Business travel recommendation method and system based on collaborative filtering and recurrent neural network

Abstract

The invention relates to the field of business travel management, in particular to a business travel recommendation method and system based on collaborative filtering and a recurrent neural network, and the method comprises the steps: obtaining an approved business travel application form, and extracting business travel information; generating an initial recommendation scheme through a user-user collaborative filtering algorithm and an article-article collaborative filtering algorithm; a user historical travel record is called, a user travel portrait is constructed, a time sequence model of user behaviors is trained through a time sequence neural network architecture, and the trained time sequence model generates a personalized travel scheme through Monte Carlo simulation; inputting the initial recommendation scheme and the personalized travel scheme into a dynamic filtering module, and outputting a final travel recommendation scheme; and pushing the final travel recommendation scheme to a user side, receiving feedback information of a user, and optimizing a collaborative filtering module and an LSTM time sequence analysis module based on the feedback information. The travel delay in emergency is effectively reduced, and the travel experience of employees is improved.

Description

Technical Field

[0001] This application relates to the field of business travel management, and specifically to a business travel recommendation method and system based on collaborative filtering and recurrent neural network. Background Art

[0002] In the field of business travel management, with the increasing frequency of corporate business travel activities and the diversified and personalized development of business travel needs, traditional business travel management methods have gradually exposed many problems and are difficult to meet the refined and intelligent needs of modern corporate business travel management. The specific problems are as follows: In the traditional business travel process, data in key links such as OA approval, employee reservation, and financial reimbursement are stored independently of each other, forming isolated information islands. According to statistics, most enterprises still adopt the model of "employees book by themselves + reimburse with pasted tickets afterwards", which makes it impossible to effectively link data between departments such as administration, finance, and travel. Even if some enterprises introduce business travel platforms, the systems are not connected. For example, the business travel platform and the financial system are separated from each other. This situation directly leads to low efficiency in the implementation of business travel policies and serious lag in budget monitoring. Once compliance risks occur, it is very difficult to conduct effective traceability.

[0003] When employees make business travel reservations, they need to jump back and forth between multiple different platforms to compare prices, consuming a lot of time and energy. Moreover, the refund and change process is cumbersome and complex. In case of emergencies, employees are extremely likely to delay their trips due to untimely refund and change, seriously affecting work arrangements and having a poor experience. Summary of the Invention

[0004] In view of the above problems, a more advanced and intelligent business travel recommendation method is needed, which can fully consider the personalized needs and time series characteristics of users, combine real-time information and user feedback, and achieve an organic integration of multiple recommendation methods, thereby improving the accuracy, practicability, and user satisfaction of business travel recommendations. Based on this, the present invention proposes a business travel recommendation method and system based on collaborative filtering and recurrent neural network.

[0005] In a first aspect, the technical solution of the present invention provides a business travel recommendation method based on collaborative filtering and recurrent neural network, including the following steps: Obtain the approved business travel application form and extract business travel information, including destination, date, rank, and business travel standard; Input the business travel information into the collaborative filtering module, and generate an initial recommendation plan through user-user collaborative filtering and item-item collaborative filtering algorithms; Input the business travel information into the LSTM time series analysis module, retrieve the user's recent M historical travel records, construct a user travel profile, train a time series model of user behavior through a time series neural network architecture, and generate a personalized travel plan through Monte Carlo simulation by the trained time series model; Input the initial recommendation plan and the personalized travel plan into the dynamic filtering module, perform secondary screening in combination with the real-time inventory of the travel suppliers, the budget constraint based on the travel standards, and the user travel portrait tags, and output the final travel recommendation plan; Push the final travel recommendation plan to the user terminal, receive the feedback information of the user, and optimize the collaborative filtering module and the LSTM time series analysis module based on the feedback information.

[0006] The combination of the collaborative filtering module and the LSTM time series analysis module can accurately analyze the travel needs and preferences of users, improve the utilization rate of contracted hotels, reduce the situation of employees choosing high-price individual traveler channels. It can effectively reduce the travel cost burden of enterprises. The automated recommendation and screening process greatly reduces the workload of manual review and reduces the financial labor cost. At the same time, accurate recommendations reduce unnecessary itinerary changes and refunds, saving a large amount of funds for enterprises. The dynamic filtering module strictly screens according to travel standards, and can effectively avoid problems such as over-standard reimbursement and non-compliant bills. Through real-time monitoring and early warning, potential violations can be discovered and processed in a timely manner, reducing the economic losses caused by illegal expenditures for enterprises.

[0007] As a further limitation of the technical solution of the present invention, the steps for the collaborative filtering module to generate the initial recommendation plan include: Calculate the user interest matching degree through cosine similarity or Pearson correlation coefficient, identify the neighbor group with similar preferences to the target user, and generate a group preference recommendation list based on the behavior data of the neighbor group; Analyze the associated behavior sequence of the user's hotel reservation and transportation tool selection, introduce a time decay factor, and construct a user-item interaction matrix; Adopt a modified cosine similarity algorithm to calculate the similarity between items and generate an associated item recommendation list; Perform weighted fusion on the recommended items of the group preference recommendation list and the associated item recommendation list to form an initial recommendation plan.

[0008] As a further limitation of the technical solution of the present invention, the steps for performing weighted fusion on the recommended items of the group preference recommendation list and the associated item recommendation list to form an initial recommendation plan include: Take the user interest matching degree as the initial weight of the recommended items in the group preference recommendation list, and take the modified cosine similarity score as the initial weight of the recommended items in the associated item recommendation list; Normalize the weights of the two types of recommendations to the same dimension; Calculate the scores of the group preference recommendation and the associated item recommendation respectively; Retain the top N items sorted by the weighted total score as the initial recommendation plan.

[0009] As a further limitation of the technical solution of the present invention, the steps for the LSTM time series analysis module to generate a personalized travel plan for the user include: Retrieve the user's historical travel records in the recent M times to generate a historical behavior sequence, and the historical travel records include structured data and unstructured note information; Parse the unstructured note information through natural language processing technology to generate user portrait tags; Jointly input the user portrait tags and the historical behavior sequence into the trained LSTM model, and capture time series features through the time decay weighted sum and attention mechanism; Generate the probability distributions of destinations, hotels, and flights at the output layer of the LSTM model; Generate a personalized travel plan through Monte Carlo simulation based on the probability distribution.

[0010] As a further limitation of the technical solution of the present invention, the steps for generating a personalized travel plan through Monte Carlo simulation based on the probability distribution include: Determine a candidate solution set according to the probability distribution output by the LSTM; Superimpose external influence factors on each candidate solution; Conduct N random samplings. Each time of sampling, randomly select a candidate solution according to the weighted probability distribution, and calculate the comprehensive score of the probability weight + difference standard compliance + real-time inventory availability of the candidate solution; Statistically analyze all sampling results, and output the top K solutions with the highest comprehensive scores among the selected candidate solutions, that is, the personalized travel plan.

[0011] As a further limitation of the technical solution of the present invention, the steps for inputting the initial recommended plan and the personalized travel plan into the dynamic filtering module, and performing secondary screening in combination with the real-time inventory of the travel supplier, the budget constraint based on the travel standard, and the user travel portrait tags to output the final travel recommendation plan include: Take the union of the initial recommended plan generated by the collaborative filtering module and the personalized travel plan generated by the LSTM time series analysis module; Obtain the real-time inventory through the supplier API interface, and retrieve the budget constraint from the enterprise travel standard database; Perform compliance filtering based on the real-time inventory, budget constraint, and user portrait tags, and exclude the items that exceed the standard; Generate the final travel recommendation plan for the selected options according to the preset weights; among them, the weight A of the personalized plan is greater than the initial recommended weight B.

[0012] As a further limitation of the technical solution of the present invention, the steps for optimizing the collaborative filtering module and the LSTM time series analysis module based on the feedback information include: Adjust the weights of the recommended items in the group preference recommendation list and the weights of the recommended items in the associated item recommendation list in the collaborative filtering module according to the user's click and reservation behaviors; Update the user portrait tags according to the user's click and reservation behaviors, and then update the attention mechanism parameters of the LSTM model in the LSTM time series analysis module.

[0013] In a second aspect, the technical solution of the present invention further provides a business travel recommendation system based on collaborative filtering and recurrent neural network, including a business travel application form receiving module, a collaborative filtering module, an LSTM time series analysis module, a dynamic filtering module, and a recommendation optimization module; The business travel application form receiving module is used to obtain the approved business travel application form and extract business travel information, including the destination, date, rank, and business travel standard; The collaborative filtering module is used to generate an initial recommendation plan based on the business travel information through user-user collaborative filtering and item-item collaborative filtering algorithms; The LSTM time series analysis module is used to retrieve the user's recent M historical travel records according to the business travel information, construct a user travel portrait, train a time series model of the user's behavior through a time series neural network architecture, and generate a personalized travel plan through Monte Carlo simulation for the trained time series model; The dynamic filtering module is used to perform secondary screening based on the initial recommendation plan and the personalized travel plan, combined with the real-time inventory of business travel suppliers, the budget constraint based on the business travel standard, and the user travel portrait tags, and output the final business travel recommendation plan; The recommendation optimization module is used to push the final business travel recommendation plan to the user side, receive the user's feedback information, and optimize the collaborative filtering module and the LSTM time series analysis module based on the feedback information.

[0014] As a further limitation of the technical solution of the present invention, the steps for the collaborative filtering module to generate an initial recommendation plan include: Calculate the user interest matching degree through cosine similarity or Pearson correlation coefficient, identify the neighbor group with similar preferences to the target user, and generate a group preference recommendation list based on the behavior data of the neighbor group; Analyze the associated behavior sequence of the user's hotel reservation and transportation tool selection, introduce a time decay factor, and construct a user-item interaction matrix; Use the modified cosine similarity algorithm to calculate the similarity between items and generate an associated item recommendation list; Perform weighted fusion on the recommended items in the group preference recommendation list and the associated item recommendation list to form an initial recommendation plan.

[0015] As a further limitation of the technical solution of the present invention, the collaborative filtering module forming the initial recommendation plan specifically includes: Use the user interest matching degree as the initial weight of the recommended items in the group preference recommendation list, and use the modified cosine similarity score as the initial weight of the recommended items in the associated item recommendation list; Normalize the weights of the two types of recommendations to the same dimension; Calculate the scores of group preference recommendation and associated item recommendation respectively; Sort by the weighted total score and retain the top N items as the initial recommendation plan.

[0016] As a further limitation of the technical solution of the present invention, the steps for the LSTM time series analysis module to generate a user personalized travel plan include: Retrieve the user's recent M historical travel records, including structured data and unstructured note information; Parse the unstructured note information through natural language processing technology to generate user portrait tags; Jointly input the user portrait tags and the historical behavior sequence into the trained LSTM model, and capture time series features through the time decay weighted sum and attention mechanism; Generate the probability distributions of destinations, hotels, and flights at the output layer of the LSTM model; Generate a personalized travel plan based on the probability distribution through Monte Carlo simulation.

[0017] As a further limitation of the technical solution of the present invention, the steps for the LSTM time series analysis module to generate a personalized travel plan based on the probability distribution through Monte Carlo simulation include: Determine the candidate solution set according to the probability distribution output by the LSTM; Overlay external influence factors on each candidate solution; Perform N random samplings. Each time a sampling is performed, randomly select a candidate solution according to the weighted probability distribution, and calculate the comprehensive score of the probability weight + difference standard compliance + real-time inventory availability of the candidate solution; Statistically analyze all sampling results, and output the top K solutions with the highest comprehensive scores among the selected candidate solutions, that is, the personalized travel plan.

[0018] As a further limitation of the technical solution of the present invention, the dynamic filtering module is specifically used to take the union of the initial recommendation plan generated by the collaborative filtering module and the personalized travel plan generated by the LSTM time series analysis module; Obtain real-time inventory through the supplier API interface, and retrieve the budget constraint from the corporate travel standard database; Perform compliance filtering based on real-time inventory, budget constraint, and user portrait tags, and exclude items that exceed the standard; Generate the final business travel recommendation plan for the selected options according to the preset weight; among them, the weight A of the personalized plan is greater than the initial recommendation weight B.

[0019] As a further limitation of the technical solution of the present invention, the recommendation optimization module is specifically configured to adjust the weights of the recommended items in the group preference recommendation list and the weights of the recommended items in the associated item recommendation list in the collaborative filtering module according to the user's click and reservation behaviors; update the user portrait tags according to the user's click and reservation behaviors, and then update the attention mechanism parameters of the LSTM model in the LSTM time series analysis module.

[0020] From the above technical solutions, it can be seen that the present application has the following advantages: The business trip recommendation method based on collaborative filtering and recurrent neural network of the present invention integrates the information of the approved business trip application forms and inputs them into different modules for processing, realizing the centralized analysis and utilization of data. The data interaction between each module breaks the traditional information silos, enabling departments such as administration, finance, and travel to share data, thereby improving the implementation efficiency of business trip policies, realizing real-time budget monitoring, and effectively tracing compliance risks.

[0021] This method provides users with a one-stop business trip recommendation service, avoiding the cumbersome process of employees jumping between multiple platforms to compare prices. At the same time, through accurate recommendations and reasonable itinerary planning, the need for itinerary changes is greatly reduced. Even if itinerary changes are required, the system can respond quickly, thus effectively reducing itinerary delays in emergency situations and enhancing the business trip experience of employees. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the present application, the drawings required for description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is a schematic flowchart of the method provided by the embodiment of the present invention.

[0024] Figure 2 It is a block diagram of the system provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] In order to make the application purpose, features, and advantages of the present application more obvious and understandable, the technical solutions protected by the present application will be clearly and completely described below by using specific embodiments and the accompanying drawings. Obviously, the embodiments described below are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0026] Such as Figure 1As shown in the figure, an embodiment of the present invention provides a business trip recommendation method based on collaborative filtering and recurrent neural network, including the following steps: S1: Obtain the approved business trip application form and extract the business trip information, including the destination, date, job level, and business trip standard; The system establishes an interface connection with the enterprise's OA approval system. When the business trip application form completes the approval process in the OA system, the OA system will send the application form data to this business trip recommendation system through the interface. After receiving the application form data, the system uses regular expressions or data parsing tools to extract information such as the destination, date, job level, and business trip standard from the application form. For example, for the date information, a date parsing library can be used to convert the text-formatted date into a date object recognizable by the system.

[0027] The parsed fields include: destination (city code), date (departure / return timestamp), job level (such as P7 / M3), and business trip standard (JSON structure, including air ticket / hotel budget limit).

[0028] S2: Input the business trip information into the collaborative filtering module, and generate an initial recommendation plan through user-user collaborative filtering and item-item collaborative filtering algorithms; After receiving the business trip information, the collaborative filtering module uses it as a screening condition to screen out user behavior data related to the current business trip information from the user behavior database. Execute user-user collaborative filtering and item-item collaborative filtering algorithms respectively.

[0029] S3: Input the business trip information into the LSTM time series analysis module, retrieve the user's recent M historical travel records, construct a user travel profile, train a time series model of user behavior through a time series neural network architecture, and generate a personalized travel plan through Monte Carlo simulation after the trained time series model; After receiving the business trip information, the LSTM time series analysis module retrieves the user's recent M historical travel records from the historical travel record database according to the user identification. Preprocess the historical travel records, including data cleaning, feature extraction, etc., and construct a user travel profile. Use a time series neural network architecture (such as LSTM) to train the time series of user behavior. After training is completed, use the Monte Carlo simulation method to generate a personalized travel plan.

[0030] In the embodiment of the present invention, query the user's recent M = 12 historical orders from a distributed database (such as HBase); parse the remarks through NLP (such as extract the "prefer quiet hotel" label by the BERT model); load the pre-trained LSTM model, input the time series features + profile labels, and output the probability distribution.

[0031] S4: Input the initial recommended plan and the personalized travel plan into the dynamic filtering module, perform secondary screening in combination with the real-time inventory of travel suppliers, budget constraints based on travel standards, and user travel profile tags, and output the final travel recommendation plan. Merge the candidate sets of the initial recommendation and the personalized plan; call the real-time inventory API of the TravelSky / Hotel PMS system to verify availability; exclude items exceeding the standard (e.g., room rate > rank limit), and sort them according to the weight A:B = 7:3.

[0032] S5: Push the final travel recommendation plan to the user side, receive the feedback information of the user, and optimize the collaborative filtering module and the LSTM time series analysis module based on the feedback information.

[0033] Send the final travel recommendation plan to the user side through message push, email, APP notification, etc. The user operates on the recommendation plan on the user side, such as clicking, booking, etc., and the system records the feedback information of the user. Optimize the collaborative filtering module and the LSTM time series analysis module according to the feedback information of the user.

[0034] In this step, record the user's click / book behavior on the recommended item (MySQL log table); update the similarity matrix of collaborative filtering and the LSTM model parameters through the daily offline task.

[0035] In some embodiments, the steps for the collaborative filtering module to generate the initial recommended plan include: S21: Calculate the user interest matching degree through cosine similarity or Pearson correlation coefficient, identify the neighbor group with similar preferences to the target user, and generate a group preference recommendation list based on the behavior data of the neighbor group. In this step, obtain the behavior data of all users from the user behavior database, including preference information for hotels, transportation means, etc. For the target user, calculate the interest matching degree between it and other users using cosine similarity or Pearson correlation coefficient. Sort according to the interest matching degree, and select the top N users with the highest interest matching degree with the target user as the neighbor group. Analyze the behavior data of the neighbor group, count their selection frequencies for different hotels, transportation means, etc., and generate a group preference recommendation list.

[0036] The embodiment of the present invention uses the Pearson correlation coefficient to calculate the matching degree, and the formula is as follows:

[0037] In the formula, respectively refer to the target user and the user to be compared, is the user 's rating for the item ; the user 's rating for the item The score of for users is the average score of all items. for users is the average score of all items. for users and the interest similarity between user ranges from [-1, 1]. 1 represents a perfect positive correlation, and -1 represents a perfect negative correlation.

[0038] Construct a user-item rating matrix R. An example of the matrix is shown in Table 1, where the rows represent users, the columns represent items, and the values are ratings. For each user, calculate the average of the rated items. Only calculate the covariance and standard deviation for the items with common ratings, and output the similarity.

[0039] Table 1: Example of User-Item Rating Matrix

[0040] For each user, calculate the average of all its rated items. Example: The average score of user_123 =(4 + 5) / 2 = 4.5; If the sim(i, j) between hotel_789 and flight_abc is 0.72, they are regarded as associated items.

[0041] S22: Analyze the associated behavior sequence of users' hotel reservations and transportation choices, introduce a time decay factor, and construct a user-item interaction matrix; Extract the associated behavior sequence of users' hotel reservations and transportation choices from the user behavior database. For example, a user first reserves a certain hotel and then selects a certain means of transportation. Introduce a time decay factor, assign a higher weight to the recent behavior data and a lower weight to the long-term behavior data. Based on the associated behavior sequence and the time decay factor, construct a user-item interaction matrix, and the elements in the matrix represent the preference degree of users for items.

[0042] S23: Use the modified cosine similarity algorithm to calculate the similarity between items and generate a recommended list of associated items; According to the user-item interaction matrix, use the modified cosine similarity algorithm to calculate the similarity between different items (such as different hotels and different means of transportation). For each item, select the top M items with the highest similarity to it and generate a recommended list of associated items.

[0043] The formula of the modified cosine similarity algorithm is as follows:

[0044] In the formula, Items for which similarity is to be calculated At the same time, for the items and the items The set of users who rate them For user The rating for the item is User The rating for the item is For user The average rating for all items For the item and the item The similarity, with a value range of [-1, 1], where 1 represents a perfect positive correlation and -1 represents a perfect negative correlation

[0045] S24: Weightedly fuse the recommended items in the group preference recommendation list and the associated item recommendation list to form an initial recommendation scheme. Specifically, it includes: S241: For each recommended item in the group preference recommendation list, use its corresponding user interest matching degree as the initial weight. For each recommended item in the associated item recommendation list, use its corresponding modified cosine similarity score as the initial weight; S242: Normalize the weights of the two types of recommendations to the same dimension; Calculate the total weights of all recommended items in the group preference recommendation list and the associated item recommendation list respectively. For each recommended item, divide its initial weight by the corresponding total weight to obtain the normalized weight.

[0046] S243: Calculate the ratings of the group preference recommendation and the associated item recommendation respectively; For each recommended item in the group preference recommendation list, calculate the rating according to its normalized weight and other relevant factors (such as the user's historical ratings for this type of item).

[0047] For each recommended item in the associated item recommendation list, calculate the rating according to its normalized weight and other relevant factors (such as the price and evaluation of the item).

[0048] S244: Sort by the weighted total score and retain the top N items as the initial recommendation scheme. Specifically, merge the recommended items in the group preference recommendation list and the associated item recommendation list, and calculate the weighted total score according to their ratings. Sort from high to low according to the weighted total score and retain the top N items as the initial recommendation scheme.

[0049] In some embodiments, the steps for the LSTM time series analysis module to generate a user's personalized travel plan include: S31: Retrieve the user's last M historical travel records, including structured data and unstructured note information. In this step, query the user's last M historical travel records from the historical travel record database according to the user identifier, including structured data such as travel date, destination, hotel selection, transportation selection, etc., and unstructured data such as note information left by the user during the travel.

[0050] S32: Parse the unstructured note information through natural language processing technology to generate user portrait tags. Use natural language processing technology (such as word segmentation, part-of-speech tagging, named entity recognition, etc.) to process the unstructured note information. Extract key information in the note information, such as the user's preferences, special needs, etc., to generate user portrait tags.

[0051] S33: Jointly input the user portrait tags and historical behavior sequences into the trained LSTM model, and capture temporal features through time decay weighted sum and attention mechanism. In this step, integrate the user portrait tags and historical behavior sequences to form input data. Introduce time decay weighting in the input data, assign higher weights to recent behavior data, and lower weights to long-term behavior data. Use the attention mechanism to make the LSTM model pay more attention to important temporal features.

[0052] S34: Generate probability distributions of destinations, hotels, and flights at the output layer of the LSTM model. After the LSTM model processes the input data, it generates probability distributions of destinations, hotels, and flights at the output layer.

[0053] S35: Generate a personalized travel plan through Monte Carlo simulation based on the probability distribution. Specifically, it includes: S351: Determine the candidate plan set according to the probability distribution output by the LSTM. According to the probability distributions of destinations, hotels, and flights output by the LSTM model, select several combinations of destinations, hotels, and flights with higher probabilities to form the candidate plan set.

[0054] S352: Superimpose external influence factors on each candidate plan. Consider the influence of external factors on the candidate plans, such as weather, holidays, etc., and superimpose the corresponding external influence factors on each candidate plan.

[0055] S353: Conduct N random samplings. Each time a sampling is performed, randomly select a candidate plan according to the weighted probability distribution, and calculate the comprehensive score of the probability weight + difference standard compliance + real-time inventory availability of the candidate plan. Perform N random samplings. For each sampling, randomly select a candidate solution from the set of candidate solutions according to the weighted probability distribution. Calculate the probability weight, difference standard compliance, and real-time inventory availability of the candidate solution, and add the three to obtain the comprehensive score.

[0056] S354: Statistically analyze all sampling results, and output the top K solutions with the highest comprehensive scores among the selected candidate solutions, that is, the personalized travel solutions.

[0057] It should be noted that the training of the LSTM model is as follows: Collect time-series data related to business travel, such as user browsing / reservation sequences, hotel / flight price fluctuations, user portraits (preferences, consumption capabilities), timestamps (seasons, holidays), etc.

[0058] According to numerical features (prices, ratings), standardization (Z-score or Min-Max) is required to eliminate the influence of dimensions. According to categorical features (user types, destinations), encoding (such as Embedding or One-Hot) is required and long sequences are sliced into input segments of a fixed length (such as 30-day behavior sequences). Divide the training set (70%), validation set (15%), and test set (15%) proportionally to ensure consistent distribution to locate model problems (overfitting or distribution shift).

[0059] The network structure is designed as follows: Input layer: LSTM(units=128, return_sequences=True) (processing sequence output).

[0060] Hidden layer: Stack LSTM layers or combine with Dense layers to extract high-order features, and the number of parameters needs to be controlled to avoid overfitting (such as reducing the number of layers when the amount of data is small). Output layer: Use linear activation for regression tasks (such as price prediction) and Softmax for classification tasks (such as user behavior prediction).

[0061] LSTM controls the information flow through the input gate, forget gate, and output gate, uses the Sigmoid function to generate gating coefficients, the Tanh function to process candidate memory units, optimizes long-term dependence learning, and uses Xavier or He initialization to avoid gradient explosion / vanishing.

[0062]

[0063] In the formula, is the weight matrix, is the bias term, is the Sigmoid function, and the output range is [0,1]. 0 means complete forgetting, and 1 means complete retention.

[0064] For regression tasks, mean squared error (MSE) or mean absolute error (MAE) is used. For classification tasks, cross-entropy loss is used. The optimizer uses Adam (adaptive learning rate), with the initial learning rate set to 0.001, combined with a learning rate decay strategy such as cosine annealing. The regularization technique uses Dropout (e.g., adding Dropout(0.3) after the LSTM layer) to prevent overfitting. L2 regularization constrains the weights, and gradient clipping (limiting the gradient norm to 5 - 15) avoids gradient explosion.

[0065] Observe the Loss curve of the training / validation set: If the validation Loss increases while the training Loss decreases, it indicates overfitting, and regularization needs to be enhanced or the data needs to be augmented. Use early stopping, and terminate the training when the validation Loss does not decrease for multiple consecutive rounds.

[0066] In some embodiments, the steps of inputting the initial recommendation plan and the personalized travel plan into the dynamic filtering module, and performing secondary screening in combination with the real-time inventory of the travel supplier, the budget constraint based on the travel standard, and the user travel portrait tags to output the final travel recommendation plan include: S41: Take the union of the initial recommendation plan generated by the collaborative filtering module and the personalized travel plan generated by the LSTM time series analysis module; Merge all the recommended items in the initial recommendation plan and the personalized travel plan, remove the duplicates, and obtain the union result.

[0067] S42: Obtain the real-time inventory through the supplier API interface, and retrieve the budget constraint from the enterprise travel standard database; Establish an API interface connection with the travel supplier, and obtain the real-time inventory information through interface calls. Query the budget constraint information based on the travel standard from the enterprise travel standard database.

[0068] S43: Perform compliance filtering based on the real-time inventory, budget constraint, and user portrait tags, and exclude the items that exceed the standard; In the travel dynamic planning system, the real-time inventory refers to the real-time available reservation resource data obtained by docking with the systems of business travel service suppliers (such as airlines, hotels, car rental companies), including flights: remaining seat numbers, cabin classes (economy class / business class), price fluctuations. Hotels: available room types, negotiated prices, occupancy rates. Other services: pick-up and drop-off vehicles, meeting venues, etc.

[0069] For each recommended item in the union result, check whether it meets the requirements of the real-time inventory, budget constraint, and user portrait tags. Exclude the recommended items that do not meet the requirements to obtain the selected options that pass the screening.

[0070] S44: Generate a final business travel recommendation plan for the options that pass the screening according to preset weights; among them, the weight A of the personalized plan is greater than the initial recommendation weight B.

[0071] Assign weights A and B (A > B) to the personalized plan and the initial recommendation plan among the options that pass the screening respectively. Calculate the scores of each option according to the weights, and sort them from high to low to generate a final business travel recommendation plan.

[0072] In some embodiments, the steps of optimizing the collaborative filtering module and the LSTM time series analysis module based on the feedback information include: S51: Adjust the weights of the recommended items in the group preference recommendation list and the weights of the recommended items in the associated item recommendation list in the collaborative filtering module according to the user's click and booking behaviors; Record the user's click and booking behaviors on the recommendation plan, and analyze the user's preference changes. According to the user's preference changes, adjust the weights of the recommended items in the group preference recommendation list and the weights of the recommended items in the associated item recommendation list. For example, if the user clicks or books a certain recommended item, increase the weight of this recommended item. User clicks on flight A → Update the rating vector of this user in User-CF; Re-calculate the item similarity matrix by SVD decomposition (offline task per week).

[0073] S52: Update the user portrait tags according to the user's click and booking behaviors, and then update the attention mechanism parameters of the LSTM model in the LSTM time series analysis module.

[0074] According to the user's click and booking behaviors, update the user portrait tags to reflect the user's latest preferences and needs. Use the updated user portrait tags to re-train the attention mechanism parameters of the LSTM model to make the model pay more attention to the user's current needs.

[0075] For example, the user books hotel B → Add this behavior to the training set, trigger incremental training; Adjust the attention mechanism parameters: Increase the time decay weight of recent behaviors.

[0076] The collaborative operation of the time series analysis module and the collaborative filtering module constructs a double guarantee system for intelligent business travel recommendations, and the intervention of the dynamic filtering module ensures that the recommendation results accurately meet the enterprise control requirements. The time series analysis module generates personalized travel plans based on the user's historical travel data (such as destination preferences, time period selection rules, etc.) using time series prediction algorithms; the collaborative filtering module mines high-quality choices in the group wisdom through cross-user behavior similarity analysis, such as identifying colleague groups highly consistent with the current user's business travel mode and extracting the transportation tools and hotels they frequently choose.

[0077] After the dual-module outputs the recommendation results, the dynamic filtering module immediately starts the differential standard rule verification process. The system automatically retrieves the user's business travel standard database, which includes hard constraints such as budget limits, cabin classes, hotel star ratings, as well as differential rules based on attributes such as position and department. For example, it opens up the flexibility of "flight upgrades to first class" for executive users, while strictly limiting the option of "only economy class" for interns.

[0078] During the filtering process, the system adopts a hierarchical decision-making mechanism: first, it excludes obvious over-standard items; second, it conducts a cost-benefit analysis on the critical value solutions; and finally, it generates a compliance score by combining the user's historical behavior data (such as past differential standard implementation). For recommendations that meet the differential standards, the system directly pushes them to the user side; if there is a risk of over-standard, it automatically triggers differential standard optimization suggestions, such as recommending flights with price lows during adjacent periods, or providing preferential packages for corporate agreement hotels, maximizing the user experience while ensuring compliance. This dynamic balance mechanism of "intelligent recommendation + rule constraint" realizes the refinement and humanization of business travel management.

[0079] For example, (1) the initial recommendation plan generated by the collaborative filtering module Data source: Based on group behavior (user-user collaborative filtering) and item association rules (item-item collaborative filtering).

[0080] Reflects common preferences (such as "economy class flights often chosen by users in the same industry"); Solves the cold start problem, but lacks in-depth mining of the user's individual sequential behavior.

[0081] Example: Recommend the "early morning flight from Beijing to Shanghai" (because 80% of similar users choose this flight).

[0082] (2) The personalized travel plan generated by the LSTM sequential analysis module Data source: Based on the user's own historical behavior sequence (such as the last 12 travel records).

[0083] Captures individual dynamics (such as "this user has recently changed flights frequently and prefers flights with high flexibility"); Combines external events (such as weather, exhibitions) to predict future demand.

[0084] Example: Recommend the "noon flight from Beijing to Shanghai" (because the user has chosen noon flights for the past 3 times).

[0085] Since the initial recommendation may ignore recent changes in the user's behavior (such as the tendency to switch from "economy class" to "business class"); it cannot respond to dynamic factors in real time (such as sudden flight price increases). The personalized plan depends on the user's historical data, and it is inaccurate for new users or low-frequency users; it may deviate from the corporate business travel policy (such as recommending five-star hotels that exceed the differential standards).

[0086] The core function of the dynamic filtering module is to exclude options that exceed the budget or have insufficient inventory in the initial recommendations or personalized solutions (such as sold-out first-class cabins). Sort the cost performance of critical value solutions (such as prices close to the upper limit of the difference standard). Filter high-risk recommendations based on user profile tags (such as "frequently cancels hotels").

[0087] Example: Initial recommendation: Flight A (economy class, group preference, price compliant); Personalized solution: Flight B (business class, individual preference, price exceeds the difference standard by 10%); Dynamic filtering result: Keep Flight A, but prompt "Flexibility in difference standard can be applied" (because the user has a tendency to upgrade recently).

[0088] As Figure 2 shown, an embodiment of the present invention also provides a business travel recommendation system based on collaborative filtering and recurrent neural network, including a business travel application form receiving module, a collaborative filtering module, an LSTM time series analysis module, a dynamic filtering module, and a recommendation optimization module; The business travel application form receiving module is used to obtain the approved business travel application form and extract business travel information, including the destination, date, rank, and business travel standard; The collaborative filtering module is used to generate an initial recommendation solution based on the business travel information through user-user collaborative filtering and item-item collaborative filtering algorithms; The LSTM time series analysis module is used to retrieve the user's recent M historical travel records according to the business travel information, construct a user travel profile, train a time series model of user behavior through a time series neural network architecture, and generate a personalized travel solution through Monte Carlo simulation by the trained time series model; The dynamic filtering module is used to perform secondary screening based on the initial recommendation solution and the personalized travel solution, combined with the real-time inventory of business travel suppliers, the budget constraint based on the business travel standard, and the user travel profile tags, and output the final business travel recommendation solution; The recommendation optimization module is used to push the final business travel recommendation solution to the user side, receive the feedback information of the user, and optimize the collaborative filtering module and the LSTM time series analysis module based on the feedback information.

[0089] In some embodiments, the steps for the collaborative filtering module to generate an initial recommendation plan include: calculating the user interest matching degree through cosine similarity or Pearson correlation coefficient, identifying a neighbor group with similar preferences to the target user, and generating a group preference recommendation list based on the behavior data of the neighbor group; analyzing the associated behavior sequence of the user's hotel reservation and transportation selection, introducing a time decay factor, and constructing a user-item interaction matrix; calculating the similarity between items using a modified cosine similarity algorithm to generate an associated item recommendation list; and performing weighted fusion on the recommended items of the group preference recommendation list and the associated item recommendation list to form an initial recommendation plan.

[0090] In some embodiments, the specific steps for the collaborative filtering module to form an initial recommendation plan include: using the user interest matching degree as the initial weight of the recommended items in the group preference recommendation list, and using the modified cosine similarity score as the initial weight of the recommended items in the associated item recommendation list; normalizing the weights of the two types of recommendations to the same dimension; calculating the scores of the group preference recommendation and the associated item recommendation respectively; and sorting by the weighted total score and retaining the top N items as the initial recommendation plan.

[0091] In some embodiments, the steps for the LSTM time series analysis module to generate a personalized travel plan for the user include: retrieving the user's last M historical travel records, including structured data and unstructured note information; parsing the unstructured note information through natural language processing technology to generate user portrait tags; jointly inputting the user portrait tags and the historical behavior sequence into a trained LSTM model, and capturing time series features through time decay weighted sum and attention mechanism; generating the probability distributions of destinations, hotels, and flights at the output layer of the LSTM model; and generating a personalized travel plan based on the probability distributions through Monte Carlo simulation.

[0092] In some embodiments, the steps for the LSTM time series analysis module to generate a personalized travel plan based on the probability distributions through Monte Carlo simulation include: determining a candidate plan set according to the probability distributions output by the LSTM; superimposing external influence factors on each candidate plan; performing N random samplings, and randomly selecting a candidate plan according to the weighted probability distribution each time, and calculating the comprehensive score of the probability weight + difference standard compliance + real-time inventory availability of the candidate plan; statistically analyzing all sampling results, and outputting the top K plans with the highest comprehensive scores among the selected candidate plans, which is the personalized travel plan.

[0093] In some embodiments, the dynamic filtering module is specifically used to take the union of the initial recommendation plan generated by the collaborative filtering module and the personalized travel plan generated by the LSTM time series analysis module; obtain real-time inventory through the supplier API interface, and retrieve budget constraints from the corporate travel standard database; perform compliance filtering based on real-time inventory, budget constraints, and user portrait tags to exclude items that exceed the standard; Generate a final business trip recommendation plan for the selected options according to the preset weights; among them, the weight A of the personalized plan is greater than the initial recommendation weight B.

[0094] In some embodiments, the recommendation optimization module is specifically configured to adjust the weights of the recommended items in the group preference recommendation list and the weights of the recommended items in the associated item recommendation list in the collaborative filtering module according to the user's click and booking behaviors; update the user portrait tags according to the user's click and booking behaviors, and further update the attention mechanism parameters of the LSTM model in the LSTM time series analysis module.

[0095] An embodiment of the present invention further provides an electronic device, which includes: a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus. The communication bus can be used for information transmission between the electronic device and the sensor. The processor can call the logical instructions in the memory to execute the following methods: S1: Obtain the approved business trip application form and extract the business trip information, including the destination, date, rank, and business trip standard; S2: Input the business trip information into the collaborative filtering module, and generate an initial recommendation plan through the user-user collaborative filtering and item-item collaborative filtering algorithms; S3: Input the business trip information into the LSTM time series analysis module, retrieve the user's recent M historical travel records, construct a user travel portrait, train the time series model of the user's behavior through the time series neural network architecture, and generate a personalized travel plan through Monte Carlo simulation after the trained time series model; S4: Input the initial recommendation plan and the personalized travel plan into the dynamic filtering module, perform secondary screening in combination with the real-time inventory of the business trip supplier, the budget constraint based on the business trip standard, and the user travel portrait tags, and output the final business trip recommendation plan; S5: Push the final business trip recommendation plan to the user terminal, receive the feedback information of the user, and optimize the collaborative filtering module and the LSTM time series analysis module based on the feedback information.

[0096] In addition, when the logical instructions in the above-mentioned memory are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: USB flash drive, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other various media that can store program codes.

[0097] An embodiment of the present invention provides a non-transitory computer-readable storage medium that stores computer instructions, and the computer instructions cause a computer to execute the method provided by the above method embodiment, for example, including: S1: Obtain a travel application form with approval completed, and extract travel information, including destination, date, rank, and travel standard; S2: Input the travel information into a collaborative filtering module, and generate an initial recommendation plan through user-user collaborative filtering and item-item collaborative filtering algorithms; S3: Input the travel information into an LSTM time series analysis module, retrieve the user's recent M historical travel records, construct a user travel profile, train a time series model of user behavior through a time series neural network architecture, and the trained time series model generates a personalized travel plan through Monte Carlo simulation; S4: Input the initial recommendation plan and the personalized travel plan into a dynamic filtering module, and perform secondary screening in combination with the real-time inventory of travel suppliers, budget constraints based on travel standards, and user travel profile tags, and output a final travel recommendation plan; S5: Push the final travel recommendation plan to the user side, receive the feedback information of the user, and optimize the collaborative filtering module and the LSTM time series analysis module based on the feedback information.

[0098] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A travel recommendation method based on collaborative filtering and recurrent neural network, characterized in that: The steps include: Obtain approved travel application forms and extract travel information, including destination, date, rank and travel standards; Input the travel information into a collaborative filtering module, and generate an initial recommendation plan through user-user collaborative filtering and item-item collaborative filtering algorithms; The travel information is input into the LSTM time series analysis module, the user's recent M historical travel records are retrieved, the user's travel profile is constructed, and the time series model of the user's behavior is trained through the time series neural network architecture. The trained time series model generates a personalized travel plan through Monte Carlo simulation; The initial recommendation plan and the personalized travel plan are input into a dynamic filtering module, and secondary screening is performed in combination with the real-time inventory of travel suppliers, the budget constraint based on travel standards, and the user travel profile label to output the final travel recommendation plan; The final travel recommendation plan is pushed to the user end, the user's feedback information is received, and the collaborative filtering module and the LSTM time series analysis module are optimized based on the feedback information.

2. The travel recommendation method based on collaborative filtering and recurrent neural network according to claim 1 is characterized in that: The steps for the collaborative filtering module to generate an initial recommendation solution include: Calculate user interest matching through cosine similarity or Pearson correlation coefficient, identify neighbor groups with similar preferences to the target user, and generate a group preference recommendation list based on the behavior data of the neighbor groups; Analyze the user's associated behavior sequence for hotel reservations and transportation selection, introduce the time decay factor, and construct the user-item interaction matrix; According to the user-item interaction matrix, the modified cosine similarity algorithm is used to calculate the similarity between items and generate a list of related item recommendations; The recommended items in the group preference recommendation list and the related item recommendation list are weighted and fused to form an initial recommendation plan.

3. The travel recommendation method based on collaborative filtering and recurrent neural network according to claim 2 is characterized in that: The steps of weighted fusion of the recommended items in the group preference recommendation list and the associated item recommendation list to form an initial recommendation plan include: The user interest matching degree is used as the initial weight of the recommended items in the group preference recommendation list, and the modified cosine similarity score is used as the initial weight of the recommended items in the associated item recommendation list; Normalize the weights of the two types of recommendations to the same dimension; Calculate the scores of group preference recommendations and related item recommendations respectively; Sort by weighted total score and retain the top N items as the initial recommended solutions.

4. The travel recommendation method based on collaborative filtering and recurrent neural network according to claim 3 is characterized in that: The steps for the LSTM time series analysis module to generate a personalized travel plan for users include: Retrieve the user's recent M historical travel records to generate a historical behavior sequence. The historical travel records include structured data and unstructured notes information. Analyze unstructured notes through natural language processing technology and generate user portrait tags; The user portrait label and historical behavior sequence are jointly input into the trained LSTM model to capture the temporal features through time decay weighting and attention mechanism; Generate the probability distribution of destinations, hotels, and flights at the output layer of the LSTM model; Generate personalized travel plans through Monte Carlo simulation based on probability distribution.

5. The travel recommendation method based on collaborative filtering and recurrent neural network according to claim 4 is characterized in that: The steps to generate personalized travel plans based on probability distribution through Monte Carlo simulation include: Determine the set of candidate solutions based on the probability distribution of LSTM output; Superimpose external impact factors on each candidate solution; Perform N random samplings, randomly select a candidate solution according to the weighted probability distribution each time, and calculate the comprehensive score of the candidate solution's probability weight + difference standard compliance + real-time inventory availability; All sampling results are counted, and the top K comprehensive scores among the selected candidate plans are output, that is, personalized travel plans.

6. The travel recommendation method based on collaborative filtering and recurrent neural network according to claim 5 is characterized in that: The steps of inputting the initial recommendation plan and the personalized travel plan into the dynamic filtering module, performing secondary screening in combination with the real-time inventory of travel suppliers, the budget constraint based on travel standards and the user travel profile label, and outputting the final travel recommendation plan include: Take the union of the initial recommendation plan generated by the collaborative filtering module and the personalized travel plan generated by the LSTM time series analysis module; Obtain real-time inventory through supplier API interfaces and retrieve budget constraints from corporate travel standard databases; Compliance filtering is performed based on real-time inventory, budget constraints, and user profile tags to exclude items that exceed the standard; The final travel recommendation plan is generated for the screened options according to the preset weights; the personalized plan weight A is greater than the initial recommendation weight B.

7. The travel recommendation method based on collaborative filtering and recurrent neural network according to claim 6 is characterized in that: The steps of optimizing the collaborative filtering module and the LSTM timing analysis module based on the feedback information include: Adjust the weights of the recommended items in the group preference recommendation list and the weights of the recommended items in the associated item recommendation list in the collaborative filtering module according to the user's click and booking behaviors; The user portrait label is updated according to the user's click and booking behavior, and then the attention mechanism parameters of the LSTM model in the LSTM timing analysis module are updated.

8. A travel recommendation system based on collaborative filtering and recurrent neural network, characterized in that: It includes travel application form receiving module, collaborative filtering module, LSTM time series analysis module, dynamic filtering module and recommendation optimization module; The travel application form receiving module is used to obtain the approved travel application forms and extract travel information, including destination, date, rank and travel standards; A collaborative filtering module, used to generate an initial recommendation scheme based on the travel information through user-user collaborative filtering and item-item collaborative filtering algorithms; LSTM time series analysis module, which is used to retrieve the user's recent M travel records based on travel information, build a user travel profile, train the time series model of user behavior through the time series neural network architecture, and generate personalized travel plans through Monte Carlo simulation of the trained time series model; The dynamic filtering module is used to perform secondary screening based on the initial recommendation plan and personalized travel plan, combined with the real-time inventory of travel suppliers, budget constraints based on travel standards, and user travel profile tags, and output the final travel recommendation plan; The recommendation optimization module is used to push the final travel recommendation plan to the user end, receive user feedback information, and optimize the collaborative filtering module and the LSTM time series analysis module based on the feedback information.

9. The travel recommendation system based on collaborative filtering and recurrent neural network according to claim 8, characterized in that: The collaborative filtering module calculates the user interest matching degree through cosine similarity or Pearson correlation coefficient, identifies neighbor groups with similar preferences to the target user, and generates a group preference recommendation list based on the behavior data of the neighbor groups; Analyze the user's related behavior sequence for hotel reservations and transportation selection, introduce the time decay factor, and construct the user-item interaction matrix; use the modified cosine similarity algorithm to calculate the similarity between items and generate a recommended list of related items; The recommended items in the group preference recommendation list and the related item recommendation list are weighted and fused to form an initial recommendation plan.

10. The travel recommendation system based on collaborative filtering and recurrent neural network according to claim 9, characterized in that: The LSTM time series analysis module retrieves the user's recent M historical travel records, including structured data and unstructured notes; parses the unstructured notes through natural language processing technology to generate user portrait labels; inputs the user portrait labels and historical behavior sequences into the trained LSTM model, and captures the time series features through time decay weighting and attention mechanism; Generate the probability distribution of destinations, hotels, and flights at the output layer of the LSTM model; Generate personalized travel plans through Monte Carlo simulation based on probability distribution.

Images