and trigger alerts during the pick-up process, transforming management actions from passive post-event handling to proactive process control.

[0082] As can be seen, this step uses a high-precision model to perform offline in-depth analysis of the complete feature set, ensuring the accuracy of the final responsibility determination. The above-mentioned dual-track parallel architecture not only meets the real-time requirements of the business, but also provides reliable data support and algorithm guarantee for the final responsibility determination decision.

[0083] S4. Feedback data generated from manual review, driver complaints, or platform arbitration is fed back to the training sample set to drive the judgment model to perform incremental learning or retraining.

[0084] Specifically, the above steps establish a closed-loop model optimization mechanism by feeding back the manual feedback data generated in the business process to the training process, enabling the accountability model to continuously adjust based on the latest business judgments, thereby mitigating the model performance degradation problem caused by changes in business rules or data distribution drift.

[0085] It is evident that this mechanism transforms human judgment experience into knowledge that the model can learn, enabling the synchronous evolution of the system's judgment capabilities and business cognition, reducing the reliance on periodic manual model updates, and improving the system's stability during long-term operation.

[0086] S5. Differentiate the weighting of liability judgment results based on the driver's historical reputation and generate interpretable natural language liability judgment basis.

[0087] Specifically, the above steps, by introducing the driver's historical reputation as a weighting factor in the judgment result, realize the transformation from uniform threshold judgment to personalized and precise governance. The above differentiated processing method enables the judgment strategy to take into account the context in which the behavior occurred, to form a positive incentive for drivers with high reputation, and to implement key supervision for drivers with low reputation.

[0088] As can be seen, this step transforms the judgment logic into a natural language description through model interpretability technology, making the decision basis transparent and readable. This provides factual basis for the judgment result, helps drivers understand the reasons for violations, thereby improving the acceptability of the result and reducing unnecessary appeals and disputes.

[0089] As an optional embodiment, step S1 in the above steps specifically includes:

[0090] The driver's trajectory data is constructed by obtaining GPS coordinate sequences, driving speed, yaw rate and network signal quality in real time through the driver's application interface.

[0091] The communication recordings were obtained through a third-party anonymous call service, and then speech recognition and semantic understanding were performed in sequence to extract semantic text containing information on whether there were leading statements and the roles in the conversation.

[0092] Call the map service and weather service application interfaces to obtain the real-time congestion index of the order route, as well as weather type, precipitation intensity, and visibility environmental data;

[0093] The driver's reporting record is formed by receiving the reporting type and reporting timestamp proactively reported by the driver through the application.

[0094] Query the estimated pick-up distance, actual pick-up time, and order metadata (whether a phone call was made) from the platform's order database;

[0095] Using a stream processing engine, the five types of data mentioned above are associated and aligned within a preset time window, using the order identifier as the key, and encapsulated into a unified pick-up event data object.

[0096] In this embodiment of the invention, the above steps systematically integrate five types of heterogeneous data: driver trajectory, communication semantics, environmental information, driver reporting, and platform orders, to construct a multi-dimensional information set covering behavior, environment, interaction, and management intent. The synchronous collection of the above multi-source data provides a raw data foundation with complete context for subsequent analysis.

[0097] As can be seen, this step uses a stream processing engine to perform spatiotemporal alignment of multi-source data on an order-by-order basis, transforming the originally scattered and asynchronous raw data into structured pick-up events. This process achieves the standardization and semanticization of data relationships, providing a unified and computable data entity for subsequent feature extraction and model inference.

[0098] As an optional embodiment, in step S2 above, the extraction of speech semantic features specifically includes:

[0099] Automatic speech recognition is performed on communication recordings to convert audio signals into raw text;

[0100] The original text is semantically analyzed using a large language model to identify whether it contains preset keywords or semantics that induce cancellation.

[0101] When a leading semantic is identified, a confidence score for that semantic is calculated.

[0102] Based on the number of conversation rounds and the content, it is determined whether the party who first made the cancellation request in the call was the driver or the passenger.

[0103] The presence of the induced semantics, the confidence score, and the information of the cancel request initiator are used to construct a speech semantic feature vector.

[0104] In this embodiment of the invention, the above steps combine speech recognition and semantic understanding to transform unstructured call audio into structured semantic information that includes intent recognition and role judgment. The above processing method overcomes the limitations of simple keyword matching, enabling the system to understand the context and true intent of the dialogue.

[0105] It is evident that by constructing a composite feature vector that integrates the existence of induced semantics, confidence level, and the responsible initiator, the liability assessment model is provided with a basis for accurately distinguishing key scenarios such as "driver-induced cancellation" and "passenger-initiated cancellation," thereby improving the model's ability to analyze complex attribution scenarios from a data perspective.

[0106] As an optional embodiment, step S3 in the above steps specifically includes:

[0107] The complete multidimensional feature vector is input into the trained gradient boosting decision tree model to calculate the probability value of the driver's violation.

[0108] The probability value is compared with preset high and low thresholds;

[0109] If the probability value is greater than or equal to the high threshold, the driver's responsibility will be automatically determined and corresponding penalties will be imposed.

[0110] If the probability value is less than the high threshold but greater than or equal to the low threshold, the current order will be marked as a questionable case and automatically pushed to the manual review queue for further processing.

[0111] If the probability value is less than the low threshold, the driver is automatically determined to be not at fault, and the process ends.

[0112] In this embodiment of the invention, the above steps use a gradient boosting decision tree model to comprehensively analyze multidimensional feature vectors and output continuous probability values ​​to quantify the risk of violations. The probability-based judgment method provides a quantifiable basis for liability determination, transforming the judgment decision from a binary judgment to a confidence-based hierarchical assessment.

[0113] As can be seen, this step divides the judgment results into three processing flows: automatic processing, manual review, and automatic case closure by setting high and low threshold ranges. Through the above-mentioned hierarchical processing mechanism, while ensuring the efficiency of handling cases with high certainty, complex situations are guided to manual judgment, forming a human-machine collaborative judgment closed loop.

[0114] As an optional embodiment, step S4 in the above steps specifically includes:

[0115] Collect and record the conclusions of manual review, successful driver appeals, and platform arbitration results, and use them as sample data with new labels;

[0116] The new label sample data is fed back to the model training sample pool in real time or near real time via a message queue component.

[0117] Based on a predetermined scheduling strategy, the expanded training sample pool is used periodically to incrementally train the judgment model in order to update the model parameters.

[0118] Continuously monitor the key performance indicators of the accountability model in the production environment, and automatically trigger a complete model retraining process when the performance degradation exceeds the preset tolerance.

[0119] In this embodiment of the invention, the above steps establish an automated feedback channel for business feedback data, transforming manual review conclusions and arbitration results into high-quality labeled samples. This mechanism effectively solves the performance degradation problem caused by changes in business rules or data distribution drift, ensuring that the judgment criteria and business understanding are updated synchronously.

[0120] As can be seen, this step, by setting a periodic incremental training strategy and an automatic retraining trigger condition based on performance monitoring, achieves continuous optimization and self-correction of model parameters, reduces the dependence on manual model maintenance, and ensures the adaptability and stability of the accountability system in long-term operation.

[0121] As an optional embodiment, the step above, generating interpretable natural language judgment criteria, includes:

[0122] The contribution of each feature in the multidimensional feature vector to the current judgment result is calculated using model interpretability techniques.

[0123] Filter out the top N key features with the highest contribution and their direction of influence;

[0124] Based on a predefined template, the key features and their direction of influence are transformed into a coherent natural language description that clearly identifies the main factors leading to the liability determination.

[0125] In this embodiment of the invention, the above steps quantify the contribution of each feature to the judgment result through model interpretability technology, and transform the decision-making process inside the model into a traceable set of key factors. This process solves the problem of algorithm black box, and transforms the basis for judgment decision from invisible numerical calculation to verifiable feature impact analysis.

[0126] As can be seen, this step transforms key features and their impact into natural language descriptions through predefined templates, providing specific factual basis for the judgment. The readable output format enables drivers to clearly understand the reasons for violations, providing clear guidance for subsequent appeals or behavior improvement, and reducing disputes caused by information asymmetry.

[0127] As an optional embodiment, the step described above, namely, differentially weighting the liability determination result based on the driver's historical reputation, includes:

[0128] Check the driver's recent service ratings and the number of recent valid complaints;

[0129] A reputation weighting factor is calculated based on the historical service ratings and number of complaints.

[0130] The original violation probability output by the accountability model is calculated with the reputation weighting factor to obtain the weighted final violation probability;

[0131] For drivers with high credit ratings, the weighting factor will be adjusted to lower the final probability of violation, thereby implementing a fault-tolerance mechanism.

[0132] For drivers with low credit ratings, the weighting factor will increase the final probability of violations, thereby implementing stricter supervision.

[0133] In this embodiment of the invention, the above steps construct a reputation-weighted factor by introducing the driver's historical service data, transforming the judgment of a single behavior into a comprehensive assessment that combines long-term performance. This mechanism enables the judgment system to distinguish between occasional behavioral deviations and habitual violation patterns, thereby improving the accuracy of identifying drivers of different risk levels.

[0134] As can be seen, this step dynamically adjusts the original probability through the reputation factor, thereby enabling error-tolerant processing for high-reputation drivers and focused supervision for low-reputation drivers. This differentiated approach allows management resources to be tilted towards high-risk individuals, while improving the experience of compliant drivers while ensuring service quality.

[0135] As an optional embodiment, the real-time lightweight risk intervention in the above steps includes:

[0136] A subset of low-computational-complexity features are extracted in real time from the pick-up event to form a lightweight feature subset, which includes instantaneous yaw status, call connection status, positioning signal quality, and service level.

[0137] The lightweight feature subset is input into a lightweight inference model deployed on an edge server;

[0138] The lightweight inference model outputs the risk probability of the current pick-up behavior within millisecond latency;

[0139] If the risk probability exceeds the intervention threshold, an alert message will be immediately sent to the driver's application or a server-side outbound call will be triggered.

[0140] In this embodiment of the invention, the above steps construct a lightweight feature subset by screening highly time-sensitive features and deploy the inference model using edge computing, thereby controlling the system response time to the millisecond level. Through this architecture design, the platform can simultaneously complete risk calculation during the pick-up process.

[0141] As can be seen, this step generates intervention instructions based on real-time risk probability, and provides immediate warnings of potential violations through driver-side reminders or server-side outbound calls. The in-process intervention mechanism moves the management node from post-processing to the behavior occurrence process, playing a role in preventing violations and reducing order cancellation rates.

[0142] Furthermore, multi-source heterogeneous data fusion mechanisms include:

[0143] 1. Synchronous acquisition and spatiotemporal alignment:

[0144] Driver trajectory data (including GPS signal quality and yaw rate)

[0145] Translated text of voice calls and its semantics (identification of intentions such as leading / negotiation / complaint)

[0146] Contextual information (weather, traffic congestion index)

[0147] Drivers proactively report records (faults, traffic control, etc.).

[0148] Order metadata (estimated / actual pick-up time, distance ratio) → forms a structured pick-up event representation, which serves as the basis for liability determination.

[0149] 2. Dual-track reasoning mechanism:

[0150] Real-time track: Uses low-latency features (whether there is a deviation, whether there is a call, signal strength) to identify risks in seconds and trigger in-process reminders (such as "Please follow the navigation") to reduce cancellation rate;

[0151] Offline track: After an order is cancelled, all fused features are invoked, and the probability and confidence of violation are output through a high-precision model to achieve accurate liability determination.

[0152] 3. Confidence rating and human-machine collaboration:

[0153] High confidence level (e.g., >0.95): Automatically enforce penalties;

[0154] Low confidence level: Automatically enters the manual review channel to protect the driver's rights;

[0155] All review / appeal results are automatically fed back as training samples, driving monthly incremental updates to the model.

[0156] 4. Explainability and business adaptability:

[0157] Use interpretable models (such as LightGBM+SHAP) to output key criteria for liability judgment (such as "due to failure to report + 30% deviation + call containing 'cancel it'");

[0158] By introducing drivers' historical credit scores to weight liability assessments, differentiated governance can be achieved, with "tolerance for high-quality drivers and strict management of high-risk drivers."

[0159] The determination of responsibility is formed by combining the following data:

[0160] 1. Eliminate ambiguity from a single data source:

[0161] Example: Deviated from the flight path + no notification + call includes "Cancel it";

[0162] →Looking at the deviation alone: ​​it could be a reasonable detour;

[0163] →Looking at the call alone: ​​This may only be a suggestion;

[0164] →The simultaneous occurrence of all three points strongly suggests a combined violation of "inducing cancellation + creating excuses for cancellation," significantly increasing the confidence level for determining liability.

[0165] 2. Construct a complete chain of evidence: behavior, context, and intention.

[0166] Example: The driver did not deviate from the route, but said during the call, "This fare is too far, you should cancel it."

[0167] →The trajectory is normal, but the traditional system cannot trigger it;

[0168] → Combining order metadata (long-distance orders) + voice and semantics (explicit inducement) + lack of reasonable reporting

[0169] →Accurately identifies "non-trajectory-based cancellation guidance," covering traditional blind spots.

[0170] 3. Noise suppression and counterfeiting:

[0171] Example: Driver falsifies "traffic control" report + did not actually take a detour.

[0172] →Looking at the filing separately: The system may be exempt from liability;

[0173] → Combining the trajectory (no detours) + environmental data (no traffic control notices) + phone calls (no mention of traffic control)

[0174] → Determine if the report is invalid to prevent abuse.

[0175] 4. Enhance generalization ability for long-tail / atypical scenarios:

[0176] Example: "Killing" in rural areas with no signal

[0177] →Trajectory interruption, no call, no network→Traditional systems cannot determine responsibility;

[0178] →Combined factors: sudden drop in signal strength + abrupt termination of the trajectory + high cancellation rate driver tag + no reporting

[0179] → Triggered a high-risk flag of "suspected malicious homing", and entered manual review.

[0180] 5. Supports interpretable and traceable liability determinations:

[0181] The system can output criteria for liability determination:

[0182] ① Yaw rate reached 42% (threshold 30%);

[0183] ② The call detected the message "It's more cost-effective for you to cancel" (LLM confidence level 0.93);

[0184] ③ No filing was submitted;

[0185] ④ There is currently no severe weather.

[0186] →This increases driver acceptance and also meets the platform's compliance audit requirements.

[0187] Furthermore, the specific implementation methods of the model and algorithm are as follows:

[0188] The core algorithm for intelligent liability determination of ride-hailing driver pick-up behavior in this invention is a supervised learning model based on Gradient Boosting Decision Tree (GBDT), specifically implemented using the LightGBM or XGBoost framework. The input, output, training, and inference process of this model are as follows:

[0189] I. Model Input (Feature Vector)

[0190] The model receives a structured feature vector X, with at least 30 dimensions, categorized into the following four types based on its source:

[0191] 1. Basic travel characteristics:

[0192] Pick-up distance (unit: meters);

[0193] The ratio of actual pick-up time to estimated pick-up time (dimensionless);

[0194] Yaw ratio = (Actual distance traveled - Navigation planned distance) / Navigation planned distance;

[0195] Did you proactively call the passenger within 2 minutes of the pick-up starting (0 indicates no, 1 indicates yes)?

[0196] 2. Contextual characteristics:

[0197] Weather phenomenon type: from third-party meteorological API (such as China Weather Network), the enumerated values ​​include "sunny, cloudy, light rain, moderate rain, heavy rain, rainstorm, snow, blizzard, fog, haze, dust storm, hail, strong wind", using integer codes (0~12);

[0198] Precipitation intensity (unit: mm / hour, continuous value);

[0199] Visibility (unit: meters, continuous value);

[0200] Wind speed (unit: meters per second, continuous value);

[0201] Weighted average congestion index of planned route (values ​​range from 0 to 10, where 0 indicates smooth traffic and 10 indicates severe congestion, sourced from Gaode or Baidu Maps API);

[0202] GPS signal-to-noise ratio (SNR, unit: dBHz, continuous value);

[0203] Network connection type (0: 4G; 1: 5G; 2: Wi-Fi; 3: No network).

[0204] 3. Speech-semantic features:

[0205] Did you detect induced cancellation semantics? (0: No; 1: Yes)

[0206] Induced intent confidence (value range [0,1], output by the large language model based on prompting engineering);

[0207] Did the passenger request cancellation first? (0: No; 1: Yes, determined by the dialogue role label).

[0208] 4. Driver behavior and credit characteristics:

[0209] Did a valid report be submitted within 2 minutes of the start of the pick-up process? (0: No; 1: Yes)

[0210] Reporting type (using unique hot-coded form: [0,0,0,1] represents "no reporting", [1,0,0,0] represents "vehicle malfunction", etc.);

[0211] Driver's service score over the past 30 days (value from 0 to 100);

[0212] Number of complaints received in the past 7 days for inducing cancellations (non-negative integers).

[0213] All of the above features are standardized or binned before being input into the model; if a feature is missing (e.g., no call records), it is filled with the default value defined by the business (e.g., voice semantic features are set to 0).

[0214] II. Model Output

[0215] The model output includes two items:

[0216] 1. Probability of violation p∈[0,1]: This represents the confidence level that the pick-up event belongs to "driver responsibility violation" (such as inducing passengers to cancel, maliciously taking a detour, or refusing to pick up without reason);

[0217] 2. SHAP Interpretability Contribution: Assign a contribution score to each input feature to generate readable explanations of the judgment criteria (e.g., "Yaw rate contribution +0.32, induced semantic confidence contribution +0.41").

[0218] The system makes a responsibility determination based on a preset threshold T (e.g., T = 0.95).

[0219] If p ≥ T: it is judged as a high-confidence violation, and business penalties will be automatically executed (such as deducting service points or suspending order taking);

[0220] If p < T: it is determined to be a low-confidence event, automatically transferred to the manual review queue, and the appeal portal is opened to the driver.

[0221] III. Key Training and Optimization Steps

[0222] 1. Sample Construction: Samples are drawn from historical ride cancellation orders, and a professional annotation team labels them according to a standardized annotation protocol (SOP) to determine "whether the driver is at fault," forming a labeled dataset {(X i ,yi)}, where yi∈{0,1};

[0223] 2. Loss function and optimization objective: A binary cross-entropy loss is adopted, and class weights are introduced to alleviate the imbalance between positive and negative samples (the proportion of illegal samples is usually <5%). The optimization objective is to maximize the precision while keeping the false positive rate (FPR) <3%, provided that the recall rate is ≥90%.

[0224] 3. Model iteration mechanism:

[0225] Each week, the results of manual review (including successful / failed appeal labels) will be used as new training samples.

[0226] Incremental training is performed once a month, updating only the leaf node split points of the model while preserving the feature importance structure;

[0227] If online monitoring detects a decrease in AUC > 0.03 or an increase in FPR > 1%, full retraining will be automatically triggered.

[0228] IV. Lightweight Real-Time Inference Submodel

[0229] To support low-latency risk alerts during the pick-up process, the system additionally deploys a simplified model whose input only includes the following 5 dimensions of time-sensitive features:

[0230] Is the ship currently yawed (0 / 1)?

[0231] Has the passenger's phone been connected (0 / 1)?

[0232] GPS signal-to-noise ratio (dBHz);

[0233] Time elapsed since order was accepted (in seconds);

[0234] Driver's current service score (0-100).

[0235] The aforementioned sub-model is deployed on edge service nodes, with an end-to-end inference latency of less than 200 milliseconds. It outputs a "high-risk behavior" warning signal to trigger a pop-up reminder on the driver's end (such as "Please follow the navigation to avoid affecting your service score").

[0236] It should be noted that this also includes the generation, application, and self-learning update mechanism of the judgment logic:

[0237] I. Generation method of judgment logic: guided by expert definition and based on data mining:

[0238] 1. Initial source of liability judgment logic

[0239] Expert experience is transformed into feature engineering and annotation standards: Business risk control experts define the criteria for judging violations such as "inducing cancellation", "malicious detours", and "abuse of calls", and formulate structured annotation SOPs accordingly (e.g.: "If a driver says 'cancel' during a call and does not report traffic problems, it is considered inducement").

[0240] Instead of directly writing down rules, it transforms expert knowledge into quantifiable features (such as yaw rate and induced semantic confidence) and high-quality labeled samples.

[0241] 2. The core accountability logic is automatically mined by the model.

[0242] Supervised learning (such as LightGBM) can automatically learn feature combinations and weights on labeled data to form a judgment boundary;

[0243] The SHAP interpretability analysis output by the model can reveal in reverse which feature combinations constitute the violation criteria. Essentially, it is an interpretable dynamic rule set (e.g., "Yaw rate > 30% + induced semantic confidence > 0.8 + no reporting → violation probability 0.96").

[0244] II. Application of the accountability logic: Confidence level classification plus dual-track execution:

[0245] 1. Automatic execution with high confidence:

[0246] When the model outputs a violation probability p≥T (e.g., T=0.95), the system automatically triggers penalties (e.g., deducting service points, restricting order acceptance) without manual intervention.

[0247] The basis for liability determination is generated in the form of SHAP contribution value and can be read by the driver (e.g., "Due to failure to report and the inclusion of leading statements in the call, liability is established").

[0248] 2. Low confidence level leads to manual review:

[0249] When p < T, the system automatically opens the appeal portal and pushes the request to the manual review queue;

[0250] Reviewers can view the integrated data view (trajectory + call summary + weather + report) and make a final decision.

[0251] III. Self-learning and updating mechanism of judgment logic: Closed-loop feedback drives model iteration:

[0252] 1. Automatic feedback data return:

[0253] All manually reviewed results (including driver appeal success / failure and arbitration conclusions) are automatically labeled as new training samples and added to the training pool;

[0254] Example: If a driver successfully appeals, the system marks the sample as a "negative sample (non-violation)" to correct the model's misjudgment of similar scenarios.

[0255] 2. Dynamically trigger model retraining:

[0256] Regular incremental training: Monthly incremental learning is performed based on new feedback data to update model parameters;

[0257] Abnormal trigger full retraining: When online monitoring metrics (such as AUC, FPR) exceed preset thresholds (such as FPR > 3%), the full retraining process is automatically started.

[0258] 3. Continuous optimization through feature engineering:

[0259] By analyzing bad cases (such as dialect inducement not being recognized, and falsified reports not being detected), we iteratively add or adjust features (such as introducing "dialect area identifier" and "report and trajectory consistency verification").

[0260] Features can only be deployed after their effectiveness has been verified through A / B testing.

[0261] The following are two typical scenarios for determining liability in ride-hailing services: "driver taking a detour", "passenger maliciously canceling", and "unclear responsibilities between driver and passenger".

[0262] Scenario 1: Driver induces passenger to cancel + malicious deviation from the flight path;

[0263] Business Background

[0264] After accepting an order, the driver was unwilling to go to a remote area, so he did not hang up the phone. After the call was connected, he said, "Cancel it, I can't get there, the road conditions are too bad," while deviating from the navigation route to create the illusion that he "cannot pick up the driver."

[0265]

[0266] Traditional methods, such as looking only at deviations or keywords, are prone to missing certain cases; manual data review is inefficient, and keyword matching for coverage is ineffective.

[0267] This invention integrates "call containing leading semantics + 32% deviation + no reporting + no severe weather", the model outputs p=0.97, automatically determines responsibility, and forms a triple chain of evidence.

[0268] Scenario 2: The passenger initiates the cancellation (responsibility unclear);

[0269] Behavior: The passenger says "I don't need it" first, and the driver agrees;

[0270] Traditional methods: The call contained the keyword "cancel," leading to misjudgment of the driver; manual listening to recordings is inefficient.

[0271] This invention identifies "passenger-initiated" responses through dialogue role labeling, combines this with trajectory compliance and historical high scores, outputs p=0.05, and waives penalties, achieving accurate attribution. The specific driver identification process is shown in Figure 4.

[0272] Example 2

[0273] Please refer to Figure 2, which is a schematic diagram of a driver pick-up behavior intelligent accountability system based on multi-source heterogeneous data fusion disclosed in an embodiment of the present invention. The driver pick-up behavior intelligent accountability system described in Figure 2 can be applied to a data processing chip, processing terminal, or processing server, and the processing server can be a local server or a cloud server; the present invention does not limit this. As shown in Figure 2, the driver pick-up behavior intelligent accountability system based on multi-source heterogeneous data fusion may include the following operations:

[0274] The data acquisition and alignment module 201 is used to collect driver trajectory data, transcribed text of communication recordings, environmental information, driver reporting records and order metadata, and perform spatiotemporal alignment according to order identifiers and preset time windows to generate structured pick-up events.

[0275] Specifically, the above steps integrate data from different sources and structures, including driver's side, passenger's side, third-party services, and the platform itself, according to a unified order identifier and time benchmark. This solves the problems of data isolation and lack of context in traditional methods, transforming the originally scattered trajectory points, call records, environmental parameters, and other information into a structured event object with complete spatiotemporal correlation.

[0276] It is evident that this structured pick-up event provides a unified and standardized input foundation for subsequent feature engineering and model analysis, encapsulating complex behavioral information into data units that can be directly processed by machines, and providing the necessary data foundation for constructing multi-dimensional feature vectors and achieving accurate identification of violations and determination of responsibility.

[0277] The feature construction and training module 202 is used to extract multi-dimensional feature vectors based on the pick-up event, including basic trip features, environmental context features, speech and semantic features, and driver behavior and reputation features, and to train a gradient boosting decision tree model as a liability judgment model based on labeled samples.

[0278] Specifically, the above steps systematically extract multi-dimensional features covering itinerary, environment, semantics, and behavioral reputation from structured events, constructing an information profile that can comprehensively describe the context of pick-up behavior. The refined feature engineering in the above steps transforms complex behavioral patterns into quantifiable feature vectors, providing an information foundation for the model to go beyond simple rules and perform deep pattern recognition.

[0279] It is evident that training labeled samples with a gradient boosting decision tree model enables the accountability model to automatically learn the complex nonlinear relationship between different dimensional features and violations. This step ensures the accuracy of model discrimination while also taking into account processing efficiency and interpretability of results, providing core algorithmic support for achieving automated and highly accurate accountability determination.

[0280] The dual-track liability determination and execution module 203 is used to perform real-time lightweight risk intervention and offline high-precision liability determination using the liability determination model.

[0281] Specifically, the above steps, by deploying a lightweight model to perform real-time reasoning on high-time-sensitivity features, build the ability to intervene in risks during the event, enabling the system to instantly identify potential violations and trigger alerts during the pick-up process, transforming management actions from passive post-event handling to proactive process control.

[0282] As can be seen, this step uses a high-precision model to perform offline in-depth analysis of the complete feature set, ensuring the accuracy of the final responsibility determination. The above-mentioned dual-track parallel architecture not only meets the real-time requirements of the business, but also provides reliable data support and algorithm guarantee for the final responsibility determination decision.

[0283] The feedback optimization and update module 204 is used to feed back feedback data generated by manual review, driver appeals or platform arbitration to the training sample set, driving the judgment model to perform incremental learning or retraining.

[0284] Specifically, the above steps establish a closed-loop model optimization mechanism by feeding back the manual feedback data generated in the business process to the training process, enabling the accountability model to continuously adjust based on the latest business judgments, thereby mitigating the model performance degradation problem caused by changes in business rules or data distribution drift.

[0285] It is evident that this mechanism transforms human judgment experience into knowledge that the model can learn, enabling the synchronous evolution of the system's judgment capabilities and business cognition, reducing the reliance on periodic manual model updates, and improving the system's stability during long-term operation.

[0286] The differentiated liability assessment output module 205 is used to differentiate the weighting of the liability assessment results based on the driver's historical reputation and generate interpretable natural language liability assessment criteria.

[0287] Specifically, the above steps, by introducing the driver's historical reputation as a weighting factor in the judgment result, realize the transformation from uniform threshold judgment to personalized and precise governance. The above differentiated processing method enables the judgment strategy to take into account the context in which the behavior occurred, to form a positive incentive for drivers with high reputation, and to implement key supervision for drivers with low reputation.

[0288] As can be seen, this step transforms the judgment logic into a natural language description through model interpretability technology, making the decision basis transparent and readable. This provides factual basis for the judgment result, helps drivers understand the reasons for violations, thereby improving the acceptability of the result and reducing unnecessary appeals and disputes.

[0289] Example 3

[0290] Please refer to Figure 3, which is a schematic diagram of another intelligent accountability system for pick-up behavior based on multi-source heterogeneous data fusion disclosed in an embodiment of the present invention. As shown in Figure 3, the device may include:

[0291] Memory 301 storing executable program code;

[0292] Processor 302 coupled to memory 301;

[0293] The processor 302 calls the executable program code stored in the memory 301 to execute some or all of the steps in the intelligent judgment method for pick-up behavior based on multi-source heterogeneous data fusion disclosed in Embodiment 1 of the present invention.

[0294] Example 4

[0295] This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute some or all of the steps in the intelligent judgment method for pick-up behavior based on multi-source heterogeneous data fusion disclosed in Embodiment 1 of this invention.

[0296] Example 5

[0297] This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of the intelligent judgment method for pick-up behavior based on multi-source heterogeneous data fusion described in Embodiment 1.

[0298] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0299] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0300] Finally, it should be noted that the intelligent accountability method and system for pick-up behavior based on multi-source heterogeneous data fusion disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent liability assessment of pick-up behavior based on multi-source heterogeneous data fusion, characterized in that, The method includes: S1. Collecting driver trajectory data, transcribed text of communication recordings, environmental information, driver reporting records, and order metadata, and aligning them spatiotemporally according to order identifiers and preset time windows to generate structured pick-up events; S2. Based on the pick-up events, extracting multi-dimensional feature vectors including basic trip features, environmental context features, speech and semantic features, and driver behavior and reputation features, and training a gradient boosting decision tree model as a liability judgment model based on labeled samples; S3. Using the liability judgment model, performing real-time lightweight risk intervention and offline high-precision liability determination, wherein the real-time lightweight risk intervention performs rapid inference based on some high-time-sensitivity features and triggers driver-side reminders, and the offline high-precision liability determination outputs the violation probability based on the complete multi-dimensional feature vectors; S4. Feedback data generated from manual review, driver appeals, or platform arbitration is fed back to the training sample set to drive the liability judgment model to perform incremental learning or retraining; S5. Differentiating the weighting of the liability judgment results according to the driver's historical reputation and generating interpretable natural language liability judgment criteria.

2. The intelligent responsibility assessment method for pick-up behavior based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The S1 step specifically includes: acquiring GPS coordinate sequences, driving speed, yaw rate, and network signal quality in real time through the driver's application interface to form driver trajectory data; acquiring communication recordings through a third-party anonymous call service, and performing speech recognition and semantic understanding to extract semantic text containing information on whether there are guiding statements and dialogue roles; calling map service and weather service application interfaces to obtain real-time congestion index, weather type, precipitation intensity, and visibility environmental data for the order route; receiving the driver's proactively reported reporting type and reporting timestamp through the application to form driver reporting records; querying the estimated pick-up distance, actual pick-up duration, and whether a phone call was made from the platform's order database; and using a stream processing engine, using the order identifier as the key, associating and aligning the above five types of data within a preset time window, and encapsulating them into a unified pick-up event data object.

3. The intelligent responsibility assessment method for pick-up behavior based on multi-source heterogeneous data fusion according to claim 1, characterized in that, In step S2, the extraction of the speech semantic features specifically includes: performing automatic speech recognition on the communication recording to convert the audio signal into raw text; using a large language model to perform semantic analysis on the raw text to identify whether it contains preset inducement cancellation keywords or semantics; when inducement semantics are identified, calculating the confidence score of the semantics; based on the dialogue rounds and content, labeling whether the party who first made the cancellation request in the call is the driver or the passenger; and constructing a speech semantic feature vector by combining the existence of the inducement semantics, the confidence score, and the information of the party initiating the cancellation request.

4. The intelligent responsibility assessment method for pick-up behavior based on multi-source heterogeneous data fusion according to claim 1, characterized in that, Step S3 specifically includes: inputting the complete multidimensional feature vector into the trained gradient boosting decision tree model to calculate the probability value of the driver's violation; comparing the probability value with preset high and low thresholds; if the probability value is greater than or equal to the high threshold, automatically determining the driver's responsibility and implementing corresponding penalties; if the probability value is less than the high threshold but greater than or equal to the low threshold, marking the current order as a questionable case and automatically pushing it to the manual review queue for further processing; if the probability value is less than the low threshold, automatically determining that the driver is not responsible, and the process ends.

5. The intelligent responsibility assessment method for pick-up behavior based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The S4 step specifically includes: collecting and recording manual review conclusions, successful driver appeal cases, and platform arbitration results, and using them as sample data with new labels; feeding the new-labeled sample data back to the model training sample pool in real time or near real time through a message queue component; periodically using the expanded training sample pool to incrementally train the accountability model based on a predetermined scheduling strategy to update the model parameters; continuously monitoring the key performance indicators of the accountability model in the production environment, and automatically triggering a complete model retraining process when performance degradation exceeds a preset tolerance.

6. The intelligent responsibility assessment method for pick-up behavior based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method for generating interpretable natural language judgment criteria includes: using model interpretability technology to calculate the contribution of each feature in the multidimensional feature vector to the judgment result; selecting the top N key features with the highest contribution and their direction of influence; and converting the key features and their direction of influence into a coherent natural language description based on a predefined template, wherein the description clearly points out the main factors leading to the judgment decision.

7. The intelligent responsibility assessment method for pick-up behavior based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The differentiated weighting of the liability assessment results based on the driver's historical reputation includes: querying the driver's recent historical service rating and the number of recent valid complaints; calculating a reputation weighting factor based on the historical service rating and the number of complaints; and performing a calculation on the original violation probability output by the liability assessment model and the reputation weighting factor to obtain the weighted final violation probability. Specifically, for drivers with high reputation, the weighting factor will lower the final violation probability to implement error tolerance; for drivers with low reputation, the weighting factor will increase the final violation probability to implement strict supervision.

8. The intelligent responsibility determination method for pick-up behavior based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The real-time lightweight risk intervention includes: extracting a subset of low-computational-complexity features from the pick-up event in real time to form a lightweight feature subset, which includes instantaneous yaw status, call connection status, location signal quality, and service level; inputting the lightweight feature subset into a lightweight inference model deployed on an edge server; the lightweight inference model outputting the risk probability of the current pick-up behavior within a millisecond delay; if the risk probability exceeds the intervention threshold, immediately sending a reminder message to the driver's application or triggering a server-side outbound call reminder.

9. A smart accountability system for pick-up behavior based on multi-source heterogeneous data fusion, used to implement the smart accountability method for pick-up behavior based on multi-source heterogeneous data fusion as described in any one of claims 1-8, characterized in that, The system includes: a data acquisition and alignment module, used to collect driver trajectory data, transcribed text of communication recordings, environmental information, driver reporting records, and order metadata, and perform spatiotemporal alignment according to order identifiers and preset time windows to generate structured pick-up events; a feature construction and training module, used to extract multi-dimensional feature vectors based on the pick-up events, including basic trip features, environmental context features, speech and semantic features, and driver behavior and reputation features, and train a gradient boosting decision tree model as a liability judgment model based on labeled samples; a dual-track liability judgment execution module, used to perform real-time lightweight risk intervention and offline high-precision liability determination using the liability judgment model; a feedback optimization and update module, used to feed back feedback data generated by manual review, driver appeals, or platform arbitration to the training sample set, driving the liability judgment model to perform incremental learning or retraining; and a differentiated liability judgment output module, used to differentiate and weight the liability judgment results according to the driver's historical reputation and generate interpretable natural language liability judgment criteria.

10. A smart accountability system for pick-up behavior based on multi-source heterogeneous data fusion, characterized in that, The system includes: a memory storing executable program code; a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the intelligent accountability method for pick-up behavior based on multi-source heterogeneous data fusion as described in any one of claims 1-8.

Images