Car rental recommendation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of car rental recommendation, in particular to a car rental recommendation method and system based on artificial intelligence, and aims to overcome the limitation of traditional rule recommendation on context understanding by fusing user historical car use records and current input requirements, constructing a multi-dimensional state vector and accurately representing user behavior habits and travel scenes. A reinforcement learning model is adopted to model a car renting behavior strategy based on a Markov decision process, a reward value is calculated in combination with real feedback such as clicking, renting conversion and scoring, a recommendation strategy is dynamically optimized through a strategy gradient algorithm, output recommendation actions cover vehicle type categories, price ranges and place screening, personalized matching of cars is achieved, and the car renting efficiency is improved. Recommendation generalization is avoided, and recommendation accuracy and user satisfaction are improved. The model has continuous learning ability, adapts to changes of different car renting scenes, and improves the intelligent level and practical value of the recommendation system.

Description

Technical Field

[0001] This invention relates to the field of car rental recommendation technology, and more particularly to a car rental recommendation method and system based on artificial intelligence. Background Technology

[0002] With the widespread adoption of ride-sharing and online car rental platforms, users face complex decision-making factors when choosing a rental vehicle, including a wide variety of models, significant price fluctuations, and diverse usage scenarios. Traditional car rental recommendation systems largely rely on static rules. However, such methods have the following problems:

[0003] Static rule systems typically filter and sort vehicles based on fixed field conditions, such as budget range, vehicle size, and pick-up location. However, these systems lack a deep understanding of the user's contextual semantics and cannot dynamically adapt to the detailed needs of users in different travel scenarios (such as commuting, road trips, and business receptions). For example, when faced with users with similar budgets but different needs for range or passenger space, static rules cannot accurately differentiate based on scenario changes, leading to severely generalized recommendation results and impacting user satisfaction. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a car rental recommendation method and system based on artificial intelligence.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An AI-based car rental recommendation method includes the following steps:

[0007] S1. Obtain the user's car usage records and user demand information input by the user, and fuse them to construct a state vector to represent the car rental recommendation environment;

[0008] S2. Input the state vector into the reinforcement learning model. The reinforcement learning model models car rental recommendation behavior based on Markov decision process and outputs the corresponding recommendation action. The recommendation action includes the type of car model to be recommended, the rental price range, and the selection of rental location.

[0009] S3. Generate a recommendation list by filtering the set of vehicles that meet the recommended action from the car rental service database;

[0010] S4. Obtain user feedback information on the recommendation list, including click behavior, whether the rental was completed, and rental experience rating data; calculate the reward value based on the feedback information, and update the reinforcement learning model using a policy gradient algorithm.

[0011] Furthermore, the user demand information includes contextual information such as the current purpose of travel, budget range, origin and destination locations, and number of passengers.

[0012] Further, S1 includes the following steps:

[0013] S11. Based on the user's historical vehicle usage records, extract long-term preference features including vehicle type preference, rental period preference, price acceptance range, and service rating preference, and construct a user preference vector;

[0014] S12. Perform structured encoding on the user's input of current travel purpose, budget range, origin and destination locations, and number of passengers to construct a context feature vector;

[0015] S13. The user preference vector and the context feature vector are concatenated and fused to construct a state vector that represents the current car rental recommendation environment.

[0016] Further, S11 includes the following steps:

[0017] The frequency of use and rating data of each vehicle type in the user's historical car rental records are statistically analyzed to generate a preference rating for each vehicle type.

[0018] Extract rental period length and payment amount from users' historical orders to construct corresponding rental period preference parameters and price acceptance range parameters;

[0019] The vehicle type preference rating vector, rental period preference parameter, and price acceptance range parameter are normalized and encoded, and then concatenated to generate a user preference vector that represents the user's long-term preferences.

[0020] Furthermore, the vehicle category includes: vehicle size, energy consumption type, vehicle brand, spare space, passenger capacity, and maximum range.

[0021] Furthermore, the reinforcement learning model is trained through the following steps:

[0022] A training dataset containing historical state vectors, historical recommended actions, and historical behavior feedback information was constructed based on historical user behavior data.

[0023] Based on the historical behavior feedback information, a multi-objective reward function reflecting immediate response and long-term satisfaction is constructed, and the corresponding reward value is calculated according to the multi-objective reward function.

[0024] The historical state vector, historical recommended actions, and corresponding reward values ​​are input into the policy network, and forward calculation and backpropagation updates are performed using a policy gradient algorithm that includes entropy regularization.

[0025] Furthermore, the multi-objective reward function is as follows:

[0026] R(s,a)=α·r1+β·r2+γ·r3·π(a|s);

[0027] Where R(s, a) is the reward value for performing the recommended action a under state vector s; r1 is the feedback score corresponding to the user's click behavior; r2 is the score for the user to complete the rental behavior; r3 is the user rating feedback value; π(a|s) is the output probability of the recommended action a under state vector s; α, β and γ are weight coefficients.

[0028] Furthermore, the feedback score r1 corresponding to the user's click behavior is calculated by normalizing the ratio of clicked recommended items to unclicked items in the recommendation list; the score r2 for the user to complete the rental behavior is a Boolean value indicating whether the recommended vehicle has been actually rented; the user rating feedback value r3 is calculated by normalizing the comprehensive rating of the vehicle service collected by the user after the rental is completed.

[0029] Further, S3 includes the following steps:

[0030] Based on the vehicle type, rental price range, and rental location filtering criteria included in the recommendation action, a set of candidate vehicles that meet the filtering criteria is retrieved from the car rental service database;

[0031] Obtain the popularity and historical user ratings for each vehicle in the candidate vehicle set, and construct a rating vector;

[0032] The vehicles in the candidate vehicle set are sorted according to the rating vector, and a recommendation list is generated based on the recommendation order priority.

[0033] An AI-based car rental recommendation system, applied to any of the aforementioned AI-based car rental recommendation methods, includes:

[0034] The data acquisition module is used to acquire users' car usage records and user demand information input by users, and integrate them to construct a state vector to represent the car rental recommendation environment;

[0035] The recommendation learning module is used to input the state vector into the reinforcement learning model and output a recommendation action, which includes the type of car model to be recommended, the rental price range, and the car rental location selection information; it is also used to calculate the reward value based on the behavioral feedback information and update the reinforcement learning model through the policy gradient algorithm;

[0036] The filtering and recommendation module is used to generate a recommendation list by filtering a set of vehicles that meet the recommendation action from the car rental service database.

[0037] The feedback information acquisition module is used to collect user feedback information on the recommended list. The feedback information includes click behavior, whether the rental is completed, and rental experience rating data.

[0038] The beneficial effects of this invention are as follows: By integrating users' historical car rental records with current input needs, this invention constructs a multi-dimensional state vector, which can more comprehensively capture users' behavioral habits and current travel scenarios, making up for the shortcomings of traditional static field matching methods in contextual understanding, thereby improving the relevance and personalization of recommendations. A car rental behavior strategy is established based on a Markov decision process using a reinforcement learning model, and reward values ​​are calculated based on real user feedback (such as clicks, rental conversions, and ratings). A policy gradient algorithm is used to continuously iterate and update the model, enabling the system to automatically adjust the recommendation strategy according to user behavior, improving its responsiveness to changes in user preferences. In addition to outputting vehicle type, it also comprehensively considers price range and location filtering to generate multi-dimensional recommendation actions, and selects the most matching vehicle set from the database, thereby avoiding generalization of recommendation results and ensuring that the recommended vehicles better match users' personalized needs, improving overall satisfaction and conversion rate. The model is trained and updated based on user behavior feedback, possessing continuous learning capabilities, and can continuously optimize the recommendation strategy under complex and ever-changing car rental scenarios, effectively solving the problem of poor adaptability of traditional static recommendation models to new scenarios. Attached Figure Description

[0039] Figure 1 This is a flowchart of the steps of an artificial intelligence-based car rental recommendation method in this invention.

[0040] Figure 2 This is a flowchart of step S1 in this invention. Detailed Implementation

[0041] Please see Figures 1-2 As shown, this invention relates to an artificial intelligence-based car rental recommendation method, which includes the following steps:

[0042] S1. Obtain the user's car usage records and user demand information input by the user, and fuse them to construct a state vector to represent the car rental recommendation environment;

[0043] S2. Input the state vector into the reinforcement learning model. The reinforcement learning model models car rental recommendation behavior based on Markov decision process and outputs the corresponding recommendation action. The recommendation action includes the type of car model to be recommended, the rental price range, and the selection of rental location.

[0044] S3. Generate a recommendation list by filtering the set of vehicles that meet the recommended action from the car rental service database;

[0045] S4. Obtain user feedback information on the recommendation list, including click behavior, whether the rental was completed, and rental experience rating data; calculate the reward value based on the feedback information, and update the reinforcement learning model using a policy gradient algorithm.

[0046] In some embodiments, the system first acquires the user's historical car rental records and current travel needs. The user's historical records include past rental vehicle types, rental durations, rental expenses, pick-up and drop-off locations, and user ratings. Long-term user preference features are extracted using data mining methods, such as frequent SUV rentals and a preference for mid-to-low-priced vehicles. Simultaneously, contextual information such as the user's current travel purpose (e.g., business trip, road trip), budget range, origin and destination, and number of passengers is converted into a vector representation using a structured encoding strategy. The system integrates these long-term preferences with contextual needs to construct a state vector, accurately depicting the current car rental recommendation environment. Subsequently, the state vector is input into a reinforcement learning model. This model models car rental recommendation strategies based on Markov Decision Processes (MDPs). The state space consists of the aforementioned state vectors, and the action space is defined by the recommended vehicle type, price range, and location selection criteria. The policy network outputs the optimal recommendation action based on the current state vector, enabling the recommendation results to dynamically adapt to the user's potential preferences in different scenarios. For example, for two users with the same budget, the system can recommend either a gasoline-powered SUV or a new energy MPV based on their travel history and purpose, thereby improving the personalization accuracy of the recommendations. After the recommendation action is generated, the system performs a multi-dimensional query in the car rental service database based on the vehicle type, price range, and location filtering conditions specified in the recommendation action, filtering out a set of candidate vehicles that meet the constraints. Further, it combines the vehicle's historical rating, usage frequency, and user preference matching to construct a ranking score, prioritizes the vehicle set, generates the final recommendation list, and pushes it to the user. During the user's interaction with the recommendation list, the system records behavioral feedback information in real time, including whether the recommended item was clicked, whether the rental was completed, and post-rental rating data. This behavioral feedback is quantified and used to construct a multi-objective reward function, reflecting the user's immediate interest, final conversion, and overall satisfaction. The reinforcement learning model's parameters are updated using a policy gradient algorithm, enabling the model to continuously optimize the recommendation strategy based on the user's actual behavior. Compared to existing car rental recommendation methods that rely on static rule matching, this invention models long-term preferences and contextual needs through state vectors, achieving semantic perception and dynamic response to complex scenarios. A reinforcement learning structure based on Markov decision processes enables the recommendation strategy to have adaptive evolution capabilities. The introduction of a multi-objective reward function balances click-through rate, conversion rate, and service satisfaction, achieving multi-dimensional collaborative optimization of recommendation performance. This method effectively solves the problems of severe recommendation generalization, insufficient scenario adaptability, and rigid strategies in existing technologies, possessing stronger intelligence and practicality.

[0047] Furthermore, the user demand information includes contextual information such as the current purpose of travel, budget range, origin and destination locations, and number of passengers.

[0048] It should be noted that the travel purpose field is identified as categories such as "commuting," "tourism," "business reception," and "airport transfer" through user selection or natural language input parsing. Different travel purposes correspond to different car rental focuses. The budget range is calculated from user input or historical consumption records and is used to constrain the price filtering strategy in the recommendation process to ensure that the recommendation results meet the user's economic expectations. The origin and destination locations are obtained through location services or user input, and the route planning algorithm calculates the travel distance and driving area type (such as urban, highway, suburban) to provide decision support for subsequent vehicle selection (such as whether to recommend new energy vehicles or whether a long-range vehicle is needed). The number of passengers directly affects the number of seats and space requirements of the required vehicle. The system uses this parameter to limit the minimum seating capacity of recommended vehicles and outputs a suitable vehicle size range based on user preferences.

[0049] Further, S1 includes the following steps:

[0050] S11. Based on the user's historical vehicle usage records, extract long-term preference features including vehicle type preference, rental period preference, price acceptance range, and service rating preference, and construct a user preference vector;

[0051] S12. Perform structured encoding on the user's input of current travel purpose, budget range, origin and destination locations, and number of passengers to construct a context feature vector;

[0052] S13. The user preference vector and the context feature vector are concatenated and fused to construct a state vector that represents the current car rental recommendation environment.

[0053] In some embodiments, in S11, a long-term preference vector is first constructed based on the user's historical car usage records. Specifically, by statistically analyzing the frequency of occurrence and average rating of each vehicle type (such as SUV, sedan, new energy, seven-seater commercial vehicle, etc.) in the user's past orders, a normalized vehicle preference distribution is formed. Simultaneously, the historical rental length distribution and average rental price range are extracted, and rental period preference parameters and price acceptance ranges are constructed through distribution fitting and mean-variance encoding. Furthermore, service ratings from the user's historical orders are extracted to construct a service preference factor, reflecting the user's sensitivity to service quality. These features are then concatenated after vector standardization to generate the user's long-term preference vector, maintaining dimensionality consistency for easy integration into subsequent models. In S12, the system performs contextual modeling based on the user's current travel needs. The travel purpose is transformed into a low-dimensional dense vector through category embedding, while the budget range and number of passengers are directly encoded as normalized numerical fields. The origin and destination locations are first converted into spatial vectors through a geolocation encoder, and then expanded by combining path length (calculated based on map API) and route type features (such as whether it passes through highways or is an urban route). The final result is a structured context feature vector with contextual semantics. In S13, the user's long-term preference vector is concatenated and fused with the context feature vector to obtain a complete state vector. To enhance the model's ability to discriminate different feature semantics, the concatenated vector is input into a trainable state encoding network (such as a feedforward perceptron) to reduce the dimensionality and perform distribution mapping on the original concatenated representation, outputting the final state vector representing the current recommendation environment. This state vector is then fed into the policy network of the reinforcement learning model for action prediction.

[0054] Further, S11 includes the following steps:

[0055] The frequency of use and rating data of each vehicle type in the user's historical car rental records are statistically analyzed to generate a preference rating for each vehicle type.

[0056] Extract rental period length and payment amount from users' historical orders to construct corresponding rental period preference parameters and price acceptance range parameters;

[0057] The vehicle type preference rating vector, rental period preference parameter, and price acceptance range parameter are normalized and encoded, and then concatenated to generate a user preference vector that represents the user's long-term preferences.

[0058] In some embodiments, the system first extracts vehicle usage frequency and rating information from the user's historical car rental records. Using vehicle type (e.g., sedan, SUV, MPV, new energy vehicle, etc.) as the basic classification dimension, the system counts the order frequency of each type of vehicle and collects the average user rating for each order. To enhance the comparability of features, the system performs Min-Max normalization on both the frequency vector and the rating vector, and then fuses the two types of features using a weighted average strategy to form a unified vehicle type preference rating vector. This vector reflects the user's overall preference for different vehicle types and can capture behavioral patterns such as "users prefer highly rated new energy vehicles" or "frequently rent large-space MPVs." The system also extracts the rental period length (in days) and corresponding payment amount for each record in the user's historical orders. Based on this, two types of statistical features are constructed: the first is the rental period preference parameter, including average rental period, standard deviation, skewness coefficient, etc., used to characterize users' habits of short-term or long-term rentals; the second is the price acceptance range parameter, which constructs a robust price range by calculating the upper and lower quartiles (Q1, Q3) of historical rental prices, and combines the mean and maximum values ​​to form a multi-scale price perception feature. All continuous parameters are uniformly normalized using Z-score to eliminate the influence of dimensions. Finally, the normalized vehicle type preference rating vector, rental period preference parameter, and price acceptance range parameter are concatenated to form a complete user preference vector. To further enhance the nonlinearity of feature representation, this concatenated vector is input into a feedforward network with an activation function (such as a fully connected layer with ReLU activation) to generate an embedded long-term preference expression, which is used for subsequent fusion with contextual features to generate a state vector.

[0059] Furthermore, the vehicle category includes: vehicle size, energy consumption type, vehicle brand, spare space, passenger capacity, and maximum range.

[0060] Specifically, this includes: vehicle size (e.g., compact, mid-size, large), energy consumption type (e.g., gasoline, plug-in hybrid, pure electric), vehicle brand (e.g., mainstream brands, luxury brands, etc.), spare space (storage capacity in liters), passenger capacity (e.g., 5-seater, 7-seater), and maximum range (especially applicable to new energy vehicles, in kilometers). In the process of constructing the model state vector, historical vehicle records matching user preferences and contextual needs are used to extract and normalize the aforementioned structural attributes, thereby strengthening the characterization of user vehicle preferences in the state vector. For example, if a user frequently selects a pure electric mid-size SUV, the system will extract feature vectors for this type of vehicle in terms of energy consumption type = pure electric, vehicle size = mid-size, passenger capacity = 5 people, and range > 500km, and assign corresponding weights to these features in the state vector.

[0061] Furthermore, the reinforcement learning model is trained through the following steps:

[0062] A training dataset containing historical state vectors, historical recommended actions, and historical behavior feedback information was constructed based on historical user behavior data.

[0063] Based on the historical behavior feedback information, a multi-objective reward function reflecting immediate response and long-term satisfaction is constructed, and the corresponding reward value is calculated according to the multi-objective reward function.

[0064] The historical state vector, historical recommended actions, and corresponding reward values ​​are input into the policy network, and forward calculation and backpropagation updates are performed using a policy gradient algorithm that includes entropy regularization.

[0065] In some embodiments, firstly, based on massive historical car rental user behavior log data, the triples (state vector, recommended action, and behavioral feedback) contained in each complete interaction record are extracted. The state vector integrates long-term user preference features (such as car model, rental period, and price range) with contextual information (such as travel purpose, budget, and number of passengers). The recommended action includes a combination of fields such as the output car model category, rental range, and recommended rental location. The behavioral feedback includes whether the recommendation was clicked, whether the rental was completed, and post-rental ratings, thus constructing a training sample set. To improve training stability, all vector data are normalized and batch-grouped before entering the training process. Subsequently, a multi-objective reward function is designed to jointly model different feedback dimensions. This function integrates immediate behavioral responses (click behavior) and long-term satisfaction indicators (rental completion rate, rating feedback) to strengthen the recommendation strategy's ability to balance different objectives. For example, the click behavior score uses a normalized value of the recommendation click-through rate; the rental completion behavior is represented by a Boolean variable indicating whether the conversion was successful; and the rating feedback is calculated using a standardized rating value. The feedback values ​​from each dimension are multiplied by preset weight coefficients (adjustable) and then summed to generate the final reward value. This reward value not only guides policy updates but also reflects the degree to which the recommendation system fits the user's true preferences. During the policy optimization phase, the constructed (state, action, reward) samples are input into a reinforcement learning model based on the policy network. In the forward propagation, the policy network maps the state vector to the action probability distribution through several layers of feedforward structure and samples recommended actions from the distribution. In the backpropagation, the model uses an entropy regularization term to smooth the policy output distribution to avoid getting trapped in local optima. The loss function adopts the truncated objective function form in PPO, combined with the reward advantage function to calculate the gradient with respect to the policy parameters, and iteratively updates the network parameters through stochastic gradient descent.

[0066] Furthermore, the multi-objective reward function is as follows:

[0067] R(s,a)=α·r1+β·r2+γ·r3·π(a|s);

[0068] Where R(s, a) is the reward value for performing the recommended action a under state vector s; r1 is the feedback score corresponding to the user's click behavior; r2 is the score for the user to complete the rental behavior; r3 is the user rating feedback value; π(a|s) is the output probability of the recommended action a under state vector s; α, β and γ are weight coefficients.

[0069] It's important to note that this reward function comprehensively considers user feedback behavior at different interaction stages, including clicks, whether an actual rental is completed, and ratings of the rental service. Specifically, when a user clicks on a vehicle listing in the recommendation list, it can be seen as an initial indication of their interest. Therefore, the system generates a click feedback score based on the normalized ratio between clicked and unclicked items. If the user completes the rental process, a higher conversion feedback score is assigned to the recommendation action. After the rental period, the system collects the user's satisfaction rating of the rental process and standardizes it into a rating feedback value. These three types of feedback are weighted and fused to form the overall reward value for the recommendation behavior, thus balancing immediate interaction performance and long-term satisfaction.

[0070] Furthermore, the feedback score r1 corresponding to the user's click behavior is calculated by normalizing the ratio of clicked recommended items to unclicked items in the recommendation list; the score r2 for the user to complete the rental behavior is a Boolean value indicating whether the recommended vehicle has been actually rented; the user rating feedback value r3 is calculated by normalizing the comprehensive rating of the vehicle service collected by the user after the rental is completed.

[0071] Specifically, firstly, the feedback score corresponding to user click behavior mainly reflects their initial interest in the recommendation list. The system records whether the user clicked on the candidate vehicle item generated by the current recommendation action and calculates the ratio between the number of clicked items and the number of unclicked items. To avoid bias caused by scale inconsistencies between samples, this ratio is further processed through min-max normalization to obtain a standardized click feedback score, ensuring its distribution is stably controlled within a preset range (e.g., [0,1]). Secondly, the score for the user completing the rental behavior is used to characterize the actual conversion capability of the recommendation results. This score is a Boolean type indicator; if the user ultimately selects a recommended vehicle to place an order and pay, the score is assigned 1, indicating a successful recommendation; if the user does not select a recommended vehicle or abandons the rental operation, the score is 0. This scoring mechanism ensures that the model strengthens high-conversion recommendation paths during training. Finally, the user rating feedback value is used to measure the user's satisfaction with the rental experience. After the rental is completed, the system guides the user to give a comprehensive rating on dimensions such as vehicle condition, service process, and platform experience, using a 5-point or 10-point scale. The scoring results are standardized and mapped to a unified score feedback value, which serves as an important component of the reinforcement learning reward function, guiding the model towards optimization of recommendation strategies with better service quality assurance.

[0072] Further, S3 includes the following steps:

[0073] Based on the vehicle type, rental price range, and rental location filtering criteria included in the recommendation action, a set of candidate vehicles that meet the filtering criteria is retrieved from the car rental service database;

[0074] Obtain the popularity and historical user ratings for each vehicle in the candidate vehicle set, and construct a rating vector;

[0075] The vehicles in the candidate vehicle set are sorted according to the rating vector, and a recommendation list is generated based on the recommendation order priority.

[0076] Specifically, first, the system receives key filtering conditions from the recommendation action, including the category of the recommended vehicle (e.g., SUV, small car, electric car, etc.), the rental price range (e.g., 100–300 yuan / day), and the rental location (e.g., the user's current location or a designated pick-up area). These conditions are used as joint search parameters to perform a query operation in the car rental service database, filtering out a set of candidate vehicles that meet the conditions. This database is typically a structured form containing fields such as basic vehicle information, real-time rental status, rental location, and historical rental records. Subsequently, for each vehicle in the candidate set, the system extracts two key evaluation indicators: "popularity" and "historical user ratings." Popularity is calculated based on a weighted average of recent page views, order volume, and page dwell time, reflecting the vehicle's level of user attention on the current platform. Historical user ratings are normalized based on the average ratings from rental users regarding the vehicle's appearance, interior, performance, and overall driving experience. These two indicators together form the vehicle's comprehensive rating vector. Finally, the system sorts the candidate vehicle set according to the score vector. The sorting strategy can employ a weighted scoring model or a learned sorting algorithm to improve the personalization and accuracy of the sorting. Based on the sorting results, a final recommendation list is constructed to ensure that top-ranked vehicles are displayed first in the user's click-through interface, thereby improving the conversion rate and satisfaction of the recommended content.

[0077] The present invention also includes an artificial intelligence-based car rental recommendation system, applied to any of the preceding artificial intelligence-based car rental recommendation methods, comprising:

[0078] The data acquisition module is used to acquire users' car usage records and user demand information input by users, and integrate them to construct a state vector to represent the car rental recommendation environment;

[0079] The recommendation learning module is used to input the state vector into the reinforcement learning model and output a recommendation action, which includes the type of car model to be recommended, the rental price range, and the car rental location selection information; it is also used to calculate the reward value based on the behavioral feedback information and update the reinforcement learning model through the policy gradient algorithm;

[0080] The filtering and recommendation module is used to generate a recommendation list by filtering a set of vehicles that meet the recommendation action from the car rental service database.

[0081] The feedback information acquisition module is used to collect user feedback information on the recommended list. The feedback information includes click behavior, whether the rental is completed, and rental experience rating data.

[0082] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A car rental recommendation method based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain the user's car usage records and user demand information input by the user, and fuse them to construct a state vector to represent the car rental recommendation environment; S2. Input the state vector into the reinforcement learning model. The reinforcement learning model models car rental recommendation behavior based on Markov decision process and outputs the corresponding recommendation action. The recommendation action includes the type of car model to be recommended, the rental price range, and the selection of rental location. S3. Generate a recommendation list by filtering the set of vehicles that meet the recommended action from the car rental service database; S4. Obtain user feedback information on the recommendation list, including click behavior, whether the rental was completed, and rental experience rating data. The reward value is calculated based on the behavioral feedback information, and the reinforcement learning model is updated using the policy gradient algorithm.

2. The car rental recommendation method based on artificial intelligence according to claim 1, characterized in that, The user demand information includes contextual information such as the current purpose of travel, budget range, origin and destination, and number of passengers.

3. The car rental recommendation method based on artificial intelligence according to claim 2, characterized in that, S1 includes the following steps: S11. Based on the user's historical vehicle usage records, extract long-term preference features including vehicle type preference, rental period preference, price acceptance range, and service rating preference, and construct a user preference vector; S12. Perform structured encoding on the user's input of current travel purpose, budget range, origin and destination locations, and number of passengers to construct a context feature vector; S13. The user preference vector and the context feature vector are concatenated and fused to construct a state vector that represents the current car rental recommendation environment.

4. The car rental recommendation method based on artificial intelligence according to claim 3, characterized in that, S11 includes the following steps: The frequency of use and rating data of each vehicle type in the user's historical car rental records are statistically analyzed to generate a preference rating for each vehicle type. Extract rental period length and payment amount from users' historical orders to construct corresponding rental period preference parameters and price acceptance range parameters; The vehicle type preference rating vector, rental period preference parameter, and price acceptance range parameter are normalized and encoded, and then concatenated to generate a user preference vector that represents the user's long-term preferences.

5. The car rental recommendation method based on artificial intelligence according to claim 4, characterized in that, The vehicle categories include: vehicle size, energy consumption type, vehicle brand, spare space, passenger capacity, and maximum range.

6. The car rental recommendation method based on artificial intelligence according to claim 1, characterized in that, The reinforcement learning model is trained through the following steps: A training dataset containing historical state vectors, historical recommended actions, and historical behavior feedback information was constructed based on historical user behavior data. Based on the historical behavior feedback information, a multi-objective reward function reflecting immediate response and long-term satisfaction is constructed, and the corresponding reward value is calculated according to the multi-objective reward function. The historical state vector, historical recommended actions, and corresponding reward values ​​are input into the policy network, and forward calculation and backpropagation updates are performed using a policy gradient algorithm that includes entropy regularization.

7. The car rental recommendation method based on artificial intelligence according to claim 6, characterized in that, The multi-objective reward function is as follows: R(s,a)=α·r1+β·r2+γ·r3·π(a|s); Where R(s, a) is the reward value for performing the recommended action a under state vector s; r1 is the feedback score corresponding to the user's click behavior; r2 is the score for the user to complete the rental behavior; r3 is the user rating feedback value; π(a|s) is the output probability of the recommended action a under state vector s; α, β and γ are weight coefficients.

8. The car rental recommendation method based on artificial intelligence according to claim 7, characterized in that, The feedback score r1 corresponding to the user's click behavior is calculated by normalizing the ratio of clicked recommended items to unclicked items in the recommendation list; the score r2 for the user to complete the rental behavior is a Boolean value indicating whether the recommended vehicle was actually rented; the user rating feedback value r3 is calculated by normalizing the comprehensive rating of the vehicle service collected by the user after the rental is completed.

9. The car rental recommendation method based on artificial intelligence according to claim 1, characterized in that, S3 includes the following steps: Based on the vehicle type, rental price range, and rental location filtering criteria included in the recommendation action, a set of candidate vehicles that meet the filtering criteria is retrieved from the car rental service database; Obtain the popularity and historical user ratings of each vehicle in the candidate vehicle set, and construct a rating vector; The vehicles in the candidate vehicle set are sorted according to the rating vector, and a recommendation list is generated based on the recommendation order priority.

10. An AI-based car rental recommendation system, applied to the AI-based car rental recommendation method described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire users' car usage records and user demand information input by users, and integrate them to construct a state vector to represent the car rental recommendation environment; The recommendation learning module is used to input the state vector into the reinforcement learning model and output a recommendation action, which includes the type of car model to be recommended, the rental price range, and the car rental location selection information; it is also used to calculate the reward value based on the behavioral feedback information and update the reinforcement learning model through the policy gradient algorithm; The filtering and recommendation module is used to generate a recommendation list by filtering a set of vehicles that meet the recommendation action from the car rental service database. The feedback information acquisition module is used to collect user feedback information on the recommended list. The feedback information includes click behavior, whether the rental is completed, and rental experience rating data.

Images