Hotel interaction sorting system and method for business travel scene

Abstract

The invention relates to the technical field of hotel systems, in particular to a hotel interactive sorting system and method for business travel scenes, and the system comprises the following processing modules: a user behavior real-time acquisition module which is used for capturing hotel searching, clicking and booking behavior data of a user; the company organization feature construction module is used for extracting dynamic data of travel policies and a historical behavior sequence of employees based on the ID of an enterprise to which the user belongs; the multi-modal feature fusion module is used for performing weighted fusion on the user personal features and the company organization features through an attention mechanism; and the deep learning sorting module is used for inputting the fusion features into a neural network model and outputting a hotel sorting list. According to the method, personal historical behaviors and enterprise travel policies are dynamically fused in real time through an attention mechanism, so that a sorting result simultaneously meets personal preferences and organization rule constraints.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hotel system, and particularly relates to a hotel interaction sorting system in a business travel scenario and a method thereof. BACKGROUND

[0002] Traditional hotel recommendation systems mainly rely on user personal historical behaviors (such as search, click, booking) for modeling and sorting. Such systems perform well in general scenarios, but have significant shortcomings in business travel scenarios:

[0003] (1) Ignoring organizational constraints: Enterprise travel policies (such as city, job level corresponding hotel difference mark, allowed hotel type / brand, internal settlement rules), budget compliance, etc. are not included in the recommendation and sorting process. The recommended results often do not meet the requirements of the enterprise, and employees need to manually filter compliant options, which reduces efficiency and is prone to illegal booking.

[0004] (2) Ignoring group synergy effect: The historical collective behavior characteristics of the employee group within the same enterprise (for example, the preference of a company to stay in certain hotels in a certain city and during a certain period, or the collective selection changes after policy updates) are not fully utilized. These group characteristics contain policy implementation and implicit preference information, which are important references for individual decision-making.

[0005] (3) Cold start problem is prominent: For new employees or employees with low travel frequency, their personal historical behavior is sparse or even missing, and traditional personalized recommendation is poor, making it difficult to provide valuable recommendations.

[0006] (4) High real-time requirement but complex processing: Business travel decisions may be affected by real-time changes in travel policies and collective behavior. Existing systems are difficult to efficiently integrate dynamic changes in personal, policy, and group features while meeting the requirement of millisecond-level response delay.

[0007] Defects of existing solutions: Current business travel management solutions usually perform simple difference mark comparison and interception / alarm in the post-processing stage of the booking process (such as after user selection). This approach is passive:

[0008] 1) Low efficiency: Users spend a lot of time searching among a large number of hotels, and finally may be told that they do not meet the policy.

[0009] 2) Experience fragmentation: Recommendation and sorting are separated from compliance checking, and the process is not smooth.

[0010] 3) Cannot guide in advance: Cannot seamlessly integrate policy compliance and group experience into the recommended results in the initial sorting phase, and cannot achieve "out-of-box compliance" intelligent sorting. SUMMARY

[0011] To solve the above technical problems, the present application provides a hotel interaction sorting system in a business travel scenario and a method thereof.

[0012] In a first aspect, the present application provides a hotel interaction sorting system in a business travel scenario, comprising the following processing modules:

[0013] A user behavior real-time collection module is configured to capture user hotel search, click, and booking behavior data.

[0014] A company organization feature construction module is configured to extract business travel policy dynamic data and employee historical behavior sequences based on the user's enterprise ID.

[0015] A multi-modal feature fusion module is configured to fuse user personal features and company organization features through an attention mechanism weighting.

[0016] A deep learning sorting module is configured to input the fused features into a neural network model and output a hotel sorting list.

[0017] Preferably, the company organization feature construction module comprises:

[0018] A difference label compliance sub-module is configured to obtain real-time feedback on difference label compliance from a business travel management platform.

[0019] A space-time sequence aggregation sub-module is configured to generate a hotel booking space-time distribution sequence of the employees of the enterprise in the last N days.

[0020] A feature update sub-module is configured to update the company dimension feature library daily.

[0021] Preferably, the multi-modal feature fusion step performs the following operations:

[0022] (1) Calculate the user behavior sparsity weight coefficient a:

[0023]

[0024] (2) Generate fused features:

[0025] Feature fusion =α·Feature user +(1-α)·Feature org

[0026] Preferably, the system further comprises a dual-channel processing architecture, comprising:

[0027] An offline training channel is configured to periodically read historical batch data and construct user-enterprise joint feature samples; perform multi-modal feature fusion to generate a training data set; and update deep learning model parameters through a stochastic gradient descent algorithm.

[0028] An online prediction channel is configured to: analyze the enterprise ID and user identity in the user search request in real time; extract pre-computed user features and enterprise dynamic features from a distributed cache; dynamically perform multi-modal feature fusion; and call a pre-trained model to generate a hotel ranking list.

[0029] Preferably, the online prediction channel further comprises:

[0030] A real-time cache management sub-module is configured to store company organization sequences within a rolling time window;

[0031] A streaming computing engine sub-module is configured to achieve millisecond-level response and end-to-end processing delay less than 100 ms;

[0032] A policy update response sub-module is configured to automatically refresh cache layer associated features when the enterprise travel policy is updated.

[0033] Preferably, the deep learning ranking module comprises:

[0034] An embedding layer processing unit is configured to encode discrete features into dense vectors;

[0035] A spatio-temporal feature extraction unit is configured to process user behavior sequences and enterprise dynamic sequences using a GRU network;

[0036] A prediction output unit is configured to generate hotel ranking scores through a fully connected layer and output a list in descending order of scores.

[0037] Preferably, the system further comprises a cold start processing module configured to:

[0038] When the number of user historical interactions is lower than a threshold, set the fusion weight coefficient α to zero; and call enterprise difference standard static features and average behavior of employees at the same level to replace missing features.

[0039] Preferably, the system further comprises an explanatory output module configured to:

[0040] Generate a company organization feature influence weight heat map; and when the user selects a hotel that exceeds the difference standard, trigger a compliance warning and mark the decision path.

[0041] In a second aspect, the application provides a hotel ranking method of the system as described in the first aspect, comprising the following steps:

[0042] S1. Real-time collection of user hotel interaction behavior and generation of a time series behavior sequence;

[0043] S2. Matching of enterprise feature libraries according to user enterprise IDs, loading of travel rules and organization behavior features;

[0044] S3. Dynamic weighted fusion of user features and enterprise features according to user behavior sparsity;

[0045] S4. The neural network outputs a hotel ranking result through the embedding layer, the GRU layer and the fully connected layer.

[0046] In a third aspect, the application provides a computer readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method according to the second aspect.

[0047] Compared with the prior art, the application has the following beneficial effects:

[0048] (1) The application dynamically fuses personal historical behavior and enterprise travel policy in real time through the attention mechanism, so that the ranking result meets both personal preferences and organizational rule constraints.

[0049] (2) The application ensures that the policy rules used by the model are the latest and contain collective wisdom through the compliance sub-module, the spatio-temporal sequence aggregation sub-module, real-time cache management and policy update response, thereby significantly reducing invalid display and post-booking compliance risks.

[0050] (3) When user behavior is sparse, the system can automatically rely on enterprise static difference marks and the mean value of behaviors of employees of the same rank to replace missing personal features, thereby providing a preliminary recommendation with reference value for users with sparse behaviors and improving the usability of new users.

[0051] (4) The application adopts a double-channel architecture: separating offline complex model training and online high-performance prediction, pre-computing and caching: the online prediction channel uses distributed cache to store pre-computed user features and dynamically updated enterprise features, and cooperates with a streaming computing engine to realize millisecond-level feature fusion and prediction response, thereby meeting the strong real-time requirements of business travel search.

[0052] (5) The application generates a company organization feature influence weight heat map, so that users or administrators can intuitively understand the influence degree of enterprise policy or group preference on a specific ranking result. When the recommendation is rejected or the selection exceeds the difference mark, decision path markers and compliance warnings are provided to assist in understanding the decision-making reasons or discovering potential problems.

[0053] (6) The embedding layer in the application effectively processes massive discrete features; the GRU layer is specially used for modeling user behavior sequences and spatio-temporal distribution sequences of enterprise employee bookings to capture dynamic trends and dependency relationships; and the fully connected layer outputs accurate ranking scores by comprehensively integrating all abstract features. This structure significantly improves the feature expression ability and nonlinear fitting ability of the model.

[0054] (7) In the application, compliance checks and group wisdom are integrated into the front end of ranking, so that users do not need to spend a lot of time on filtering non-compliant items, the top options of the search results are highly relevant and have high compliance, the booking process is greatly simplified, and the satisfaction is improved.

[0055] (8) In the present application, offline training and feature update mechanism, periodic use of new data to train the model, feature update sub-module to ensure that the company dimension feature library is refreshed daily. Policy update triggers automatic refresh cache. These mechanisms ensure that the system can continuously learn new data, adapt to policy changes and user behavior evolution, and maintain long-term effectiveness. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 Offline training data flow and module interaction flow diagram;

[0057] Figure 2 Real-time processing link flow diagram for online prediction; DETAILED DESCRIPTION

[0058] The technical solutions of the present application will be described in detail below in conjunction with specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0059] Embodiment 1

[0060] The present embodiment provides a hotel interaction sorting system and method in a business travel scenario, as shown in Figure 1 and Figure 2

[0061] The present embodiment provides a hotel interaction sorting system in a business travel scenario, including the following processing modules:

[0062] User behavior real-time collection module: capture user interaction behavior data such as hotel search, click, booking, etc. This module is deployed in a distributed log system (such as Kafka stream), which listens to user behavior events in real time (for example, a user triggers a search request through a mobile app or web interface), extracts key fields (user ID, hotel ID, timestamp, behavior type). Behavior data is stored in time series form for subsequent feature construction.

[0063] Company organization feature construction module: based on the user's enterprise ID (obtained from the user identity authentication system), dynamically extract travel policy data and employee historical behavior sequence. This module includes three sub-modules:

[0064] Difference compliance sub-module: interact with enterprise travel management platform (such as SAP Concur) through API interface, and obtain real-time difference compliance feedback (such as hotel star rating limit, price cap). Difference data includes static rules (such as enterprise preset policies) and dynamic updates (such as policy change notifications).

[0065] ​Temporal-spatial sequence aggregation sub-module: generate the temporal-spatial distribution sequence of hotel booking of the employee of the enterprise in the last N days (N = 7 days, configurable). Specifically, this sub-module aggregates data from the behavior database to form sequence features (for example, the geographical position distribution and time distribution of booking hotels) and updates through a time window (executed every morning).

[0066] Feature update sub-module: automatically update the company dimension feature library (stored in Redis cache) every day to ensure the timeliness of the features. The feature library includes enterprise ID, employee behavior sequence mean, difference standard compliance status, etc.

[0067] Multi-modal feature fusion module: use attention mechanism to weight and fuse user personal features (such as personal historical click rate) and company organizational features (such as enterprise travel dynamics). The fusion logic performs the following operations: (1) calculate the user behavior sparsity weight coefficient a: based on the number of user historical interactions (obtained by counting statistics), when the number is less than the threshold (default threshold = 5 times), a = 0 (indicating dependence on company features); otherwise, (where k is the adjustment parameter, and the default k = 0.1).

[0068] (2) generate fusion features: use the formula F

[0069] Feature fusion = a · Feature user + (1-a) · Feature org

[0070] where F user is the user personal feature vector (such as the embedded click sequence), and F company is the company organizational feature vector (such as the sequence output processed by GRU).

[0071] Deep learning ranking module: input the fusion features into the neural network model to output the hotel ranking list. This module includes:

[0072] Embedding layer processing unit: encode discrete features (such as hotel ID, city code) into dense vectors (dimension = 64) to reduce sparsity.

[0073] Temporal-spatial feature extraction unit: use GRU (Gated Recurrent Unit) network to process user behavior sequence and enterprise dynamic sequence to extract time sequence dependence (for example, learn recent booking patterns).

[0074] Prediction output unit: generate hotel ranking scores through a fully connected layer (activation function is ReLU) and output the Top-K hotel list (K = 10) in descending order of scores.

[0075] The system also integrates a dual-channel processing architecture to realize the cooperation of offline model training and online real-time prediction.

[0076] Offline training channel: periodically (daily) read historical batch data (such as user behavior logs stored in HDFS), construct user-enterprise joint feature samples; perform multi-modal feature fusion to generate training data set; update deep learning model parameters through stochastic gradient descent algorithm (optimizer is Adam).

[0077] Online prediction channel: real-time analysis of user search requests (analysis of enterprise ID and user identity); extract pre-computed user features and enterprise dynamic features from distributed cache (such as Redis); dynamically perform multi-modal feature fusion; call pre-trained model to generate hotel ranking list.

[0078] The data flow and module interaction of offline training is as shown in Figure 1 The real-time processing link of online prediction is as shown in Figure 2 .

[0079] In addition, the system also includes the following extension modules:

[0080] Cold start processing module: when the number of user historical interactions is less than the threshold (default = 2 times), the module sets the fusion weight coefficient α to zero; at the same time, call enterprise difference standard static features and the same job level employee behavior mean to replace missing features.

[0081] Explanatory output module: generate company organization feature influence weight heat map, visualize feature contribution; when the user selects a hotel that exceeds the difference standard, trigger compliance warning (push message to user end) and mark the decision path.

[0082] The embodiment provides a method for ranking system of hotel interaction in business travel scenario, including the following steps,

[0083] S1. Real-time collection of user hotel interaction behavior: the user initiates a search through the front-end interface, and the system captures behavior data (such as clicking "Shenzhen hotel") through the user behavior real-time collection module, and generates a time series behavior sequence (stored in the time series database InfluxDB).

[0084] S2. Match enterprise feature library: according to the user enterprise ID (parsed from the request header), the company organization feature construction module queries the enterprise feature library, loads the travel rules (such as difference standard price limit) and organization behavior features (such as the hot area of employee booking in the last 7 days).

[0085] S3. Dynamic weighted fusion features: the multi-modal feature fusion module calculates the user behavior sparsity weight α (such as user interaction times = 3 times, α = 0.7), and generates the fusion features according to the formula

[0086] Feature fusion = a * Feature user + (1 - a) * Feature org

[0087] The fusion process is performed in real time (millisecond level response) in the online prediction channel.

[0088] S4. Neural network ranking output: the deep learning ranking module processes the fusion features: the embedding layer encodes the discrete features, the GRU layer extracts the sequence dependency, and the full connection layer outputs the scores. Finally, the ranking list is returned.

[0089] The embodiment provides a method for supporting computer readable storage media: program code (Python / TensorFlow) is stored in an SSD, and the above steps are triggered when a processor is executed, ensuring high concurrency processing (actual QPS>1000).

[0090] In summary, the embodiment clearly and comprehensively embodies the practical application value and innovative achievements of the intelligent travel service system based on MCP, and verifies the strong advantages and broad application prospects of the system in improving the intelligent level, personalized service quality and service safety of tourism travel.

[0091] The specific embodiments further describe the present application, but it should be understood that the specific description herein should not be construed as limiting the spirit and scope of the present application. Various modifications made by those skilled in the art after reading the specification belong to the scope of protection of the present application.

Claims

1. A sorting system for hotel interactions in a business travel scenario, characterized in that, Includes the following processing modules: The real-time user behavior collection module is used to capture user hotel search, click, and booking behavior data. The company organizational feature construction module is used to extract dynamic data on travel policies and historical behavior sequences of employees based on the user's company ID. The multimodal feature fusion module is used to weightedly fuse user personal features and company organizational features through an attention mechanism; The deep learning ranking module is used to input fused features into a neural network model and output a hotel ranking list.

2. The system according to claim 1, characterized in that, The company organizational feature construction module includes: The travel compliance submodule is used to obtain real-time feedback on travel compliance from the enterprise travel management platform. The spatiotemporal sequence aggregation submodule is used to generate the spatiotemporal distribution sequence of hotel bookings made by the company's employees over the past N days; The feature update submodule is used to update the company's feature library daily.

3. The system according to claim 1, characterized in that, The multimodal feature fusion step performs the following operations: (1) Calculate the sparsity weighting coefficient α of user behavior: (2) Generate fusion features: Fe nature fusion =a·Feature user +(1-a)·Feature org 。 4. The system according to claim 1, characterized in that, It also includes a dual-channel processing architecture, package include: The offline training channel is used for: periodically reading historical batch data and constructing user-enterprise joint feature samples; performing multimodal feature fusion to generate training datasets; and updating deep learning model parameters using the stochastic gradient descent algorithm. The online prediction channel is used for: real-time parsing of enterprise IDs and user identities in user search requests; extraction of pre-computed user features and enterprise dynamic features from distributed cache; dynamic execution of multimodal feature fusion; and calling pre-trained models to generate a hotel ranking list.

5. The system according to claim 4, characterized in that, The online prediction channel also includes: The real-time cache management submodule is used to store the company organization sequence within a rolling time window; The streaming computing engine submodule is used to achieve millisecond-level response and end-to-end processing latency of less than 100ms; The policy update response submodule is used to automatically refresh the cache layer associated features when the enterprise travel policy is updated.

6. The system according to claim 1, characterized in that, The deep learning ranking module includes: Embedding layer processing units are used to encode discrete features into dense vectors; The spatiotemporal feature extraction unit uses a GRU network to process user behavior sequences and enterprise dynamic sequences; The prediction output unit generates hotel ranking scores through a fully connected layer and outputs a list in descending order of scores.

7. The system according to claim 1, characterized in that, It also includes a cold start processing module, used for: When the number of historical user interactions is below the threshold, the fusion weight coefficient α is set to zero; the missing features are replaced by the enterprise differential static features and the mean behavior of employees at the same job level.

8. The system according to claim 1, characterized in that, It also includes an interpretable output module for: Generate a heatmap showing the impact of company organizational characteristics on weights; when a user selects a hotel that exceeds the acceptable range, trigger a compliance alert and mark the decision path.

9. A hotel sorting method based on the system of any one of claims 1-8, characterized in that, Includes the following steps: S1. Real-time collection of user hotel interaction behaviors and generation of time-series behavior sequences; S2. Match the enterprise feature library based on the user's enterprise ID, and load travel rules and organizational behavior characteristics; S3. Dynamically weight and fuse user features and enterprise features based on the sparsity of user behavior; S4. The hotel ranking results are output through a neural network consisting of an embedding layer, a GRU layer, and a fully connected layer.

10. A computer-readable storage medium storing a computer program, characterized in that, When the program is executed by the processor, it implements the steps of the method of claim 9.

Images