Tobacco delivery prediction method based on time sequence large language model
By using the time series-based large language model Time-LLaMA to perform feature extraction and preprocessing of tobacco delivery information, the problem of limited prediction accuracy in the prior art is solved, and higher prediction accuracy and adaptability are achieved.
Patent Information
- Application Number
- CN202411568479.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-05
- Publication Date
- 2025-02-18
AI Technical Summary
The existing tobacco delivery prediction methods have limited prediction accuracy and are difficult to cope with real-time dynamic changes in complex markets.
Time-LLaMA, a large language model based on time series, is used to obtain information in the tobacco system, trading software and map software, extract and preprocess it, and enter the Time-LLaMA model for training, so as to predict tobacco delivery.
It improves the accuracy of tobacco delivery prediction, can better adapt to inputs of different types of features, and enhances the adaptability and prediction capabilities of the model.
Smart Images
Figure CN119477397A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of tobacco management, and in particular to a tobacco placement prediction method based on a time series large language model. Background Art
[0002] The tobacco market is a highly competitive and strictly regulated market. In recent years, with the continuous changes in market demand and consumer preferences, tobacco merchants need to more accurately predict the release volume of tobacco specifications in order to optimize inventory management and improve sales efficiency.
[0003] Traditional forecasting methods rely on historical sales data and empirical judgment, which makes it difficult to cope with the real-time dynamic changes in complex markets. Therefore, using big data and machine learning technology to predict the release volume of tobacco products has become an innovative solution. However, the existing tobacco release volume forecasting methods on the market mainly include time series analysis based on historical data and simple statistical models. These methods have problems such as insufficient data integration, limited prediction accuracy and poor real-time performance. Summary of the invention
[0004] In view of this, it is necessary to provide a tobacco delivery prediction method based on a time series large language model to solve the problem of limited prediction accuracy of existing tobacco delivery prediction methods.
[0005] In order to solve the above problems, the present invention provides a tobacco delivery prediction method based on a time series large language model, comprising: Acquire first information; the first information includes tobacco information in the tobacco system, tobacco transaction information in the transaction software, and location information in the map software; Extracting features from the first information to obtain numerical features, classification features, and static features; Inputting the numerical features, categorical features and static features into a time series-based large language Time-LLaMA model to train the Time-LLaMA model; The tobacco delivery situation is predicted using the trained Time-LLaMA model.
[0006] Optionally, the tobacco information in the tobacco system includes: basic information of the target merchant, tobacco specification information, tobacco order information and violation information of the target merchant; the tobacco transaction information in the transaction software includes: basic information of the target merchant and transaction information of the target merchant; the location information in the map software includes: location information of each merchant in a preset area around the target merchant; then the feature extraction of the target information to obtain numerical features, classification features and static features includes: Feature extraction is performed on the first information to obtain target features; the target features include: business district features, business features and standard features; wherein the business district features include: the first average monthly customer volume, the tier where the average monthly transaction number is located and the tier where the average monthly traffic volume is located of each merchant in a preset area around the target merchant; the business features include: the ratio of the second average monthly customer volume of the target merchant to the first average monthly customer volume, the tier where the monthly customer traffic volume of the target merchant is located, the tier where the monthly transaction number is located, the tier where the monthly transaction amount is located, the average monthly transaction number, the tier where the average monthly traffic volume of customers of each age group is located, the average monthly transaction amount of each customer, the order volume, the number of order product types and the order amount; the standard features include: the amount of illegal commodities and the types of illegal commodities; According to the target features, numerical features, classification features and static features are generated.
[0007] Optionally, the extracting features from the first information to obtain target features includes: Classifying the first information by labels; the labels include: descriptive labels, fact labels, product specification labels, and violation labels; Feature extraction is performed on the first information after label classification to obtain target features.
[0008] Optionally, extracting features from the first information includes: Preprocessing the first information to obtain second information; the preprocessing includes: abnormal information processing, missing information processing, repeated information processing, information enhancement processing and information normalization processing; Perform feature extraction on the second information.
[0009] Optionally, preprocessing the first information to obtain the second information includes: When detecting that duplicate data exists in the first information, determining whether the duplicate data has business significance; If the duplicate data has business significance, retain the duplicate data; If the duplicate data has no business significance, the duplicate data are integrated and then retained, or one of the duplicate data is selectively retained.
[0010] Optionally, the method further includes: During the training of the Time-LLaMA model, the Time-LLaMA model is fine-tuned by a preset low-order adaptive LoRA fine-tuning model.
[0011] Optionally, the method further includes: The prediction results are smoothed, and outliers and extreme values of the prediction results are eliminated.
[0012] The beneficial effects of the present invention are: The first information is obtained; the first information includes tobacco information in the tobacco system, tobacco transaction information in the transaction software, and location information in the map software; the first information is feature extracted to obtain numerical features, classification features, and static features; the numerical features, classification features, and static features are input into the large language Time-LLaMA model based on the time series to train the Time-LLaMA model; and the tobacco delivery situation is predicted by the trained Time-LLaMA model. The Time-LLaMA model has a better processing effect on time series information. The present invention integrates and extracts tobacco-related information, inputs it into the Time-LLaMA model, and trains the Time-LLaMA model, thereby predicting the tobacco delivery situation in the future time through the trained Time-LLaMA model, which can improve the prediction accuracy. In addition, during the training process, the first information is decomposed and extracted into three independent tensors (numerical features, classification features, and static features) and then input into the model, so that the model can learn the characteristics of different categories respectively. This differentiated processing provides more refined data control and helps the model better adapt to the input of different types of features. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 A schematic flow chart of an embodiment of a tobacco placement prediction method based on a time series large language model provided by the present invention; Figure 2 A schematic diagram of a missing value filling method provided by the present invention; Figure 3 A schematic diagram of a flow chart of a duplicate data processing method provided by the present invention; Figure 4 A schematic diagram of a model prediction process provided by the present invention; Figure 5 Another model prediction process diagram provided by the present invention; Figure 6 A schematic diagram of a prediction effect provided by the present invention; Figure 7 Another schematic diagram of prediction effect provided by the present invention. DETAILED DESCRIPTION
[0014] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.
[0015] In the description of the embodiments of the present invention, unless otherwise specified, "multiple" means two or more. The descriptions of "first", "second", etc. involved in the embodiments of the present invention are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the technical features defined as "first" and "second" may explicitly or implicitly include at least one of the features.
[0016] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.
[0017] Reference Figure 1 , shows a flow chart of an embodiment of a tobacco delivery prediction method based on a time series large language model provided by the present invention, the method comprising: S101, obtaining first information; the first information includes tobacco information in the tobacco system, tobacco transaction information in the transaction software, and location information in the map software.
[0018] Tobacco information in the tobacco system may include: basic information of target merchants, tobacco product specification information, tobacco order information and target merchant violation information. Target merchants are merchants that sell tobacco; basic information of target merchants may include: retailer code, customer stall, customer regional information, whether the customer is a seasonal customer and customer business format. Customer regional information may include the name of the marketing department, urban and rural classification code, and market type. Tobacco product specification information may include: product code, specification, model, whether it is a slim cigarette, whether it is a medium cigarette, whether it is a short cigarette, whether it is a popping bead, etc. Tobacco order information may include: actual order quantity, demand quantity, and supply quantity. Target merchant violation information may include: violation type, number of violations, and violation amount.
[0019] The tobacco transaction information in the transaction software includes: basic information of the target merchant and transaction information of the target merchant; the location information in the map software includes: location information of each merchant in the preset area around the target merchant. The transaction software can be Alipay, WeChat, or a bank APP.
[0020] The location information in the map software may include: location information of each merchant in a preset area around the target merchant. The map software may be Amap or the like.
[0021] The first information may also include holiday information and the like.
[0022] S102, extracting features from the first information to obtain numerical features, classification features, and static features.
[0023] Numerical features are features that can be represented by numerical values. Categorical features are features that can be represented by categories or labels. Static features are features that remain unchanged throughout the dataset. They usually describe the inherent properties of the data and do not change over time or other conditions.
[0024] S103, inputting the numerical features, the categorical features and the static features into a time series-based large language Time-LLaMA model to train the Time-LLaMA model.
[0025] The Time-LLaMA model is a Transformer-based time series forecasting model. The Time-LLaMA model can accept three independent input tensors: numerical features, categorical features, and static features. Decomposing the first information into three independent tensors and then inputting them into the model enables the model to learn the characteristics of different categories separately. This differentiated processing provides more refined data control and helps the model better adapt to different types of inputs, especially when dealing with complex data with multiple labels. The Time-LLaMA model averages the categorical and static feature embeddings and combines them with the numerical features to form a tensor with shape, (batch_size, window_size, embedding_size), ready for the Transformer block. The Transformer block extracts sequential information and then aggregates the resulting tensor along the time dimension and passes it into the MLP to generate the final forecast (batch_size, forecast_length). The root mean square logarithmic error (RMSLE) is used as the evaluation metric during the training process.
[0026] S104, predicting tobacco delivery conditions using the trained Time-LLaMA model.
[0027] Deploying the trained Time-LLaMA model to the production system can achieve automatic data prediction. The prediction system can receive updated data and output the prediction of cigarette brand release volume.
[0028] Acquire the first information; the first information includes tobacco information in the tobacco system, tobacco transaction information in the transaction software, and location information in the map software; extract features from the first information to obtain numerical features, classification features, and static features; input the numerical features, classification features, and static features into a large language Time-LLaMA model based on a time series to train the Time-LLaMA model; and predict tobacco delivery through the trained Time-LLaMA model. The Time-LLaMA model has a better processing effect on time series information. The present invention integrates and extracts features from tobacco-related information, inputs the information into the Time-LLaMA model, and trains the Time-LLaMA model, thereby predicting tobacco delivery in the future through the trained Time-LLaMA model, which can improve prediction accuracy.
[0029] In one embodiment, S102 may specifically include: extracting features from the first information to obtain target features; the target features include: business district features, operating features and standard features; wherein the business district features include: the first average monthly customer volume, the average monthly transaction times and the average monthly traffic volume of each merchant in a preset area around the target merchant; the operating features include: the ratio of the second average monthly customer volume to the first average monthly customer volume, the target merchant's monthly customer traffic volume, the monthly transaction times, the monthly transaction amount, the average monthly transaction times, the average monthly traffic volume of customers of each age group, the average monthly transaction amount of each customer, the order volume, the number of order product types and the order amount; the standard features include: the amount of illegal goods and the types of illegal goods.
[0030] Specifically, the target features are: 1. Business characteristics of merchants - the monthly customer traffic classification of target merchants L1|…|L12; where L1|…|L12 represents the most recent 12 months.
[0031] -Source: Alipay face-to-face payment merchant information and transaction details.
[0032] -Function: Reflect the changes in store customer flow in each month of the year and help analyze the store's customer flow trends in different seasons.
[0033] 2. Merchant operating characteristics - the tier in which the target merchant L1|…|L12’s monthly transaction count is classified.
[0034] -Source: Alipay face-to-face payment merchant information and transaction details.
[0035] -Function: Displays the business activity of stores in different seasons by grading the number of transactions per month.
[0036] 3. Merchant operating characteristics - the monthly transaction amount classification of target merchants L1|…|L12.
[0037] -Source: Alipay face-to-face payment merchant information and transaction details.
[0038] -Function: Displays the store’s business performance at different times of the year by grading the monthly transaction amount.
[0039] 4. Merchant operating characteristics - the average number of transactions per customer per month of the target merchant in the past 12 months.
[0040] -Source: Alipay face-to-face payment merchant information and transaction details.
[0041] -Function: By analyzing the average number of transactions per customer, the transaction frequency and customer stickiness of the store in a year can be evaluated.
[0042] 5. Merchant operating characteristics - the target merchant’s order volume in the past month.
[0043] -Source: Tobacco system order details.
[0044] -Function: Use the order volume in the last month to evaluate the store’s short-term business performance and market demand.
[0045] 6. Merchant operating characteristics - the number of product types ordered by the target merchant in the past month.
[0046] -Source: Tobacco system order details.
[0047] -Function: Evaluate the store’s performance in terms of product diversity and customer selection based on the number of product types and orders in the past month.
[0048] 7. Merchant operating characteristics - order amount of the target merchant in the past month.
[0049] -Source: Tobacco system order details.
[0050] -Function: Use the order amount in the past month to evaluate the store’s sales status and profitability.
[0051] 8. Merchant operating characteristics - the target merchant's customer age and average monthly customer inflow in the past 12 months.
[0052] -Source: C-end customer information and face-to-face payment transaction details in Alipay face-to-face payment.
[0053] -Function: Analyze the traffic of different customer age groups in the past 12 months and predict the store’s possible product demand and customer groups.
[0054] 9. Merchant operating characteristics - average monthly transaction amount per customer of target merchants in the past 12 months.
[0055] -Source: Alipay face-to-face payment transaction details.
[0056] -Function: By analyzing the average transaction amount per customer, the store’s market positioning and consumption level can be determined.
[0057] 10. Merchant Standard Characteristics - The amount of illegal goods sold by the target merchant in the past six months.
[0058] -Source: Tobacco system violation data.
[0059] -Function: Assess the compliance and operating risks of stores by tracking the amount of illegal goods sold in stores in the past six months.
[0060] 11. Product Standard Characteristics - Types of illegal products of the target merchant in the past six months.
[0061] -Source: Tobacco system violation data.
[0062] -Function: Analyze the types of illegal products in the store in the past six months to evaluate the store’s response to specific market demands and risk management.
[0063] 12. Business district operating characteristics - the average number of customers of acquiring merchants in the business district within 1000m of the target merchant in the past 12 months (10,000).
[0064] -Source: Alipay face-to-face payment transaction details, Amap data.
[0065] -Function: Analyze the average number of customers of all merchants within the 1,000-meter business district where the merchant is located in the past 12 months, and evaluate the customer flow density and market vitality of the business district.
[0066] 13. Business district operating characteristics - the ratio of the target merchant’s average monthly customer number in the past 12 months to the acquiring merchant’s average monthly customer number in the past 12 months within 1000m of the merchant’s business district.
[0067] -Source: Alipay face-to-face payment transaction details, Amap data, statistics on the percentage of customers within 1000m of the business district where the merchant is located in the past 12 months.
[0068] -Function: Evaluate the market competitiveness of a store by comparing the number of customers of the store with other businesses within a 1,000-meter business district.
[0069] 14. Business district operating characteristics - the tier in which the average number of transactions of acquiring merchants in the business district within 1000m of the target merchant is ranked in the past 12 months.
[0070] -Source: Alipay face-to-face payment transaction details, Amap data.
[0071] -Function: Evaluate the transaction activity of the target merchant's business district based on the average monthly transaction number within the 1,000-meter business district.
[0072] 15. Business characteristics of merchant business district - the classification of the average monthly traffic volume of acquiring merchants in the business district within 1000m of the target merchant in the past 12 months.
[0073] -Source: Alipay face-to-face payment transaction details, Amap data.
[0074] -Function: By analyzing the average monthly passenger flow within the 1,000-meter business district where the merchant is located, the passenger flow density and potential market size of the business district can be evaluated.
[0075] In one embodiment, the step of extracting features from the first information to obtain target features specifically includes: classifying the first information with labels; the labels include: characterization labels, fact labels, product specification labels, and violation labels; extracting features from the first information after label classification to obtain target features.
[0076] Characteristic tags are mainly used to describe the characteristics, attributes or emotional tendencies of content. They are usually subjective, qualitative, and may vary from person to person. Factual tags are used to describe objective facts or data in content. They are usually objective, quantitative, and universal.
[0077] Before extracting features from the first information, the first information is classified into characterization labels, fact labels, product specification labels and violation labels. This not only improves the structuring of the information, but also enables the feature extraction model to more accurately understand and process specific information by specifically labeling different features and contents in the information.
[0078] In one embodiment, the step of extracting features from the first information specifically includes: preprocessing the first information to obtain second information; the preprocessing includes: abnormal information processing, missing information processing, repeated information processing, information enhancement processing and information normalization processing; and extracting features from the second information.
[0079] After the first information is collected, the first information may be preprocessed. The preprocessing may specifically include: 1. Data cleaning: During the data cleaning stage, the original data (first information) is strictly checked and processed to ensure the integrity and accuracy of the data.
[0080] 2. Outlier removal: Clustering-based methods can use clustering algorithms (such as K-means, DBSCAN, etc.) to divide training samples into several classes. If the number of samples in a class is very small and the distance between the class center and all other classes is very far, then the samples in this class are very likely to be abnormal feature samples.
[0081] 3. Missing value filling: Use reasonable filling methods (such as mean, median, mode or interpolation) to fill the missing values in the original data to ensure the continuity and integrity of the data. Figure 2 , showing a flow chart of a missing value filling method provided by the present invention. When data missing is detected in the original data, it is determined whether the missing data is a normal missing; if it is a normal missing, it is determined whether the missing data needs to be filled; if it needs to be filled, the filling data can be generated according to the mode or mean of the adjacent data of the missing data; if the missing data is not a normal missing, it is determined whether the missing data is recoverable; if it is recoverable, it is recovered; if it is not recoverable, it is discarded.
[0082] 4. Data normalization: Use normalization methods (such as Min-Max normalization and Z-score normalization) to scale the original data so that the data values are distributed within a reasonable range and improve the model training effect.
[0083] 5. Data enhancement: In order to increase the data diversity and robustness of model training, data enhancement technology is applied to expand and optimize the original data.
[0084] 6. Time series expansion: For time series data, add new features (such as lag terms, moving averages) or generate derivative data to enrich the dimensions of the data set.
[0085] 7. Abnormal data point deletion: The algorithm automatically identifies and deletes extremely abnormal data points to improve the stability and accuracy of the model.
[0086] 8. Duplicate data processing: The data generated in the business process may sometimes be stored in duplicate to maintain data integrity or for system performance considerations. We need to analyze the business meaning of such duplicate data to determine how to handle the duplicate data. Figure 3 , showing a flow chart of a duplicate data processing method provided by the present invention. When duplicate data is detected in the first information, it is determined whether the duplicate data has business meaning; if the duplicate data has business meaning, the duplicate data is retained; if the duplicate data has no business meaning, the duplicate data is integrated and retained, or one of the duplicate data is retained. Even if the duplicate data has no obvious effect in the current business logic, they may have an impact on the training and verification of the model. Therefore, even if the duplicate data has no business meaning, it is necessary to summarize and analyze these duplicate data to evaluate whether they may introduce noise or deviation, rather than simply retaining or deleting them.
[0087] In one embodiment, structured data can be obtained by analyzing and preprocessing the original data, and then the data needs to be designed and processed to generate features. Feature binning refers to segmenting the features of continuous attributes into discrete intervals. There are four common feature binning methods: equal frequency binning, equal interval binning, chi-square binning, and decision tree binning. After a single feature is discretized into multiple bins, the bins are numerically converted to improve the expressive power of the linear model. For example, WOE (Weight of Evidence) encoding is commonly used in supervised encoding. Specifically, the logarithm of the ratio of the probability of each category Y=1 to the probability of Y=0 in the feature is used to replace the feature value of each category. Calculation formula:
[0088] Among them B i is the number of samples with label 1 in category i, G i is the number of samples with label 0 in category i, B T is the number of samples with label 1 in all samples, G T is the number of samples with label 0 among all samples.
[0089] WOE is used to analyze the predictive ability of each feature item for the target feature, and IV can be used to reflect the overall predictive ability of the feature. The calculation formula of IV is:
[0090] Among them, Good i / Good T is the proportion of “good” samples in the ith bin, Bad i / Bad T The proportion of "bad" samples in the ith bin.
[0091] In the model, we generally believe that feature IV less than 0.02 means that the feature has almost no predictive ability; when IV is in the interval [0.02, 0.1), we believe that the feature predictive ability is weak; when IV is in the interval [0.1, 0.3), we believe that the feature predictive ability is medium; when IV is in the interval [0.3, 0.5), we believe that the feature predictive ability is strong; when IV is greater than or equal to 0.5, we believe that the feature predictive ability is too strong and we need to confirm whether the feature logic is normal.
[0092] In one embodiment, the tobacco placement prediction method further includes: during the training of the Time-LLaMA model, fine-tuning the Time-LLaMA model by using a preset low-order adaptive LoRA fine-tuning model.
[0093] The LORA algorithm helps the model to adjust the learning speed appropriately according to the error rate at different training stages by adjusting the learning rate, avoiding the instability caused by too high a learning rate or the slow convergence caused by too low a learning rate. This is especially important when processing complex data, which helps to speed up the convergence of the model and improve stability.
[0094] In one embodiment, the tobacco delivery prediction method further includes: smoothing the prediction results and eliminating abnormal values and extreme values of the prediction results. According to the optimized prediction results, a refined tobacco product delivery strategy is formulated, and the market response strategy and inventory management can be dynamically adjusted.
[0095] Reference Figure 4 , showing a schematic diagram of a model prediction process provided by the present invention. The first information is obtained, and then the first information is preprocessed such as data cleaning to obtain the second information. Then, feature extraction is performed on the second information, and the extracted features are input into the Time-LLaMA model through a decoder to pre-train the Time-LLaMA model. During the training process, the Time-LLaMA model is fine-tuned through LoRA. Finally, the trained and adjusted model is put into use. The trained and adjusted model can be used by tobacco manufacturers to predict tobacco production or the amount of tobacco put on the market; it can also be used by tobacco sellers to predict tobacco purchases.
[0096] Reference Figure 5 , showing another model prediction process diagram provided by the present invention. Acquire data from a data source, clean and process the data, train the model based on the cleaned and processed data, and finally predict the release volume of non-best-selling tobacco products and the release volume of new tobacco products through the trained model.
[0097] Reference Figure 6 , showing a schematic diagram of a prediction effect provided by the present invention. Compared with the historical sales cycle, the sales cycle of merchants has been improved, and the sales cycle has been shortened by about one week.
[0098] Reference Figure 7 , showing another prediction effect diagram provided by the present invention. The sales cycle of merchants near scenic spots has been greatly shortened compared with the historical situation. According to the survey data, more than 20% of the merchants have improved their sales cycle compared with the historical situation, and the sales cycle has been shortened by about one week. It can be analyzed that the combination of business district location and other characteristics can effectively improve.
[0099] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.
[0100] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A tobacco delivery prediction method based on a time series large language model, characterized in that: include: Acquire first information; the first information includes tobacco information in the tobacco system, tobacco transaction information in the transaction software, and location information in the map software; Extracting features from the first information to obtain numerical features, classification features, and static features; Inputting the numerical features, categorical features and static features into a time series-based large language Time-LLaMA model to train the Time-LLaMA model; The tobacco delivery situation is predicted using the trained Time-LLaMA model.
2. The tobacco placement prediction method based on time series large language model according to claim 1 is characterized in that: The tobacco information in the tobacco system includes: basic information of the target merchant, tobacco product specification information, tobacco order information and violation information of the target merchant; the tobacco transaction information in the transaction software includes: basic information of the target merchant and transaction information of the target merchant; the location information in the map software includes: location information of each merchant in a preset area around the target merchant; the feature extraction of the target information to obtain numerical features, classification features and static features includes: Feature extraction is performed on the first information to obtain target features; the target features include: business district features, business features and standard features; wherein the business district features include: the first average monthly customer volume, the tier where the average monthly transaction number is located and the tier where the average monthly traffic volume is located of each merchant in a preset area around the target merchant; the business features include: the ratio of the second average monthly customer volume of the target merchant to the first average monthly customer volume, the tier where the monthly customer traffic volume of the target merchant is located, the tier where the monthly transaction number is located, the tier where the monthly transaction amount is located, the average monthly transaction number, the tier where the average monthly traffic volume of customers of each age group is located, the average monthly transaction amount of each customer, the order volume, the number of order product types and the order amount; the standard features include: the amount of illegal commodities and the types of illegal commodities; According to the target features, numerical features, classification features and static features are generated.
3. The tobacco placement prediction method based on time series large language model according to claim 2 is characterized in that: The extracting features from the first information to obtain target features includes: Classifying the first information by labels; the labels include: descriptive labels, fact labels, product specification labels, and violation labels; Feature extraction is performed on the first information after label classification to obtain target features.
4. The tobacco placement prediction method based on time series large language model according to claim 1 is characterized in that: The extracting features from the first information includes: Preprocessing the first information to obtain second information; the preprocessing includes: abnormal information processing, missing information processing, repeated information processing, information enhancement processing and information normalization processing; Perform feature extraction on the second information.
5. The tobacco placement prediction method based on time series large language model according to claim 4 is characterized in that: The preprocessing of the first information to obtain the second information includes: When detecting that duplicate data exists in the first information, determining whether the duplicate data has business significance; If the duplicate data has business significance, retain the duplicate data; If the duplicate data has no business significance, the duplicate data are integrated and then retained, or one of the duplicate data is selectively retained.
6. The tobacco placement prediction method based on time series large language model according to claim 1 is characterized in that: The method further comprises: During the training of the Time-LLaMA model, the Time-LLaMA model is fine-tuned by a preset low-order adaptive LoRA fine-tuning model.
7. The tobacco placement prediction method based on time series large language model according to claim 1 is characterized in that: The method further comprises: The prediction results are smoothed, and outliers and extreme values of the prediction results are eliminated.