An intelligent retail dynamic pricing system based on customer portrait

Abstract

This invention discloses a smart retail dynamic pricing system based on customer profiles, belonging to the field of smart retail information technology. The system includes a profile initialization module, a gating decision module, a detection execution module, a utilization execution module, a profile update module, and a threshold adaptation module. The profile initialization module acquires the customer's elasticity belief distribution and uncertainty measure. The gating decision module compares the dynamic threshold with the target price to trigger detection or utilization modes. In the detection mode, the detection price is selected by maximizing information gain, and the effective elasticity observation value is calculated by collecting response signals. In the utilization mode, the optimal profit price is generated based on the expected elasticity value. The profile update module updates the belief distribution with differentiated weights, and the threshold adaptation module determines the price based on sliding window quantiles. This system constructs price elasticity profiles and achieves personalized pricing even with limited historical customer data.

Description

Technical Field

[0001] This invention relates to the field of intelligent retail information technology, and more specifically, to an intelligent retail dynamic pricing system based on customer profiles. Background Technology

[0002] This invention relates to the field of dynamic pricing technology in smart retail. In retail scenarios, dynamic pricing systems generate personalized prices based on customers' sensitivity to changes in product prices. Existing systems typically rely on customers' historical transaction data to construct price sensitivity profiles, such as calculating RFM indicators based on historical purchase records or inferring customer preferences through collaborative filtering, and then inputting the profile features into a pricing model to output prices.

[0003] For first-time customers or those making anonymous visits (cold start), their historical interaction records are sparse or missing. In this situation, traditional profiling methods struggle to obtain effective individual price elasticity characteristics, and the pricing system degenerates into adopting a group averaging strategy or preset rules, thus limiting the degree to which the generated price matches the customer's individual price elasticity characteristics.

[0004] In addition, some systems use random price probing to explore customer demand curves. In this type of approach, the choice of trial price is not linked to the efficiency of information acquisition, the amount of flexible information that can be obtained from a single trial is limited, and the trial results are not solidified into customer profiles that can be transferred across products. Therefore, a smart retail dynamic pricing system based on customer profiles is proposed to address the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an intelligent retail dynamic pricing system based on customer profiles. The problems to be solved include: the difficulty in quickly obtaining the price elasticity characteristics of individual customers during cold starts in the prior art; the lack of adaptability of the threshold used by the pricing system to switch between the two states of information exploration and profit acquisition; the single dimension of customer response signals to price changes, which makes it difficult to fully reflect the strength of purchasing tendency; and the failure to distinguish between the observation data obtained from different interaction modes in profile updates.

[0006] To achieve the above objectives, the present invention provides an intelligent retail dynamic pricing system based on customer profiles, including a profile initialization module, a gate control decision module, a detection execution module, a utilization execution module, a profile update module, and a threshold adaptive module.

[0007] In response to a pricing request for a target customer, the profile initialization module obtains the target customer's current elasticity belief distribution and corresponding uncertainty measure. The elastic belief distribution is a Beta distribution, determined by its shape parameter. and Determined, the uncertainty measure Let be the variance of the Beta distribution. Using the Beta distribution to characterize resilient beliefs and measuring uncertainty with variance allows the system's perceived state of customer price sensitivity to be quantified.

[0008] For target customers who are appearing for the first time and have no historical elasticity belief distribution records, the profile initialization module is initialized using a pre-trained hierarchical Bayesian model: The target customer's subgroup is determined based on their static attributes, and the posterior distribution parameters corresponding to that subgroup are used as the initial values ​​for the elastic belief distribution. and Values, and calculate the initial uncertainty measure. The static attributes include at least one of age group, gender, and city of residence level.

[0009] Hierarchical Bayesian models transmit elasticity distribution information at the group level to individuals through subgroup posterior parameters, enabling cold-start customers to obtain an initial elasticity estimate based on the statistical regularity of similar groups during their first interaction, while also giving them a high initial uncertainty to trigger subsequent exploration.

[0010] The gating decision module measures the uncertainty. With a dynamic threshold When comparing, Greater than When the detection mode is triggered, Not greater than The exploitation mode is triggered at any time.

[0011] Furthermore, when triggering the probe mode, the system must also ensure that the number of probes for the target customer within the same customer session does not exceed a preset upper limit, or that the time interval between two probes is not less than a preset minimum interval. This constraint prevents the system from frequently performing price tests on the same customer within a short period of time, thus controlling the impact of probe behavior on user experience.

[0012] The threshold adaptive module maintains a sliding window to store the most recent... Uncertainty Measurement for Sub-Pricing Decisions The value is calculated within the sliding window each time a gating decision is made. The value of the first Percentile as the dynamic threshold , For the preset window length, This is the preset percentile value.

[0013] This mechanism enables the triggering frequency of the detection mode to be automatically adjusted according to the overall cognitive state of the system: when the proportion of customers with high uncertainty in the system is high, the threshold is automatically raised to encourage detection; as the profile accumulates and the overall uncertainty decreases, the threshold is automatically lowered to reduce unnecessary detection.

[0014] In probe mode, the probe execution module selects a probe price from a set of candidate prices. Show the product to the target customers. The candidate price set is based on the product's benchmark price. A discrete set of prices is generated centered on a preset non-zero floating ratio group, wherein the magnitude of the floating ratio is limited to a preset range.

[0015] Detection price The selection method is as follows: For each candidate price in the candidate price set, calculate the expected reduction in the elastic belief distribution entropy before and after implementing the candidate price based on the current elastic belief distribution, and select the candidate price that maximizes the expected reduction as the probe price. .

[0016] The expected reduction was calculated using Monte Carlo sampling: sampling from the current elastic belief distribution. For each elastic sample value, a pre-defined response probability model is used to calculate the probability of a purchase inclination towards the candidate price. Based on the probabilities corresponding to each elastic sample value, the marginal response probability and posterior distribution entropy are estimated. The expected reduction amount is then obtained using the marginal response probability and posterior distribution entropy. The preset sampling number is used. Monte Carlo sampling transforms the calculation of expected information gain into a statistical estimate based on multiple samplings of the elastic belief distribution, making the computational complexity of the information gain criterion controllable in real-time pricing scenarios.

[0017] The detection execution module collects data from the target customer at the detected price. The system includes multiple response signals. These signals include whether the page dwell time after price display exceeds the category reference threshold, whether product details are viewed, whether the product is added to the shopping cart, and whether the purchase is finally completed. Each response signal has an indicator value of 1 or 0, representing yes or no, respectively.

[0018] A response intensity score is obtained by linearly weighting and summing the indicated values ​​of each response signal with preset weighting coefficients. The preset weighting coefficients include , , , Each of these corresponds to a fixed value for a specific response signal. The multi-layered response signals cover multiple behavioral stages from price perception to final purchase, allowing the system to obtain partial information about customer resilience even when a purchase event has not occurred.

[0019] The detection execution module is based on the detection price. Compared with the benchmark price of goods The differences and the response intensity score Calculate effective elasticity observations The calculation method is as follows: ,in , , This serves as the baseline response strength. The calculation, based on the definition of price elasticity, uses the ratio of the relative change in response strength to the relative change in price as the effective elasticity observation.

[0020] In the utilization mode, the execution module generates a profit-optimal price based on the expected elasticity value of the elastic belief distribution and displays it to the target customer. The generation process is as follows: first, a base optimal price is calculated based on the expected elasticity value of the elastic belief distribution; then, a correction amount is subtracted from the base optimal price to obtain the final execution price. The correction amount is related to the uncertainty measure. The correction amount is proportional to the square root of the uncertainty, and increases with the square root of the uncertainty, so that the system automatically gives a more conservative price to customers whose cognition is not completely certain.

[0021] The profile update module uses the first effective sample weight in detection mode. In the utilization mode, the second effective sample weight is used. Perform a Bayesian update on the elastic belief distribution. Greater than Assigning higher update weights to detection modes allows high-information-density observation data to play a greater role in belief updates, which helps accelerate the convergence of resilient profiles.

[0022] Through the collaborative work of the above modules, the system dynamically switches pricing behavior between exploration mode and utilization mode, enabling the construction of elasticity profiles and personalized pricing to be completed synchronously in the same session, and allowing for the rapid acquisition of individual price elasticity characteristics even when customer historical data is scarce.

[0023] The technical effects and advantages of this invention are as follows: (1) To address the problem of acquiring individual elasticity characteristics of cold-start customers, the profile initialization module uses a pre-trained hierarchical Bayesian model to determine the subgroup to which the customer belongs based on static attributes such as age group, gender, and tier of the city where they reside. The posterior distribution parameters corresponding to this subgroup are then used as the initial elasticity belief distribution. and This value enables the system to have an initial basis for personalized pricing during the customer's first interaction, which helps to shorten the data accumulation cycle required for profile building.

[0024] (2) To address the adaptive problem of mode switching threshold, the threshold adaptation module maintains a sliding window to store the most recent Uncertainty Measurement for the Pricing Decision The value is calculated within the window each time a gating decision is made. The value of the first Percentile as a dynamic threshold This mechanism enables the triggering frequency of the detection mode to be automatically adjusted according to the overall cognitive state of the system, which helps to improve the system's adaptability in different operational phases.

[0025] (3) To address the issue of a single dimension in response signal acquisition, the detection execution module displays the detection price. Then, multi-level response signals are collected, including whether the page dwell time exceeds the category reference threshold, whether product details are viewed, whether the item is added to the cart, and whether the purchase is finally completed. The signal indication values ​​are then compared with preset weight coefficients. to The response intensity score is obtained by linear weighted summation. This improves the amount of flexible information that can be extracted in a single interaction.

[0026] (4) To address the issue of insufficient differentiation of information value in profile updates, the profile update module uses the first effective sample weight in the detection mode. In the utilization mode, the second effective sample weight is used. Perform a Bayesian update on the elastic belief distribution, and Greater than This allows observational data acquired from different models to receive weights that match the information density during belief updates, which helps accelerate the convergence of resilient profiles. Attached Figure Description

[0027] Figure 1 This is a system module framework diagram of the present invention. Detailed Implementation

[0028] 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.

[0029] Example 1 As attached Figure 1The system illustrates a customer profile-based intelligent retail dynamic pricing system. This system includes a profile initialization module, a gating decision module, a detection execution module, a utilization execution module, a profile update module, and a threshold adaptation module. These modules exchange parameters and statuses via data interfaces, collectively constructing price elasticity profiles for cold-start customers and making personalized pricing decisions.

[0030] The profile initialization module is used to respond to a pricing request for a target customer by obtaining the target customer's current elasticity belief distribution and corresponding uncertainty measure. The elastic belief distribution adopts a Beta distribution form, determined by its shape parameter. and This distribution characterizes the sensitivity of target customers to changes in product prices. Uncertainty measure. The variance of this Beta distribution is calculated as follows: .

[0031] For target customers appearing for the first time without any historical elasticity belief distribution records, the profile initialization module uses a pre-trained hierarchical Bayesian model for initialization. This hierarchical Bayesian model has a three-layer structure: the first layer models the price elasticity distribution of all customers, the second layer divides the customers into subgroups according to static attributes and learns the offset of each subgroup relative to the global distribution, and the third layer is the individual layer.

[0032] The hierarchical Bayesian model uses the static attributes and corresponding elasticity coefficients of all historical customers as training data and employs Gibbs sampling for parameter estimation. The elasticity coefficients of historical customers in the training data are estimated using the rate of price change and the rate of change in purchase quantity during their historical transactions. The estimation formula is as follows: ,in The change in price Original price This represents the change in purchase quantity. This is the original purchase quantity.

[0033] After training, the model outputs the posterior parameters of the Beta distribution corresponding to each subgroup, and then compares each subgroup with its corresponding parameters. and Stored in the subgroup mapping table.

[0034] During online initialization, the profile initialization module takes the static attributes of the target customer as input, searches for the matching subgroup in the subgroup mapping table, and obtains the corresponding... and The static attributes of target customers include all three of the following: age group, gender, and city of residence.

[0035] The image initialization module will Assigned value ,Will Assigned value This serves as the initial elasticity belief distribution for the target customer, and the initial uncertainty measure is calculated based on the variance calculation method described above. If the target customer already has a history, the profile initialization module directly reads the stored data. and Value and calculate the current uncertainty measure. .

[0036] The threshold adaptive module is used to maintain a sliding window queue of fixed length. queue length The preset window length is used. After each pricing decision is made, regardless of whether the decision occurs in probe mode or exploitation mode, the customer's uncertainty measure at the time of that decision is... The values ​​are all pushed into the queue. The tail; If pushed into the front queue Length reached Then delete the queue first. The first element at the head is then pushed in. The threshold adaptation module acquires the queue before the gating decision module performs the comparison operation. All currently stored in The values ​​are sorted in ascending order, and the index of the sorted values ​​is taken. The value corresponding to the percentile is used as the current dynamic threshold. , This is the preset percentile value.

[0037] The gating decision module is used to measure uncertainty. With dynamic threshold The module performs a comparison. It maintains a counter for the number of probes made by the target customer in the current session. and the time of the most recent probe. And calculate the current time. and time interval ,like If it does not exist The value can be infinity.

[0038] When both conditions are met Greater than , Less than the preset upper limit under this implementation method ,and Greater than or equal to the preset minimum interval under this implementation method At that time, the gating decision module triggers the detection mode and... Increase the value by 1, Updated to ; If any condition is not met, the exploitation mode is triggered.

[0039] A session is defined as the continuous interaction period of a target customer from the moment they identify themselves until they leave the store or exit the application. In offline scenarios, a session begins when the customer scans their membership code to enter the store and ends when the customer checks out or leaves the store, or when there is no continuous interaction for a preset duration. In online scenarios, a session begins when the customer launches the application or opens a webpage and ends when the customer closes the application or when there is no continuous operation for a preset duration. (The last sentence appears to be a repetition of the previous one and can be omitted.) and All reset.

[0040] The probe execution module is used to select a probe price from a set of candidate prices in probe mode. Showcase to target customers. The candidate price set is based on the product's benchmark price. A discrete set of prices is generated centered on a set of preset non-zero floating ratios. The floating ratios are a set of preset percentage values, and the range of each floating ratio is limited to a preset range.

[0041] Detection price The selection method is as follows: For each candidate price in the candidate price set, calculate the expected reduction in the elastic belief distribution entropy before and after implementing the candidate price based on the current elastic belief distribution, and select the candidate price that maximizes the expected reduction as the probe price. .

[0042] The expected reduction is calculated using a Monte Carlo sampling method, specifically: from the current elasticity belief distribution Beta( , The inverse transform sampling method is used to generate the data. One elastic sample value , This is the preset sampling number. For each elastic sample value... The probability of a user exhibiting a purchase inclination towards a particular candidate price is calculated using a pre-defined response probability model. ,in This represents the fluctuation percentage corresponding to the candidate price. and It is a constant pre-fitted based on all historical transaction data. and Before system deployment, historical transaction data was used to fit the data through logistic regression, with the price fluctuation ratio in historical transactions used as the fitting factor. And customer elasticity coefficient Features are defined by whether a purchase behavior occurs, and the historical customer elasticity coefficient is estimated by the rate of change in price and the rate of change in purchase quantity in the customer's historical transactions; after training... and As a fixed constant embedded system.

[0043] Calculate the marginal response probability .

[0044] Calculate the current prior distribution entropy ,in For Beta functions, This is the digamma function.

[0045] Regarding the response results Calculate the posterior distribution parameters for both the 1 and 0 values: when hour , ; when hour , , This is a preset scaling constant used to map elastic observations to a scale that matches the updates to the Beta distribution parameters.

[0046] For each elastic sample value Calculate the corresponding posterior distribution entropy according to the above formula. and Substitute into the calculation respectively and , and ; Then calculate For all The arithmetic mean, For all The arithmetic mean of the expected posterior entropy. The expected reduction corresponding to this candidate price is .

[0047] The probe execution module iterates through all candidate prices in the candidate price set, compares their respective expected reduction amounts, and selects the candidate price with the largest expected reduction amount as the probe price. .

[0048] The detection execution module is also used to collect data on target customers' detection prices. The multilayer response signal includes: First response signal This indicates whether the time spent on the page after the price display exceeds the category reference threshold; Second response signal This indicates whether you want to view the product details; Third response signal This indicates whether to add the item to the shopping cart; Fourth response signal This indicates whether the purchase has been completed. Each response signal has an indication value of 1 or 0, where 1 represents yes and 0 represents no. The category reference threshold is determined based on the historical average dwell time of the product's category and is stored in the product category configuration table.

[0049] The detection execution module linearly weights and sums the indication values ​​of each response signal with preset weighting coefficients to obtain the response intensity score. ,in , , , The fixed weight values ​​for each response signal are pre-set, and the weight coefficients are determined by regression analysis of the correlation strength between the response signal and the final purchase in historical transaction data.

[0050] The detection execution module obtains the baseline response strength. If the target customer has a history of interactions with similar products, then Take the response intensity score from each historical interaction. The arithmetic mean, if there are no historical records. Take a preset constant , The average response intensity of all historical customers is statistically determined and stored as a global configuration parameter.

[0051] The detection execution module calculates the effective elasticity observations. The method is as follows: , , .

[0052] The execution module is used to generate the most profitable price for displaying to target customers based on the expected elasticity value of the elasticity belief distribution in the utilization mode. The execution module first calculates the expected elasticity value of the current elasticity belief distribution. And obtain the current cost price of the product. .

[0053] like If the value is greater than 1, then the basic optimal price is... ;like If not greater than 1, then The preset upper limit of the product price is taken. This upper limit is the upper limit of the suggested retail price of the product in the corresponding sales channel, and is stored in the product information database in advance.

[0054] Calculate the correction amount using the execution module ,in Preset risk factor. Final execution price. Using the execution module to The price will be displayed as the most profitable price.

[0055] The profile update module is used in probe mode with the first effective sample weight. In the utilization mode, the second effective sample weight is used. A Bayesian update is performed on the elastic belief distribution, where Greater than .

[0056] In probe mode, the update rule is: , In the utilization mode, if the target customer ultimately completes a purchase, the updated rule is: , ,in This is a preset constant, which is obtained by pre-statistically analyzing the effective elasticity observations corresponding to purchasing events in historical detection patterns. The mean value is determined and stored as a global configuration parameter; if the purchase is not completed, the update will not be performed.

[0057] After the update is complete, the profile update module recalculates the uncertainty measure. and will , and Stored in association with the customer's identifier.

[0058] The following section uses an offline physical retail store scenario as a specific implementation method to explain the working process of this system in detail.

[0059] In this implementation, the system is deployed on a central server within the retail enterprise and communicates with electronic shelf label systems, smart shelf sensors, membership identification terminals, and POS systems deployed in each store. Customers complete identification by scanning their membership QR code at the store entrance, and the system creates a session record for that customer, maintaining a counter for the number of scans in the session record. and the most recent detection time , The initial value is 0. Initially empty. When a customer stops in front of a smart shelf and looks at a product for more than the preset attention time, the smart shelf sends a pricing request to the system, which includes the customer's identifier and the product's identifier.

[0060] In response to the pricing request, the system first uses the profile initialization module to retrieve records from the elastic belief distribution storage using the customer identifier as an index. If the search result is empty, the profile initialization module calls the hierarchical Bayesian model, taking the target customer's age group, gender, and resident city tier as input, and searches for a matching subgroup in the subgroup mapping table to obtain the corresponding... and ,Will Assigned value , Assigned value And calculate the initial uncertainty measure. If the search results are not empty, then the stored data will be read directly. and Value and calculate the current .

[0061] Subsequently, the threshold adaptive module obtains the sliding window queue. Stored in Value sequence, sorted in ascending order and then taking the first value. The percentile value is used as a dynamic threshold. In this embodiment The value is 500. The value is 75.

[0062] Gating decision module acquisition and Calculate the time interval This implementation method presets an upper limit. The value is 3, which is the preset minimum interval. The value is 30 seconds.

[0063] when and 3 and When all three conditions are met simultaneously within 30 seconds, the detection mode is triggered, and... Add 1, and Updated to the current time If any condition is not met, the exploitation mode is triggered.

[0064] In probe mode, the probe execution module obtains the current product's base price. A candidate price set is generated according to a preset floating ratio group. In this embodiment, the floating ratio group is {-10%, -6%, -2%, +2%, +6%, +10%}. The detection execution module starts from Beta ( , The inverse transform sampling method is used to generate the data. One elastic sample value, The value is set to 100. The probability of purchase tendency under each sample value is calculated using the response probability model. The expected reduction of each candidate price is calculated in the aforementioned manner, and the candidate price with the largest expected reduction is selected as the probe price. and will The electronic price tags sent to the shelf are refreshed.

[0065] exist Within the preset observation time window after the display, the observation time window duration is 60 seconds. The countdown begins when the product is displayed and ends when the customer leaves the shelf area or completes their purchase, whichever occurs first. Customer departure from the shelf area is determined by a human detection algorithm from the smart shelf's cameras. If a customer is detected leaving and the departure continues for more than the confirmed duration, the countdown will continue. It was determined to be leaving at that time. The value is 5 seconds.

[0066] Within this window, the detection execution module collects multi-layer response signals.

[0067] First response signal The system uses smart shelf cameras and dwell time timers to track the duration of customer dwell time in front of the shelves, comparing it to a category reference threshold for that product. If the threshold is exceeded, then... Select 1 if the value is 1, otherwise select 0.

[0068] Second response signal The shelf gravity sensor detects whether the product has been picked up. If it has been picked up, then... Select 1 if the value is 1, otherwise select 0.

[0069] Third response signal The system uses sensors attached to the shopping basket or cart to detect whether an item has been added to the cart. If an addition action is detected, then... Select 1 if the value is 1, otherwise select 0.

[0070] Fourth response signal The system uses a data interface with the POS system to determine whether the customer has completed their purchase. If the customer completes the payment transaction for the item within the current session, then... Select 1 if the customer does not produce any response signal within the observation time window, then all... to All values ​​are 0.

[0071] The detection execution module reads the preset weight coefficients. , , , Calculate the response intensity score .

[0072] Obtain the reference response strength If the target customer has a history of interactions with similar products, then the history will be used. The arithmetic mean of the two values ​​is used; otherwise, a preset constant is used. .

[0073] Calculate effective elasticity observations : , , .

[0074] In utilization mode, the execution module calculates the expected elasticity value. Obtain the cost price of goods Calculate the basic optimal price Calculate the correction amount In this embodiment Values Final execution price Using the execution module to The electronic price tags sent to the shelf are refreshed.

[0075] After each interaction, the threshold adaptive module will set the threshold value at the time of decision-making. Value pushed into queue The profile update module performs a Bayesian update.

[0076] In detection mode, the weight of the first effective sample is used. Update , .

[0077] In the utilization model, if a customer ultimately completes a purchase, the second valid sample weight is used. Update , , The value is 0.5; if the purchase is not completed, the update will not be performed.

[0078] The system learns of the purchase completion event through the transaction callback of the POS system. Upon receiving the callback, it triggers the profile update module to perform the update in the exploitation mode. The value is set to 20. Recalculation will be performed after the update. and will , and Stored in association with the customer's identifier.

[0079] As another implementation, this system can be applied to online e-commerce retail platform scenarios. In this implementation, the system is deployed on the backend server cluster of the e-commerce platform and communicates with the frontend web page service, mobile application interface, and order system. Customers access the e-commerce platform through a browser or mobile application, and the system reads the device identifier of the customer's terminal device as a temporary customer identifier; If a customer is already logged into their member account, the member ID is used as the official customer identifier and associated with the device ID. When a customer enters a product details page, the front end sends a pricing request to the system.

[0080] The main differences between this implementation method and offline physical retail scenarios are as follows.

[0081] In the profile initialization module, for anonymous customers who are visiting for the first time and have not logged in, their static attributes are obtained as follows: the city level of residence is obtained by resolving the IP address, the age group is inferred by the device model and operating system version, and the gender tendency is inferred by the time of visit.

[0082] The inference results are determined using a pre-built mapping rule table. This table is generated offline based on statistical analysis of the correlation between device features and customer static attributes in historical data and is stored in the system configuration database. It is retrieved directly from the table during online runtime. The inference results are then used as input to the hierarchical Bayesian model for initialization.

[0083] In the gating decision module, The number of probes performed on this customer during the current session. This represents the time interval between the current moment and the last detection moment. (Preset upper limit) The value is 5, and the preset minimum interval is... The value is 15 seconds. Greater than and Less than 5 and The detection mode is triggered when all three conditions—greater than or equal to 15 seconds—are met simultaneously; otherwise, the exploitation mode is triggered. (Threshold adaptive module) The value is 500. The value is 75.

[0084] In the detection execution module, the price is detected. The price display on the product details page is replaced with a new one by the front-end rendering. The observation window is 60 seconds long. The countdown begins when the product is displayed and ends when the customer leaves the product details page or completes the purchase, whichever occurs first. Customer departure from the product details page is determined by the front-end page visibility state application programming interface and the page unload event. A customer is considered to have left when the page becomes invisible for more than 5 seconds.

[0085] The acquisition method for multi-layer response signals is adapted to online interactive environments: First response signal The duration of the dwell time is obtained by monitoring the page visibility status and dwell timer through the front-end code and compared with the category reference threshold. Second response signal This can be determined by listening to whether the customer scrolls to the details description area or clicks the details tab; Third response signal Determine the issue by listening for API call events related to adding items to the shopping cart; Fourth response signal The determination is made through the order creation callback event in the order system. The indicator values ​​of each response signal are linearly weighted and summed with their respective weighting coefficients to obtain the result. The method is the same as the offline implementation method, with the same weighting coefficient. , , , If no response signal is generated by the customer within the observation time window, then all to All values ​​are 0. The method for determining this is the same as the offline implementation method. The value is 0.5. The category reference threshold is determined based on the historical average dwell time of the product category and is stored in the product category configuration table. The candidate price set adopts the same floating ratio group {-10%, -6%, -2%, +2%, +6%, +10%} as the offline implementation method.

[0086] In the execution module, the final execution price The display method involves replacing the displayed price of the product on the product details page, subsequent shopping cart pages, and checkout pages. The system receives the purchase completion event through order system callbacks, and upon receiving the callback, triggers the profile update module to execute the update in the exploitation mode.

[0087] The specific implementation methods of the remaining modules, including the threshold adaptive module, are as follows. and Get the value, detect the execution module and Value retrieval, portrait update module , and Value retrieval, utilizing the execution module's... The values, as well as all calculation formulas and variable definitions, are implemented in the same way as in offline physical retail scenarios.

[0088] Through the two implementation methods described above, this system can complete the process of building a flexible profile of a customer from completely unknown to one that can support personalized pricing within a single customer session, thereby achieving accurate and dynamic pricing for cold-start customers.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the scope of the technical concept of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart retail dynamic pricing system based on customer profiles, characterized in that, include: The profile initialization module, in response to a pricing request for a target customer, obtains the target customer's current elasticity belief distribution and corresponding uncertainty measure. The elastic belief distribution is a Beta distribution, determined by its shape parameter. and Determined, the uncertainty measure The variance of the Beta distribution; The gating decision module measures the uncertainty. With a dynamic threshold When comparing, Greater than When the detection mode is triggered, Not greater than Time-triggered exploitation mode; The probe execution module selects a probe price from a set of candidate prices in probe mode. The selection method, as shown to the target customer, is as follows: For each candidate price in the candidate price set, the expected reduction in the elastic belief distribution entropy before and after implementing that candidate price is calculated based on the current elastic belief distribution; the candidate price that maximizes this expected reduction is selected as the probe price. ; Collect data from the target customers at the detected price The multi-layered response signals are linearly weighted and summed with preset weighting coefficients to obtain a response intensity score. Based on the detection price A benchmark price for a commodity The differences and the response intensity score Calculate effective elasticity observations ; Using the execution module, in utilization mode, a profit-optimal price is generated based on the expected elasticity value of the elasticity belief distribution and displayed to the target customer; The profile update module, in detection mode, uses the first effective sample weight. In the utilization mode, the second effective sample weight is used. Perform a Bayesian update on the elastic belief distribution. Greater than ; The threshold adaptive module maintains a sliding window to store the most recent... Uncertainty Measurement for Sub-Pricing Decisions The value is calculated within the sliding window each time a gating decision is made. The value of the first Percentile as the dynamic threshold , For the preset window length, This is the preset percentile value.

2. The intelligent retail dynamic pricing system based on customer profiles according to claim 1, characterized in that, For target customers appearing for the first time without historical elastic belief distribution records, the profile initialization module uses a pre-trained hierarchical Bayesian model for initialization: it determines the subgroup to which the target customer belongs based on the target customer's static attributes, and uses the posterior distribution parameters corresponding to the subgroup as the initial elastic belief distribution. and Values, and calculate the initial uncertainty measure. The static attributes include at least one of age group, gender, and city of residence level.

3. The intelligent retail dynamic pricing system based on customer profiles according to claim 1, characterized in that, The candidate price set is based on the benchmark price of the commodity. A discrete set of prices is generated centered on a preset non-zero floating ratio group, wherein the magnitude of the floating ratio is limited to a preset range.

4. The intelligent retail dynamic pricing system based on customer profiles according to claim 1 or 3, characterized in that, When the detection execution module calculates the expected reduction, it uses Monte Carlo sampling: sampling from the current elastic belief distribution. For each elastic sample value, a pre-defined response probability model is used to calculate the probability of a purchase inclination towards the candidate price. Based on the probabilities corresponding to each elastic sample value, the marginal response probability and posterior distribution entropy are estimated. The expected reduction amount is then obtained using the marginal response probability and posterior distribution entropy. This is the preset number of samples.

5. The intelligent retail dynamic pricing system based on customer profiles according to claim 1, characterized in that, The multi-layer response signals include whether the page dwell time after price display exceeds the category reference threshold, whether product details are viewed, whether the product is added to the shopping cart, and whether the purchase is finally completed. Each response signal has an indicator value of 1 or 0, representing yes or no, respectively. The preset weighting coefficients for the linear weighted summation include... , , , Each of these corresponds to a fixed value for a given response signal.

6. The intelligent retail dynamic pricing system based on customer profiles according to claim 1, characterized in that, The detection execution module calculates the effective elasticity observation value. The method is as follows: ,in , , This is the baseline response strength.

7. The intelligent retail dynamic pricing system based on customer profiles according to claim 1, characterized in that, When the gated decision module triggers the detection mode, it also satisfies the following conditions: the number of times the target customer is detected within the same customer session does not exceed a preset upper limit, or the time interval between two detections is not less than a preset minimum interval.

8. The intelligent retail dynamic pricing system based on customer profiles according to claim 1, characterized in that, When generating the profit-optimal price using the execution module, the basic optimal price is first calculated based on the expected elasticity value of the elasticity belief distribution. Then, a correction amount is subtracted from the basic optimal price to obtain the final execution price. The correction amount is related to the uncertainty measure. It is proportional to the square root.

Images