An AI-based hot-selling product trend discovery and intelligent restocking recommendation method

Abstract

The application discloses a hot-selling product trend discovery and intelligent inventory recommendation method based on AI, which comprises the following steps: collecting and preprocessing multi-source sales related data; multi-dimensionally extracting trend signals, calculating signal confidence and deviation intensity by adopting a grouping confrontation arbitration mechanism, quantifying arbitration coefficients to determine the hot-selling trend; matching a supply chain product library according to the trend identification result and evaluating; performing demand prediction by adopting a component decomposition model, decomposing the inventory quantity into a basic inventory quantity and a delivery coverage quantity, synthesizing the recommended inventory quantity by calculating the predicted demand in the replenishment period and the predicted demand in the supplier delivery period; converting the inventory scheme into a purchase execution instruction, collecting actual sales and prediction deviation, and updating the model in an incremental learning mode. The application solves the problem of multi-dimensional signal conflict arbitration, overcomes the defects of traditional inventory models that are not suitable for non-stationary trend demand, forms a full-link closed-loop feedback, and effectively improves the hot-selling product identification accuracy and the scientific nature of inventory decision-making.

Description

Technical Field

[0001] This invention belongs to the field of e-commerce and artificial intelligence technology, specifically relating to an AI-based method for discovering hot-selling product trends and intelligent inventory recommendation. Background Technology

[0002] With the rapid development of global e-commerce, cross-border e-commerce has become an important form of international trade, with platforms such as Shopify, TikTok Shop, Amazon, and Temu providing sellers with diversified sales channels. However, sellers still face core pain points in actual operation, such as a lack of foresight in product selection, a disconnect between identifying best-selling trends and making inventory preparation decisions, and a lack of scientific basis for inventory quantity.

[0003] Existing technologies already include patents related to intelligent product selection and inventory management. Chinese patent application CN120543081A discloses a "method for intelligent product selection and dynamic inventory management for e-commerce platforms," ​​which collects internal platform transaction data and external environmental data (such as order flow, user behavior, competitor prices, social media trending terms, etc.), generates a standardized feature matrix through cross-modal feature fusion, and makes joint decisions on product selection and inventory management based on this matrix. It also calculates dynamic inventory weights and short-term sales forecasts by combining product lifecycle stages, generating a priority queue for warehouse allocation. Chinese patent CN121563398B discloses an "AI-based sales forecasting and intelligent replenishment method and system," which constructs and optimizes a sales forecasting model using historical sales and product characteristics, calculates the confidence interval boundary of the forecast value and a dynamic safety stock benchmark, and determines replenishment demand by adding compensation when the basic gap is lower than the safety stock, and initiates parallel replenishment from multiple suppliers based on supplier delay coefficients.

[0004] The aforementioned existing technologies still have the following shortcomings:

[0005] First, there is a lack of arbitration mechanisms for conflicting multi-dimensional trend signals. In trend identification, signals from different dimensions, such as sales trends, social media buzz, and competitive landscape, may have different sources, but they all point to the same business judgment question—whether the product is about to become a bestseller. When the signals from different dimensions contradict each other (such as sales growth while social media buzz drops sharply), existing technologies either simply use weighted averaging, which dilutes strong signals and makes it impossible to make a clear judgment; or they process each dimension independently and then integrate them ex-post at the decision-making level, lacking a unified arbitration framework.

[0006] Secondly, existing inventory models are not suitable for the non-stationary demand characteristics of trending commodities. The safety stock framework (confidence interval boundaries and dynamic safety stock benchmark) adopted by CN121563398B is essentially based on the statistical stationarity assumption of demand fluctuations, which is suitable for the periodic replenishment of mature commodities. However, the demand for trending commodities in an upward channel shows accelerated growth, and models based on the stationarity assumption will seriously underestimate the inventory level, causing systemic shortages.

[0007] Third, product selection decisions and inventory preparation are disconnected in existing technologies. CN120543081A only addresses the joint decision-making process between product selection and inventory, without involving post-decision tracking and feedback optimization. While CN121563398B involves replenishment decisions, its feedback is limited to adjusting the parameters of the replenishment model itself, without feeding the execution results back to the trend identification stage. Overall, there is a lack of a fully automated closed loop from trend discovery, supply chain matching, inventory calculation to execution feedback, resulting in a long decision implementation cycle. Summary of the Invention

[0008] This invention provides an AI-based method for discovering hot-selling product trends and intelligent inventory recommendation. It extracts trend signals from multiple dimensions, adopts a group adversarial arbitration mechanism, quantifies arbitration coefficients to determine hot-selling trends, matches and evaluates supply chain product libraries based on trend recognition results, uses a component decomposition model for demand forecasting, transforms inventory plans into procurement execution instructions, collects the deviation between actual sales and forecasts, and updates the model using incremental learning to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for discovering best-selling product trends and intelligent inventory recommendation based on AI, comprising the following steps:

[0010] S1: Collect multi-source data related to product sales from at least one data source and perform preprocessing;

[0011] S2: Extract trend judgment signals from multiple dimensions, arbitrate conflicts between signals from different dimensions, identify the trend of best-selling products, and output the trend identification results;

[0012] S3: Based on the trend identification results, match the corresponding supply resources in the supply chain product database and evaluate the matched suppliers;

[0013] S4: For products identified as having a hot-selling trend, calculate the inventory quantity based on demand forecast results and generate an inventory recommendation plan;

[0014] S5: Convert the recommended inventory preparation plan into a procurement execution instruction, collect execution result data, and perform incremental learning and update the model based on the deviation between the execution result and the prediction result.

[0015] As a further option, in step S2, the multiple dimensions include at least two of the following dimensions: the sales trend dimension reflecting the dynamic changes in the sales volume of the product itself; the demand heat dimension reflecting consumers' active attention and purchase intentions; the social dissemination dimension reflecting the product's spontaneous dissemination ability on social platforms; and the competitive landscape dimension reflecting the market competition situation of the product category. The arbitration of conflicts between the signals of each dimension specifically includes: dividing the signals of each dimension into positive signal groups and negative signal groups according to their signal direction; and calculating the comprehensive evidence strength of the positive signal group and the negative signal group respectively.

[0016] The arbitration coefficient is calculated based on the comprehensive evidence strength, and the arbitration coefficient is calculated as R = S1 / (S1+S2), where S1 is the comprehensive evidence strength of the positive signal group and S2 is the comprehensive evidence strength of the negative signal group; the trend identification result is output based on the comparison result of the arbitration coefficient and the preset threshold.

[0017] As a further option, the comprehensive evidence strength is calculated based on the confidence level and deviation strength of the signals in each dimension; the confidence level is determined by the product of the historical contribution factor and the current anomaly modulation factor, the historical contribution factor reflects the statistical contribution of the dimension to the prediction result in historical data, and the current anomaly modulation factor is dynamically adjusted according to the degree of deviation of the current signal of the dimension from the historical mean.

[0018] As a further option, the historical contribution factor is automatically extracted from the machine learning model, specifically by obtaining the proportion of the number of times the feature corresponding to this dimension was used to split nodes in all decision trees of the machine learning model out of the total number of splits.

[0019] As a further option, the demand forecasting described in step S4 uses a component decomposition model that decomposes the time series into horizontal, trend, and seasonal components.

[0020] As a further option, the calculation of inventory quantity in step S4 specifically includes: decomposing the inventory requirement into two independent components: basic inventory quantity and delivery coverage quantity. Recommended inventory quantity = D f (T) × α + D f (D lead ) × β, where D f (T) represents the projected demand during the replenishment period T, α represents the safety stock factor, and D f (D lead (D) represents the supplier's delivery cycle. leadThe basic inventory level is determined based on the predicted demand within the replenishment cycle and the safety stock factor; the delivery coverage level is determined based on the predicted demand within the supplier's delivery cycle and the delivery coverage factor; the basic inventory level and the delivery coverage level are added together to obtain the recommended inventory level.

[0021] As a further option, the safety stock coefficient is negatively correlated with the supplier fulfillment rate, and the delivery coverage coefficient is negatively correlated with the stability of the supplier delivery cycle.

[0022] As a further option, the incremental learning update in step S5 specifically includes: extracting new training samples containing real results from the recent data at a preset period, merging them with the historical training set, fine-tuning the existing model parameters, and verifying the performance of the updated model on the validation set. If the performance degrades, the model is automatically rolled back to the previous version.

[0023] As a further option, the method is deployed in a client-server architecture, with the server side including a hot-selling trend identification engine, an intelligent matching engine, an inventory recommendation engine, and an execution feedback engine. The output of the hot-selling trend identification engine drives the intelligent matching engine, the output of the intelligent matching engine drives the inventory recommendation engine, and the output of the inventory recommendation engine drives the execution feedback engine. The execution feedback engine sends the deviation signal between the execution result and the prediction result back to the hot-selling trend identification engine and the inventory recommendation engine, forming a complete closed loop.

[0024] As a further option, in step S4, when there are multiple products identified as having a hot-selling trend, the system calculates a comprehensive priority based on the hot-selling probability score, expected profit per unit, and urgency of inventory preparation for each product, and generates an inventory preparation queue in descending order of the comprehensive priority.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. A group-based adversarial signal arbitration mechanism is adopted. After grouping multi-dimensional trend signals by direction, the comprehensive evidence strength is calculated and the arbitration coefficient is quantified. This effectively solves the problem that strong signals are diluted due to simple weighted averaging when signals conflict, and improves the accuracy and robustness of hot-selling trend identification.

[0027] 2. The confidence level calculation integrates historical contribution factors and current anomaly modulation factors, enabling the signal weights of each dimension to adaptively adjust with market fluctuations. This preserves the statistical reliability of historical experience while enhancing the sensitivity to sudden market signals.

[0028] 3. The inventory preparation model decomposes demand into two independent components: basic inventory and delivery coverage. These components address the uncertainty of demand-side growth trends and the uncertainty of supply-side delivery cycles, respectively. Compared with the traditional safety stock method based on the assumption of stability, this model is more in line with the non-stable growth characteristics of trending commodities and effectively reduces the risk of stockouts.

[0029] 4. Construct a fully automated closed loop of "trend identification → supply chain matching → inventory recommendation → execution feedback". Execution deviation signals are synchronously transmitted back to the trend identification and inventory recommendation modules. Through incremental learning, the model is continuously iterated and evolved to achieve adaptive improvement in decision quality.

[0030] 5. In multi-product scenarios, a priority queue is generated by combining the probability of best-selling products, unit profit and the urgency of inventory preparation. The safety factor is dynamically adjusted in combination with the supplier fulfillment rate and delivery stability to ensure that limited resources are tilted towards high-value and high-urgency products, thereby improving overall operational efficiency. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0032] Figure 2 This is a schematic diagram of the workflow of the feature conflict arbitration module in step S2 of the present invention;

[0033] Figure 3 This is a flowchart illustrating the process of optimizing the inventory quantity calculation in step S4 of the present invention.

[0034] Figure 4 This is a schematic diagram of the system module deployment and data flow architecture of the present invention;

[0035] Figure 5 This is a schematic diagram of the feedback and incremental learning closed-loop process of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1: As Figure 1 , Figure 4As shown, this invention discloses an AI-based method for discovering best-selling product trends and intelligent inventory recommendation. The method is deployed in a client-server architecture. The server side includes a best-selling trend recognition engine (including a feature conflict arbitration module), an intelligent matching engine, an inventory recommendation engine, and an execution feedback engine, which interface with e-commerce platforms and supply chain management systems via standardized APIs.

[0038] The data flow between the engines forms a complete closed loop: the output of the hot-selling trend recognition engine (the list of trending products and their hot-selling probability, and the expected burst time window) drives the intelligent matching engine; the output of the matching engine (related SKUs, supplier information, and cost data) drives the inventory recommendation engine; the output of the recommendation engine (inventory quantity and purchase suggestion) drives the execution feedback engine; and the execution feedback engine sends the deviation signal between actual sales and forecasts back to the trend recognition engine and the inventory recommendation engine, triggering incremental learning.

[0039] Step S1: Multi-source data acquisition and preprocessing

[0040] The system collects two types of data through e-commerce platform APIs: first, order data, reflecting actual purchasing behavior and forming the basis for sales trend analysis; second, product performance data (page views, unique visitors, and add-to-cart counts), reflecting potential purchasing intentions and enabling the early detection of demand signals before sales show significant growth. It also collects topic content quantity and interaction rates through social media platform interfaces, reflecting the degree of spontaneous user dissemination and penetration of products. Search popularity indices are collected through keyword popularity tools. Finally, the system collects product database data and supplier historical fulfillment data through supply chain system APIs, providing supply-side constraints for subsequent matching and inventory preparation.

[0041] The collection frequency and timeliness of the above-mentioned data types vary (orders are collected daily, searches weekly, and social media data in real-time). After collection, they need to be aligned using a daily time axis. Test orders and cancelled orders must be removed, outliers must be detected and removed using statistical methods, and missing values ​​must be appropriately imputed. After preprocessing, all data is stored in partitions according to a unified format.

[0042] Step S2: Identifying Hot-Selling Trends and Arbitrating Conflicts

[0043] We extract trend signals with commercial significance from raw data from multiple dimensions and solve the problem of comprehensive judgment when these signals conflict at the conclusion level.

[0044] The system extracts trend judgment signals from the following four dimensions. These four dimensions describe the likelihood of a product becoming a best-seller from different perspectives, and they are complementary and independent of each other:

[0045] Sales trend signals reflect the dynamic changes in a product's sales volume. The calculation method involves calculating the month-on-month growth rate of sales over multiple sliding time windows, and simultaneously using univariate linear regression to fit the slope of sales volume changes over time. The implication of this signal is that when the sales growth rate is consistently positive and the slope value increases significantly, it indicates that the product is in a period of rising demand.

[0046] Demand intensity signals reflect consumers' proactive attention to products and their purchase intentions. They are comprehensively evaluated based on the synchronized trends of three indicators: product page views, unique visitors, and the number of items added to cart. The implication of this signal is that consumers typically browse first, then add to cart, and finally place an order. Changes in page views and add-to-cart rates usually appear before actual sales figures, allowing for the perception of accumulating potential demand even before sales data shows significant changes.

[0047] Social media dissemination signals reflect a product's ability to spread spontaneously on social media platforms. This involves collecting data on the growth rate of content related to product keywords and user interaction rates. The implication is that in the current cross-border e-commerce environment, user-generated content on social media platforms has an increasingly significant impact on consumer decisions. When a large number of users spontaneously create content and interact around a particular product, that product is likely on the verge of a surge in popularity.

[0048] Competitive landscape signals reflect the market competition situation of a product category. They are assessed based on the number of active sellers and the concentration of price distribution within the same category. The implication of this signal is that when there are few competing products and the price range is stable, new entrants face less market resistance and have a longer profit window after initial success.

[0049] There may be conflicts among these four dimensions—for example, rapid sales growth but a sharp drop in social media engagement, or strong demand but a saturated competitive landscape. When a conflict occurs, the feature conflict arbitration module of this invention intervenes to resolve it.

[0050] like Figure 2 As shown, the feature conflict arbitration module adopts a "group adversarial" architecture. Since all dimensions point to the same question ("Will this product become a best-selling item?"), the dimensions should first be divided into two groups according to their signal direction—a positive signal group supporting "becoming a best-selling item" and a negative signal group opposing "becoming a best-selling item". Then, the two groups are allowed to adversarially challenge each other based on their respective strength of evidence, and an arbitration coefficient is calculated to quantify which side's evidence is more sufficient. This architecture ensures that strong signals are not diluted by weak signals—because the signals are not fused into a fuzzy average, but compared in a group adversarial manner.

[0051] In order to make the group confrontation based on evidence, the system first needs to quantify the "confidence" of each signal dimension—that is, how much say should the signal be given in the current decision-making.

[0052] For each dimension d, its confidence level C d Determined by the product of two factors:

[0053] C d = W d × M d

[0054] W d (Historical Contribution Factor) This factor reflects the statistical contribution of this dimension to sales forecasting in historical data. This factor is not assigned by human experience, but is automatically learned from the XGBoost model—after model training, it represents the proportion of times each feature is used for splitting nodes in all decision trees out of the total number of splits. Specifically, let f be the total number of times feature d is used for splitting in all decision trees. d The total number of splits for all features is F = Σf d Then W d = f d / F, and W of all features d The sum equals 1. This method is used because it reflects the actual influence of the feature on the prediction results across a large number of historical samples, rather than human judgment.

[0055] M d (Current Anomaly Modulation Factor) enables the weights of dimensions to dynamically change according to the anomaly level of the current signal. Its calculation method is as follows:

[0056] First, calculate the standardized deviation score of the current signal:

[0057] z d = (s d - μ d ) / σ d

[0058] Where s d μ is the current signal value. d and σ d These represent the mean and standard deviation of the signal over a past period, respectively. d A positive value indicates a signal stronger than normal, while a negative value indicates a signal weaker than normal.

[0059] Then z is used through the sigmoid function d Mapped to modulation factor:

[0060] M d = 1 + 0.5 × (2 / (1 + e^{-k ×z d}) - 1)

[0061] Where k is an adjustment parameter (ranging from 0.5 to 2.0), controlling the modulation sensitivity, and can be determined through validation set optimization. The properties of this function are:

[0062] When z d When = 0 (signal is at a historical normal level), M d = 1, the basic weight remains unchanged;

[0063] When z d When M > 0 and increases, d When the value approaches 1.5, the weight is amplified (the signal receives more attention when it is exceptionally strong).

[0064] When z d When M < 0 and decreases, d When the value approaches 0.5, the weight is compressed (the influence decreases when the signal is abnormally weak).

[0065] When a signal dimension shows an unusually strong change, regardless of its historical contribution, it indicates that the dimension is currently undergoing a noteworthy change and should therefore be given greater influence. Conversely, if a signal is negative but only shows a slight decrease within the normal fluctuation range, its negative impact should not be exaggerated.

[0066] Confidence level C for each dimension d Once confirmed, the system initiates the group-based adversarial process:

[0067] The first step is to categorize each dimension into either a positive or negative signal group based on its signal direction. The categorization is determined by whether the current signal is stronger than its historical average: if s d >μ d If s d ≤μ d It is classified into the negative signal group.

[0068] The second step is to calculate the weighted combined evidence strength for the positive signal group and the negative signal group separately:

[0069] S1 = Σ(C d × | z d |) / Σ(C d )

[0070] S2= Σ(C d × | z d |) / Σ(C d )

[0071] Among them, the confidence levels C of each dimension within the group of molecules d Multiply by deviation intensity | z d The summation is then performed, and the denominator is the total confidence level of the group. Using |z dThe logic, rather than simply using the positive or negative sign of the signal, lies in: | z d | Reflects the degree of deviation from the historical mean. The greater the deviation, the stronger the evidence provided by the signal. A signal with a slight deviation should not have the same weight of evidence as a signal with a strong deviation.

[0072] The third step is to calculate the arbitration coefficient:

[0073] R = S1 / (S1 + S2)

[0074] The R-value is between 0 and 1, which intuitively means the percentage of evidence supporting the claim that the product "became a best-selling item" out of all valid evidence.

[0075] The fourth step is to make a three-tiered decision based on the R value:

[0076] When R > 0.6, output "Trend established";

[0077] When R < 0.4, output "Trend not established";

[0078] When 0.4 ≤ R ≤ 0.6, output "Signal conflict to be observed", marking it as requiring manual review.

[0079] The aforementioned thresholds can be determined through validation set optimization, and different combinations of thresholds can be used for different business scenarios. This three-level output mechanism differs from the existing approach of outputting a single fusion result. Its design intention is to avoid giving an unreliable conclusion when signal conflicts are severe and the system lacks confidence in its judgment, but rather to clearly state "to be observed" to avoid misleading the user.

[0080] Above the feature conflict arbitration module, the system uses XGBoost as the base classification model. The model outputs a best-selling probability score P. hot and the expected outbreak time window T w Provides input for subsequent modules.

[0081] The XGBoost model is trained by constructing a four-dimensional signal feature vector for each day and each product in the historical data, with the label indicating whether the product will actually become a best-selling item within a specified future time window. FocalLoss is used as the loss function, enabling the model to focus on learning boundary samples even in cases of extreme imbalance between positive and negative samples (the proportion of best-selling items is typically less than 5%). The model is fine-tuned through time-series cross-validation, and inference is performed daily on all active products at set intervals.

[0082] Step S3: Supply Chain Resource Matching

[0083] Product Matching

[0084] Trending products are represented as feature vectors, including category codes, semantic codes for core attribute text, price ranges, and other dimensions. A corresponding feature vector is also extracted for each product in the supply chain product library. Cosine similarity is used to calculate the matching degree.

[0085] Similarity = (V t · V s ) / (||V t || × || V s ||)

[0086] Where V t and V s These are the normalized feature vectors for trending products and products in the product catalog, respectively. "·" represents the vector dot product, and "||||" represents the L2 norm. The closer the cosine similarity value is to 1, the more similar the two products are in the feature space. A successful match associates the supplier's SKU, procurement cost, and inventory status; a failed match generates a product search request, which is then handled by the supply chain operations team.

[0087] Supplier evaluation

[0088] Supplier evaluation comprehensively considers three dimensions: fulfillment rate (R) f (Number of orders shipped on time / Total number of orders), Delivery cycle score S3 (based on average delivery cycle D) lead The calculation includes a shorter delivery cycle (higher score) and a quality score of S4 (1 - after-sales problem rate). The overall score is S... supplier = w1×R f + w2×S3 + w3×S4, where w1, w2, and w3 are weighting coefficients (configurable, with a balanced distribution across dimensions by default). The supplier with the highest overall score is recommended as the first choice, and its delivery cycle D lead and fulfillment rate R f This will be used as an input parameter for calculating subsequent inventory levels.

[0089] Step S4: Smart Inventory Recommendation

[0090] For trending goods that have been identified as being in an upward trend, their demand is showing signs of accelerated growth. Traditional inventory models that are suitable for the maturity stage are no longer applicable, and specialized calculation methods are required.

[0091] Demand Forecast

[0092] Demand forecasting is performed using the Holt-Winters three-parameter exponential smoothing model. This model decomposes the time series into three components: the level term L... t (Benchmark value), Trend term T t (Direction of monotonic change), seasonal term S t (Periodic fluctuations).

[0093] The prediction formula is:

[0094] Horizontal item update: L t = α × (Y t - S {t-m} ) + (1-α) × (L {t-1} + T {t-1} )

[0095] Trend update: T t = β × (L t - L {t-1} ) + (1-β) ×T {t-1}

[0096] Seasonal update: S t = γ × (Y t - L t ) + (1-γ) ×S {t-m}

[0097] Predicted value: F {t+k} = L t + k ×T t + S {t-m+k}

[0098] Where α, β, and γ are smoothing parameters (range 0~1), controlling the update speed of the horizontal, trend, and seasonal terms, respectively; m is the seasonal period (7 days, in weeks); Y t F represents the actual sales volume at time t; {t+k} This is the predicted sales volume at time t+k. The reason for choosing this model is that it does not require the time series to meet the stationarity assumption, can handle the trend growth and periodic fluctuations common in trending commodities, and has high computational efficiency and strong interpretability.

[0099] The model is initialized as follows: when sufficient historical data is available for the product, the initial value L0 for the horizontal term is the average sales volume of the previous two weeks; the initial value T0 for the trend term is the slope of the univariate linear regression of sales volume of the previous two weeks; and the initial value S0 for the seasonal term is the average proportion of sales volume on the same day of each week that deviates from the mean. The values ​​of the three smoothing parameters are determined by minimizing the mean squared error on the training set.

[0100] When a product lacks sufficient historical data, cold start prediction is performed using transfer learning based on similar product categories. This involves mining products from historical sales data that are most similar to the new product in terms of category, price range, target customer group, etc., and adjusting them by adding a category seasonality coefficient based on their life cycle sales curve.

[0101] The confidence interval for the predicted value is calculated using the standard deviation of the historical prediction errors:

[0102] Upper Boundary D u = F {t+k} + λ × σ e

[0103] Lower Boundary D l = F {t+k} - λ ×σ e

[0104] Where σ e λ is the standard deviation of the historical forecast residuals, and λ is the confidence coefficient (corresponding to the required confidence level).

[0105] like Figure 3 As shown, the inventory quantity calculation of the present invention adopts a dual-component independent control structure.

[0106] The traditional safety stock formula SS = Z × σ × √(L) is based on the assumptions of stable demand and a normal distribution of fluctuations. Its derivation presupposes that the randomness of demand changes is statistically stable. The dynamic safety stock benchmark method used in CN121563398B essentially still belongs to this framework—determining safety stock by calculating the confidence interval boundaries of forecast values. However, goods identified as trending commodities mean that their demand does not meet this premise—it is in an accelerating upward trend, and the growth in demand is not random but a deterministic trend. Therefore, the inventory level model needs a fundamental reconstruction.

[0107] This invention decomposes inventory requirements into two components with different properties:

[0108] Recommended stock level = Basic stock level + Delivery coverage

[0109] Basic inventory level = D f (T) × α

[0110] Where D f (T) represents the projected demand within the replenishment cycle T (which can be configured by the seller according to their own replenishment schedule). This component addresses the question of "how much sales can be met within the normal replenishment cycle" and is the main component of inventory preparation. α is a safety stock coefficient greater than 1, and its value is negatively correlated with the supplier fulfillment rate—the lower the fulfillment rate, the larger the required safety buffer. The intention behind this coefficient is that when the supplier's historical fulfillment rate is low, trending goods may face greater supply uncertainty during the replenishment period, thus requiring a larger safety buffer to cope with situations where demand growth exceeds expectations.

[0111] Delivery coverage = D f (D lead ) × β

[0112] Where Df (D lead (D) represents the supplier's delivery cycle. lead The projected demand is calculated based on the amount of goods or services sold. This component addresses the question of "how much will be sold while waiting for replenishment to arrive." β is the delivery coverage coefficient (0 < β ≤ 1), and its value is negatively correlated with the stability of the supplier's delivery cycle—the more stable the delivery cycle, the lower the coverage ratio can be; the greater the fluctuation in the delivery cycle, the more full coverage is required. The intention behind this coefficient is that when the supplier's delivery cycle is unstable, it is impossible to accurately predict the replenishment arrival time, and a more conservative coverage strategy (β approaching 1) must be used to prevent stockouts caused by delayed arrivals.

[0113] These two components correspond to two different types of risk: α addresses the uncertainty of demand-side growth trends (how fast will the commodity growth rate be?); β addresses the uncertainty of supply-side delivery (when will the replenishment arrive?). Controlling these two risks independently is more precise than using a single coefficient for a general approach. This is fundamentally different from the traditional safety stock formula, which packages all uncertainties into a single standard deviation σ.

[0114] Verification and adjustment of inventory preparation plan

[0115] The system calculates the out-of-stock risk rate:

[0116] R risk = D u / Q r

[0117] Where D u To predict the upper bound of the confidence interval, Q r Recommended stock level. If R risk If the stock level exceeds a preset threshold (e.g., 1.2), a stockout warning will be triggered, and it will be recommended to increase the stock level. Simultaneously, if the total purchase amount of the recommended stock level exceeds the seller's available budget, the system will calculate the maximum stock level under budget constraints and indicate the shortfall, ensuring the plan is feasible under actual financial conditions.

[0118] When multiple trending products exist simultaneously, the system calculates a comprehensive priority based on the probability of best-selling, expected profit per unit, and urgency of inventory replenishment.

[0119] Priority=w hot ×P hot + w profit × P n + w u × U n

[0120] Where P hot To score the probability of it becoming a bestseller, w hot P represents the weighting coefficient corresponding to the probability of a best-selling product.n w is the normalized value of expected profit per unit. profit The weighting coefficient corresponding to the unit expected profit; U n The normalized value for the urgency of stockpiling (calculated based on the expected outbreak time window; the closer the outbreak, the higher the urgency), w u A weighting coefficient is assigned to the urgency of inventory preparation. This weighting coefficient can be configured according to the seller's business strategy. Inventory preparation queues are generated by sorting items by Priority from highest to lowest.

[0121] like Figure 5 As shown, step S5: Solution execution and feedback optimization

[0122] The inventory preparation decision is translated into actual procurement actions, and the execution results are fed back in a closed loop to drive continuous model optimization.

[0123] Execution trace

[0124] After the seller confirms the inventory preparation plan, the system pushes the purchase order to the supplier through the supply chain API, tracking four key nodes: supplier confirmation, shipment, logistics pickup, and warehousing completion, ensuring full visibility of the procurement execution process.

[0125] Bias-driven incremental learning

[0126] After the goods have been in stock and sold normally for a period of time, the system calculates the deviation rate between the actual sales volume and the predicted sales volume:

[0127] E s = |D a - D f | / D f

[0128] And the deviation between the actual outbreak time and the predicted outbreak time. These deviation data are the most valuable feedback signals—they reflect the gap between model predictions and reality, indicating the direction in which the model needs to be optimized.

[0129] The system performs incremental learning at preset intervals: it extracts new training samples containing real results from recent data, merges them with the historical training set, and then fine-tunes the model parameters based on the previous version. Fine-tuning uses a smaller learning rate and fewer iterations to absorb new knowledge while avoiding forgetting learned general patterns. The updated model is then validated on recent data; if its performance drops significantly, it is automatically rolled back.

[0130] Through this mechanism, the model achieves an adaptive evolution of "prediction → decision-making → execution → feedback → re-prediction", and the prediction accuracy continues to improve with the continuous accumulation of data.

[0131] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for discovering best-selling product trends and intelligent inventory recommendation based on AI, characterized in that, Includes the following steps: S1: Collect multi-source data related to product sales from at least one data source and perform preprocessing; S2: Extract trend judgment signals from multiple dimensions, arbitrate conflicts between signals from different dimensions, identify the trend of best-selling products, and output the trend identification results; S3: Based on the trend identification results, match the corresponding supply resources in the supply chain product database and evaluate the matched suppliers; S4: For products identified as having a hot-selling trend, calculate the inventory quantity based on demand forecast results and generate an inventory recommendation plan; S5: Convert the recommended inventory preparation plan into a procurement execution instruction, collect execution result data, and perform incremental learning and update the model based on the deviation between the execution result and the prediction result.

2. The method for discovering hot-selling product trends and intelligent inventory recommendation based on AI according to claim 1, characterized in that, In step S2, the multiple dimensions include at least two of the following dimensions: the sales trend dimension reflecting the dynamic changes in the sales volume of the product itself; the demand heat dimension reflecting consumers' active attention and purchase intentions; the social dissemination dimension reflecting the product's spontaneous dissemination ability on social platforms; and the competitive landscape dimension reflecting the market competition situation of the product category. The arbitration of conflicts between the signals of each dimension specifically includes: dividing the signals of each dimension into positive signal groups and negative signal groups according to their signal direction; and calculating the comprehensive evidence strength of the positive signal group and the negative signal group respectively. The arbitration coefficient is calculated based on the comprehensive evidence strength, and the arbitration coefficient is calculated as R = S1 / (S1+S2), where S1 is the comprehensive evidence strength of the positive signal group and S2 is the comprehensive evidence strength of the negative signal group; the trend identification result is output based on the comparison result of the arbitration coefficient and the preset threshold.

3. The method for discovering best-selling product trends and intelligent inventory recommendation based on AI according to claim 2, characterized in that, The overall evidence strength is calculated based on the confidence level and deviation strength of the signals in each dimension; the confidence level is determined by the product of the historical contribution factor and the current anomaly modulation factor, the historical contribution factor reflects the statistical contribution of the dimension to the prediction result in historical data, and the current anomaly modulation factor is dynamically adjusted according to the degree of deviation of the current signal of the dimension from the historical mean.

4. The method for discovering best-selling product trends and intelligent inventory recommendation based on AI according to claim 3, characterized in that, The historical contribution factor is automatically extracted from the machine learning model. Specifically, it is obtained that the feature corresponding to this dimension is used to split nodes in all decision trees of the machine learning model out of the total number of splits.

5. The method for discovering hot-selling product trends and intelligent inventory recommendation based on AI according to claim 1, characterized in that, The demand forecasting described in step S4 uses a component decomposition model that decomposes the time series into horizontal, trend, and seasonal components.

6. The method for discovering best-selling product trends and intelligent inventory recommendation based on AI according to claim 5, characterized in that, Step S4, specifically calculating the inventory requirement, involves decomposing the inventory demand into two independent components: the basic inventory requirement and the delivery coverage requirement. The recommended inventory requirement is calculated as follows: D f (T) × α + D f (D lead ) × β, where D f (T) represents the projected demand during the replenishment period T, α represents the safety stock factor, and D f (D lead (D) represents the supplier's delivery cycle. lead The basic inventory level is determined based on the predicted demand within the replenishment cycle and the safety stock factor; the delivery coverage level is determined based on the predicted demand within the supplier's delivery cycle and the delivery coverage factor; the basic inventory level and the delivery coverage level are added together to obtain the recommended inventory level.

7. The method for discovering hot-selling product trends and intelligent inventory recommendation based on AI according to claim 6, characterized in that, The safety stock coefficient is negatively correlated with the supplier fulfillment rate, and the delivery coverage coefficient is negatively correlated with the stability of the supplier's delivery cycle.

8. The method for discovering hot-selling product trends and intelligent inventory recommendation based on AI according to claim 1, characterized in that, The incremental learning update in step S5 specifically includes: extracting new training samples containing real results from the recent data at a preset period, merging them with the historical training set, fine-tuning the existing model parameters, and verifying the performance of the updated model on the validation set. If the performance degrades, the model is automatically rolled back to the previous version.

9. A method for discovering best-selling product trends and intelligent inventory recommendation based on AI according to any one of claims 1 to 8, characterized in that, The method is deployed in a client-server architecture, with the server side including a hot-selling trend recognition engine, an intelligent matching engine, an inventory recommendation engine, and an execution feedback engine; The output of the hot-selling trend recognition engine drives the intelligent matching engine, the output of the intelligent matching engine drives the inventory recommendation engine, the output of the inventory recommendation engine drives the execution feedback engine, and the execution feedback engine sends the deviation signal between the execution result and the prediction result back to the hot-selling trend recognition engine and the inventory recommendation engine, forming a complete closed loop.

10. The method for discovering hot-selling product trends and intelligent inventory recommendation based on AI according to claim 1, characterized in that, In step S4, when there are multiple products identified as having a hot-selling trend, the system calculates a comprehensive priority based on the hot-selling probability score, expected unit profit, and urgency of inventory preparation for each product, and generates an inventory preparation queue in descending order of the comprehensive priority.

Images