E-commerce platform commodity sales data analysis system and method

Abstract

The invention provides an e-commerce platform commodity sales data analysis system and method, relates to the field of data analysis, and effectively improves the efficiency and accuracy of e-commerce platform commodity sales data analysis through operations such as hierarchical coding, incremental matrix optimization and price sensitivity hierarchical recommendation. Through hierarchical coding and dynamic compression, data storage and processing are optimized, the system burden is reduced, and the real-time performance of data processing is improved. The incremental matrix optimization method ensures that the system can quickly respond to changes of new users and new commodities, and the cold start problem is effectively solved. And the hierarchical dynamic recommendation module performs personalized recommendation for price-sensitive users, so that the recommendation accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data analysis, and more particularly, to an e-commerce platform commodity sales data analysis system and method. BACKGROUND

[0002] Currently, e-commerce platforms are booming worldwide, and the collection, analysis and application of sales data have become the core link to improve user experience and platform operation efficiency. With the progress of big data technology, e-commerce platforms can optimize product recommendation strategies based on user behavior data and improve conversion rates. However, due to the characteristics of high dimensionality, strong timeliness and non-uniform distribution of e-commerce sales data, how to build an efficient analysis system in a large-scale data environment has become an industry problem. Traditional e-commerce data analysis methods mostly use batch processing, that is, periodically updating the product recommendation list through offline computing, but such methods are difficult to meet the real-time needs of market changes. In recent years, hybrid recommendation methods combining deep learning and incremental optimization have gradually become mainstream, but there are still many challenges in data storage, computing resource allocation and real-time optimization.

[0003] The existing e-commerce data analysis technology mainly has the following deficiencies. First, the data storage is relatively redundant, affecting the computing efficiency. Traditional methods usually store sales data with fixed granularity, without fully considering the timeliness characteristics of e-commerce data, resulting in a sharp increase in computing load under high concurrency. Second, the sparsity of user-item interaction matrix affects the accuracy of recommendation. Due to the small number of historical ratings of some products, traditional collaborative filtering algorithms have difficulty in establishing effective recommendation rules, resulting in limited personalized recommendation capabilities. Third, the recommendation system lacks adaptability to price sensitivity. Existing recommendation algorithms often rely on static user preference modeling and fail to fully consider the reactions of different user groups to price changes, making it difficult to balance between accurate recommendation and marketing strategies.

[0004] Therefore, how to optimize the data storage structure, improve the matrix calculation efficiency, and conduct personalized recommendation combined with price sensitivity has become a key factor affecting the effectiveness of e-commerce platform sales data analysis. SUMMARY

[0005] In order to solve the above technical problems, the present application is proposed. The present application provides an e-commerce platform commodity sales data analysis system and method.

[0006] According to one aspect of the present application, an e-commerce platform commodity sales data analysis system is provided, which comprises: a real-time data hierarchical coding module, which codes sales data according to time response demand, and dynamically adjusts data collaborative compression channels between layers according to system load; a user-commodity interaction matrix construction module, which constructs a sparse interaction matrix based on user behavior records in the hierarchical coding, fills in missing values based on average ratings of similar category commodities for non-interaction data items, and forms a complete user-commodity rating matrix; an incremental matrix optimization module, which obtains a latent feature matrix by alternating least squares decomposition of the user-commodity rating matrix; a hierarchical dynamic recommendation module, which divides user groups based on price sensitivity level, generates recommendation ratings based on the latent feature matrix, and increases discount commodity weight coefficients for high-sensitivity user groups to form a final recommendation list.

[0007] Further, the hierarchical coding comprises: dividing sales data into millisecond-level real-time layers, minute-level aggregation layers and hour-level archiving layers according to time response demand, storing real-time layer data using adjacent record difference value coding, and reducing dimensionality compression of archiving layer data by tensor decomposition.

[0008] Further, the dynamic adjustment comprises: collecting system load state parameters, judging the pressure interval of data flow between layers, and determining a data compression channel adjustment strategy; when the system load drops to a set range, restoring the original compression channel parameters.

[0009] Further, the data compression channel adjustment strategy comprises: increasing the compression ratio of the archiving layer under high load; dynamically adjusting the data transfer rate from the real-time layer to the aggregation layer to reduce data storage pressure under high load; reducing the storage duration of real-time layer data; and switching adjacent record difference value coding to a higher compression ratio prediction coding mode.

[0010] Further, the construction of the sparse interaction matrix comprises: extracting effective interaction behaviors from hierarchical coding data; setting weights for different types of interaction behaviors to generate interaction weight scores of users on commodities; filling normalized interaction scores of users on commodities into the sparse interaction matrix, and keeping non-interaction items as null values.

[0011] Further, the missing value filling is filling based on average ratings of similar category commodities for non-interaction data items.

[0012] Further, the obtaining of the potential feature matrix comprises: approximating and decomposing a user-goods rating matrix into two potential feature matrices, including a user feature matrix and a goods feature matrix, by using alternating least squares decomposition; in the first round of optimization, keeping all parameters of the goods feature matrix unchanged, and only optimizing and calculating the feature vector of each user, for each user, based on the rating data of the user on multiple goods in the rating matrix, calculating the feature vector of the user to minimize the rating error; after optimizing the user feature matrix, keeping the user feature matrix unchanged, and calculating the feature vector of each goods, according to the user data of all users who have rated the goods, adjusting the feature vector of the goods to minimize the rating error under the constraint of the current user feature matrix; repeating the optimization of the user feature matrix and the optimization of the goods feature matrix until the decline amplitude of the rating error is lower than a set threshold, or the maximum optimization number is reached.

[0013] Further, the incremental matrix optimization module further comprises generating an initial user feature vector according to the first interaction goods set of the new user when the new user joins, and completing the incremental update of the user feature matrix in the cold start state.

[0014] Further, the generation of the recommendation score comprises: the basic score of the user on the goods is obtained by inner product calculation of the user feature vector and the goods feature vector; a price elasticity matrix is set, wherein the price elasticity coefficient is calculated based on historical sales data of the goods, and reflects the influence of the price change of the goods on the purchase probability of the user; different user price sensitivity weights are divided according to different price sensitivity levels; the product of the user price sensitivity weight and the price elasticity coefficient is added to the basic score to obtain the recommendation score.

[0015] According to another aspect of the present application, an e-commerce platform goods sales data analysis method is provided, which comprises: layering and coding the sales data according to time response demand, dynamically adjusting the data collaborative compression channel between layers according to system load; constructing a sparse interaction matrix based on the user behavior records in the layering coding and filling in the missing values to form a complete user-goods rating matrix; obtaining potential feature matrices by alternating least squares decomposition on the user-goods rating matrix; dividing user groups based on price sensitivity levels, and generating recommendation scores with the potential feature matrices, and increasing discount goods weight coefficients for the medium and high sensitivity user groups to form a final recommendation list.

[0016] Compared with the prior art, the e-commerce platform commodity sales data analysis system and method provided by the present application effectively improves the efficiency and accuracy of e-commerce platform commodity sales data analysis through layered coding, incremental matrix optimization, and price sensitivity layered recommendation operations. Through layered coding and dynamic compression, data storage and processing are optimized, system burden is reduced, and data processing real-time performance is improved. The incremental matrix optimization method ensures that the system can quickly respond to changes in new users and new commodities, effectively solving the cold start problem. The layered dynamic recommendation module provides personalized recommendations for price-sensitive users, improving recommendation accuracy. Compared with the prior art, the present application can achieve more efficient system resource utilization and more accurate personalized recommendations, significantly improving the operating efficiency and user conversion rate of e-commerce platforms. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. In the drawings: Figure 1 The system block diagram of the e-commerce platform commodity sales data analysis system according to the embodiment of the present application.

[0018] Figure 2 The block diagram of the layered dynamic recommendation module 400 in the e-commerce platform commodity sales data analysis system according to the embodiment of the present application.

[0019] Figure 3 The flowchart of the e-commerce platform commodity sales data analysis method according to the embodiment of the present application. DETAILED DESCRIPTION

[0020] In the following, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described here.

[0021] As mentioned in the background section, existing e-commerce data analysis technologies suffer from several shortcomings. First, high data storage redundancy impacts computational efficiency. Traditional methods typically store sales data at a fixed granularity, failing to fully consider the time-sensitive nature of e-commerce data, leading to a sharp increase in computational load under high concurrency. Second, the sparsity of the user-product interaction matrix affects recommendation accuracy. Due to limited historical ratings for some products, traditional collaborative filtering algorithms struggle to establish effective recommendation rules, limiting personalized recommendation capabilities. Third, recommendation systems lack adaptability to price sensitivity. Existing recommendation algorithms often rely on static user preference modeling, failing to adequately consider the responses of different user groups to price changes, making it difficult to strike a balance between accurate recommendations and marketing strategies.

[0022] Therefore, optimizing data storage structure, improving matrix calculation efficiency, and combining price sensitivity for personalized recommendations have become key factors affecting the effectiveness of sales data analysis on e-commerce platforms.

[0023] Figure 1 This is a system block diagram of an e-commerce platform product sales data analysis system according to an embodiment of the present invention. Figure 1 As shown, the e-commerce platform product sales data analysis system includes: a real-time data layered encoding module 100, which encodes sales data in layers according to time response requirements and dynamically adjusts the data collaborative compression channels between layers based on system load; a user-product interaction matrix construction module 200, which constructs a sparse interaction matrix based on user behavior records in the layered encoding, and fills missing values ​​for non-interactive data items using the average rating of similar products as a benchmark to form a complete user-product rating matrix; an incremental matrix optimization module 300, which performs alternating least squares decomposition on the user-product rating matrix to obtain a latent feature matrix, and generates an initial user feature vector based on the set of products the user first interacts with when a new user joins, completing the incremental update of the matrix in the cold start state; and a layered dynamic recommendation module 400, which divides user groups based on price sensitivity levels and generates recommendation ratings with the latent feature matrix, and adds a weight coefficient for discounted products to medium- and high-sensitivity user groups to form a final recommendation list.

[0024] In this embodiment of the invention, the real-time data layered encoding module 100 specifically includes dividing sales data into a millisecond-level real-time layer, a minute-level aggregation layer, and an hour-level archive layer according to time response requirements. The real-time layer data is stored using adjacent record difference value encoding, and the archive layer data is compressed by tensor decomposition. The data collaborative compression channel between layers is dynamically adjusted according to the system load.

[0025] In a data storage and processing system, time layering of sales data is a common strategy aimed at balancing data access speed and storage efficiency. In traditional e-commerce platforms, sales data is usually stored in a unified format and optimized for access efficiency through indexing when queried. However, this approach can lead to query response delays, storage resource waste, and increased computing pressure when the data volume is large. Therefore, the present invention proposes a time granularity-based hierarchical encoding scheme that divides sales data into the following three levels: a millisecond-level real-time layer, which is mainly used for real-time queries and high-frequency data analysis within a short time, stores the latest sales data, and has a very high data update frequency, suitable for scenarios that require second-level decision-making, such as dynamic inventory management, real-time promotion optimization, etc.; a minute-level aggregation layer, which is used for data analysis within a short time window, stores sales data within the past few minutes, and performs partial aggregation calculations, such as calculating product sales volume, user browsing volume, etc. Compared to the real-time layer, this layer has a longer data storage duration but a relatively lower update frequency, mainly serving business operation-level decision-making; an hour-level archiving layer, which is mainly used for long-term data storage and deep data mining, stores sales data within a range of past hours or longer. Since the access frequency of archived data is low, efficient compression techniques are used for storage to reduce storage occupancy and improve data analysis and calculation efficiency.

[0026] To address the resource waste caused by static setting methods, such as when system load is low, maintaining a high compression ratio can lead to idle computing resources, while in high-load situations, fixed storage strategies can cause data access bottlenecks, affecting system real-time performance. Therefore, the present invention introduces a mechanism for dynamically adjusting hierarchical data flow, with a key step being to collect system load state parameters for subsequent adjustment of data storage and compression strategies, as follows: Collect the current computing resource occupancy, obtain CPU usage, memory occupancy, and data storage read / write speed to form system load state parameters; based on the system load state parameters, determine the pressure interval of hierarchical data flow, and determine the data compression channel adjustment strategy; when the system load drops to a set range, restore the original compression channel parameters to return the data storage and compression strategies of the real-time layer, aggregation layer, and archiving layer to normal, improving data access efficiency.

[0027] In the implementation of the present application, when the system load state parameter collection is completed, the system will judge the current data flow pressure interval based on these parameters, and adopt appropriate adjustment strategy to balance the data storage pressure and access efficiency. Specifically, the judgment logic of system load state is as follows: low load state (system resource occupation is low): at this time, the system allows real-time layer data to be stored for a longer time, thereby reducing the overhead of frequent data transfer and improving query availability; normal load state (system runs in an acceptable range): at this time, the data flow transfer maintains the default strategy, that is, the real-time layer data is transferred to the aggregation layer according to the set frequency, and the aggregation layer data is archived to the archive layer according to the plan; high load state (system resource occupation exceeds the set threshold): at this time, the system takes a series of optimization measures to reduce the calculation and storage pressure, such as increasing the compression ratio of the archive layer data, reducing the real-time layer data storage time, etc.

[0028] For example, the data compression channel adjustment strategy includes: when the system load exceeds the set threshold, that is, in high load, the data compression ratio of the archive layer is increased (such as from ordinary Huffman coding to higher level arithmetic coding) to reduce the storage occupation; dynamically adjusting the data transfer rate from the real-time layer to the aggregation layer to reduce the real-time storage pressure in high load state, and allowing the real-time layer data to be stored for a longer time in low load state to improve the query response ability; in high load, the storage time of real-time layer data is reduced, such as from millisecond level to smaller window, and in low load, the storage time is prolonged to improve data query availability; according to the load condition, the data compression format is adjusted, such as using more compact block storage format in high load and standard storage format in low load to optimize data decoding efficiency.

[0029] The core goal of the above strategy is to optimize data storage and calculation efficiency under different load states, to ensure that the system can still run stably in high load, and at the same time, to fully utilize the calculation and storage resources in low load to improve query performance.

[0030] Optionally, after the system load recovers to the set range, the present application allows the data storage and compression strategy to return to normal to ensure the long-term stability of data query and storage. Specifically, the system will perform the following steps: restore the default data storage time window, that is, in low load, the real-time layer data storage time will be appropriately prolonged to improve query availability; restore the default data compression strategy, that is, after the load falls back, the archive layer data compression will return to normal to reduce the calculation burden of data decompression without affecting the storage resources; adjust the data flow rate, that is, after the load recovers, the system allows data to transfer according to the original plan to maintain the consistency of data access. This recovery mechanism ensures that the system can dynamically adapt to load changes and recover to the best running state after load fluctuation, improving the long-term stability and data availability of the system.

[0031] In the real-time data hierarchical encoding module 100, by performing millisecond, minute and hour level hierarchical storage on the sales data, and using difference value encoding and tensor decomposition for data compression, the system can efficiently store and retrieve sales data in different time dimensions. At the same time, in order to optimize the data flow process, the data collaborative compression channel between the layers is dynamically adjusted to cope with different system load conditions. However, the storage and compression of sales data cannot be directly used for product recommendation or user behavior analysis, and further meaningful information needs to be extracted from the stored data.

[0032] In the embodiment of the application, the user-goods interaction matrix construction module 200 specifically includes constructing a sparse interaction matrix based on the user behavior records in the hierarchical encoding data, filling in missing values using the average rating of similar goods as a reference for non-interaction data items, and forming a complete user-goods rating matrix.

[0033] The user-goods interaction matrix is a key data structure that describes the relationship between user behavior preferences and goods. Specifically, the user-goods interaction matrix is composed of a user set and a goods set , where and are the number of elements in the user set and the goods set, respectively, and the matrix element represents the interaction data of user to goods . The specific construction process is as follows: extract valid interaction behaviors (such as clicks, collections, purchases, etc.) from the hierarchical encoding data and remove noise data (such as abnormal short-time access); set weights for different types of interaction behaviors to generate interaction weight scores of users to goods; fill the normalized interaction scores of users to goods into . For users with few interaction behaviors, combine similar user behavior patterns to reduce data over-sparse conditions and improve matrix computability using local filling methods.

[0034] It should be noted that due to differences in user behavior habits, such as some users may frequently browse but rarely collect, while others only browse for a short time before purchasing, it is necessary to normalize the interaction data. Normalization can use Min-Max Scaling or Z-score standardization, so that the interaction data of different users can be calculated on the same scale, avoiding bias caused by individual behavior habit differences.

[0035] Since users cannot rate all products, the user-product interaction matrix is ​​typically sparse. In subsequent recommendation calculations, missing rating data needs to be appropriately imputed to improve model stability and computational accuracy. Specific imputation strategies are as follows: For the non-interactive data items in the user-product interaction matrix M The complete user-product rating matrix is ​​obtained by calculating the average rating of similar products and then filling it in: a set of product categories is constructed based on product category tags. , the goods Map to the corresponding category ,in, The number of elements in the product category set; for products Category Calculate the mean of all rated items within this category: ; in, For category The following is a collection of all rated products. For category The number of items that have been rated. For users For goods The rating, For category The average rating of all rated products in the category, i.e., the average rating of all products in this category.

[0036] for Fill value If there is no valid rating data within a category, the global average will be used to fill the gap. If the filled value exceeds the range of the user's personalized rating (e.g., users generally give low ratings to high-priced products), the deviation will be corrected by combining the user's historical rating average to improve the personalization matching.

[0037] For items that have not been interacted with, keep them empty to avoid forcibly filling in data that has not been interacted with, which could affect the accuracy of subsequent calculations.

[0038] Optionally, since some users interact less, the matrix may become too sparse, reducing computational efficiency. Therefore, for users with fewer interactions, local imputation methods (such as collaborative imputation based on user similarity) can be used to reduce data sparsity and improve the computability of the matrix, by combining the behavioral patterns of similar users. Specifically, the similarity between the user and other users can be calculated, and missing items can be imputed by referring to the interaction behaviors of similar users, making the user's behavioral pattern more complete.

[0039] The module converts the original user behavior data into a structured interaction matrix, providing computable data input for subsequent recommendation systems. The module ensures data accuracy and integrity through data cleaning, normalization, and matrix filling. The missing value filling strategy combines product category mean, global mean, and personalized score correction to ensure reasonable filling and personalized matching. Through this module, the sparsity of user-item interaction data can be significantly reduced.

[0040] In the user-item interaction matrix construction module 200, the user-item rating matrix has been constructed, and the data integrity has been improved by filling in missing values. However, the rating matrix is still a high-dimensional sparse matrix, and direct use for calculation may cause problems such as high data dimension and low calculation efficiency. In addition, in actual scenarios, user and item interaction behavior is dynamic, with new users and items constantly joining, resulting in frequent updates to the rating matrix. Therefore, further processing of the rating matrix is required to approximate the rating matrix into two low-dimensional latent feature matrices, namely the user feature matrix and the item feature matrix, through alternating least squares decomposition. This can reduce the dimensionality of the data, improve the efficiency of recommendation calculation, and extract the latent features of users and items, enabling the recommendation system to more accurately predict user preferences.

[0041] In the embodiment of the present application, the incremental matrix optimization module 300 specifically includes alternating least squares decomposition of the user-item rating matrix to obtain latent feature matrices, and when a new user joins, generating an initial user feature vector based on the first interaction item set to complete matrix incremental update in a cold start state.

[0042] The obtaining of the latent feature matrix includes: using alternating least squares decomposition to approximately decompose the user-item rating matrix into two latent feature matrices, including a user feature matrix and an item feature matrix; in the first round of optimization, keeping all parameters of the item feature matrix unchanged and only optimizing the calculation of the feature vector of each user, for each user, based on the rating data of multiple items in the rating matrix, calculating the feature vector of the user to minimize the rating error; after optimizing the user feature matrix, keeping the user feature matrix unchanged and calculating the feature vector of each item, adjusting the feature vector of the item based on the user data of all users who have rated the item to minimize the rating error under the constraint of the current user feature matrix; repeating the optimization of the user feature matrix first and then the optimization of the item feature matrix until the decline amplitude of the rating error is lower than the set threshold or the maximum optimization number is reached.

[0043] For example, the basic idea of ALS method is to approximate the user-item rating matrix M into two low-dimensional latent feature matrices, i.e., the user feature matrix and the item feature matrix , such that the approximation of the rating matrix is as follows: ; where denotes the user feature matrix, is the number of users, is the latent feature dimension; denotes the item feature matrix n is the number of items.

[0044] ALS algorithm minimizes the objective function by alternatingly optimizing and , and the detailed optimization steps are as follows: Set the latent feature dimension and the regularization parameter , initialize the user feature matrix and the item feature matrix by random initialization or pre-training based on historical data; for each user , solve the user feature vector by minimizing the following objective function according to its rating data: ; where is the set of items rated by the user. This optimization problem is a standard ridge regression which can be solved by minimizing the user rating error. In the calculation process, by fixing the item feature matrix, the user feature vector is optimized according to the user's historical rating data so that it can fit the known ratings as much as possible under the constraint condition. At the same time, a regularization term is introduced to prevent overfitting and improve the generalization ability of the model.

[0045] Similarly, for each item, based on the data of all users who have rated the item, the item feature vector is optimized. The user feature matrix and the item feature matrix are alternately optimized until the rating error decreases below a set threshold or reaches the maximum number of iterations.

[0046] Further, the incremental matrix optimization module further comprises generating an initial user feature vector according to a first interaction commodity set of the new user when the new user joins, and completing incremental updating of the user feature matrix in a cold start state. It can be known that in the traditional matrix decomposition method, the joining of the new user needs to retrain the entire model, and the calculation cost is high. Therefore, the module supports incremental updating, that is, when the new user joins, only the feature vector of the user is calculated, and the existing user and commodity feature matrix are not affected.

[0047] For example, the first interaction commodity set of the new user is set, that is, the commodity set purchased or browsed by the user; the mean of the commodity feature vectors interacted by the user is used as the initial user feature vector, which can ensure that the initial feature vector of the new user is consistent with the features of the commodities interacted, thereby improving the accuracy of cold start recommendation. On the basis of initialization, the Ridge regression is used to further optimize the new user feature vector to minimize the rating error.

[0048] For example, for the cold start problem of the new user, the first interaction commodity set of the user is first set. The set is usually composed of commodities purchased, browsed, collected or clicked by the user when using the system for the first time. In the traditional recommendation system, the cold start user is often difficult to model accurately due to the lack of historical interaction data, and therefore the invention adopts an initialization strategy based on the mean of commodity features.

[0049] Specifically, first, the feature vectors of all commodities interacted by the user are obtained, and the mean of the commodity features is calculated to serve as the initial feature vector of the user. The core idea of the method is that the interest of the new user is likely to be related to the features of the commodities initially selected, and therefore using the mean as the initialization vector can effectively ensure that the features are consistent with the features of the interacted commodities.

[0050] After initialization, the feature vector is further optimized to improve the accuracy of personalized recommendation. The optimization process minimizes the rating error to enable the user's feature vector to better fit the rating behavior. For example, after the user interacts with more commodities, the feature vector is continuously updated, thereby gradually improving the adaptability of the recommendation system. In order to prevent the feature vector from overfitting to a small number of commodities, a regularization mechanism is introduced in the optimization process to ensure that the user features remain reasonable and balanced, and to avoid recommendation bias caused by individual commodity ratings being too high or too low.

[0051] The advantage is that since the feature vector of the new user is consistent with the features of the interactive commodity, the recommendation result can better meet the real interests of the user; the method is only based on the interactive data of the user itself, without relying on additional user portrait or social information, and is suitable for scenes with high data privacy requirements; in the subsequent interaction process, the feature vector of the user is continuously adjusted, so that the recommendation system can more accurately capture the change of the user's interest.

[0052] It can be seen that the core target of the module 300 of the application is to decompose and optimize the user-commodity score matrix, so that the score matrix can be reduced in dimension and potential features are extracted, improving the calculation efficiency and recommendation accuracy of the recommendation system. Meanwhile, for the new user cold start problem, a user feature initialization scheme based on the first interactive commodity set is provided, and combined with the incremental optimization strategy, the system can adapt to the behavior mode of the new user in real time, improving the adaptability and personalization ability of the recommendation system to the new user.

[0053] In the embodiment of the application, the hierarchical dynamic recommendation module 400 specifically includes: dividing the user group based on the price sensitivity level, weighting and fusing the potential feature matrix and the price elasticity matrix to generate a recommendation score, and increasing the weight coefficient of the discounted commodity for the medium and high sensitivity user group to form a final recommendation list, as shown in Figure 2 As shown, it includes a user group division unit 410: based on the price sensitivity level, the user group is divided, the price sensitivity of the user is determined according to the historical purchase behavior of the user, the influence of the price change of the commodity on the decision of the user and the consumption habit, and the user is classified into a low sensitivity group, a medium sensitivity group and a high sensitivity group; a recommendation score calculation unit 420: for the obtained potential feature matrix, the price elasticity matrix of the commodity is weighted and fused according to the price sensitivity of different user groups. The price elasticity matrix is used to measure the influence of the price change of the commodity on the purchase tendency of the user, and the final score of each user is calculated by linear combination of the potential feature score and the price elasticity coefficient, wherein the weight of the price elasticity factor of the user with higher price sensitivity is larger; a discount weight adjustment unit 430: for the medium and high sensitivity users, the weight coefficient of the discounted commodity is additionally increased based on the fused recommendation score, so that the sorting priority of the discounted commodity in the recommendation list is improved. The weight adjustment mode is dynamically set according to the price sensitivity level of the user, and the adjustment range of the discounted commodity weight of the high sensitivity user is higher than that of the medium sensitivity user, so as to enhance the personalized matching degree of the recommendation. According to the fusion score result, a plurality of commodities with the highest score are selected as the recommendation result by arranging the commodities in descending order, and a proper proportion of high-score commodities and discounted commodities are ensured in the recommendation list to balance the personalized needs of the user and the promotion effect. The final recommendation list can be used for front-end display or personalized push to improve the purchase conversion rate of the user.

[0054] In this embodiment of the invention, the user group segmentation unit 410 specifically includes: During the incremental matrix optimization process of module 300, the user's feature vector has been obtained, and the user-product rating matrix has been optimized. However, in actual recommendation scenarios, users' purchasing decisions are not only influenced by their latent features but also closely related to price sensitivity. Therefore, in this unit, users will be segmented according to their price sensitivity levels for targeted optimization in subsequent recommendation processes. The execution result of this unit will directly affect the subsequent recommendation rating calculation and the generation of the final recommendation list.

[0055] In this invention, price sensitivity measures the degree to which a user reacts to changes in the price of a product. Traditional methods primarily classify price sensitivity based on the distribution of price ranges in a user's purchase history, but this method has significant limitations, such as failing to consider changes in user behavior across different scenarios. Therefore, this invention proposes a price sensitivity classification method that comprehensively considers multiple factors.

[0056] First, we collect data on users' purchase history, browsing history, and items added to their cart but not purchased, and then perform price range analysis on this data. For each user, we statistically analyze the price distribution of their purchased items, calculate their main consumption range, and consider their response to price fluctuations. For example, if a user has repeatedly changed their purchasing decisions due to discounts and promotions in their past purchase history, they can be considered to have high price sensitivity.

[0057] Secondly, a time dimension is introduced to analyze changes in users' price sensitivity at different times. For example, some users may have increased price sensitivity during major promotional periods (such as Singles' Day), while their price sensitivity is relatively low in daily purchases. To address this, a hierarchical clustering method (such as K-Means or hierarchical clustering) is used to divide users into three levels: high sensitivity, medium sensitivity, and low sensitivity, so that different weight coefficients can be used for personalized optimization in the subsequent recommendation stage.

[0058] As can be seen, this unit combines data from multiple dimensions to conduct price sensitivity analysis, avoiding the limitations of relying solely on historical purchase prices and improving classification accuracy; it introduces the time factor to dynamically adjust users' price sensitivity, enabling the recommendation system to adapt to changes in users' consumption behavior and improve the effectiveness of recommendations; and it uses machine learning methods for automatic stratification, rather than the traditional fixed price range division, to ensure more accurate price sensitivity stratification.

[0059] The recommendation rating calculation unit 420 specifically includes: after completing the user's price sensitivity level classification, it is necessary to introduce price factors into the recommendation rating calculation. Module 300 has already optimized the user-product rating matrix based on the alternating least squares method, obtaining the latent feature matrices of users and products. However, this rating matrix does not consider price factors. Therefore, in this unit, by constructing a price elasticity matrix and weighting and fusing it with the latent feature matrix, the recommendation rating not only depends on user preferences but also reflects the user's sensitivity to price changes.

[0060] The price elasticity matrix is ​​constructed based on users' price sensitivity levels. For each user, their responsiveness to price changes of different goods is calculated, and the price elasticity matrix is ​​constructed accordingly. Traditional price elasticity calculation methods are mainly based on the price-sales relationship, that is, measuring price elasticity by analyzing the impact of price changes on sales volume. However, this method is mainly applicable to overall market analysis and is difficult to accurately model the price sensitivity of individual users. Therefore, this invention proposes a price elasticity calculation method based on individual user behavior: First, the historical price elasticity coefficient of each user is calculated. For each user, their historical purchase records are analyzed to statistically analyze the fluctuations in their purchase decisions with price changes. For example, if a user purchases goods significantly more frequently during discount periods than during non-discount periods, it indicates that the user has higher price elasticity. Second, adjustments are made based on the price elasticity data of users at the same price sensitivity level. Since some users have fewer purchase records, price elasticity data from users at the same price sensitivity level is introduced for smoothing to improve the stability of the calculation. Finally, weights are assigned based on the price sensitivity level, and the price elasticity matrix and the latent feature matrix are weighted and fused. For highly sensitive users, the weight of price factors is larger, while for low-sensitive users, the weight of price factors is smaller, thus ensuring that the recommendation score accurately reflects the user's actual purchasing tendency.

[0061] The discount weight adjustment unit 430 specifically includes: In the previous unit, the weighted fusion of the price elasticity matrix and the latent feature matrix was completed, and a recommendation score was generated. However, for users with high price sensitivity, the recommendation strategy needs to be further optimized by increasing the weight of discounted products to improve the conversion rate of recommendations. Therefore, the goal of this unit is to prioritize the recommendation of discounted products in the final recommendation list for users with medium to high price sensitivity.

[0062] The weighting adjustment of discounted products mainly involves the following key steps: The product database is monitored in real time to determine which products are currently discounted. For some platforms, discount information may not be directly disclosed, so historical price fluctuation data can be used to infer the true discount situation. For each medium-to-high price elasticity user, a weighting factor for discounted products is set based on their price elasticity coefficient. For example, for users with a high price elasticity coefficient, the rating weight of discounted products is increased, making them more likely to appear at the top of the recommendation list. When generating the final recommendation list, the user's personalized preferences and the weighting adjustment factor of discounted products are combined to ensure that highly sensitive users see discounted products first, while the recommendation results for low-sensitive users are still mainly based on product matching.

[0063] Optionally, the difference between the actual conversion rate of recommended products and that before the recommendation can be monitored. Based on the difference, matrix factorization parameter adjustments and data stratification time window optimization instructions can be triggered synchronously, forming a closed-loop feedback adjustment mechanism. It should be noted that this rating recommendation may not perfectly match actual user purchasing behavior. Due to factors such as changes in user preferences, market trend fluctuations, and changes in product inventory, the model may exhibit bias. Therefore, a real-time monitoring and dynamic adjustment closed-loop feedback mechanism can be set up to ensure the accuracy and robustness of the recommendation system.

[0064] In recommendation systems, conversion rate is a crucial metric for measuring recommendation effectiveness. Conversion rate typically refers to the percentage of users who actually click, favorite, or purchase a product after a recommendation. A low overall conversion rate for a recommendation list indicates that the system is failing to effectively match user needs. Therefore, monitoring conversion rate is a core aspect of optimizing a recommendation system. If the conversion rate of recommended products significantly decreases compared to before the recommendation, it may indicate issues such as: user needs changing over time, previous recommendations becoming obsolete, the emergence of new products, or competitors adjusting prices, all of which can affect product attractiveness. This suggests a bias in the user or product feature matrix, leading to inaccurate recommendation scoring. Therefore, appropriately increasing the regularization coefficient can prevent the model from overfitting historical data, and adjusting the learning rate can accelerate matrix updates and improve adaptability. For users with declining conversion rates, recalculating their feature vectors to better reflect current behavioral patterns, updating product features using the latest user interaction data, and improving recommendation accuracy create a closed-loop feedback adjustment mechanism, ensuring the accuracy, adaptability, and stability of the recommendation system.

[0065] In summary, the e-commerce platform product sales data analysis system based on embodiments of the present invention has been clarified. Through operations such as hierarchical encoding, incremental matrix optimization, and price-sensitivity-based hierarchical recommendation, it effectively improves the efficiency and accuracy of e-commerce platform product sales data analysis. Hierarchical encoding and dynamic compression optimize data storage and processing, reduce system load, and improve the real-time performance of data processing. The incremental matrix optimization method ensures that the system can quickly respond to changes in new users and new products, effectively solving the cold start problem. The hierarchical dynamic recommendation module provides personalized recommendations for price-sensitive users, improving recommendation accuracy. Compared with existing technologies, the present invention achieves more efficient system resource utilization and more accurate personalized recommendations, significantly improving the operational efficiency and user conversion rate of e-commerce platforms.

[0066] Figure 3 This is a flowchart of a product sales data analysis method for an e-commerce platform according to an embodiment of the present invention. Figure 3 As shown, the e-commerce platform product sales data analysis method includes: S1, hierarchically encoding sales data according to time response requirements, and dynamically adjusting the data collaborative compression channels between layers according to system load; S2, constructing a sparse interaction matrix based on user behavior records in the hierarchical encoding and filling in missing values ​​to form a complete user-product rating matrix; S3, performing alternating least squares decomposition on the user-product rating matrix to obtain a latent feature matrix; S4, dividing user groups based on price sensitivity levels, generating recommendation ratings with the latent feature matrix, and adding weight coefficients for discounted products to medium- and high-sensitivity user groups to form a final recommendation list.

[0067] Here, those skilled in the art will understand that the specific operations of each step in the above-mentioned e-commerce platform product sales data analysis method have been referenced above. Figures 1 to 2 The description of the e-commerce platform's product sales data analysis system is detailed here, and therefore, its repeated description will be omitted.

[0068] In summary, the e-commerce platform product sales data analysis method based on the embodiments of the present invention has been clarified. Through operations such as hierarchical encoding, incremental matrix optimization, and price-sensitivity-based hierarchical recommendation, it effectively improves the efficiency and accuracy of e-commerce platform product sales data analysis. Hierarchical encoding and dynamic compression optimize data storage and processing, reduce system load, and improve the real-time performance of data processing. The incremental matrix optimization method ensures that the system can quickly respond to changes in new users and new products, effectively solving the cold start problem. The hierarchical dynamic recommendation module provides personalized recommendations for price-sensitive users, improving recommendation accuracy. Compared with existing technologies, the present invention achieves more efficient system resource utilization and more accurate personalized recommendations, significantly improving the operational efficiency and user conversion rate of e-commerce platforms.

Claims

1. A product sales data analysis system for an e-commerce platform, characterized in that, include: The real-time data layered encoding module encodes sales data in layers according to time response requirements and dynamically adjusts the data collaborative compression channels between layers based on system load. The user-product interaction matrix construction module constructs a sparse interaction matrix based on user behavior records in the hierarchical coding. For non-interactive data items, it uses the average rating of similar products as a benchmark and fills in missing values ​​to form a complete user-product rating matrix. The incremental matrix optimization module performs alternating least squares decomposition on the user-product rating matrix to obtain the latent feature matrix. The hierarchical dynamic recommendation module divides user groups based on price sensitivity levels and generates recommendation scores with the latent feature matrix. For users with medium to high sensitivity, the module adds weight coefficients for discounted products to form the final recommendation list.

2. The e-commerce platform product sales data analysis system according to claim 1, characterized in that, The hierarchical encoding includes: dividing sales data into a millisecond-level real-time layer, a minute-level aggregation layer, and an hour-level archive layer according to time response requirements; storing real-time layer data using adjacent record difference value encoding; and compressing archive layer data through tensor decomposition for dimensionality reduction.

3. The e-commerce platform product sales data analysis system according to claim 2, characterized in that, The dynamic adjustment includes: collecting system load status parameters, determining the pressure range of data flow between layers, and determining the data compression channel adjustment strategy; when the system load drops to the set range, restoring the original compression channel parameters.

4. The e-commerce platform product sales data analysis system according to claim 3, characterized in that, The data compression channel adjustment strategy includes: increasing the compression ratio of the archive layer under high load conditions; dynamically adjusting the data transfer rate from the real-time layer to the aggregation layer to reduce data storage pressure under high load conditions; reducing the storage time of real-time layer data; and switching the encoding of adjacent record difference values ​​to a predictive encoding method with a higher compression ratio.

5. The e-commerce platform product sales data analysis system according to claim 1, characterized in that, The construction of the sparse interaction matrix includes: extracting effective interaction behaviors from hierarchical encoded data; assigning weights to different types of interaction behaviors to generate user interaction weight scores for products; filling the sparse interaction matrix with the normalized user interaction scores for products, while keeping non-interactive items as null values.

6. The e-commerce platform product sales data analysis system according to claim 5, characterized in that, The missing value filling is performed by filling in the missing values ​​for non-interactive data items using the average rating of similar products as a benchmark.

7. The e-commerce platform product sales data analysis system according to claim 1, characterized in that, Obtaining the latent feature matrix includes: approximating the user-product rating matrix into two latent feature matrices using alternating least squares decomposition, including a user feature matrix and a product feature matrix; in the first round of optimization, keeping all parameters of the product feature matrix unchanged, and only optimizing the feature vector for each user. For each user, based on their rating data for multiple products in the rating matrix, the user's feature vector is calculated to minimize the rating error; after optimizing the user feature matrix, keeping the user feature matrix unchanged, the feature vector for each product is calculated, and the product's feature vector is adjusted based on all user data that have rated the product, so that it minimizes the rating error under the constraints of the current user feature matrix; repeating the process of first optimizing the user feature matrix and then optimizing the product feature matrix until the decrease in rating error is lower than a set threshold, or the preset maximum number of optimizations is reached.

8. The e-commerce platform product sales data analysis system according to claim 6, characterized in that, The incremental matrix optimization module also includes generating an initial user feature vector based on the set of products the user first interacts with when a new user joins, thus completing the incremental update of the user feature matrix in the cold start state.

9. The e-commerce platform product sales data analysis system according to claim 1, characterized in that, The generation of the recommendation score includes: the user's basic score for the product is calculated by the inner product of the user feature vector and the product feature vector; a price elasticity matrix is ​​set, wherein the price elasticity coefficient is calculated based on the product's historical sales data, reflecting the impact of product price changes on the user's purchase probability; different user price sensitivity weights are assigned according to different price sensitivity levels; the product of the user's price sensitivity weight and the price elasticity coefficient plus the basic score is the recommendation score.

10. A method for analyzing product sales data on an e-commerce platform, characterized in that, include: Sales data is coded in layers according to time response requirements, and the data collaborative compression channels between layers are dynamically adjusted based on system load. Based on the user behavior records in the hierarchical encoding, a sparse interaction matrix is ​​constructed and missing values ​​are filled to form a complete user-product rating matrix; the user-product rating matrix is ​​then subjected to alternating least squares decomposition to obtain a latent feature matrix. User groups are segmented based on price sensitivity levels, and recommendation scores are generated using the latent feature matrix. For users with medium to high price sensitivity, the weight coefficient of discounted products is increased to form the final recommendation list.

Images