A hotel price change prediction method and system based on information entropy and a large language model

Abstract

The present disclosure belongs to the technical field of computer learning, and particularly relates to a hotel price change prediction method and system based on information entropy and a large language model. The method comprises the following steps: obtaining hotel price change data from different data sources; extracting data features of the hotel price change data; wherein the hotel price change data comprises price change days, daily price change days and the farthest price change day; the data features comprise information entropy of the price change days; constructing a natural language prompt word according to the data features; calling a large language model for reasoning according to the natural language prompt word to generate a query plan containing a query date, a time window and a priority; and performing hotel price change prediction through the query plan. The present disclosure realizes intelligent conversion from data features to strategy generation through a large language model, does not need manual rule making, and can automatically adapt to complex and changeable hotel price change modes.

Description

Technical Field

[0001] This disclosure belongs to the field of computer learning technology, and specifically relates to a method and system for predicting hotel price changes based on information entropy and large language models. Background Technology

[0002] As a core infrastructure of the global tourism industry, the hotel distribution sector faces the dual challenges of integrating massive amounts of heterogeneous data and managing real-time price changes. Currently, the industry commonly uses hotel aggregation platforms as data hubs, acquiring salable hotel resources by calling application programming interfaces (APIs) from multiple vendors (such as Central Reservation System (CRS), Hotel Management System (PMS), and Global Distribution System (GDS). These platforms standardize and transform the scattered hotel data before storing it in local databases, building an electronic pipeline connecting hotels (suppliers) and distribution channels (demand-sides) to achieve automatic matching and flow of transaction data. With the industry's expansion, mainstream platforms need to manage hundreds of thousands to millions of hotel resources and process tens of millions of price change events daily, placing extremely high demands on system architecture performance and stability.

[0003] Hotel search, a core function of the platform, involves multi-dimensional filtering based on hotel name, address, keywords, and other criteria. To achieve millisecond-level response times, the platform needs to keep key index information such as hotel codes, supplier codes, and channel codes resident in memory. However, existing filtering algorithms have significant drawbacks: first, they are inefficient, experiencing response delays during high-concurrency queries; second, they are heavily reliant on external systems, with over 70% of algorithms relying on remote database or Redis calls, and network communication accounting for 45% of the time spent; third, they lack scalability, with vertical scaling costs increasing exponentially, requiring a complete redesign of the architecture for scaling to millions of hotels; and fourth, data inconsistency leads to 10%-30% or more discrepancies in query results, with accuracy generally below 90%. Hotel search typically includes multi-hotel list searches, displaying hotel overview information and single-hotel detail searches, generally providing specific room types and real-time pricing information. The difference in data acquisition strategies between the two leads to a severe gap in user experience. Multi-hotel searches generally use local data caching mechanisms to improve response speed, while single-hotel queries call supplier interfaces in real time to obtain accurate prices. When the frequency of price changes exceeds the cache update frequency, price differences between pages can exceed 10%, even reaching 30%. The "price trap" scenario, where users click on a low-priced hotel in a multi-hotel list but are then redirected to a high-priced single hotel's details page, directly leads to a transaction cancellation rate of around 20%.

[0004] Currently, hotel-related systems in the industry generally attempt to solve this problem through two approaches: One is a fixed time window strategy, such as using a uniform cache update frequency (e.g., every 10 minutes). This approach cannot adapt to the characteristics of different hotels; for example, business hotel prices fluctuate mainly between 9:00-11:00 on weekdays, while resort hotel prices change frequently between 20:00-22:00 on weekends. This results in high resource waste, and many key price changes remain uncaptured. The second approach is traditional machine learning models, mainly relying on algorithms such as Gradient Boosting Tree (GBDT) and Recurrent Neural Networks (RNN). This approach has three limitations: 1. It can only process structured numerical features and cannot parse key event information in hotel description text; 2. Feature engineering is costly, and the feature extraction process needs to be redesigned when new data sources are added; 3. The model has weak generalization ability, and its prediction accuracy drops sharply when facing new markets (such as emerging tourist cities). Summary of the Invention

[0005] To address the aforementioned issues, this disclosure provides a hotel price change prediction method based on information entropy and a large language model, which has the advantage of high efficiency in price monitoring. The method includes: Obtain hotel price change data from different data sources; Extract data features from the hotel price change data; wherein, the hotel price change data includes: the number of days with price changes, the number of days with price changes per day, and the furthest number of days with price changes; the data features include: the information entropy of the number of days with price changes; Construct natural language prompt words based on the data features; Based on the natural language prompts, a large language model is invoked to perform inference and generate a query plan that includes the query date, time window, and priority. The hotel price change is predicted using the query plan.

[0006] Furthermore, obtain hotel price change data from different data sources, including: Price change data is obtained from multiple data sources, including OTA platforms, GDS systems, and direct hotel connections, via API. The price change data includes parameters such as hotel code, supplier code, channel code, price change time, original price, and new price.

[0007] Furthermore, the data features of the hotel price change data are extracted, including: A two-dimensional time distribution matrix is ​​constructed based on the hotel price change data; wherein, the rows and columns of the two-dimensional time distribution matrix represent the number of days until check-in and the time of day, respectively, and the elements in the two-dimensional time distribution matrix represent the number of price changes at the corresponding time points; Extract the data features from the two-dimensional time distribution matrix.

[0008] Furthermore, the calculation formula for the information entropy includes:

[0009] in, The information entropy represents the number of days the price changed; Indicates the first The proportion of daily price changes to the total number of price changes.

[0010] Furthermore, before constructing natural language prompts based on the data features, the method further includes: Remove data features whose price change frequency is below a threshold from the data features.

[0011] Furthermore, based on the natural language prompts, a large language model is invoked to perform inference, generating a query plan that includes the query date, time window, and priority, including: The query plan generation is performed in parallel by batching and grouping using a thread pool. The target query plan is obtained by merging several query plans generated in parallel.

[0012] Furthermore, after predicting hotel price changes using the aforementioned query plan, the process also includes: The price capture rate and omission rate of the query results corresponding to the target query plan are used as feedback parameters to modify the working parameters, including natural language prompts, large model parameters, and data feature weights. The target query plan is then iteratively optimized based on the modified working parameters.

[0013] This disclosure also proposes a hotel price change prediction system based on information entropy and a large language model, including: The data acquisition module is used to acquire hotel price change data from different data sources; The feature extraction module is used to extract data features from the hotel price change data; wherein, the hotel price change data includes: the number of days with price changes, the number of days with price changes per day, and the furthest number of days with price changes; the data features include: the information entropy of the number of days with price changes; The prompt word construction module is used to construct natural language prompt words based on the data features; The model invocation module is used to invoke the large language model for inference based on the natural language prompts and generate a query plan that includes the query date, time window and priority. The price change prediction module is used to predict hotel price changes based on the query plan.

[0014] This disclosure also proposes a computer-readable storage medium storing a computer program or instructions, which, when executed by a processor, are at least used to implement the above-described hotel price prediction method based on information entropy and a large language model.

[0015] This disclosure also proposes a computer program product stored in a computer-readable storage medium, which, when executed by a processor, is used to at least implement the above-described hotel price prediction method based on information entropy and a large language model.

[0016] Compared with existing technologies, this disclosure has the following advantages: it realizes intelligent transformation from data features to strategy generation through a large language model, without the need for manual rule formulation, can automatically adapt to complex and ever-changing price change patterns, and the information entropy quantification index accurately depicts the regularity of price changes. The hierarchical feature engineering fully explores the value of historical data, and the reasoning ability of the large language model ensures the rationality of the strategy.

[0017] Other features and advantages of this disclosure will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A hotel price change prediction method based on information entropy and a large language model is illustrated according to an embodiment. Figure 2 A hierarchical feature extraction method according to an embodiment is illustrated; Figure 3 The process of information entropy calculation and entropy classification according to an embodiment is shown; Figure 4 A query strategy generation scheme according to an embodiment is shown; Figure 5 A parallel query plan generation scheme according to an embodiment is shown; Figure 6 A parallel query plan generation scheme according to another embodiment is shown; Figure 7A hotel price change prediction method based on a feedback mechanism is illustrated according to an embodiment; Figure 8 A method for structured transformation of variable price time distribution features according to an embodiment is shown; Figure 9 The hotel price change prediction process according to an embodiment is shown; Figure 10 A hotel price change prediction system based on information entropy and a large language model is illustrated according to an embodiment. Figure 11 The collaborative relationships of the hotel price change prediction system modules according to an embodiment are illustrated; Figure 12 A computer-readable storage medium according to an embodiment is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0021] Figure 1 The hotel price change prediction method based on information entropy and large language model proposed in this disclosure includes: Obtain hotel price change data from different data sources; Extract data features from the hotel price change data; wherein, the hotel price change data includes: the number of days with price changes, the number of days with price changes per day, and the furthest number of days with price changes; the data features include: the information entropy of the number of days with price changes; Construct natural language prompt words based on the data features; Based on the natural language prompts, a large language model is invoked to perform inference and generate a query plan that includes the query date, time window, and priority. The hotel price change is predicted using the query plan.

[0022] According to some embodiments of this disclosure, the hotel price change prediction method based on information entropy and large language model proposed in this disclosure mainly includes the following steps: (1) Multi-source data acquisition and standardized processing; (2) Construct a hierarchical feature extraction system and gradually abstract high-level features from the original data: construct a two-dimensional time distribution matrix for each hotel-supplier-channel combination, extract daily statistical features based on the two-dimensional time distribution matrix, and finally calculate the information entropy of the data features as the core quantitative indicator.

[0023] (3) Query strategy generation based on large language model: engineering design of prompt words based on extracted feature information, calling and reasoning of large language model and strategy parsing and conversion; (4) Strategy supplementation and integrity guarantee: Multiple mechanisms are implemented to ensure the integrity of the strategy, including: automatically supplementing the strategy based on historical patterns for time periods when the model has not generated a strategy; identifying and specially handling hotels with frequent price fluctuations; merging queries from multiple channels with the same discount to reduce redundant operations, etc.

[0024] Furthermore, obtain hotel price change data from different data sources, including: Price change data is obtained from multiple data sources, including OTA platforms, GDS systems, and direct hotel connections, via API. The price change data includes parameters such as hotel code, supplier code, channel code, price change time, original price, and new price.

[0025] According to some embodiments of this disclosure, multi-source data acquisition and standardization processing mainly includes: Price change data is acquired from multiple data sources, including OTA platforms, GDS systems, and direct hotel connections, via an Application Programming Interface (API). The raw data includes fields such as hotel code, vendor code, channel code, price change timestamp, original price, and new price. The collected raw data undergoes cleaning and standardization processes, including removing spaces and special characters, standardizing the time format, handling missing values, and detecting and correcting outliers. The key derived field TIME_DIS_HOURS is calculated, representing the difference in hours from check-in date 22:00 to the time of the price change, for subsequent time-dimensional analysis.

[0026] Furthermore, the data features of the hotel price change data are extracted, including: A two-dimensional time distribution matrix is ​​constructed based on the hotel price change data; wherein, the rows and columns of the two-dimensional time distribution matrix represent the number of days until check-in and the time of day, respectively, and the elements in the two-dimensional time distribution matrix represent the number of price changes at the corresponding time points; Extract the data features from the two-dimensional time distribution matrix.

[0027] Furthermore, the calculation formula for the information entropy includes:

[0028] in, The information entropy represents the number of days the price changed; Indicates the first The proportion of daily price changes to the total number of price changes.

[0029] According to some embodiments of this disclosure, such as Figure 2 As shown, a hierarchical feature extraction system is constructed to gradually abstract high-level features from raw data, mainly including: The first layer involves constructing a time distribution matrix. A two-dimensional time distribution matrix, `time_distribution`, is built for each hotel-supplier-channel combination. Rows represent the number of days remaining until check-in (0-29 days), and columns represent the hours of the day (0-23 hours). Each element in the matrix records the number of price changes at the corresponding time point.

[0030] Second layer: Daily statistical feature extraction change_days: A list of the number of days since the price change, such as [0, 7, 14, 21, 28]. `day_change_counts`: A dictionary of the number of price changes per day, such as `{0: 15, 7: 12, 14: 10}`. max_change_day: The furthest number of days since the price change, determining the monitoring time range. The third layer: pattern quantification feature calculation, calculating information entropy as the core quantification indicator:

[0031] in, The information entropy represents the number of days the price changed; Indicates the first The proportion of daily price changes to the total number of price changes. Entropy quantifies the regularity of price change patterns: low entropy (0-2.0) indicates that price changes are concentrated in a specific number of days, while high entropy (>3.0) indicates that price changes are scattered and random.

[0032] Figure 3 The flowchart illustrates the process of calculating and classifying entropy based on information entropy. After obtaining the data on the number of days of price changes, the total number of price changes and the probability of price changes per day are calculated. The information entropy of the number of days of price changes is calculated based on the proportion of the number of price changes to the total number of price changes. Based on the range of entropy values, it is determined whether it belongs to low entropy, medium entropy, high entropy, or extremely high entropy. Different query strategies are then selected based on the different entropy values.

[0033] Furthermore, before constructing natural language prompts based on the data features, the method further includes: Remove data features whose price change frequency is below a threshold from the data features.

[0034] Furthermore, based on the natural language prompts, a large language model is invoked to perform inference, generating a query plan that includes the query date, time window, and priority, including: The query plan generation is performed in parallel by batching and grouping using a thread pool. The target query plan is obtained by merging several query plans generated in parallel.

[0035] According to some embodiments of this disclosure, such as Figure 4 As shown, the query strategy generation methods include: Prompt word engineering design: The extracted features are converted into structured natural language prompts, including three parts: task description, feature description, and output format requirements. The feature data is streamlined, retaining only the data from the 15-20 days with the most frequent price changes, reducing the length of the model input while preserving key information.

[0036] Model Invocation and Inference: This embodiment uses the DeepSeek-R1 large language model for policy generation. A temperature parameter is set to 0.3 to improve output determinism, requiring the return of a structured query policy in JSON format. Retry mechanisms and exception handling are implemented to ensure service stability.

[0037] Strategy parsing and conversion: The JSON query strategies output by the model are converted into query plans that the system can execute. Each strategy includes information such as: query date (number of days relative to check-in date), query time window (start and end hours), and query priority.

[0038] According to some embodiments of this disclosure, such as Figure 5 As shown, the query plan generation method proposed in this disclosure, which uses a thread pool to perform parallel execution in batches, includes: Calculate the total query plan and generate resource requirements based on the query strategies for all hotels; Determine whether the total query plan generation resource requirement exceeds the threshold. If not, directly execute the query strategy corresponding to each hotel. If not, assign weights to each hotel based on the information entropy of the number of days of price change for each hotel, generate corresponding query strategies, and execute them. Hotel price data is collected based on the query strategy.

[0039] According to some embodiments of this disclosure, such as Figure 6 As shown, to improve the processing efficiency of query plans, a parallel processing architecture based on ThreadPoolExecutor was implemented. The hotel list is grouped by batch size (default 50), and a thread pool is created to process each batch in parallel. Each batch independently performs feature extraction, model inference, and strategy generation, and finally the results of all batches are merged.

[0040] Furthermore, after predicting hotel price changes using the aforementioned query plan, the process also includes: The price capture rate and omission rate of the query results corresponding to the target query plan are used as feedback parameters to modify the working parameters, including natural language prompts, large model parameters, and data feature weights. The target query plan is then iteratively optimized based on the modified working parameters.

[0041] According to some embodiments of this disclosure, such as Figure 7 As shown, after predicting hotel price changes based on the query plan, a strategy effectiveness evaluation system is established to collect indicators such as price capture rate and omission rate during actual implementation. Feature weights are adjusted, prompt word templates are optimized, and strategy generation parameters are updated based on feedback data. For example, A / B testing is used to verify the optimization effect, forming a closed loop of continuous improvement.

[0042] According to some embodiments of this disclosure, such as Figure 8 As shown, this disclosure proposes a scheme to structurally transform the price change time distribution features (change_windows list) according to window type and splitting granularity, generating executable daily query strategies (daily_strategies), providing standardized input for subsequent large-scale model strategy optimization. This scheme includes: For the extracted change_windows price change time window list, single-hour windows and continuous range windows are distinguished by window type: for single-hour windows, a corresponding single query window is directly generated; for continuous range windows, three splitting strategies—fine-grained, medium-grained, and coarse-grained—are selected based on the price change pattern to generate sub-windows of corresponding granularity; after all windows are validated by end>start, they are integrated into standardized daily_strategies daily query strategies, providing structured input for subsequent large language model strategy generation, and achieving accurate conversion of price change time features into executable query strategies.

[0043] like Figure 9 As shown, the hotel price change prediction process proposed in this disclosure is as follows: First, multi-source hotel price data is collected and preprocessed. Through a hierarchical feature extraction architecture, daily statistical features are extracted layer by layer from the original timestamp data, and then the price change information entropy is calculated to quantify the regularity of the price change pattern. Based on the entropy value range, the price change pattern is classified, and the classification results are input into a large language model to generate personalized query strategies. After strategy optimization and resource allocation, price monitoring is executed. Finally, a closed-loop feedback mechanism is constructed through effect evaluation. If optimization is required, the data is fed back to iterate the feature extraction and strategy generation logic, continuously improving the price capture rate and reducing redundant queries, thereby achieving intelligent, efficient, and adaptive optimization of hotel price change monitoring.

[0044] Based on the same technological concept, such as Figure 10As shown, this disclosure also proposes a hotel price prediction system based on information entropy and a large language model, including: The data acquisition module is used to acquire hotel price change data from different data sources; The feature extraction module is used to extract data features from the hotel price change data; wherein, the hotel price change data includes: the number of days with price changes, the number of days with price changes per day, and the furthest number of days with price changes; the data features include: the information entropy of the number of days with price changes; The prompt word construction module is used to construct natural language prompt words based on the data features; The model invocation module is used to invoke the large language model for inference based on the natural language prompts and generate a query plan that includes the query date, time window and priority. The price change prediction module is used to predict hotel price changes based on the query plan.

[0045] According to some embodiments of this disclosure, such as Figure 11 As shown, the collaborative relationships among the modules of the hotel price prediction system based on information entropy and large language model proposed in this disclosure are as follows: 1. Feature Processing Stage (corresponding to Feature Extraction Module + Prompt Word Construction Module): Taking the hotel price feature data (including core features such as number of days with price changes, frequency of price changes, and information entropy) output by the feature extraction module as input, feature simplification and transformation are performed, retaining only the key number of days with the most frequent price changes to compress the input length. Then, structured prompt words are created to provide standardized feature input for prompt word construction.

[0046] 2. Hint Construction Phase (corresponding to Hint Word Construction Module): Based on structured features, configure the system role hints (defining the large language model as a hotel pricing strategy expert), task description (clarifying the core objective of generating the optimal query strategy), and output format definition (requiring the return of a JSON-formatted structured strategy). Finally, assemble the complete hints to complete the construction of natural language hint words, preparing for model inference.

[0047] 3. API call phase (corresponding to the model call module): Call the large language model API to perform inference. First, determine whether the response is successful: if successful, parse the JSON result and extract the query plan generated by the model; if it fails, trigger the retry mechanism. When the number of retries exceeds the threshold, the default strategy is used as a fallback to ensure service stability. Finally, the model inference result is output.

[0048] 4. Result Processing Stage (corresponding to the model calling module + price prediction module): The strategy output by the model is validated to ensure its completeness and rationality. Then, the effectiveness of the strategy is determined: if invalid, the strategy is corrected; if effective, it is converted to the system format, transforming the JSON strategy into an executable query plan (including query date, time window, and priority). Finally, the final strategy is output for the price prediction module to execute and predict hotel price changes.

[0049] like Figure 12 As shown, this disclosure also proposes a computer-readable storage medium storing a computer program or instructions, which, when executed by a processor, are at least used to implement the above-described hotel price prediction method based on information entropy and a large language model.

[0050] This disclosure also proposes a computer program product stored in a computer-readable storage medium, which, when executed by a processor, is used to at least implement the above-described hotel price prediction method based on information entropy and a large language model.

[0051] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for predicting hotel price changes based on information entropy and a large language model, characterized in that, include: Obtain hotel price change data from different data sources; Extract data features from the hotel price change data; wherein, the hotel price change data includes: the number of days with price changes, the number of days with price changes per day, and the furthest number of days with price changes; the data features include: the information entropy of the number of days with price changes; Construct natural language prompt words based on the data features; Based on the natural language prompts, a large language model is invoked to perform inference and generate a query plan that includes the query date, time window, and priority. The hotel price change is predicted using the query plan.

2. The hotel price prediction method based on information entropy and large language model as described in claim 1, characterized in that, Obtain hotel price change data from different data sources, including: Price change data is obtained from multiple data sources, including OTA platforms, GDS systems, and direct hotel connections, via API. The price change data includes parameters such as hotel code, supplier code, channel code, price change time, original price, and new price.

3. The hotel price prediction method based on information entropy and large language model as described in claim 1, characterized in that, Extracting data features from the hotel price change data includes: A two-dimensional time distribution matrix is ​​constructed based on the hotel price change data; wherein, the rows and columns of the two-dimensional time distribution matrix represent the number of days until check-in and the time of day, respectively, and the elements in the two-dimensional time distribution matrix represent the number of price changes at the corresponding time points; Extract the data features from the two-dimensional time distribution matrix.

4. The hotel price prediction method based on information entropy and large language model as described in claim 3, characterized in that, The information entropy, and the corresponding calculation formula, includes: in, The information entropy represents the number of days the price changed; Indicates the first The proportion of daily price changes to the total number of price changes.

5. The hotel price prediction method based on information entropy and large language model as described in claim 1, characterized in that, Before constructing natural language prompts based on the data features, the process also includes: Remove data features whose price change frequency is below a threshold from the data features.

6. The hotel price prediction method based on information entropy and large language model as described in any one of claims 1-5, characterized in that, Based on the natural language prompts, a large language model is invoked to perform inference, generating a query plan that includes the query date, time window, and priority, including: The query plan generation is performed in parallel by batching and grouping using a thread pool. The target query plan is obtained by merging several query plans generated in parallel.

7. The hotel price prediction method based on information entropy and large language model as described in claim 6, characterized in that, After predicting hotel price changes using the aforementioned query plan, the process also includes: The price capture rate and omission rate of the query results corresponding to the target query plan are used as feedback parameters to modify the working parameters, including natural language prompts, large model parameters, and data feature weights. The target query plan is then iteratively optimized based on the modified working parameters.

8. A hotel price change prediction system based on information entropy and a large language model, characterized in that, include: The data acquisition module is used to acquire hotel price change data from different data sources; The feature extraction module is used to extract data features from the hotel price change data; wherein, the hotel price change data includes: the number of days with price changes, the number of days with price changes per day, and the furthest number of days with price changes; the data features include: the information entropy of the number of days with price changes; The prompt word construction module is used to construct natural language prompt words based on the data features; The model invocation module is used to invoke the large language model for inference based on the natural language prompts and generate a query plan that includes the query date, time window and priority. The price change prediction module is used to predict hotel price changes based on the query plan.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program or instructions, which, when executed by a processor, are at least used to implement the hotel price prediction method based on information entropy and a large language model as described in any one of claims 1-7.

10. A computer program product, said computer program product being stored in a computer-readable storage medium, characterized in that, When the computer program product is executed by a processor, it is used to implement at least the hotel price prediction method based on information entropy and large language model as described in any one of claims 1-7.

Images