E-commerce order data processing method
Patent Information
- Application Number
- CN202510760118.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-09
- Publication Date
- 2025-09-26
Smart Images
Figure CN120707236A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing, and in particular to a method for processing data of e-commerce orders. Background Art
[0002] E-commerce order data often exhibits significant temporal correlation, seasonal fluctuations, and complex anomaly patterns. Time series decomposition methods are widely used in the industry to process and analyze raw order data. Traditional time series decomposition methods can break down order data into trend components, seasonal components, and random fluctuation components. The trend component can reflect long-term sales growth or decline, providing important insights for formulating long-term marketing strategies and forecasting future order volumes. The seasonal component reveals cyclical fluctuations, helping to optimize inventory management, rationalize human resource allocation, and precisely design promotional campaigns.
[0003] However, in practical applications, traditional time series decomposition methods face the serious problem of being unable to accurately identify and distinguish complex anomalies. Existing technologies have the following major limitations: First, traditional methods typically rely solely on simple numerical thresholds to identify anomalies, failing to distinguish the dynamic processes and underlying mechanisms that cause them. For example, sudden anomalies caused by system failures and gradual anomalies caused by promotional activities may appear to have similar magnitudes in data, but their causes and business significance are completely different, requiring different handling strategies. Second, the causes of anomalies in e-commerce order data are diverse and vary in severity, including system failures, cyberattacks, promotional activities, and seasonal fluctuations. Traditional time series decomposition algorithms treat all outliers equally during the fitting process, either overly influencing them and leading to fitting bias, or overly filtering out anomaly data containing important business information. Furthermore, traditional methods typically use kernel functions to adjust weights based on the size and distance of the outliers when dealing with anomaly data. This mechanical approach makes it difficult to distinguish anomalies with different business significance, resulting in the incorrect filtering or inappropriate downweighting of some anomaly information with important decision-making value.
[0004] For example, related technical document CN116703534B discloses a method for intelligent management of e-commerce order data, including: obtaining a series of amounts and a series of statuses; constructing a three-dimensional time series space based on the series of amounts and statuses; recording each point in the three-dimensional time series space as a spatial data point, and determining the degree of amount anomaly based on the difference in the amount values of the spatial data points; obtaining a point to be fitted, and determining the clustering factor between the spatial data point and the point to be fitted based on the difference in the amount value and the time series value, as well as the state value; obtaining Class I and Class II data points based on the degree of amount anomaly, and combining the Class I and Class II data points with the clustering factor to determine the fitting weight; and completing intelligent management of e-commerce order data based on the fitting weight of each spatial data point. However, this solution only determines the degree of anomaly based on static amount value differences and cannot identify the dynamic evolution of order data. Summary of the Invention
[0005] In the analysis of e-commerce order data, traditional time series decomposition methods only rely on numerical thresholds to judge anomalies and cannot effectively distinguish the formation process of anomalies. This application provides an e-commerce order data processing method. By analyzing e-commerce order data based on phase features and multi-scale decomposition, the dynamic evolution characteristics and abnormal gradient areas of order data are identified, and sudden anomalies (such as system failures and network attacks) and gradual anomalies (such as promotion preheating and seasonal fluctuations) are effectively distinguished.
[0006] The present application provides a data processing method for e-commerce orders, comprising: obtaining e-commerce order data, and extracting an order amount sequence A(t), an order status sequence S(t) and an order timestamp sequence T(t) from the e-commerce order data, performing denoising on the order amount sequence to obtain a preprocessed order amount sequence A(t)'; mapping the preprocessed order amount sequence A(t)', the order status sequence S(t) and the order timestamp sequence T(t) to a three-dimensional time series space, performing phase transformation on the data points in the three-dimensional time series space, extracting the phase features of the data and calculating the phase The method is to filter the phase difference Δφ(t) to obtain a smooth phase difference Δφ(t)' sequence, and decompose the phase difference Δφ(t)' sequence into multiple sub-band sequences of different time scales through multi-scale decomposition; calculate the time rate of change of the sub-band sequence to obtain the velocity component v(t) of the order change, accumulate the velocity component, and identify abnormal gradient areas; divide the data points into different types of areas according to the abnormal gradient areas, and set weight coefficients for different types of areas; use the weight coefficients to perform weighted calculation on the order amount sequence to obtain the order amount prediction value.
[0007] Furthermore, the data's phase characteristics are extracted and the phase difference Δφ(t) is calculated. This involves constructing a three-dimensional time series space using the preprocessed order amount sequence A(t)' as the first-dimensional coordinate value, the order status sequence S(t) as the second-dimensional coordinate value, and the order timestamp sequence T(t) as the third-dimensional coordinate value. Specifically, the design for mapping the order data into the three-dimensional space (A(t)', S(t), T(t)) is based on multidimensional signal processing theory. This mapping expands the original one-dimensional time series into a three-dimensional trajectory, allowing the dynamic evolution of the order to form a specific trajectory pattern in space.
[0008] Construct a time-weighted composite signal s(t) = A(t) + j × f(S(t), T(t)), where A(t) is the order amount sequence, f(S(t), T(t)) is a function mapping that considers both order status and timestamp, and j is an imaginary unit.
[0009] f(S(t),T(t))=α×S(t)×exp(-λ×(T now -T(t))), where α is the state weight coefficient, ranging from 0.1 to 5; λ is the time decay factor, T now is the current timestamp, and j is an imaginary unit. Specifically, the real part A(t) represents the economic value dimension of the order, and the imaginary part f(S(t), T(t)) represents the state-time composite dimension of the order.
[0010] Perform Hilbert transform on the composite signal s(t) to obtain the analytical signal z(t) = s(t) + j × H[s(t)], where H[s(t)] is the Hilbert transform of s(t); calculate the instantaneous phase based on the analytical signal z(t) The instantaneous phase difference Δφ(t)=φ(t+1)-φ(t) between adjacent time points is calculated as the phase feature of the order change.
[0011] In particular, first of all, the limitation of the traditional Hilbert transform applied to a single real signal is manifested in the e-commerce scenario as the inability to simultaneously process multi-dimensional heterogeneous data. This application integrates multi-dimensional order features into the complex signal processing framework, realizing the transition from single numerical analysis to multi-dimensional phase analysis. Secondly, the phase feature is naturally robust to amplitude noise. In traditional analysis, fluctuations in small orders may be misjudged as abnormalities, while phase analysis focuses on the overall evolution characteristics of the signal and can effectively filter out such interference. Furthermore, the instantaneous phase It can capture the inherent dynamic characteristics of the signal. Even if the order amount fluctuates only slightly, if the underlying state transition pattern changes, the phase feature can still keenly capture this change. Finally, the phase difference Δφ(t) = φ(t+1) - φ(t) serves as a "speed indicator" of system dynamics, detecting subtle changes in system behavior before anomalies fully manifest. For example, a promotional activity may manifest as an increase in the frequency of order state transitions before a significant increase in the amount, and this change will be directly reflected in the phase difference sequence.
[0012] Furthermore, the phase difference Δφ(t)' sequence is decomposed into multiple sub-band sequences of different time scales through multi-scale decomposition, including: filtering the phase difference Δφ(t) with a low-pass filter to remove high-frequency noise and obtain a smooth phase difference Δφ(t)' sequence; performing multi-scale decomposition on the smooth phase difference Δφ(t)' sequence using a wavelet transform method, selecting a Daubechies wavelet basis function, and decomposing the sequence to 3 to 5 levels to obtain wavelet coefficients of multiple frequency scales; reconstructing the wavelet coefficients of each scale to obtain a sub-band sequence φ at the corresponding scale. k (t), k=1,2,.....,K, K is the number of decomposition scales; according to the energy distribution of the sub-band sequence at each scale, the weight coefficient λ is assigned k , so that λ k =E k / ∑E i , where E k Represents the energy of the kth subband sequence; synthesizes the smooth phase difference sequence Ensure λ k The value of And 0<λ k <1.
[0013] Furthermore, the abnormal gradient region is identified, including: according to each subband sequence φ k (t), calculate the corresponding time rate of change Indicates the rate of change of order data at the corresponding time scale; in particular, by the phase sequence φ k (t) Derivative, the system realizes the transformation from "phase domain" to "velocity domain", which enables the changing characteristics of the signal to be directly quantified; the change rate is calculated for each sub-band sequence separately, ensuring that the system can capture the change rate at different time scales, similar to monitoring the short-term, medium-term and long-term motion states at the same time; in addition, the derivative operation is extremely sensitive to the instantaneous changes of the signal, and can amplify the tiny change signals that may be ignored in traditional analysis, providing a mathematical basis for early anomaly detection.
[0014] The time change rate v at each time scale k (t) Perform weighted synthesis to obtain the speed component of order change Among them, w k represents the synthesis weight of the kth scale; w k =E k / ∑E i , where E k represents the kth subband sequence φ k The energy value of (t) is calculated as i ranges from 1 to the total number of scales K;
[0015] Cumulative integration of the velocity component v(t) in the time dimension The cumulative change C(t) is obtained, where t0 represents the starting time of the accumulation. The integral operation is equivalent to a low-pass filter (1 / jω) in the frequency domain, which can naturally suppress high-frequency noise and make regional judgments more stable. Abnormal events usually manifest as long-tail phenomena in statistical distributions. The cumulative integral difference can effectively amplify and identify this long-tail effect, making anomaly detection more sensitive. Traditional amplitude-based methods are easily affected by transient noise, while the integral operation can automatically smooth these disturbances. The cumulative integral particularly emphasizes continuous unidirectional changes and can distinguish between random fluctuations and directional abnormal developments. By selecting different integral starting points t0, the system can flexibly adjust the time range of interest to meet the monitoring needs of different business scenarios.
[0016] In the time window [t1, t2] to be detected, when the cumulative integral difference C(t2)-C(t1) exceeds the preset threshold θ, the corresponding time window is identified as an abnormal gradient area. The abnormal gradient area reflects sudden changes, periodic fluctuations or abnormal trading patterns in the order data. Unlike static judgment at a fixed time point, this method can adaptively detect abnormal patterns of different durations. Through the size of the cumulative integral difference, the system can distinguish between sudden anomalies (high speed in a short time) and gradual anomalies (medium and low speed over a long period of time).
[0017] In particular, traditional time series decomposition methods have significant limitations in e-commerce order analysis. This is primarily due to their reliance on simple numerical thresholds to identify anomalies, which prevents them from effectively distinguishing the formation process and type of anomalies (e.g., sudden versus gradual anomalies). This application uses cumulative integral differences to assess the cumulative effect of data changes, enabling it to distinguish between sudden and gradual anomalies, addressing the inability of traditional methods to identify the formation process of anomalies.
[0018] Furthermore, the data points are divided into different types of regions according to the abnormal gradient region, including: according to the abnormal gradient region, the data points in the three-dimensional time series space are classified into stable region, fluctuating region and abnormal region; wherein: the stable region is the region where the cumulative integral difference C(t2)-C(t1) is less than the first threshold θ1, indicating that the order data changes smoothly, corresponding to the normal daily trading mode, the order volume and state transition are in line with the historical pattern, and the system behavior is predictable; the fluctuating region is the region where the cumulative integral difference C(t2)-C(t1) is greater than or equal to the first threshold θ1 and less than the second threshold θ2, indicating that there is a fluctuation in the order data. Fixed fluctuations correspond to controlled business fluctuations, such as small promotions, weekend effects, and beginning / end-of-month effects. The abnormal area is the area where the cumulative integral difference C(t2)-C(t1) is greater than or equal to the second threshold θ2, indicating abnormal fluctuations in order data, reflecting abnormal transactions or market emergencies, which may be large-scale promotions, system failures, network attacks, etc.; and θ1 is less than θ2; θ1=μ+σ, where μ is the mean of the cumulative integral difference C(t2)-C(t1) and σ is its standard deviation; θ2=μ+2.5σ; μ+σ and μ+2.5σ correspond to confidence intervals of approximately 68% and 99%, respectively.
[0019] Set weight coefficients for different types of regions, including:
[0020] Weight coefficient of stable region Among them, β1 is the adjustment parameter of the stable region;
[0021] Weight coefficient of fluctuation area Among them, β2 is the adjustment parameter of the fluctuation area;
[0022] Weight coefficient of abnormal area Among them, β3 is the adjustment parameter of the abnormal area.
[0023] Where w1>w2>w3. The value range of β1 is [2.0, 5.0], the value range of β2 is [1.0, 3.0], and the value range of β3 is [0.5, 1.5].
[0024] Furthermore, the order amount sequence is weighted by using the weight coefficient to obtain the order amount forecast value, including: constructing the weight matrix W(t) according to the weight coefficients w1, w2, w3 of the stable area, the fluctuating area and the abnormal area, and each element in W(t) corresponds to the weight value of the time point t; calculating the weighted forecast benchmark value P(t) according to the order amount sequence A(t)' and the weight matrix W(t), Here, τ represents a historical time point within the time window [tT, t], and T is the length of the prediction window. The setting of the time window [tT, t] ensures that the prediction is based only on the most recent relevant historical data, avoiding possible interference from distant data. The weighting mechanism ensures that the prediction benchmark value P(t) mainly reflects the data characteristics of the stable area, while moderately considering the information of the fluctuating area to minimize the impact of the abnormal area.
[0025] According to the weighted prediction reference value P(t) and the time attenuation factor λ, the future time point t is predicted n+1 The order amount A(t n+1 )=P(t n )×exp(-λ×Δt)+η×C(t n ), where t n represents the current time point, Δt represents the prediction time interval, i.e. t n+1 -t n , η represents the cumulative change influence coefficient with a range of [0.1, 0.5], C(t n ) represents the period from the cumulative starting time point t0 to the current time point t n The cumulative change of The value range of λ is 0.01 to 0.5. Specifically, P(t n )×exp(-λ×Δt) reflects the time decay term based on the weighted prediction benchmark value. The exponential decay function exp(-λ×Δt) simulates the natural decay characteristics of the order amount over time, which is consistent with the time sensitivity of e-commerce order data; η×C(t n ) reflects the adjustment item based on the cumulative change, the cumulative change C(t n ) captures the time from the start time t0 to the current time t n The order change trend, cumulative change C(t n ) enables the system to capture the gradual process of order changes and distinguish between sudden anomalies and gradual anomalies.
[0026] In particular, traditional time series forecasting methods (such as ARIMA, exponential smoothing, etc.) are based on the assumption of stationarity and assign the same processing strategy to all historical data points. However, e-commerce order data is highly non-stationary and contains multiple heterogeneous components. This application uses a three-region differentiated weight mechanism to achieve accurate identification and processing of different data features: stable region w1: retain the predictive value of the normal business model; fluctuating region w2: moderate weight reduction to avoid overfitting; abnormal region w3: intelligent processing to distinguish between valuable anomalies and noise. It fundamentally solves the problem of "outlier pollution" and enables the prediction model to maintain robustness in complex environments.
[0027] Compared with the existing technology, the advantages of this application are:
[0028] By mapping order amounts, status, and timestamps into a three-dimensional time series space and applying the Hilbert transform to extract phase features, the system can capture subtle trends that are difficult to detect using traditional numerical analysis methods. Furthermore, phase domain analysis reduces reliance on fixed numerical thresholds, making anomaly detection more adaptive and maintaining stable detection performance across diverse business scenarios and data sizes, reducing the complexity of parameter tuning. Furthermore, phase features effectively characterize the nonlinear dynamic behavior of order data, enabling the system to identify complex order patterns that are invisible to traditional methods, such as those caused by seasonality and promotional effects, and the overlay of multiple periods. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The present application will be further described in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same numbers represent the same structures, wherein:
[0030] Figure 1 is an exemplary flow chart of a data processing method for an e-commerce order according to some embodiments of the present application;
[0031] Figure 2 is an exemplary flow chart of gradient division of abnormal regions according to some embodiments of the present application;
[0032] Figure 3 This is a comparison diagram of the phase analysis warning effect shown in some embodiments of the present application;
[0033] Figure 4 It is a schematic diagram of cumulative points according to some embodiments of the present application.
[0034] Description of the numbers in the figure: DETAILED DESCRIPTION
[0035] The method and system provided in the embodiments of the present application are described in detail below with reference to the accompanying drawings.
[0036] like Figure 1As shown, e-commerce order data is obtained, and the order amount sequence A(t), order status sequence S(t) and order timestamp sequence T(t) are extracted from the e-commerce order data. The order amount sequence is denoised to obtain the preprocessed order amount sequence A(t)'; the preprocessed order amount sequence A(t)', order status sequence S(t) and order timestamp sequence T(t) are mapped to a three-dimensional time series space, and phase transformation is performed on the data points in the three-dimensional time series space to extract the phase features of the data and calculate the phase difference Δφ(t); the phase difference Δφ(t) is filtered to obtain a smooth phase difference Δφ(t)' sequence, and the phase difference Δφ(t)' sequence is decomposed into multiple sub-band sequences of different time scales through multi-scale decomposition; the time change rate of the sub-band sequence is calculated to obtain the velocity component v(t) of the order change, the velocity component is accumulated and calculated, and the abnormal gradient area is identified; the data points are divided into different types of areas according to the abnormal gradient area, and weight coefficients of different types of areas are set; the order amount sequence is weighted using the weight coefficient to obtain the order amount prediction value.
[0037] The system extracted order data from the transaction database for 37 days, from May 20 to June 25, 2024, at an hourly granularity, forming a continuous sequence of 888 time points. Each raw record contains key fields such as order ID, order amount, creation time, current status, and product category.
[0038] The data collection process uses an incremental extraction strategy, with hourly batch synchronization combined with a 5-minute real-time data stream to ensure the analysis system has timely access to the latest data. The data volume is massive (nearly 100,000 orders per hour during peak periods), so a distributed processing framework is used for pre-processing.
[0039] Three key sequences are constructed from the original data: Order amount sequence A(t): The order amount at each time point is aggregated. For example, the total order amount at 0:00 on June 1 was 1,120,000 yuan (698 orders), while the order amount at 20:00 on June 18 surged to 12,850,000 yuan (8,950 orders).
[0040] Order Status Sequence S(t): Order status is quantified into a value from 1 to 7, where 1 = unpaid, 2 = paid and awaiting shipment, 3 = shipped and in transit, 4 = received, 5 = refund requested, 6 = refunded, and 7 = completed. A weighted average is used to calculate the status value at each point in time. For example, the calculation at 0:00 on June 1st is: (1 × 50 + 2 × 280 + 3 × 200 + 4 × 150 + 5 × 10 + 6 × 8 + 7 × 0) / 698 = 2.59.
[0041] Order timestamp sequence T(t): records the Unix timestamp of each time point for subsequent time decay calculations. For example, at 0:00 on May 25th, the Unix timestamp is 1716566400, and at 20:00 on June 18th, the Unix timestamp is 1718726400.
[0042] Applying the Daubechies-4 (db4) wavelet basis to the order amount sequence A(t) through a four-layer decomposition and denoising process, for example, the standard deviation of the data was reduced from the original 135,000 yuan to 108,000 yuan, effectively eliminating random noise while preserving the true trend. For example, the original value of 950,000 yuan at a certain point in time on June 1st was processed to 942,500 yuan, correcting the data anomaly caused by the temporary rush to buy.
[0043] The system designed a specialized data structure, OrderPoint, to store three-dimensional coordinate information. Each order point is represented as: P(t) = [A(t)', S(t), T(t)]. To optimize the discriminability of status codes, the system introduced the status duration as a sub-code, forming a composite code in the "St" format. For example, "2.4" means "paid pending shipment status lasted for 4 hours." Orders during the 618 shopping festival formed a unique trajectory in three-dimensional space: the pre-sale period (June 1-15) showed a spiral upward pattern, while the explosive growth period (June 16-18) showed a vertical burst.
[0044] Calculate various parameters and divide all products on the platform into three value levels: high, medium, and low. Through two-dimensional grid search (α range [0.1-5.0], λ range [0.01-0.5]), determine the optimal α value: high-value products: α = 3.8, such as mobile phones, digital products, and high-end home appliances; medium-value products: α = 2.5, such as mid-range clothing, beauty and skin care products; low-value products: α = 1.2, such as daily necessities and low-priced food.
[0045] Through autocorrelation analysis, we determined that λ = 0.05, reducing the contribution of order information from 72 hours ago to less than 10%. Dynamic adjustments were made during each phase of 618: In the early pre-sale period (June 1-6.5), λ was reduced to 0.04 to enhance historical reference; in the late pre-sale period (June 13-6.15), λ was increased to 0.06 to reduce historical interference; during the outbreak period (June 16-6.18), λ was increased to 0.08 to focus on the latest status. For each 6-hour window, the specific implementation of the composite signal formula s(t) = A(t) + j×f(S(t), T(t)) is:
[0046] f(S(t),T(t))=α×S(t)×exp(-λ×(T now -T(t))), which fuses the three-dimensional information of amount, status and time into a unified signal.
[0047] The Hilbert transform is implemented by the FFT-IFFT method, and the calculation formula is:
[0048] H[s(t)] = IFFT(i × sgn(f) × FFT(s(t))), where sgn(f) is the frequency sign function. Through parallel computing, the Hilbert transform time for a single 6-hour window is controlled within 0.4 seconds, ensuring that the system can handle the data flow of more than 1,800 orders per minute during the peak period of June 18th.
[0049] The system uses the four-quadrant inverse tangent function atan2(H[s(t)],s(t)) to calculate the instantaneous phase and applies a phase unwrapping algorithm to ensure continuous phase variation. Phase difference calculation employs a multi-timescale strategy, simultaneously calculating 10-minute, 30-minute, and 60-minute phase differences to improve early warning sensitivity.
[0050] The early warning system sets three thresholds: Yellow warning: |Δφ'(t)|>2σ for more than 3 points; Orange warning:
[0051] |Δφ'(t)|>3σ for more than 2 points; Red warning: |Δφ'(t)|>4σ or |Δφ'(t)|>3σ for more than 4 points. Figure 3 As shown in the implementation results, phase-based early warning far outperforms traditional methods: When the yellow alert was first triggered at 12:20 AM on June 1st, order volume only increased by 3.2%. A red alert was triggered at 4:10 AM, resulting in a 12.3% increase in order volume. Traditional methods, on the other hand, could only detect changes at 10:30 AM (a 35% increase in order volume). Phase analysis detected trend changes an average of 6 to 9 hours in advance, providing valuable time for inventory adjustments.
[0052] A fourth-order Butterworth low-pass filter with a cutoff frequency of 0.05 is applied to the original phase difference sequence Δφ(t) to remove high-frequency noise. To avoid over-smoothing, a peak protection mechanism is designed: when the phase difference exceeds three standard deviations, the filter strength is automatically reduced by 50%, ensuring that important signals are not weakened.
[0053] The Daubechies-5 (db5) wavelet basis was selected for a four-level decomposition. This combination performs well for both spikes and gradual patterns in e-commerce data. The four decomposition scales correspond to different time periods: Scale 1 (6-hour period) captures short-term intraday fluctuations; Scale 2 (24-hour period) captures daily cyclical changes; Scale 3 (72-hour period) captures medium-term trend changes; and Scale 4 (168-hour period) captures long-term weekly patterns. The actual data decomposition results show that when pre-sales began on June 1st, the coefficients of Scales 1 and 2 first showed significant changes, while Scale 4 remained almost unchanged, indicating that the initial pre-sale period mainly affected short-term purchasing behavior, and long-term patterns remained unchanged.
[0054] By calculating the energy distribution of each subband sequence, the optimal weight configuration was determined: scale 1 (high frequency): E1 = 0.0856, energy accounting for 7.8%, and allocation weight λ1 = 0.078; scale 2 (sub-high frequency): E2 = 0.1892, energy accounting for 17.2%, and allocation weight λ2 = 0.172; scale 3 (sub-low frequency): E3 = 0.3685, energy accounting for 33.5%, and allocation weight λ3 = 0.335; scale 4 (low frequency): E4 = 0.4567, energy accounting for 41.5%, and allocation weight λ4 = 0.415. This energy-based weight allocation ensures maximum information retention. The smoothed phase difference sequence Δφ'(t) comprehensively reflects the dynamic characteristics of multiple time scales, improving signal quality by 22% compared with single-scale analysis.
[0055] In order to improve the accuracy of derivative estimation, the system uses the central difference method to calculate the time rate of change of each sub-band sequence: This method is more stable than forward differencing. The rates of change of the four scales reflect dynamic characteristics of different periods: the high-frequency subband v1(t) captures short-term disturbances; the low-frequency subband v4(t) reflects long-term trends. Analysis shows that starting on June 1st, v3(t) and v4(t) simultaneously showed sustained positive values for the first time, indicating a shift in medium- to long-term purchasing intentions.
[0056] The system synthesizes the rate of change of each scale using the same weight as the energy distribution:
[0057] v(t) = 0.078 × v1(t) + 0.172 × v2(t) + 0.335 × v3(t) + 0.415 × v4(t); this weighted synthesis preserves the characteristic information of each scale while highlighting the important scale with a large energy share. Analysis of speed characteristics at different stages: During the regular period, the average speed is approximately 0.02, indicating slow and stable changes; during the pre-sale period, the average speed increases to 0.15, reflecting accelerated growth; during the explosive period, the average speed reaches 0.85, indicating extremely rapid changes. After the pre-sale launch on June 1st, v(t) remained above 0.1 for the first time for 24 consecutive hours, approximately 15 days in advance of the peak order volume, demonstrating the method's early warning capability.
[0058] The system selects May 20 as the cumulative starting point t0 and uses the trapezoidal integration method to calculate the cumulative change: The cumulative values of key time nodes and their business significance are as follows: June 1: C(t) = 0.48, marking the start of the pre-sale period; June 8: C(t) = 2.85, the cumulative effect of the mid-term pre-sale; June 15: C(t) = 3.25, indicating that the explosive period is about to begin; June 18: C(t) = 8.76, the cumulative effect of the entire promotion period reaches its peak; Figure 4As shown, the cumulative integral diagram clearly shows the dynamic characteristics of each stage of 618: the curve is flat from May 20 to May 31; the slope increases from June 1 to June 15; and the slope is steep from June 16 to June 18, forming a typical "S-shaped" curve.
[0059] Through statistical analysis, the system determined the threshold parameters for anomaly detection: the mean of the cumulative integral difference within all 24-hour windows is μ = 0.35; the standard deviation is σ = 0.42, the first threshold θ1 = μ + σ = 0.77, and the second threshold θ2 = μ + 2.5σ = 1.4. Based on these thresholds, the system successfully identified three key anomaly gradient areas: the pre-sale launch area (June 1-6.3): a gradual anomaly with an average cumulative difference of 0.82; the pre-sale acceleration area (June 10-6.12): a fluctuating anomaly with an average cumulative difference of 1.12; and the main outbreak area (June 16-6.18): a sudden anomaly with a maximum cumulative difference of 2.75. This detailed anomaly area identification provides a foundation for subsequent differentiated processing, especially the successful identification of the "pre-sale acceleration area" starting on June 10, a key turning point that is easily overlooked by traditional methods.
[0060] Based on the cumulative integral difference threshold, the system accurately divides the data during the entire 618 period into three areas: stable area (C difference <0.77): May 20 to May 31, June 4 to June 9, June 19 to June 25; fluctuating area (0.77≤C difference <1.40): June 1 to June 3, June 10 to June 15; abnormal area (C difference ≥1.40): June 16 to June 18; the three areas show significantly different order characteristics: stable area: order status conversion follows the normal pattern, and the amount fluctuates within the range of ±15%; fluctuating area: the conversion from state 1 to state 2 is accelerated by 30%, and the amount fluctuation expands to ±35%; abnormal area: the state conversion speed is increased by more than 60%, and the amount fluctuation reaches 3-5 times that of normal days; this refined area division enables the system to adopt differentiated processing strategies for different data characteristics.
[0061] The system sets different adjustment parameters for the three types of areas, reflecting the concept of differentiated processing: Stable area: β1 = 4.5, ensuring that normal data receives high weight; Fluctuating area: β2 = 2, moderately reducing the weight of fluctuating data; Abnormal area: β3 = 1, significantly reducing the weight of abnormal data; Specific weight calculation example: Stable area (June 5, C difference = 0.25): Fluctuation range (June 12, C difference = 1.05): Abnormal area (June 17, C difference = 1.85): The key concept of weight design is that "stable data is more reliable". Through the relative relationship of w1>w2>w3, it ensures that the prediction is mainly based on stable data while not completely ignoring the abnormal period information.
[0062] The system integrates the regional classification and weight calculation results into a time-varying weight matrix W(t). Each time point receives a corresponding weight based on the region type to which it belongs. To ensure a smooth transition, the weight values at the region boundaries are processed using a ±6-hour Gaussian smoothing process. Actual matrix examples: May 30 (stable region): W(t) = 0.79; June 2 (fluctuating region): W(t) = 0.68; June 17 (abnormal region): W(t) = 0.52. This dynamic weight matrix design overcomes the limitations of traditional fixed weights and can adaptively process different types of data points.
[0063] Set different forecast window lengths T for different product categories: Fast Moving Consumer Goods: T = 168 hours (7 days); Durable Goods: T = 336 hours (14 days); Seasonal Goods: T = 504 hours (21 days). Taking the forecast from 20:00 on June 17 to 20:00 on June 18 as an example, select T = 168 hours and calculate the weighted forecast benchmark value: τ∈[t-168,t]. Specific step-by-step calculation: Extract the order amount sequence A(τ)' within the window [8:00 PM, June 10th, to 8:00 PM, June 17th]; apply the corresponding weight matrix W(τ); and calculate the weighted average to obtain P(t) = 5,875,000 yuan. This baseline value comprehensively reflects recent order trends and avoids the excessive influence of outliers through differentiated weights.
[0064] The key parameter settings for predicting the order amount at 20:00 on June 18 are: time decay factor λ = 0.12, reflecting the rapid changes during the 618 period; cumulative impact coefficient η = 0.35, indicating moderate reliance on historical trends; prediction time interval Δt = 24 hours.
[0065] Final prediction calculation process: Exponential decay term: 5875000×exp(-0.12×1)=5875000×0.887=5211,125 yuan; cumulative change: current C(t)=7.25; momentum correction term: 0.35×7.25×1000000=2537500 yuan; final prediction value: 5211125+2537500=7748,625 yuan; considering the 618 peak amplification effect, historical data shows that the peak period on 618 was 1.88 times the prediction benchmark. After applying this coefficient, the final prediction is: 7748625×1.88=14567415 yuan.
[0066] The system automatically converts the prediction results into specific business decision-making indicators: Inventory preparation: Calculate the required inventory amount based on the predicted order amount, adding a 10% safety factor; Logistics resources: Deploy delivery personnel and vehicles in advance based on order volume forecasts; Customer service configuration: Forecast customer service demand during peak periods and arrange human resources reasonably; Actual business application effect on June 18: The actual order amount on June 18 was 14,985,000 yuan, with a prediction error of only 2.8%; The inventory satisfaction rate increased to 98.5%, an increase of 12 percentage points over previous years; This application captures early signals of order changes through phase characteristics, and detects trend changes 12 to 24 hours in advance on average; It processes order patterns at different time scales at the same time, avoiding the limitations of single-scale analysis.
[0067] The invention of the present application and its implementation methods are described schematically above. This description is not restrictive. Without departing from the spirit or basic features of the present application, the present application can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention of the present application, and the actual structure is not limited to this. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and embodiment similar to the technical solution are designed without creativity, which should all fall within the scope of protection of the present application. In addition, the word "including" does not exclude other elements or steps, and the word "one" before an element does not exclude the inclusion of "multiple" elements. Words such as first and second are used to indicate names and do not indicate any specific order.
Claims
1. A data processing method for e-commerce orders, characterized in that: include: Obtain e-commerce order data, and extract the order amount sequence A(t), order status sequence S(t), and order timestamp sequence T(t) from the e-commerce order data. Perform denoising on the order amount sequence to obtain the preprocessed order amount sequence A(t)'. Map the preprocessed order amount sequence A(t)', order status sequence S(t), and order timestamp sequence T(t) into a three-dimensional time series space, perform phase transformation on the data points in the three-dimensional time series space, extract the phase features of the data, and calculate the phase difference Δφ(t); The phase difference Δφ(t) is filtered to obtain a smooth phase difference Δφ(t)' sequence, and the phase difference Δφ(t)' sequence is decomposed into multiple sub-band sequences of different time scales through multi-scale decomposition; Calculate the time rate of change of the sub-band sequence to obtain the velocity component v(t) of the order change, perform cumulative calculation on the velocity component, and identify the abnormal gradient area; Divide the data points into different types of areas according to the abnormal gradient area, and set the weight coefficients of different types of areas; The order amount sequence is weighted by using the weight coefficient to obtain the order amount forecast value.
2. The method for processing e-commerce order data according to claim 1, characterized in that: Extract the phase characteristics of the data and calculate the phase difference Δφ(t), including: The pre-processed order amount sequence A(t)' is used as the first dimension coordinate value, the order status sequence S(t) as the second dimension coordinate value, and the order timestamp sequence T(t) as the third dimension coordinate value to construct a three-dimensional time series space; Construct a time-weighted composite signal s(t) = A(t) + j × f(S(t), T(t)), where A(t) is the order amount sequence, f(S(t), T(t)) is a function mapping that considers both order status and timestamp, and j is an imaginary unit. Performing Hilbert transform on the composite signal s(t) yields the analytical signal z(t)=s(t)+j×H[s(t)], where H[s(t)] is the Hilbert transform of H[s(t)]; Calculate the instantaneous phase based on the analytical signal z(t) The instantaneous phase difference Δφ(t)=φ(t+1)-φ(t) between adjacent time points is calculated as the phase feature of the order change.
3. The method for processing e-commerce order data according to claim 2, characterized in that: f(S(t),T(t))=α×S(t)×exp(-λ×(T now -T(t))), where α is the state weight coefficient; λ is the time decay factor, T now is the current timestamp, and j is the imaginary unit.
4. The method for processing e-commerce order data according to claim 2, characterized in that: The phase difference Δφ(t)' sequence is decomposed into multiple sub-band sequences of different time scales through multi-scale decomposition, including: Among them, λ k is the weight coefficient of the kth scale, φ k (t), k = 1, 2, ....., K is the subband sequence at the kth scale.
5. The method for processing e-commerce order data according to claim 4, characterized in that: Identify areas of anomalous gradients, including: According to each subband sequence φ k (t), calculate the corresponding time rate of change Indicates the rate of change of order data in the corresponding time scale; The time change rate v at each time scale k (t) Perform weighted synthesis to obtain the speed component of order change Among them, w k represents the synthesis weight of the kth scale; w k =E k / ∑E i , where E k represents the kth subband sequence φ k The energy value of (t) is calculated as i ranges from 1 to the total number of scales K; Cumulative integration of the velocity component v(t) in the time dimension Obtain the cumulative change C(t); where t0 represents the starting time point of accumulation; In the time window to be detected [t1, t2], when the cumulative integral difference C(t2)-C(t1) exceeds the preset threshold θ, the corresponding time window is identified as an abnormal gradient area. The abnormal gradient area reflects sudden changes, periodic fluctuations or abnormal trading patterns in the order data.
6. The method for processing e-commerce order data according to claim 5, characterized in that: The data points are divided into different types of areas according to the abnormal gradient areas, including: According to the abnormal gradient area, the data points in the three-dimensional time series space are classified into stable area, fluctuating area and abnormal area; in: The stable region is the region where the cumulative integral difference C(t2)-C(t1) is less than the first threshold θ1, indicating that the order data changes steadily, reflecting a normal trading pattern; The fluctuation area is the area where the cumulative integral difference C(t2)-C(t1) is greater than or equal to the first threshold θ1 and less than the second threshold θ2, indicating that there is a certain fluctuation in the order data; The abnormal area is the area where the cumulative integral difference C(t2)-C(t1) is greater than or equal to the second threshold θ2, indicating abnormal fluctuations in order data, reflecting abnormal transactions or market emergencies; and θ1 is less than θ2; θ1 = μ + σ, where μ is the mean of the cumulative integral difference C(t2)-C(t1) and σ is its standard deviation; θ2 = μ + 2.5σ; The weight coefficients of stable area, fluctuating area and abnormal area are set respectively.
7. The method for processing e-commerce order data according to claim 6, characterized in that: Set weight coefficients for different types of regions, including: Weight coefficient of stable region Among them, β1 is the adjustment parameter of the stable region; Weight coefficient of fluctuation area Among them, β2 is the adjustment parameter of the fluctuation area; Weight coefficient of abnormal area Among them, β3 is the adjustment parameter of the abnormal area; Among them, w1>w2>w3.
8. The method for processing e-commerce order data according to claim 7, characterized in that: The value range of β1 is [2.0, 5.0], the value range of β2 is [1.0, 3.0], and the value range of β3 is [0.5, 1.5].
9. The method for processing e-commerce order data according to claim 7, characterized in that: The order amount series is weighted by using the weight coefficient to obtain the order amount forecast value, including: Based on the weight coefficients w1, w2, w3 for stable, fluctuating, and abnormal regions, a weight matrix W(t) is constructed. Each element in W(t) corresponds to the weight value at time point t. The weight matrix W(t) allows the system to differentiate data from different types of regions, making predictions more dependent on data from stable regions. According to the order amount sequence A(t)' and the weight matrix W(t), the weighted prediction benchmark value P(t) is calculated. Where τ represents the historical time point in the [tT, t] time window, and T is the length of the prediction window; According to the weighted prediction reference value P(t) and the time attenuation factor λ, the future time point t is predicted n+1 The order amount A(t n+1 )=P(t n )×exp(-λ×Δt)+η×C(t n ), where t n represents the current time point, Δt represents the prediction time interval, i.e. t n+1 -t n , η represents the cumulative change influence coefficient with a range of [0.1, 0.5], C(t n ) represents the period from the cumulative starting time point t0 to the current time point t n The cumulative change of 10. The method for processing e-commerce order data according to claim 9, characterized in that: The value range of λ is 0.01 to 0.5.