Commodity sales volume prediction method and system based on artificial intelligence

Abstract

The invention provides a commodity sales volume prediction method and system based on artificial intelligence, and relates to the field of data processing and intelligent prediction.The method comprises the steps that hot topic data are obtained, keyword tags in the hot topic data are extracted to be matched with commodity class tags of an e-commerce platform, and a hot associated commodity list is generated; and dynamic modeling of a user interest conversion path is combined, a user behavior sequence mode is analyzed, and a commodity popularity value of a real purchase intention is generated. The method comprises the following steps: carrying out time sequence sampling on popularity values, setting differential popularity threshold values based on commodity category features, dynamically identifying potential hot commodities, and calculating a stepped sales prediction value in future time in combination with a historical popularity value change curve and a sales conversion rate. In this way, the real-time performance and accuracy of sales volume prediction can be improved, the e-commerce platform can find potential hot commodities in time, inventory management is optimized, operation efficiency is improved, and prediction requirements in a complex and changeable market environment are met.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing and intelligent prediction, and more specifically, to a method and system for predicting commodity sales volume based on artificial intelligence. Background Art

[0002] With the rapid development of e-commerce and the increasing maturity of its business models, commodity sales forecasting technology has gradually become an important tool for e-commerce platforms to optimize operations and improve efficiency. Traditional commodity sales forecasting methods mainly rely on historical sales data for regression analysis or time series modeling, such as using techniques like ARIMA models and exponential smoothing methods. These methods perform well when the commodity life cycle is long and the environment changes relatively smoothly. However, with the explosive growth of commodity categories on e-commerce platforms and the intensification of market competition, relying solely on historical data for forecasting can no longer meet the actual needs. In the modern e-commerce environment, the changes in commodity popularity and sales volume are often affected by multi-dimensional factors such as real-time market dynamics, diverse user behaviors, and the spread of hot topics. The complexity and timeliness of these factors pose higher requirements for existing forecasting technologies. At the same time, the rise of external data sources such as social media and short video platforms has made the driving effect of hot topics on commodity sales volume more obvious, and the purchasing behaviors of users have gradually shown short-term volatility and non-linear trends, which further increases the difficulty of sales volume forecasting.

[0003] However, there are many deficiencies in the current related technologies when dealing with these complex scenarios. On the one hand, traditional forecasting methods generally rely on static historical data and are less sensitive to real-time dynamic data, making it difficult to quickly respond to changes in market hotspots. On the other hand, existing methods usually adopt single-dimensional static analysis in user behavior modeling and fail to fully explore the deep logic behind the conversion path of user interests and the behavior sequence pattern. In addition, existing technologies lack support for commodity category characteristics and personalized weights in the modeling of commodity popularity, resulting in deviations in the heat evaluation and prediction accuracy of different category commodities; and the identification of potential hot-selling commodities also mostly stays at simple threshold judgments and fails to conduct dynamic analysis in combination with the change characteristics of time series. More critically, the current mainstream sales forecasting methods often ignore the association between hot topics and commodity categories and fail to effectively use the keyword tags in hot topics to drive the identification of potential hot-selling commodities and sales volume forecasting. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention is proposed. The present invention provides a method and system for predicting commodity sales volume based on artificial intelligence.

[0005] According to one aspect of the present invention, there is provided a method for predicting commodity sales volume based on artificial intelligence, which includes: Obtain hot topic data, extract keyword tags from the hot topic data, match them with the commodity category tags of the e-commerce platform, and generate a list of hot-related commodities; Based on the list of hot-related commodities, identify the user interest conversion path, analyze the user's behavior sequence pattern, adjust the behavior index weights of commodities in different categories, and generate the commodity heat value of real purchase intention; Perform time series sampling on the commodity heat value, record the heat value change curve in hours, set different heat thresholds based on the commodity category characteristics, and mark the commodities exceeding the corresponding thresholds as potential hot-selling commodities; According to the historical heat value change curve and historical sales conversion rate of the potential hot-selling commodities, calculate the stepped sales prediction value for future time in combination with the current heat value change.

[0006] Furthermore, calculate the transition probability between different user behavior patterns based on the behavior sequence pattern, and generate the commodity heat value of real purchase intention; The calculation of the transition probability uses an improved Markov chain calculation method to calculate the transition probability between states. The improved method introduces a time decay factor, and through the time decay factor, it ensures that the recent transition data has higher reference value.

[0007] Furthermore, the extraction of the keyword tags is through a three-layer relationship traversal of the hot topic data; During the first-layer traversal, visit the synonym nodes directly connected to the core keyword; During the second-layer traversal, expand to the hyponyms connected to the first-layer nodes; During the third-layer traversal, expand outward to reach the more macroscopic hypernym nodes; During the traversal process, record the access path and cumulative weight of each related word in real time to ensure that when the same word is accessed multiple times through different paths, the optimal semantic association path and corresponding weight can be selected.

[0008] Furthermore, obtain the commodity category system from the e-commerce platform, split the category names into tags by level, and extract the key attributes in the commodity titles and descriptions as supplementary tags to obtain the commodity category tags; Based on the commodity category tags and the keyword tags, perform hierarchical matching, and sort all the successfully matched commodities according to the association strength; Within each association strength level, further consider the real-time performance indicators of the commodities, and obtain the comprehensive score of each commodity through weighted calculation, and preferentially select the commodities with higher scores to enter the association list; And by analyzing the commodity purchase association relationships in the historical sales data, identify the commodity combinations that are often purchased together to form a complete list of hot commodity associations.

[0009] Further, based on the commodity popularity value of the true purchase intention, potential hot-selling commodities are identified through time series analysis; The time series analysis identifies commodities with rapidly rising popularity by calculating the change speed of popularity; at the same time, the change acceleration of popularity is calculated to predict whether the upward trend of popularity will continue; And a category-differentiated judgment criterion is constructed; If the current popularity of a commodity exceeds the category standard, the popularity shows a stable upward trend, and the popularity fluctuation is within a reasonable range, a time series prediction model is used to predict the popularity performance in a future period of time to discover possible hot-selling commodities in advance.

[0010] Further, the time series prediction model takes into account that the commodity popularity is affected by both historical data and real-time influencing factors. Starting from time series analysis, a basic popularity function is constructed to reflect the influence of historical data, and an exponential decay term is introduced to adjust the basic popularity in the time dimension; Moreover, since the commodity popularity shows periodic fluctuations and is affected by various external factors, an adjustment function is constructed to describe the periodic changes, and an influence factor combination term integrates the influences of various external factors. The result is standardized through a normalization coefficient to obtain a complete popularity prediction model.

[0011] Further, the historical sales conversion rate is obtained by establishing a mapping relationship between the historical popularity change curve and the actual sales volume and analyzing the existing typical patterns; When it is observed that the popularity of a potential hot-selling commodity begins to rise, its popularity change trajectory is captured in real time, and pattern matching is performed with the typical patterns to find the closest historical trajectory pattern and judge the conversion rate change rule.

[0012] Further, the pattern matching adopts the dynamic time warping algorithm to calculate the similarity between the current trajectory and the historical pattern; The Euclidean distance between each time point is calculated through the dynamic time warping algorithm to construct an accumulated distance matrix. The minimum accumulated distance from the starting point to each point is calculated through dynamic programming. Starting from the lower right corner of the matrix, the optimal matching path is found. Based on the accumulated distance on the optimal matching path, the initial similarity score is calculated to find the closest historical trajectory pattern.

[0013] According to another aspect of the present invention, a commodity sales volume prediction system based on artificial intelligence is provided, which includes: A generation module for obtaining hot topic data, extracting keyword tags from the hot topic data, matching them with the commodity category tags of the e-commerce platform, and generating a list of hot topic-related commodities; An adjustment module, based on the hot-spot associated product list, identifies the user interest conversion path, analyzes the user's behavior sequence pattern, adjusts the behavior index weights of products in different categories, and generates the product heat value of the true purchase intention. An analysis module performs time series sampling on the product heat value, records the heat value change curve in hours, sets different heat thresholds based on product category characteristics, and marks the products exceeding the corresponding thresholds as potential hot-selling products. A prediction module calculates the stepped sales prediction value for future time according to the historical heat value change curve and historical sales conversion rate of the potential hot-selling products, combined with the current heat value change.

[0014] Compared with the prior art, the present invention accurately predicts the future sales performance of products by capturing the real-time heat change of products, matching the historical similar trajectories, and establishing a phased prediction model. The dynamic time warping algorithm is used for trajectory matching, combined with the sliding time window mechanism to process real-time data, and accurate prediction is achieved through the phased prediction strategy and prediction correction mechanism. It significantly improves the accuracy and real-time performance of product sales prediction, helps enterprises grasp market opportunities in a timely manner, optimize inventory management decisions, and improve operation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings: Figure 1 FIG. is a flowchart of a method for predicting product sales based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments of the present invention. It should be understood that the present invention is not limited by the exemplary embodiments described herein.

[0017] As mentioned in the above background art, there are mainly three prominent problems in the prior art: First, the traditional sales volume prediction method overly relies on static historical data and lacks responsiveness to real-time market dynamics and hot topics, making it difficult to quickly respond to market changes. Second, a single-dimensional static analysis is adopted in user behavior modeling, failing to fully explore the conversion path of user interests and the pattern of behavior sequences. Third, there is a lack of differential treatment for the characteristics of product categories, resulting in deviations in the heat evaluation and prediction accuracy of products in different categories. In response to these technical pain points, our invention proposes a product sales volume prediction method based on artificial intelligence. Through the correlation analysis between hot topic data and product categories, combined with the dynamic modeling of the conversion path of user interests, potential hot-selling products are identified based on time series sampling and differential thresholds, achieving more accurate sales volume prediction. This method effectively solves the problems of insufficient utilization of hot data, weak ability to model user behavior sequences, and insufficient adaptability of product category characteristics in the prior art.

[0018] Figure 1 FIG. is a flowchart of a product sales volume prediction method based on artificial intelligence according to an embodiment of the present invention. As Figure 1 shown, in a product sales volume prediction method based on artificial intelligence, it includes: S1: Obtain hot topic data from the interface of the social media platform, extract the keyword tags in the hot topic data, and match them with the product category tags of the e-commerce platform to generate a list of hot-related products.

[0019] Real-time obtain hot topic data through the open API interface of the social media platform, specifically including the top 50 topics on the Sina Weibo hot search list, topics on the Douyin hot list with a popularity of over 1 million within 24 hours, the title content of popular notes on Xiaohongshu within the past 48 hours, and the video titles and descriptions on the Bilibili soaring list of the day, etc. If the obtained topic content contains sensitive words or illegal content, it will be automatically filtered. And monitor the quantity and update frequency of topic acquisition in real time to ensure continuous data update.

[0020] Perform Chinese word segmentation on the obtained hot topic data, automatically remove stop words such as "de", "le", "a", etc., and focus on extracting keywords containing effective parts of speech such as nouns, adjectives, and brand words. When the part of speech of a certain word is not clear, its part of speech is judged through context analysis. If the number of extracted keywords exceeds the preset upper limit, screening is performed according to word frequency and part of speech importance. The extracted keywords will be counted for word frequency, and the top 10 words in terms of word frequency are selected as core keywords, and at the same time, a relationship traversal with a depth of 3 is performed on these keywords.

[0021] For example, when a core keyword like "thermos cup" is detected, starting from this node, explore related entries layer by layer along the predefined semantic relationship path. In the first-level traversal, visit the synonym nodes directly connected to the core keyword, such as "thermal flask" and "thermal water cup", which have the strongest semantic association with the original keyword and are given the highest weight value of 1.0.

[0022] In the second-level traversal, expand to the hyponyms connected to the first-level nodes, such as "stainless steel thermos cup" and "vacuum thermos cup". Although these words still maintain a strong semantic connection with the core keyword, due to the presence of an intermediate node interval, their weight values will be appropriately reduced.

[0023] In the third-level traversal, continue to expand outward to reach more macroscopic hypernym nodes, such as "water cup" and "drinking utensils", which have the farthest semantic distance from the core keyword and reflect more generalized concepts, so they are given smaller weight values.

[0024] Through the hierarchical progressive traversal method, comprehensively capture the semantic association network between words, and reflect the strength of semantic association through the gradual change of weights. During the traversal process, record the access path and cumulative weight of each related word in real time to ensure that when the same word is accessed multiple times through different paths, the optimal semantic association path and corresponding weight can be selected.

[0025] At the same time, obtain the standardized product category system from the e-commerce platform, such as the category link "Home Life -> Dining Utensils -> Thermos Cup". Split the category name into tags layer by layer, and extract the key attributes in the product title and description as supplementary tags to establish a many-to-many association relationship between products and tags. When the product category changes, automatically update the related tag relationship. If the product is taken off the shelf or becomes invalid, the method will automatically remove the related tag association to ensure the real-time accuracy of the tag system.

[0026] More specifically, after obtaining the standardized commodity category system, the category link "home life->dining utensils->thermos cup" is first split into levels. The method designs the commodity table, category table, label table and the association table between them in the relational database. Each commodity record is simultaneously associated with all hierarchical nodes in its complete category path. For example, a thermos cup commodity is not only associated with the subcategory label "thermos cup", but also with the parent category labels "dining utensils" and "home life". At the same time, the key attribute features are extracted from the commodity title and description text through natural language processing technology, such as descriptive words of dimensions such as material, capacity, and function, and these attribute features are converted into standardized attribute labels. In this way, a commodity has a label set of two dimensions, namely, category label and attribute label, forming a one-to-many association relationship. The same label can also be used by multiple different commodities, thus building a many-to-many association network between commodities and labels.

[0027] In order to ensure the dynamic update of label relationships, when the product category changes, the update process will be triggered: first, the product is associated with all labels on the original category path, and the historical records are saved for traceability; then the product is associated with the labels on the new category path, including the new parent category label; the title and description text of the product are re-parsed to identify whether there are new attribute features, and the attribute label set is updated accordingly. If the changed category has specific attribute requirements, the necessary attribute labels are automatically added. In this process, a label usage frequency statistic is maintained. When the number of associated products of a label is lower than the threshold, it is marked as pending cleanup, and the operator will decide whether to keep it after confirmation.

[0028] When the status of a product changes, different processing strategies are adopted. For products that are temporarily removed from the shelves, their tag association is retained but marked as inactive, so that the product can be quickly restored when it is put back on the shelves; for products that are permanently removed from the shelves or deleted, their tag association is completely removed, but a historical record is retained in the database for data analysis and optimization of the tag system. The validity of tags is checked regularly, and tags that have not been used for a long time or have too few associated products are cleaned up to ensure that the tag system is streamlined and efficient.

[0029] Furthermore, the hot keyword tags are matched with the product tags for similarity. When the similarity exceeds the preset threshold, the product is marked as a highly relevant product. If the correlation between a hot topic and a product is particularly strong, the correlation relationship is processed first. During the matching process, if it is found that certain keywords are prone to mismatching, they are added to the blacklist and the similarity algorithm is continuously optimized.

[0030] During the matching process, a hierarchical matching strategy is adopted. In the first layer, fast matching is carried out, mainly targeting the core tags of products and the main keywords of hot topics. Through the inverted index, a set of products that may be relevant is quickly screened out. In the second layer, precise matching is carried out, and a detailed similarity calculation is performed on the preliminary screening results, considering tag features in more dimensions. In the third layer, the association strength is evaluated. By comprehensively considering factors such as the similarity score, the popularity value of the hot topic, and the historical matching success rate of the product, the final association strength is calculated.

[0031] All successfully matched products are sorted by association strength into three priority levels: strong association, medium association, and weak association. Within each association strength level, the real-time performance indicators of the products are further considered, including dimensions such as recent sales volume, inventory status, profit margin, and user ratings. Through weighted calculation, the comprehensive score of each product is obtained. Products with higher scores are preferentially selected to enter the association list, while ensuring that the category distribution of the products in the list is relatively balanced, avoiding the situation of excessive concentration of a single category.

[0032] By analyzing the product purchase association relationships in historical sales data, product combinations that are often purchased together are identified. When a certain hot topic is associated with multiple highly complementary products at the same time, these products are used as a combined recommendation plan to provide a better user selection space. For example, when the hot topic is highly related to thermos cups, not only the thermos cups themselves will be recommended, but also supporting products such as cleaning brushes and cup sleeves will be associated and recommended to form a complete list of associated hot products.

[0033] S2: Based on the list of hot topic associated products, identify the user interest conversion path, analyze the user's behavior sequence pattern, adjust the behavior index weights of products in different categories, and generate the popularity value of products with real purchase intention.

[0034] Based on the obtained list of hot topic associated products, a comprehensive user behavior data collection and analysis process is initiated. First, through a real-time data interface, the user interaction data of target products on the e-commerce platform is continuously collected, including multi-dimensional behavior data such as the number of visits to the product detail page, the duration of stay, the depth of page browsing, favorite operations, and shopping cart addition records. These raw data are organized according to the user ID and timestamp to construct a user-product interaction event stream. Different initial weights are set for each type of behavior data to reflect the purchase intention intensity of different behaviors.

[0035] More specifically, first, a time-series-based user behavior data storage structure is established. Each user behavior record contains attributes such as user ID, product ID, behavior type, behavior timestamp, behavior duration, and behavior depth. These original behavior data are sorted according to user ID and timestamp to form a complete behavior sequence for each user on each product. By analyzing these sequences, all possible transfer paths between six basic states (initial browsing, in-depth browsing, collection, adding to cart, generating an order, and completing payment) of users are identified. At the same time, a time interval attribute is attached to each transfer path to record the time required for state transfer and relevant behavior characteristics during the transfer process, such as browsing depth and interaction times.

[0036] At the same time, analyze the user behavior data in the past 90 days, segment and count according to different time windows. Within each time window, calculate the transfer frequency matrix between states. The transfer frequency matrix records the number of times users transfer from one state to another within a specific time window. By normalizing the transfer frequency matrix, a standard transfer probability matrix is obtained.

[0037] An improved Markov chain calculation method is used to calculate the transfer probability between states. The traditional Markov chain only considers the direct transfer probability between states. The improved method introduces a time decay factor. The time decay factor ensures that recent transfer data has higher reference value.

[0038] By analyzing a large amount of historical data, calculate the transfer probability from different states to completing payment. The formula is expressed as: ; Among them, is the historical conversion rate of each state, , is the corresponding state weight, is the time interval from the occurrence of user behavior to the current time, is the time decay coefficient, represents the total number of conversion states, represents the initial conversion probability from the current state to completing payment.

[0039] Based on the user's transfer probability, combine it with the product association list to calculate the product purchase intention heat value, specifically including: For each product, calculate the initial popularity value. The initial popularity value first considers the basic popularity of the product itself. This part of the data comes from the generated list of associated hot products, which includes information such as the basic traffic data and historical sales situation of the product. Combine the purchase intentions of all users who have taken actions on this product. The purchase intention of each user is reflected by the calculated transfer probability, and different weights are assigned to different users. For example, the weight of a paid member is higher than that of an ordinary user, and the weight of a user with strong historical purchasing power is higher than that of a user with weak purchasing power. Combine these two parts of data through a weight coefficient to obtain the popularity value of the product with real purchase intention. The configuration of the weight needs to be determined according to the actual operation data.

[0040] S3: Perform time series sampling on the product popularity value, record the popularity value change curve in hours, set different popularity thresholds based on the characteristics of the product category, and mark the products that exceed the corresponding thresholds as potential hot-selling products.

[0041] After obtaining the real-time popularity value of the product with purchase intention, identify potential hot-selling products through time series analysis. Specifically include: Taking hours as the basic unit, continuously record the popularity value of each product. At the same time, the basic data collection will be updated in real time to ensure that the latest popularity changes can be obtained.

[0042] Based on the collected time series data, first calculate the change speed of the popularity, that is, the change amount of the popularity value per hour. Through the change speed, identify the products with rapidly rising popularity. At the same time, calculate the change acceleration of the popularity, that is, the change situation of the change speed, which is used to predict whether the rising trend of the popularity will continue. Combining these two dynamic indicators can more accurately judge the development trend of the product popularity.

[0043] Further calculate statistical characteristic indicators. Include the average popularity value in the past 24 hours, which reflects the basic popularity level of the product; the standard deviation of the popularity value, which represents the degree of popularity fluctuation; and the peak situation of the popularity, etc. These statistical indicators complement the calculated dynamic indicator data and jointly constitute a complete popularity characteristic system. For example, a product may currently have a very fast change speed of popularity, but if its basic popularity level is too low, or the popularity fluctuation is too large, then this rising trend may not be reliable enough.

[0044] Furthermore, construct a category-differentiated judgment standard. Products in different categories have different hot-selling characteristics, and corresponding judgment standards need to be set according to the characteristics of each category. This standard includes three dimensions: basic popularity requirements, popularity change requirements, and popularity stability requirements. For the fast-moving consumer goods category, pay attention to the rapid rise of popularity; for the durable goods category, value the persistence of popularity. And these standards will also be dynamically adjusted according to time characteristics (such as time periods, weeks, holidays, etc.).

[0045] Compare the popularity characteristics of the products with the standards of the corresponding categories. Products that meet the following conditions simultaneously will be marked as potential bestsellers: the current popularity exceeds the category standard, the popularity shows a stable upward trend, and the popularity fluctuation is within a reasonable range. For these products, use a time series prediction model to predict the popularity performance in the future period and discover possible best-selling products in advance; The time series prediction model takes into account that the product popularity is affected by both historical data and real-time influencing factors. Therefore, starting from time series analysis, a basic popularity function is constructed to reflect the influence of historical data. On this basis, considering the characteristic that popularity will naturally decay over time, an exponential decay term is introduced to adjust the basic popularity in the time dimension. Since the product popularity usually shows periodic fluctuations and is affected by various external factors, an adjustment function is further constructed, which includes a trigonometric function term to describe the periodic changes and an influencing factor combination term to integrate the influence of various external factors. To ensure the comparability and practicality of the prediction results, the results are finally standardized through a normalization coefficient, and the complete popularity prediction formula is obtained, expressed as: ; ; ; where, is the predicted popularity value at time is the basic popularity function, is the time decay coefficient, is the interval from the current time, is the popularity adjustment function, is the normalization coefficient, is the popularity value at the th historical time point, is the time weight of the th historical point, n is the number of historical data sampling points, is the periodic fluctuation intensity, is the period length, is the phase shift, is the adjustment factor weight, is the intensity of the th influencing factor, is the adjustment coefficient of the

[0046] Specifically, , including marketing activity factors, inventory level factors, competitor activity factors, holiday factors, category seasonality factors, and price change factors.

[0047] Regularly update the category judgment criteria and retrain the prediction model every week. At the same time, continuously record the comparison between the prediction results and the actual sales situation, evaluate the effectiveness of the method through indicators such as accuracy and recall rate, and continuously optimize the model parameters and judgment criteria based on this data.

[0048] S4: According to the historical heat value change curve and historical sales conversion rate of potential hot-selling products, calculate the stepped sales prediction value for future time in combination with the current heat value.

[0049] Based on the predicted hot-selling products, further predict the product heat and sales conversion. The core is to establish the mapping relationship between the historical heat change curve and the actual sales volume. First, construct a heat trajectory archive of products. The archive contains the heat change trajectories of a large number of successfully hot-selling products in history. For each hot-selling product, record its complete heat change curve, including specific data points in the rising period, peak period, and decline period. The heat trajectories are normalized and converted into standard time series data. Each heat trajectory will be marked with key features, such as the rising rate, duration, fluctuation amplitude, periodic characteristics, etc. At the same time, collect the actual sales volume data of these products at different heat stages and establish a heat-sales correspondence database.

[0050] When analyzing historical data, it is found that the heat change of products usually shows some typical patterns. For example, the heat curve of quickly explosive products shows a steep rising section and a relatively rapid falling section; while continuously hot-selling products show a slow rise and then maintain a plateau at a relatively high level. By extracting and classifying these typical patterns, a heat trajectory pattern library is established. Each pattern corresponds to different sales conversion characteristics. For example, the conversion rate of quickly explosive products is usually high during the rising period of heat but has a short duration; while continuously hot-selling products have a relatively stable but slightly lower conversion rate. This pattern information is used to optimize the accuracy of sales prediction.

[0051] When it is observed that the heat of a potential hot-selling product begins to rise, capture its heat change trajectory in real time and match it with the patterns in the historical trajectory archive. The dynamic time warping algorithm is used in the matching process to calculate the similarity between the current trajectory and the historical patterns, and find the several most similar historical trajectory patterns. The historical conversion data of these similar trajectories will be used as an important reference for prediction. For example, if the heat trajectory of the current product is highly similar to the early trajectory of a certain historical bestseller, pay special attention to the conversion rate change rule of this historical bestseller in the similar stage.

[0052] More specifically, the dynamic time warping algorithm is used to calculate the similarity between the current trajectory and the historical patterns, including: Obtain the time series data of the observed product's popularity and match it with the data in the historical trajectory library. By using the dynamic time warping algorithm, construct a distance matrix of size n×m, where n is the number of time points in the current trajectory and m is the number of time points in the historical trajectory. Calculate the Euclidean distance between each pair of time points and fill the distance matrix D. Then construct the cumulative distance matrix C and calculate the minimum cumulative distance from the starting point to each point through dynamic programming. Start backtracking from the lower right corner of the matrix to find the optimal matching path, which represents the best correspondence between the two trajectories. Based on the cumulative distance on the optimal matching path, calculate the initial similarity score and find the several most similar historical trajectory patterns. The historical conversion data of these similar trajectories will be used as an important reference for prediction. For example, if the popularity trajectory of the current product is highly similar to the early trajectory of a certain historical bestseller, pay special attention to the conversion rate change rule of the historical bestseller in the similar stage.

[0053] Furthermore, the judgment of the optimal matching path includes: If the coordinates of the current position in the matrix C are (i, j), it is necessary to judge the previous positions based on the principle of dynamic programming. When starting the judgment, three possible previous positions will be considered simultaneously: the left position (i, j - 1), the lower position (i - 1, j), and the lower left diagonal position (i - 1, j - 1).

[0054] If you want to judge the optimal previous position, first calculate the transfer cost from each possible previous position to the current position. When considering the left position, check whether C(i, j) is equal to C(i, j - 1) plus the corresponding distance value; when considering the lower position, verify whether C(i, j) is equal to C(i - 1, j) plus the corresponding distance value; if considering the diagonal position, then it is necessary to confirm whether C(i, j) is equal to C(i - 1, j - 1) plus the corresponding distance value. Among these three equations, the position that satisfies the condition is the possible previous position.

[0055] In the actual execution process, if it is found that multiple previous positions satisfy the conditions, the position with the minimum cumulative distance will be preferentially selected as the previous node of the path. When encountering the situation where the cumulative distances are equal, the diagonal movement will be preferentially selected. Diagonal movement usually provides a more natural time correspondence. If the diagonal direction cannot be selected, according to the preset priority rules, horizontal or vertical movement will be selected.

[0056] When the judgment and selection of a position are completed, the selected previous position will be immediately used as the new current position, and the same judgment process will be repeated at the new position. During this iterative process, a path from the lower right corner to the upper left corner is continuously constructed. When the path reaches the upper left corner of the matrix, it indicates that the backtracking process of the optimal matching path is completed. Each position on this path represents the optimal matching relationship between the corresponding points in the two time series.

[0057] After finding the several most similar historical trajectory patterns, sales forecasting is carried out. First, calculate the heat change rate within the recent time window to determine the heat stage of the product. If the heat change rate is positive and the acceleration is positive, it is determined to be the rising period; if the heat change rate is close to zero and the acceleration is negative, it is determined to be the peak period; if the heat change rate is negative, it is determined to be the decline period.

[0058] After determining the heat stage, process the similar trajectory data. Align the time axis of each trajectory with the current product, extract the heat value and actual sales data of each trajectory in the corresponding time period to form a data sample set. If a certain trajectory lacks data at a specific time point, interpolate and supplement it with data from adjacent time points.

[0059] When calculating the conversion rate using the hierarchical processing method. First, calculate the conversion rate at each time point of each similar trajectory. After obtaining the conversion rates of all trajectories, calculate the weights according to the trajectory similarity. If the heat change patterns of two trajectories are more similar, a higher weight is assigned, and the initial conversion rate is obtained through weighted average.

[0060] In different heat stages, corresponding prediction strategies are adopted, including: If it is in the rising period, calculate the current heat growth rate, screen out the historical data with the same growth rate interval from the similar trajectories, extract the corresponding relationship between the heat value and sales volume in these data, establish the mapping relationship between the heat growth rate and the sales volume growth rate, and predict the sales volume change in the next time period according to the current heat growth rate.

[0061] When entering the peak period, extract the peak period characteristic parameters of the similar trajectories, calculate the average duration of the peak period, analyze the stability characteristics of the sales volume in the peak period, predict the duration according to the current sales volume level and stability characteristics, and combine the duration to predict the overall sales volume performance in the peak period If entering the decline period, calculate the current heat decline rate, extract the data with the same decline rate interval from the similar trajectories, analyze the sales volume decay law in the decline period, and predict the sales volume change trend in the subsequent time period Based on the prediction values in different periods, generate the prediction interval for the current product, specifically including: calculating the expected sales volume at the current conversion rate as the most likely value, extracting the conversion rate fluctuation range of the similar trajectories in the corresponding stage, determining the upper and lower limits of the prediction interval according to the fluctuation range, calculating the confidence level of the prediction interval, and generating a prediction report including the most likely value, the upper and lower limits of the interval, and the confidence level.

[0062] Continuously run the prediction, record the data of the prediction deviation, analyze the reasons for the deviation, optimize the adjustment strategy of the prediction parameters, and improve the prediction accuracy of the method.

[0063] In summary, the method for predicting commodity sales volume based on artificial intelligence according to the embodiments of the present invention is elucidated. By capturing the real-time heat changes of commodities, matching historical similar trajectories, and establishing a phased prediction model, the future sales performance of commodities can be accurately predicted. The dynamic time warping algorithm is used for trajectory matching, and the sliding time window mechanism is combined to process real-time data. Accurate prediction is achieved through the phased prediction strategy and the prediction correction mechanism. In this way, the accuracy and real-time performance of commodity sales volume prediction are significantly improved, helping enterprises to promptly seize market opportunities, optimize inventory management decisions, and enhance operational efficiency.

Claims

1. A commodity sales forecasting method based on artificial intelligence, characterized in that: include: Obtain hot topic data, extract keyword tags from the hot topic data, match them with product category tags on the e-commerce platform, and generate a list of hot topic-related products; Based on the hotspot-related product list, identify the user interest conversion path, analyze the user's behavior sequence pattern, adjust the behavior indicator weights of products in different categories, and generate the product heat value of real purchase intention; Perform time series sampling on the popularity value of the product, record the popularity value change curve in hours, set differentiated popularity thresholds based on product category characteristics, and mark products exceeding the corresponding thresholds as potential hot-selling products; According to the historical popularity value change curve and historical sales conversion rate of the potential hot-selling product, combined with the current popularity value change, the step-by-step sales forecast value for the future time is calculated.

2. The commodity sales forecasting method based on artificial intelligence according to claim 1 is characterized in that: Calculate the transition probability between different user behavior patterns based on the behavior sequence pattern, and generate the commodity heat value of the real purchase intention; The calculation of the transfer probability adopts an improved Markov chain calculation method to calculate the transfer probability between states. The improved method introduces a time decay factor, which ensures that recent transfer data has a higher reference value.

3. The commodity sales forecasting method based on artificial intelligence according to claim 1 is characterized in that: The keyword tags are extracted by traversing three layers of relationships on the hot topic data; When traversing the first level, synonym nodes directly connected to the core keywords are visited; When traversing the second level, it expands to the hyponyms connected to the nodes in the first level; When traversing the third level, it expands outward to reach more macroscopic hypernym nodes; During the traversal process, the access path and cumulative weight of each related word are recorded in real time to ensure that when the same word is accessed multiple times through different paths, the optimal semantic association path and corresponding weight can be selected.

4. The commodity sales forecasting method based on artificial intelligence according to claim 3 is characterized in that: Obtain a product category system from an e-commerce platform, split the category name into labels according to the hierarchy, and extract key attributes in the product title and description as supplementary labels to obtain the product category label; Perform hierarchical matching based on the product category tags and the keyword tags, and prioritize all successfully matched products according to their association strength; Within each level of association strength, the real-time performance indicators of the products are further considered, and the comprehensive score of each product is obtained through weighted calculation. Products with higher scores are given priority to enter the association list; By analyzing the purchase associations of goods in historical sales data, we can identify combinations of goods that are often purchased together and form a complete list of hot goods associations.

5. The commodity sales forecasting method based on artificial intelligence according to claim 2 is characterized in that: Based on the commodity popularity value of the real purchase intention, potential hot-selling commodities are identified through time series analysis; The time series analysis identifies products with rapidly increasing popularity by calculating the speed of change in popularity. At the same time, it calculates the acceleration of change in popularity to predict whether the increasing trend in popularity will continue. And build judgment criteria for category differentiation; If the current popularity of a product exceeds the category standard, the popularity shows a steady upward trend, and the fluctuation of popularity is within a reasonable range, the time series prediction model is used to predict the popularity performance in the future and discover possible hot-selling products in advance.

6. The commodity sales forecasting method based on artificial intelligence according to claim 5 is characterized in that: The time series prediction model takes into account that the popularity of a product is affected by both historical data and real-time factors. Starting from time series analysis, it constructs a basic popularity function to reflect the influence of historical data, and introduces an exponential decay term to adjust the basic popularity in the time dimension. Moreover, the popularity of commodities fluctuates periodically and is affected by a variety of external factors. Therefore, an adjustment function is constructed to describe the periodic changes, and an influencing factor combination item is used to integrate the influence of various external factors. The results are standardized through the normalization coefficient to obtain a complete popularity prediction model.

7. The commodity sales forecasting method based on artificial intelligence according to claim 6 is characterized in that: The historical sales conversion rate is achieved by establishing a mapping relationship between the historical popularity change curve and the actual sales volume, and analyzing the existing typical patterns; When you observe that the popularity of a potential hot-selling product begins to rise, capture its popularity change trajectory in real time, match it with the typical pattern, find the most similar historical trajectory pattern, and determine the conversion rate change pattern.

8. The commodity sales forecasting method based on artificial intelligence according to claim 7 is characterized in that: The pattern matching uses a dynamic time warping algorithm to calculate the similarity between the current trajectory and the historical pattern; The Euclidean distance between each time point is calculated through the dynamic time warping algorithm, and the cumulative distance matrix is ​​constructed. The minimum cumulative distance from the starting point to each point is calculated through dynamic programming. The optimal matching path is found by backtracking from the lower right corner of the matrix. Based on the cumulative distance on the optimal matching path, the initial similarity score is calculated to find the most similar historical trajectory pattern.

9. A commodity sales forecasting system based on artificial intelligence, characterized in that: include: A generation module is used to obtain hot topic data, extract keyword tags in the hot topic data, match them with product category tags on the e-commerce platform, and generate a list of hot topic-related products; An adjustment module, based on the hotspot-related product list, identifies the user interest conversion path, analyzes the user's behavior sequence pattern, adjusts the behavior indicator weights of products in different categories, and generates a product heat value of real purchase intention; The analysis module performs time series sampling on the popularity value of the product, records the popularity value change curve in hours, sets differentiated popularity thresholds based on product category characteristics, and marks products exceeding the corresponding thresholds as potential hot-selling products; The prediction module calculates the step-wise sales forecast value for the future time based on the historical popularity value change curve and historical sales conversion rate of the potential hot-selling product and the current popularity value change.

Images