CN-121998202-A - AI-based business trip scene user demand prediction and service recommendation method

Abstract

The invention provides an AI-based business trip scene user demand prediction and service recommendation method, which relates to the technical field of AI recommendation and comprises the steps of obtaining user time sequence workflow and enterprise trip history data, constructing a multi-mode dataset, constructing a workflow continuity loss coefficient WCLI model by utilizing a two-stage hybrid strategy training model and a user dynamic knowledge map, quantitatively evaluating potential impact of trip on work efficiency, generating a differential intelligent recommendation scheme according to WCLI risk level, providing workflow compensation and relief suggestions for high-risk trips, automatically executing workflow compensation services such as schedule adjustment bound with a selected scheme, outputting a comprehensive decision support report and returning user feedback to the model.

Inventors

• WANG JUN
• KONG QINGHE
• WANG FENG
• ZHENG JIA
• Feng Dezhang

Assignees

• 北京悦途出行科技(集团)股份有限公司

Dates

Publication Date

20260508

Application Date

20260228

Claims (10)

1. 1. An AI-based business trip scene user demand prediction and service recommendation method is characterized by comprising the following steps: s10, acquiring original workflow data of a user as model input data, and combining desensitized enterprise travel history data to collect and construct a multi-mode training and reasoning data set covering user behaviors, schedules, a collaboration network and an external environment according to data association rules; S20, setting a two-stage hybrid training strategy, reading samples from a data set, and performing off-line pre-training and on-line fine tuning on a core prediction model to obtain the prediction model; S30, deploying the prediction model on a real-time data processing platform, connecting an enterprise schedule system, a travel management system and a third party data source through a standardized interface, constructing a user personal dynamic knowledge graph and a real-time calculation engine, and realizing continuous perception and analysis of travel intention embedded in workflow; S40, constructing a quantitative evaluation model of workflow continuity influence, comprehensively evaluating the correlation features extracted by using a knowledge graph and the output of a prediction model, combining the core work conflict degree, the cooperative network interruption degree and the recovery load prediction quantity related to the journey to form a workflow continuity loss coefficient WCLI, and performing risk evaluation and scheme previewing in a simulation environment and a real scene; And S50, generating a candidate travel service scheme by adopting a differentiated intelligent recommendation and negotiation strategy according to different risk intervals in which the numerical value WCLI is located, and automatically generating workflow compensation and relief suggestions for a high-risk trip.
2. 2. The method of claim 1, wherein the method further comprises S60 executing an automated workflow compensation service that binds with the end user selection scheme, including automatically adjusting a schedule, generating a work delivery summary, and enabling a post-return attention mode; S70, outputting a comprehensive decision support report comprising a demand prediction basis, a risk assessment process, a recommendation scheme and the executed compensation measures, and feeding back user feedback to a model training and knowledge graph updating process.
3. 3. The method of claim 1, wherein step S10 comprises: S101, synchronizing structured calendar event data in a preset time period from an enterprise office system of a user through an authorization interface, wherein data fields comprise event titles, start-stop time stamps, places, participant lists, event category labels and belonging item IDs; S102, collecting user working context data through a security agent and a privacy computing middleware, wherein the user working context data comprises the steps of carrying out natural language processing on an email title and a meeting invitation text sent and received by a user in a latest preset time period, extracting high-frequency project names, customer names and product terms as keywords, acquiring the current state, the last updating time and the next milestone date of the responsible projects in a collaborative tool, and acquiring a list of affiliated team members and common collaborative colleagues of the user from an enterprise HR system.
4. 4. The method of claim 1, wherein step S20 comprises: S201, setting an offline pre-training stage, namely using a record of completed travel in a historical data set as a positive sample, using all relevant schedule, mail and project state data in a preset number of days before a travel date as an input feature sequence, using travel order details as labels, and training a time sequence converter network as a basic prediction model; S202, setting an online incremental learning stage, namely taking all interaction behaviors predicted and recommended by a user to a system and subsequent actual travel results as incremental training samples with reward signals, and finely adjusting a basic prediction model every time a preset number of new samples are accumulated or every preset number of days; Step S30 includes: S301, deploying the prediction model obtained in the S20 into a real-time micro-service, automatically running in a preset period, and inputting a latest calendar event sequence, a latest work context snapshot and a real-time refreshed knowledge graph feature vector for a user; S302, correlating the prediction result output by the model with flight, hotel inventory and price information queried in real time from a global distribution system; s303, constructing and maintaining a user personal dynamic knowledge graph based on graph database technology, wherein graph nodes comprise users, cities, airports, contacts, projects and companies.
5. 5. The method of claim 1, wherein the workflow continuity loss factor WCLI is calculated in step S40 by weighting and summing a core work conflict level, a collaborative network interrupt level and a recovery load prediction amount, wherein the core work conflict level is used for quantifying a severity of a time overlap between a predicted travel time window and a key calendar event existing in a user calendar; The collaborative network outage degree is used for quantifying the risk of collaborative work outage generated to a team or a close collaborative network where the user leaves during the prediction travel; The recovery load pre-measurement is used for quantifying the extra workload which is expected to be input for processing backlog transactions after the user finishes business trip return; Based on the values of the workflow continuity loss coefficients WCLI, the defined risk level thresholds include: when WCLI is less than the first threshold, determining a low workflow continuity risk; When WCLI is greater than or equal to the first threshold and less than the second threshold, determining that the medium workflow continuity risk; when WCLI is greater than or equal to the second threshold, a high workflow continuity risk is determined.
6. 6. The method of claim 1, wherein step S50 comprises: if WCLI is smaller than a first threshold, an active seamless recommendation mode is adopted, namely, taking the shortest total journey time consumption and the lowest journey pressure index as multiple targets, solving an optimal integrated package by utilizing an optimization algorithm on the premise of conforming to the business travel policy, and automatically generating a fine adjustment suggestion of the schedule with the highest contribution to WCLI; if WCLI is greater than or equal to the first threshold, starting a risk early warning and negotiation mode, namely pushing a high risk warning to a user, simultaneously providing alternatives including a local alternative scheme, a travel splitting scheme and a compensation front-end scheme, waiting for the user to confirm the requirement clearly, and providing limited travel scheme selection.
7. 7. The method of claim 2, wherein step S60 comprises: When the user confirms and completes the reservation of the travel scheme, the system automatically triggers a preset compensation rule executor, wherein the rule comprises automatically adjusting the state in the enterprise calendar to be a meeting of a preset category on the first day after the user arrives at the destination city and returns to the residence if the journey is a long-distance international flight; automatically extracting key update abstracts from a project management system and a mail system of the user at a preset time before the user returns to the day and before the user returns to the day, generating a working brief during off-duty and sending the working brief to a mailbox of the user; And automatically activating a do-not-disturb mode of the instant messaging software of the user in a preset time period of the first working day after the user returns.
8. 8. The method of claim 2, wherein step S60 comprises the system automatically generating a comprehensive decision support report after service closure, the report including demand forecast backtracking, risk assessment logs, recommendation comparisons, user decisions and feedback, and compensation measures performed: the demand prediction traceability shows the core schedule event, project state and cooperation relation of trigger prediction, the risk assessment log details the calculation process and the numerical value of each sub-term of WCLI, the recommended scheme is compared with all provided schemes and key parameters thereof, the user decision and feedback are the final selection and the evaluation of the recommendation of the user, the executed compensation measures are all workflow compensation operation lists automatically completed by the system, and the user feedback is returned to the model training and knowledge map updating process.
9. 9. The method of claim 1, further comprising obtaining multi-source data and constructing a user workflow hypergraph model, wherein nodes in the workflow hypergraph represent work task units, and superedges represent workflows, and connect to the task unit nodes to which the nodes belong; The method further comprises the step of carrying out joint intention recognition and influence deduction based on a mixed expert network, wherein the mixed expert network comprises a gating network, a trip requirement triggering expert, a workflow toughness assessment expert and a joint strategy generation expert, and the gating network is used for receiving a state characteristic vector of a current workflow hypergraph and outputting a weight distribution vector to dynamically allocate the contribution degree of a downstream expert sub-network; The trip demand triggering expert is used for identifying task nodes which are in the workflow hypergraph, have the space elasticity coefficient Se lower than a preset threshold and contain external entities in the core collaborative person set, judging the necessity and urgency of triggering the offline trip demand and outputting a potential trip event set; The workflow toughness evaluation expert is used for evaluating the affected range and bearing capacity of the original workflow hypergraph of the user if the user leaves in the time window aiming at the potential trip event; the combined strategy generation expert is used for triggering the output of the expert and the workflow toughness evaluation expert according to the comprehensive travel requirement and actively generating a preliminary workflow self-adaptive adjustment strategy set; The method further comprises the steps of executing dynamic conjugate optimization, generating a business trip-workflow combined action plan, defining a travel variable group X and a workflow adjustment variable group Y which are conjugate with each other, constructing an optimization problem aiming at minimizing a global disturbance index GPI, wherein the global disturbance index GPI is a comprehensive index obtained by carrying out weighted summation calculation based on the total currency cost of the business trip, the total in-transit time of a user, the total delay days of a workflow key path and the estimated number of new coordination sessions which need to be initiated due to adjustment; The method further comprises the steps of carrying out path dependent intensity analysis and comprehensive risk grading, wherein the step of calculating path dependent intensity index PDS, wherein PDS is based on the central change of the medium number in the workflow hypergraph before and after adjustment and the importance weight of the node task in a task node set changed due to workflow adjustment; the method further comprises the steps of initiating joint interrogation and dynamic revision of man-machine cooperation, displaying a complete joint action plan interaction view comprising a travel scheme, workflow adjustment comparison, risk analysis and cooperation state to a user, and restarting local optimization in response to a modification instruction of the user.
10. 10. The method of claim 2, further comprising performing automated collaboration and workflow injection, automatically performing travel reservations after the user approves the joint action plan, automatically performing approved calendar changes, meeting mode conversions, and task reassignment operations through the API of the enterprise collaboration tool, sending change notifications, inserting buffer and resume task events in the user's calendar according to the plan; the method further comprises the steps of generating a joint action traceability report and model evolution, wherein the report elaborates an optimization decision process, and feeding back actual execution effect data fed back by a user and planned to the mixed expert network of the S2 as a reinforcement learning signal.

Description

AI-based business trip scene user demand prediction and service recommendation method Technical Field The invention relates to the technical field of AI recommendation, in particular to an AI-based business trip scene user demand prediction and service recommendation method. Background In the business trip service field, iteration of information technology is always developed around the promotion of reservation efficiency and service individuation, the industry realizes online instant retrieval and transaction of business trip services such as air tickets, hotels and vehicles by building a full-class product database and a transaction engine, a price comparison tool and a regularized recommendation function which are subsequently emerging further help users to quickly screen options meeting preset conditions such as price, time and position, a core foundation of modern online business trip services is laid together, but as business activities become more complex, the demands of users on business trip services are achieved from simple information query and transaction, the auxiliary experience of deep adaptation of working flow and full-flow intelligent adaptation is turned to, the limitation of the prior art scheme is increasingly highlighted, firstly, the service mode of passive reaction is adopted, the current main flow business trip platform takes the active query initiated by the users as a service starting point, even if the preference recommendation based on historical order is introduced, the static induction of the past behavior is still carried out, and the active pre-judging capability on the future journey demands of the users is lacking. The system can record the hotel and flight preference selected by the user and display the preference preferentially, but can not be combined with the latest work schedule, project node and business travel policy change of the user, sense the un-started reservation demand in advance, and can not dynamically generate a travel scheme adapting to the whole scene, so that the user still needs to invest a great deal of effort to complete the journey planning and multiparty comparison, and the intelligent stay on the surface layer only. Secondly, the understanding depth and breadth of the prior art on the user demands are insufficient, most recommendation models only depend on structured historical transaction data, but complex and dynamic unstructured background factors behind business trips cannot be captured, and deep trip intentions hidden behind selection are difficult to read. Meanwhile, the business trip is not an isolated event, but is embedded into an important link of a user continuous workflow, and the purpose of single trip, the staff of the same person, subsequent meeting arrangement and the like jointly form a complete context of decision making, but the prior art always looks at each booking behavior in an isolated way, the priority and constraint condition of the user decision cannot be judged based on the continuous context, and real accurate adaptation is difficult to realize. In addition, the existing service recommendation dimension is single, cross-scene resource integration and global optimization cannot be realized, modules such as air tickets, hotels and traffic in the traditional recommendation engine are mutually independent, recommendation results are simply spliced by local optimal solutions, instead of a global optimal scheme based on the whole travel experience of a user, the user has to independently and comprehensively manage multi-link information, and repeatedly switch and check among a plurality of platforms, if the service optimization focused on 'traveling', interference of business travel behaviors on work continuity in non-travel periods before and after traveling is easily ignored, hidden efficiency loss caused by the business travel behaviors is easily ignored, and low-price red-eye flights recommended by the system can cause the influence on the work efficiency of a core meeting of the next day due to insufficient rest while the optimal traveling cost is realized, and the hidden cost is far higher than saved ticket cost. Therefore, a brand new business trip scene user demand prediction and service recommendation method based on AI is urgently needed in the market. Disclosure of Invention The invention aims to provide an AI-based business trip scene user demand prediction and service recommendation method, which solves the problems that in the prior art, a traditional business trip service platform only can passively process dominant trip requests and cannot actively predict potential trip arrangements embedded in continuous workflows when responding to user demands, deep trip intention and work continuity interference risks under complex business backgrounds are difficult to identify, and effective perception and predictive relief measures are lacking in non-trip period work efficiency hidden loss and recovery demands caused b