Trade article selection recommendation method based on multi-source data analysis

Abstract

The invention discloses a trade article selection recommendation method based on multi-source data analysis, and the method comprises the steps: collecting multi-source trade data, carrying out the combination and collection according to a commodity-market, and generating a time sequence; constructing a trade behavior state space, identifying a state transition relationship, and generating an evolution path; detecting path consistency, calculating data credibility, and fusing to form a reconstruction time sequence; identifying a market event, and calculating a response displacement value of each index to the event; dynamically translating the time axis according to the response displacement, and aligning the multi-source time sequence; and constructing multivariable input, predicting opportunities by using the improved Times FM, and generating a recommendation result. According to the invention, by constructing a multi-source data credibility reconstruction mechanism, event-driven time alignment processing and an improved time sequence prediction model, accurate recommendation and foresight evaluation of a commodity-market combination in a trade commodity selection process are realized.

Description

Technical Field

[0001] This invention relates to the fields of data intelligence and e-commerce technology, and in particular to a trade product selection and recommendation method based on multi-source data analysis. Background Technology

[0002] With the rapid development of cross-border e-commerce and digital trade, trading entities are increasingly relying on data-driven decision-making in product selection and market layout. Existing trade product selection technologies typically analyze and evaluate product performance in different markets by collecting transaction records, user behavior data, supply chain information, and some market environment data, thereby assisting in product selection decisions. These technologies often rely on historical sales statistics, keyword popularity analysis, or simple time series forecasting methods, which can reflect a product's past sales performance and short-term trends to a certain extent, and have been widely deployed in e-commerce platforms and trade management systems in practice.

[0003] However, existing technologies generally have limitations when processing multi-source trade data. On the one hand, different data sources vary significantly in terms of collection criteria, update frequency, and data quality. Existing methods typically use fixed weights or simple concatenation to fuse multi-source data, making it difficult to accurately identify the validity of each data source at the commodity-market dimension and susceptible to interference from anomalous or locally distorted data. On the other hand, existing product selection methods often assume that different data sources are synchronous in the time dimension, ignoring the objective fact that indicators such as search behavior, transaction behavior, and inventory changes respond to market events in advance or with lag. This leads to unreasonable alignment of multi-source time series, thereby affecting the accuracy and stability of product selection results.

[0004] For product selection scenarios involving new commodities or new markets, existing technologies typically rely on historical data from a single market or commodity for prediction, lacking systematic modeling of multi-source information and temporal evolution characteristics. This makes it difficult to reliably assess future trade opportunities under data sparsity conditions. Existing rule-based or traditional prediction model-based product selection methods lag behind in response to frequently changing market environments, making it difficult to balance the foresight and robustness of product selection decisions.

[0005] Therefore, how to provide a trade product selection recommendation method based on multi-source data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a trade product selection recommendation method based on multi-source data analysis. This invention introduces a unified collection and analysis mechanism for multi-source trade data, constructs a trade behavior state space, and generates trade behavior evolution paths along the commodity-market dimension. Based on this, it achieves the judgment and reconstruction of the credibility of multi-source data. Furthermore, it combines event-driven time series dynamic alignment and an improved time series prediction model to comprehensively evaluate the future trade opportunities of commodities in different markets, thereby completing the trade product selection recommendation. This invention fully utilizes multi-source data fusion analysis, time series modeling, and intelligent prediction technologies, effectively reducing the impact of abnormal data and time response differences on product selection results. It possesses advantages such as high accuracy in product selection decisions, timely response to market changes, and strong adaptability to new commodities and new markets.

[0007] A trade product selection recommendation method based on multi-source data analysis according to an embodiment of the present invention includes: Collect multi-source trade data related to trade product selection, and aggregate the multi-source trade data according to the commodity-market combination to form corresponding multi-source time series data; Based on multi-source time series data, a trade behavior state space is constructed, multiple indicators reflecting trade activities are mapped as state variables, the ordered state transition relationships between each state variable are identified, and the trade behavior evolution path corresponding to the commodity-market combination is generated based on the state variables and transition relationships. Based on the evolution path of trade behavior, consistency detection is performed on the evolution paths corresponding to different data sources, and data credibility weights of each data source under the commodity-market combination are generated. Multi-source time series data are weighted and fused to obtain credible reconstructed multi-source trade time series data. Based on multi-source trade time series data, we identify market events that affect commodity-market combination trade activities, and calculate the response shift of corresponding indicators relative to the time of occurrence of market events for each data source. Based on response displacement, dynamic time series shifting is performed on multi-source trade time series data to form event-driven aligned multi-source trade time series data. Based on multi-source trade time series data aligned by event-driven methods, a multivariate trade time series input at the commodity-market level is constructed. An improved TimesFM model is used to predict the multivariate trade time series input to obtain the trade opportunity assessment results of the commodity-market combination. Based on the trade opportunity assessment results, trade product selection recommendation results are generated.

[0008] Optionally, the multi-source trade data includes transaction data, market behavior data, supply chain fulfillment data, and market environment data. The transaction data includes historical sales volume, transaction amount, order quantity, and return rate of goods in the target market. The market behavior data includes user search volume, click volume, add-to-cart volume, favorites volume, and number of reviews. The supply chain fulfillment data includes changes in inventory quantity, replenishment cycle, delivery time, and logistics costs. The market environment data includes exchange rate change information, tariff or regulatory policy change information, promotional activity information, and holiday information.

[0009] Optionally, the aggregation of multi-source trade data according to the commodity-market combination specifically includes: Using product identifiers and target market identifiers as a joint index, trade data from different data sources are matched and associated. Data with the same product identifier and target market identifier are divided into the same data set, and the data in the data set are sorted and organized according to the timestamp to form multi-source time series data of corresponding product-market combinations.

[0010] Optionally, the generation of the trade behavior evolution path corresponding to the commodity-market combination based on state variables and transition relationships includes: For the same commodity-market combination in each data source, the multi-source time series data are organized in chronological order within a preset analysis time window to determine several indicators to characterize the trade activities of the commodity-market combination and form a continuous indicator sequence. Based on the indicator sequence, the first layer of basic indicator state space of trade behavior is constructed. For each indicator, according to the preset numerical range threshold and change direction, the value of the indicator at each time point is labeled as several discrete basic indicator states, generating a basic indicator state sequence that changes with time. Based on the basic indicator state sequence, a second behavioral stage state layer is constructed in the trade behavior state space. At the same time point, according to the preset mapping rules, multiple basic indicator states are combined and mapped to a single behavioral stage state corresponding to the time point, forming a behavioral stage state sequence that changes over time. Based on the state sequence of behavioral stages, a third evolutionary pattern state layer is constructed in the state space of trade behavior. The duration and state switching frequency of each behavioral stage state are statistically analyzed within a sliding time window. The distribution of behavioral stage states within each time window is classified into preset evolutionary pattern states, generating an evolutionary pattern state sequence that changes over time. Based on the state sequence of basic indicators, the state sequence of behavioral stages, and the state sequence of evolution patterns, the three-layer state corresponding to each time point of the same commodity-market combination in the same data source within the analysis time window is connected sequentially in chronological order to generate the trade behavior evolution path.

[0011] Optionally, obtaining the reliable reconstructed multi-source trade time series data includes: For the same commodity-market combination, trade behavior evolution paths are obtained from various data sources, and each trade behavior evolution path is associated with the corresponding data source identifier to form a set of trade behavior evolution paths of the commodity-market combination under different data sources. Based on the set of trade behavior evolution paths, the analysis time window is divided into several continuous time segments. Within each time segment, the state sequence of the trade behavior evolution path corresponding to each data source is statistically analyzed at three levels: basic indicator state, behavior stage state, and evolution mode state. The state sequences of different data sources within the same time segment are compared, and time segments with consistent state sequences and time segments with inconsistent state sequences are marked. Based on the consistency marking results of each data source in each time segment, the number of consistent time segments and inconsistent time segments of the data source within the analysis time window are statistically analyzed. According to the ratio between the number of consistent time segments and the total number of time segments, a corresponding data credibility weight is generated for each data source. Based on data credibility weights, a weighted calculation is performed on the multi-source time series data corresponding to the same commodity-market combination at each time point. The values ​​of the same indicator from different data sources are weighted and superimposed according to the data credibility weights to obtain the weighted fusion indicator value at the time point. The weighted fusion index values ​​obtained at each time point within the analysis time window are arranged in chronological order to generate reliable reconstructed multi-source trade time series data for the corresponding commodity-market combination.

[0012] Optionally, the calculation of the response shift of the corresponding indicator relative to the time of the market event for each data source specifically involves: Based on the reliably reconstructed multi-source trade time series data, a response analysis time window is set for each commodity-market combination. Within the response analysis time window, an indicator sequence related to changes in the market environment is selected for the identification of market events. Within the response analysis time window, the magnitude and frequency of changes in the indicator series related to exchange rate changes, tariff or regulatory policy changes, promotional activities and holidays are detected. When the magnitude or frequency of change at a time point meets the preset event judgment conditions, the time point is marked as the market event occurrence time of the corresponding commodity-market combination, and the time point is associated with the corresponding event type. For each commodity-market combination, within a preset time range before and after the occurrence of the market event, extract indicator sequences of transaction data, market behavior data, and supply chain fulfillment data corresponding to each data source from the reliably reconstructed multi-source trade time series data, and determine the start time, peak time, or inflection point time of change relative to the baseline level for each indicator sequence. The time difference between the determined start time, peak time, or inflection point time of change and the corresponding market event occurrence time is calculated to obtain the response time difference of each indicator in each data source relative to the market event occurrence time. The response time difference is recorded as the response displacement of the indicator relative to the market event. Statistical processing is performed on the response displacements of each indicator in the same data source to generate summary response displacement information of the data source under the corresponding commodity-market combination.

[0013] Optionally, the multi-source trade time series data that forms the event-driven aligned data includes: Obtain reliable reconstructed multi-source trade time series data, as well as aggregated response displacement information generated for each data source for the same commodity-market combination, and read the time displacement corresponding to the data source in the aggregated response displacement information of each data source; For the same commodity-market combination, within a unified analysis time window, based on the reliable reconstructed multi-source trade time series data, the indicator sequences of transaction data, market behavior data, and supply chain fulfillment data of each data source are selected respectively. The time displacement corresponding to the data source is applied to the indicator sequence, and the time axis is shifted for each time point in the indicator sequence according to the time displacement to obtain the trade time series data within the data source after time shifting. After completing the time axis shifting process within each data source, for the same commodity-market combination, time point alignment processing is performed on all time-shifted trade time series data within the data source on a unified time coordinate. Values ​​from different data sources that correspond to the same time point and the same indicator are merged into a unified time index, forming a multi-source aligned indicator sequence on a unified time axis. For time gaps that occur during time axis translation and time point alignment, interpolation is used to fill the gaps based on the effective index values ​​of adjacent time points to mark the missing data. The multi-source aligned indicator sequences formed on a unified time axis will be aggregated according to the commodity-market combination to generate multi-source trade time series data that is event-driven aligned with the commodity-market combination.

[0014] Optionally, generating trade product recommendation results based on the trade opportunity assessment results includes: Based on multi-source trade time series data aligned by event-driven processes, preset indicators are extracted according to commodity-market combinations. The values ​​of each indicator for the same commodity-market combination on a unified time axis are arranged in order to construct a multivariate trade time series input set corresponding to the commodity-market combination. The multivariate trade time series input set is preprocessed by performing missing value imputation, outlier correction and numerical normalization. The preprocessed series is then segmented according to the historical observation window and the prediction window to generate the input series of the improved TimesFM model. An improved TimesFM model is used to predict the input sequence. This improved TimesFM model, from bottom to top, includes a basic time series encoding layer, a cross-market correlation attention layer, a multi-scale seasonal adapter layer, and an opportunity score output layer, wherein: The basic time series coding layer has the same structure as the original TimesFM coding layer and is used to perform initial feature encoding on the input sequence; The cross-market association attention layer is located after the basic time series encoding layer. It is used to perform cross-series attention operations on the same product sequences from different markets within the same batch. The cross-market association attention layer is connected to the upstream encoding layer through a residual connection. The multi-scale seasonal adapter layer is located after the cross-market association attention layer and is used to adjust the seasonal pattern of the encoding results at three time scales: weekly, monthly and quarterly. The multi-scale seasonal adapter layer is connected to the cross-market association attention layer through layer normalization. The opportunity score output layer is located after the multi-scale seasonal adapter layer and is used to generate a multi-step forecast sequence of the target indicator and output a trade opportunity score for the commodity-market combination. The training of the improved TimesFM model includes a pre-training phase and a fine-tuning phase. The pre-training phase uses historical aligned data divided by commodity-market combination to perform self-supervised learning, and the fine-tuning phase uses labeled target metrics of commodity-market combination to perform supervised learning and update the TimesFM model parameters. The multi-step forecast sequence is weighted and aggregated in the time dimension and adjusted in combination with risk parameters to generate the corresponding commodity-market combination trade opportunity assessment results; The trade opportunity assessment results of each commodity-market combination are used as the ranking criteria to rank and screen all candidate commodity-market combinations, and output the recommended trade products.

[0015] The beneficial effects of this invention are: This invention unifies the collection and analysis of multi-source trade data within a commodity-market dimension, introducing a method for constructing a trade behavior state space and trade behavior evolution path. This enables an effective characterization of the consistency of trade behavior reflected by different data sources, thereby accurately identifying the validity of multi-source data and reconstructing its credibility during product selection decisions. Compared to existing product selection methods that rely on fixed weights or simple splicing, this invention reduces the interference of outlier, noisy, or locally distorted data on product selection results, improving the stability and reliability of multi-source data fusion results.

[0016] This invention identifies market events and calculates the response displacement of each data source to these events, performing event-driven dynamic alignment processing on multi-source trade time series. This enables data from different sources to truly reflect the order of responses to the same market changes in the time dimension, avoiding the time series mismatch problem caused by simple timestamp alignment in existing technologies. This allows trend analysis and forecasting to be based on time series that better conform to the actual evolution of trade, improving the speed and accuracy of product selection analysis results in responding to market changes.

[0017] Based on reliable and time-aligned multi-source data, this invention employs an improved time series forecasting model to assess the future trade performance of commodities in target markets. It can provide stable trade opportunity judgments even with limited historical data, effectively enhancing product selection capabilities for new commodities or new market scenarios. Through the synergistic effect of the aforementioned technical means, this invention not only improves the foresight and robustness of trade product selection recommendations but also enhances the adaptability of product selection decisions in complex and ever-changing market environments, demonstrating significant practical application value. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a trade product selection recommendation method based on multi-source data analysis proposed in this invention; Figure 2 This is a schematic diagram of the trade opportunity assessment structure based on the improved TimesFM model, which is a trade product selection recommendation method based on multi-source data analysis proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2A trade product selection recommendation method based on multi-source data analysis includes: Collect multi-source trade data related to trade product selection, and aggregate the multi-source trade data according to the commodity-market combination to form corresponding multi-source time series data; Based on multi-source time series data, a trade behavior state space is constructed, multiple indicators reflecting trade activities are mapped as state variables, the ordered state transition relationships between each state variable are identified, and the trade behavior evolution path corresponding to the commodity-market combination is generated based on the state variables and transition relationships. Based on the evolution path of trade behavior, consistency detection is performed on the evolution paths corresponding to different data sources, and data credibility weights of each data source under the commodity-market combination are generated. Multi-source time series data are weighted and fused to obtain credible reconstructed multi-source trade time series data. Based on multi-source trade time series data, we identify market events that affect commodity-market combination trade activities, and calculate the response shift of corresponding indicators relative to the time of occurrence of market events for each data source. Based on response displacement, dynamic time series shifting is performed on multi-source trade time series data to form event-driven aligned multi-source trade time series data. Based on multi-source trade time series data aligned by event-driven methods, a multivariate trade time series input at the commodity-market level is constructed. An improved TimesFM model is used to predict the multivariate trade time series input to obtain the trade opportunity assessment results of the commodity-market combination. Based on the trade opportunity assessment results, trade product selection recommendation results are generated.

[0021] In this embodiment, the multi-source trade data includes transaction data, market behavior data, supply chain fulfillment data, and market environment data. The transaction data includes historical sales volume, transaction amount, order quantity, and return rate of goods in the target market. The market behavior data includes user search volume, click volume, add-to-cart volume, favorites volume, and number of reviews. The supply chain fulfillment data includes changes in inventory quantity, replenishment cycle, delivery time, and logistics costs. The market environment data includes exchange rate change information, tariff or regulatory policy change information, promotional activity information, and holiday information.

[0022] In this embodiment, the collection of multi-source trade data according to the commodity-market combination specifically includes: Using product identifiers and target market identifiers as a joint index, trade data from different data sources are matched and associated. Data with the same product identifier and target market identifier are divided into the same data set, and the data in the data set are sorted and organized according to the timestamp to form multi-source time series data of corresponding product-market combinations.

[0023] In this embodiment, the generation of the trade behavior evolution path corresponding to the commodity-market combination based on state variables and transition relationships includes: For the multi-source time series data of the same commodity-market combination in each data source, the data is organized in chronological order within a preset analysis time window to determine several indicators used to characterize the trade activities of the commodity-market combination and form a continuous indicator sequence. The preset analysis time window is a continuous 30-day time window, with the start time being 30 days before the benchmark date for product selection evaluation of the commodity-market combination and the end time being the day before the benchmark date. Based on the indicator sequence, the first layer of basic indicator state space is constructed in the trade behavior state space. For each indicator, according to the preset numerical range threshold and direction of change, the value of the indicator at each time point is labeled as several discrete basic indicator states, generating a basic indicator state sequence that changes over time. The preset numerical range threshold is: For each indicator, the historical values ​​of the indicator within the preset analysis time window are used as statistical samples. The indicator values ​​are divided into three numerical ranges according to the quantile method. The low value range is the range that is no higher than the 30th percentile of the historical value of the indicator, the middle value range is the range that is higher than the 30th percentile but no higher than the 70th percentile, and the high value range is the range that is higher than the 70th percentile. The direction of change is determined based on the difference between the values ​​of the indicator at two adjacent time points. When the change in the value at the current time point relative to the previous time point is greater than the positive change threshold, it is determined to be an upward direction. When the change is less than the negative change threshold, it is determined to be a downward direction. All other cases are determined to be a stable direction. The positive change threshold and the negative change threshold are both set to 5% of the average value of the indicator within a preset analysis time window. Based on the basic indicator state sequence, a second behavioral stage state layer is constructed in the trade behavior state space. At the same point in time, according to a preset mapping rule, multiple basic indicator states are combined and mapped to a single behavioral stage state corresponding to that point in time, forming a behavioral stage state sequence that changes over time. The preset mapping rule is as follows: When the basic indicator corresponding to search volume is in a low or medium range and is trending upward, while the basic indicator corresponding to order volume is in a low or medium range and is trending steadily, the time point will be mapped to the stage of market attention accumulation. When the basic indicator corresponding to search volume is in the middle or high range and is on the rise, and the basic indicator corresponding to order volume is in the middle or high range and is on the rise, the time point is mapped to the demand amplification stage. When the basic indicator status corresponding to the order volume is in the high range and is stable or rising, and the basic indicator status corresponding to the inventory change is in the medium range or low range and is falling, the time point is mapped to the order conversion and fulfillment feedback stage. Other time points that do not meet the above conditions are uniformly mapped to the transition stage, forming a behavioral stage state sequence that changes over time. Based on the sequence of behavioral stages and states, a third evolutionary pattern state layer is constructed in the trade behavior state space. Within a sliding time window, the duration and state transition frequency of each behavioral stage state are statistically analyzed. The distribution of behavioral stage states within each time window is categorized into preset evolutionary pattern states, generating a time-varying sequence of evolutionary pattern states. The preset evolutionary pattern states are: When the duration of the demand amplification stage and the order conversion and fulfillment feedback stage exceeds the preset proportion threshold within the sliding time window, and the switching frequency of the behavior stage status is low, the distribution of the behavior stage status corresponding to the time window is classified as a continuous growth mode. When the duration of the market attention accumulation phase or transition phase exceeds the preset threshold within the sliding time window, and the order conversion and fulfillment feedback phases occur less frequently, the distribution of the behavioral phases corresponding to the time window will be classified as a continuous decline mode. When different behavioral stage states alternate within a sliding time window, and the state switching frequency is within a preset frequency range, the distribution of behavioral stage states corresponding to the time window is classified as a periodic fluctuation pattern. When a behavior stage state appears rapidly and quickly switches to another behavior stage state within a short period of time within a sliding time window, and the state switching frequency is significantly higher than the preset frequency threshold, the distribution of behavior stage states corresponding to the time window is classified as a sudden abnormal mode, and an evolutionary mode state sequence that changes over time is generated. Based on the state sequence of basic indicators, the state sequence of behavioral stages, and the state sequence of evolution patterns, the three-layer state corresponding to each time point of the same commodity-market combination in the same data source within the analysis time window is connected sequentially in chronological order to generate the trade behavior evolution path.

[0024] In this embodiment, obtaining the reliable reconstructed multi-source trade time series data includes: For the same commodity-market combination, trade behavior evolution paths are obtained from various data sources, and each trade behavior evolution path is associated with the corresponding data source identifier to form a set of trade behavior evolution paths of the commodity-market combination under different data sources. Based on the set of trade behavior evolution paths, the analysis time window is divided into several continuous time segments. Within each time segment, the state sequence of the trade behavior evolution path corresponding to each data source is statistically analyzed at three levels: basic indicator state, behavior stage state, and evolution mode state. The state sequences of different data sources within the same time segment are compared, and time segments with consistent state sequences and time segments with inconsistent state sequences are marked. Based on the consistency marking results of each data source in each time segment, the number of consistent time segments and inconsistent time segments of the data source within the analysis time window are statistically analyzed. According to the ratio between the number of consistent time segments and the total number of time segments, a corresponding data credibility weight is generated for each data source. Based on data credibility weights, a weighted calculation is performed on the multi-source time series data corresponding to the same commodity-market combination at each time point. The values ​​of the same indicator from different data sources are weighted and superimposed according to the data credibility weights to obtain the weighted fusion indicator value at the time point. The weighted fusion index values ​​obtained at each time point within the analysis time window are arranged in chronological order to generate reliable reconstructed multi-source trade time series data for the corresponding commodity-market combination.

[0025] In this embodiment, the calculation of the response shift of the corresponding indicator relative to the time of the market event for each data source specifically involves: Based on the reliably reconstructed multi-source trade time series data, a response analysis time window is set for each commodity-market combination. Within the response analysis time window, an indicator sequence related to changes in the market environment is selected for the identification of market events. Within the response analysis time window, the magnitude and frequency of changes in indicator sequences related to exchange rate fluctuations, tariff or regulatory policy changes, promotional activities, and holidays are detected. When the magnitude or frequency of change at a given time point meets preset event determination criteria, the time point is marked as the market event occurrence time for the corresponding commodity-market combination, and the time point is associated with the corresponding event type. The preset event determination criteria are as follows: Within the response analysis time window, when any indicator related to exchange rate changes, tariff or regulatory policy changes, promotional activities or holidays changes by more than 10% relative to the average level of the indicator within the analysis time window at a single point in time, or when there are unidirectional changes in three consecutive adjacent points in time and the cumulative change exceeds 15%, the point in time that meets the event determination criteria is marked as the market event occurrence time of the corresponding commodity-market combination. The point in time is associated with an exchange rate event, policy event, promotional event or holiday event according to the type of indicator that triggered the change. For each commodity-market combination, within a preset time range before and after the occurrence of the market event, indicator sequences of transaction data, market behavior data, and supply chain fulfillment data corresponding to each data source are extracted from the reliably reconstructed multi-source trade time series data. For each indicator sequence, the start time, peak time, or inflection point time of change relative to the baseline level is determined. The preset time range is centered on the occurrence time of the market event, with 14 consecutive days before the event as the observation interval before the event and 14 consecutive days after the event as the observation interval after the event, thus forming an analysis time range covering a total of 28 days before and after the occurrence of the market event, which is used to extract the start time, peak time, or inflection point time of change of each indicator relative to the baseline level. The time difference between the determined start time, peak time, or inflection point time of change and the corresponding market event occurrence time is calculated to obtain the response time difference of each indicator in each data source relative to the market event occurrence time. The response time difference is recorded as the response displacement of the indicator relative to the market event. Statistical processing is performed on the response displacements of each indicator in the same data source to generate summary response displacement information of the data source under the corresponding commodity-market combination.

[0026] In this embodiment, the formation of event-driven aligned multi-source trade time series data includes: Obtain reliable reconstructed multi-source trade time series data, as well as aggregated response displacement information generated for each data source for the same commodity-market combination, and read the time displacement corresponding to the data source in the aggregated response displacement information of each data source; For the same commodity-market combination, within a unified analysis time window, based on the reliable reconstructed multi-source trade time series data, the indicator sequences of transaction data, market behavior data, and supply chain fulfillment data of each data source are selected respectively. The time displacement corresponding to the data source is applied to the indicator sequence, and the time axis is shifted for each time point in the indicator sequence according to the time displacement to obtain the trade time series data within the data source after time shifting. After completing the time axis shifting process within each data source, for the same commodity-market combination, time point alignment processing is performed on all time-shifted trade time series data within the data source on a unified time coordinate. Values ​​from different data sources that correspond to the same time point and the same indicator are merged into a unified time index, forming a multi-source aligned indicator sequence on a unified time axis. For time gaps that occur during time axis translation and time point alignment, interpolation is used to fill the gaps based on the effective index values ​​of adjacent time points to mark the missing data. The multi-source aligned indicator sequences formed on a unified time axis will be aggregated according to the commodity-market combination to generate multi-source trade time series data that is event-driven aligned with the commodity-market combination.

[0027] In this embodiment, generating trade product recommendation results based on the trade opportunity assessment results includes: Based on multi-source trade time series data aligned by event-driven processes, preset indicators are extracted according to commodity-market combinations. The values ​​of each indicator for the same commodity-market combination on a unified time axis are arranged in order to construct a multivariate trade time series input set corresponding to the commodity-market combination. The multivariate trade time series input set is preprocessed by performing missing value imputation, outlier correction and numerical normalization. The preprocessed series is then segmented according to the historical observation window and the prediction window to generate the input series of the improved TimesFM model. An improved TimesFM model is used to predict the input sequence. This improved TimesFM model, from bottom to top, includes a basic time series encoding layer, a cross-market correlation attention layer, a multi-scale seasonal adapter layer, and an opportunity score output layer, wherein: The basic time series coding layer has the same structure as the original TimesFM coding layer and is used to perform initial feature encoding on the input sequence; The cross-market association attention layer is located after the basic time series encoding layer. It is used to perform cross-series attention operations on the same product sequences from different markets within the same batch. The cross-market association attention layer is connected to the upstream encoding layer through a residual connection. The multi-scale seasonal adapter layer is located after the cross-market association attention layer and is used to adjust the seasonal pattern of the encoding results at three time scales: weekly, monthly and quarterly. The multi-scale seasonal adapter layer is connected to the cross-market association attention layer through layer normalization. The opportunity score output layer, located after the multi-scale seasonal adapter layer, is used to generate a multi-step forecast sequence for the target indicator and output the trade opportunity score for the commodity-market combination. Specifically, the process for generating the multi-step forecast sequence for the target indicator and outputting the trade opportunity score for the commodity-market combination is as follows: The time series features output by the multi-scale seasonal adapter layer are input into the opportunity score output layer. Multi-step prediction results covering a preset prediction window are generated by time step expansion. The multi-step prediction results correspond to the target indicator prediction values ​​of the commodity-market combination at multiple future time points. The multi-step prediction results are summarized and processed in the time dimension, and the changing trends and relative strengths of the target indicators corresponding to each prediction time point are comprehensively considered to form a comprehensive prediction result that reflects the overall future performance level of the commodity-market combination. Based on the comprehensive forecast results, a trade opportunity score for the corresponding commodity-market combination is generated, and the trade opportunity score is used as a quantitative output to characterize the trade potential of the commodity-market combination within the forecast time window. The training of the improved TimesFM model includes a pre-training phase and a fine-tuning phase. The pre-training phase uses historical aligned data divided by commodity-market combination to perform self-supervised learning, and the fine-tuning phase uses labeled target metrics of commodity-market combination to perform supervised learning and update the TimesFM model parameters. The multi-step forecast sequences are weighted and aggregated over time and adjusted in conjunction with risk parameters to generate trade opportunity assessment results for the corresponding commodity-market combinations. These risk parameters include demand volatility risk parameters, contract fulfillment stability risk parameters, and data reliability risk parameters, wherein: Demand volatility risk parameter is used to characterize the degree of volatility of the multi-step forecast sequence within the forecast time window; fulfillment stability risk parameter is used to characterize the stability level of inventory changes and delivery timeliness of the corresponding commodity-market combination within the historical time window; data credibility risk parameter is used to characterize the overall level of the commodity-market combination in terms of multi-source data credibility weights; after weighting and summarizing the multi-step forecast sequence in the time dimension, the summarization results are adjusted according to the risk parameters to generate the trade opportunity assessment results of the corresponding commodity-market combination; The trade opportunity assessment results of each commodity-market combination are used as the ranking criteria to rank and screen all candidate commodity-market combinations, and output the recommended trade products.

[0028] The improved TimesFM model is a modular enhancement based on the original TimesFM time series encoding structure. Its overall architecture, from bottom to top, consists of a basic time series encoding layer, a cross-market association attention layer, a multi-scale seasonal adapter layer, and an opportunity score output layer connected in series. The basic time series encoding layer performs initial feature encoding on the constructed and preprocessed commodity-market-level multivariate trade time series input. A cross-market association attention layer is inserted after the encoding result, enabling cross-series attention calculations for similar commodity series from different markets within the same batch. This layer retains the original time series representation while introducing cross-market association information through residual connections with the upstream encoding layer. A multi-scale seasonal adapter layer is added after the cross-market association attention layer, employing a layered regression approach. The model connects with the previous layer's output and applies seasonal pattern adjustment to the encoded features across three time scales: weekly, monthly, and quarterly, to adapt to the multi-period fluctuations common in trade scenarios. Finally, the opportunity score output layer receives the seasonally adjusted features, expands them over time to generate a multi-step prediction sequence of the target indicator covering the prediction window, and summarizes the multi-step prediction results over time to form a comprehensive prediction result, which in turn outputs the trade opportunity score of the commodity-market combination. The improved model's training process includes a pre-training stage using historical aligned data for self-supervised learning and a fine-tuning stage using labeled target indicators for supervised learning. While maintaining TimesFM's general time-series modeling capabilities, it strengthens cross-market transfer and multi-period adaptation capabilities and directly addresses the trade opportunity score output.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a cross-border e-commerce platform. This platform primarily serves small and medium-sized foreign trade enterprises, with product categories mainly consisting of household goods, daily consumer goods, and small electronic accessories. In its daily product selection process, the platform needs to comprehensively consider historical transaction data, user behavior data, warehousing and logistics data, as well as market environment data such as exchange rates and promotions. However, due to the diverse data sources and inconsistent update frequencies, different data sources exhibit significant differences in their response time to market changes. Traditional product selection methods are prone to biases, especially in the case of new products or new market scenarios, resulting in a low success rate.

[0030] In this application scenario, the platform deploys the trade product selection recommendation method based on multi-source data analysis proposed in this invention into the product selection decision-making system. The system first collects historical transaction data and user behavior data from within the platform, including daily sales volume, order volume, search volume, and click volume of products in different markets. Simultaneously, it collects inventory changes, delivery timeliness, and logistics cost data from the warehousing system, and obtains exchange rate fluctuation information and promotional activity information through third-party interfaces. The aforementioned multi-source trade data is uniformly aggregated according to product identifiers and target market identifiers, forming multi-source time-series data in the product-market dimension.

[0031] During the data processing phase, the system constructs a trade behavior state space based on multi-source time series data. Indicators such as changes in search behavior, transaction behavior, and fulfillment feedback are mapped to basic indicator states, which are further combined to form behavior stage states and evolutionary pattern states, thereby generating trade behavior evolution paths corresponding to each data source. By detecting the consistency of evolution paths from different data sources, the system can identify data sources with abnormal fluctuations or delayed responses under specific commodity-market combinations, and accordingly reduce their influence weight in the data fusion process, thus obtaining reliable reconstructed multi-source trade time series data.

[0032] Building upon this foundation, the system further identifies market events that influence product sales performance, such as platform promotional activities and periodic exchange rate fluctuations, and calculates the response shifts of various indicators from different data sources relative to the time of these market events. Based on the obtained response shift information, the system performs time-axis shifting and alignment processing on the reliably reconstructed multi-source trade time series data, ensuring that different data sources can reflect the true impact process of the same market event in the time dimension, avoiding time series deviations caused by simple timestamp alignment.

[0033] After completing credibility reconstruction and event-driven alignment, the system constructs the processed multi-source trade time series data into a commodity-market level multivariate time series input, which is then fed into the improved TimesFM model for prediction processing. The improved TimesFM model, based on the basic time series encoding structure, introduces a cross-market correlation processing mechanism and a multi-scale seasonality adaptation structure, enabling the model to refer to the time evolution characteristics of other similar markets and adapt to weekly and monthly sales fluctuation patterns when predicting the performance of a commodity in a particular market. The multi-step prediction results output by the model are aggregated to form the trade opportunity assessment results of the commodity-market combination, which serve as the basis for product ranking and screening.

[0034] Table 1 Comparison of the application effects of the method of the present invention and traditional product selection methods

[0035] As shown in Table 1, while maintaining a consistent number of evaluated products, the method of this invention outperforms traditional methods in several key product selection performance indicators. The average sales volume of products recommended using this method increased from 212 to 247 units within 30 days of launch, and the median sales volume increased from 198 to 231 units. This demonstrates that the product selection approach based on multi-source data credibility reconstruction and dynamic time-series alignment can more accurately identify products with market potential, resulting in more stable overall sales performance, rather than relying on a few high-selling products to inflate the average.

[0036] Regarding the proportion of slow-moving goods, the method of this invention reduced the proportion of goods with sales of less than 50 units in 30 days from 18.7% to 11.9%, indicating that by constructing a trade behavior evolution path and performing consistency checks on multi-source data, the impact of abnormal data and misjudgment factors on product selection decisions is effectively reduced, thereby lowering the probability of product selection failure from the source. Simultaneously, the average inventory turnover days decreased from 42.5 days to 36.8 days, reflecting a higher degree of matching between product selection results and actual market demand, which helps alleviate inventory backlog problems and improve supply chain operational efficiency.

[0037] In terms of the success rate of new product selection, the method of this invention improved from 63.4% to 72.6%, demonstrating its significant advantage in scenarios with limited historical data or for new products. Through event-driven time series alignment and multi-step prediction based on an improved TimesFM model, this invention can make a relatively stable assessment of future trade opportunities for commodities even in the absence of sufficient historical transaction data, thereby increasing the success probability of new product selection. This further verifies the foresight and practicality of the method of this invention in actual trade product selection applications.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A trade product selection recommendation method based on multi-source data analysis, characterized in that, include: Collect multi-source trade data related to trade product selection, and aggregate the multi-source trade data according to the commodity-market combination to form corresponding multi-source time series data; Based on multi-source time series data, a trade behavior state space is constructed, multiple indicators reflecting trade activities are mapped as state variables, the ordered state transition relationships between each state variable are identified, and the trade behavior evolution path corresponding to the commodity-market combination is generated based on the state variables and transition relationships. Based on the evolution path of trade behavior, consistency detection is performed on the evolution paths corresponding to different data sources, and data credibility weights of each data source under the commodity-market combination are generated. Multi-source time series data are weighted and fused to obtain credible reconstructed multi-source trade time series data. Based on multi-source trade time series data, we identify market events that affect commodity-market combination trade activities, and calculate the response shift of corresponding indicators relative to the time of occurrence of market events for each data source. Based on response displacement, dynamic time series shifting is performed on multi-source trade time series data to form event-driven aligned multi-source trade time series data. Based on multi-source trade time series data aligned by event-driven methods, a multivariate trade time series input at the commodity-market level is constructed. An improved TimesFM model is used to predict the multivariate trade time series input to obtain the trade opportunity assessment results of the commodity-market combination. Based on the trade opportunity assessment results, trade product selection recommendation results are generated.

2. The trade product selection recommendation method based on multi-source data analysis according to claim 1, characterized in that, The multi-source trade data includes transaction data, market behavior data, supply chain fulfillment data, and market environment data. The transaction data includes historical sales volume, transaction amount, order quantity, and return rate of goods in the target market. The market behavior data includes user search volume, click volume, add-to-cart volume, favorites volume, and number of reviews. The supply chain fulfillment data includes changes in inventory quantity, replenishment cycle, delivery time, and logistics costs. The market environment data includes exchange rate change information, tariff or regulatory policy change information, promotional activity information, and holiday information.

3. The trade product selection recommendation method based on multi-source data analysis according to claim 1, characterized in that, The aggregation of multi-source trade data according to the commodity-market combination is specifically as follows: Using product identifiers and target market identifiers as a joint index, trade data from different data sources are matched and associated. Data with the same product identifier and target market identifier are divided into the same data set, and the data in the data set are sorted and organized according to the timestamp to form multi-source time series data of corresponding product-market combinations.

4. The trade product selection recommendation method based on multi-source data analysis according to claim 1, characterized in that, The evolution path of trade behavior corresponding to the commodity-market combination generated based on state variables and transition relationships includes: For the same commodity-market combination in each data source, the multi-source time series data are organized in chronological order within a preset analysis time window to determine several indicators to characterize the trade activities of the commodity-market combination and form a continuous indicator sequence. Based on the indicator sequence, the first layer of basic indicator state space of trade behavior is constructed. For each indicator, according to the preset numerical range threshold and change direction, the value of the indicator at each time point is labeled as several discrete basic indicator states, generating a basic indicator state sequence that changes with time. Based on the basic indicator state sequence, a second behavioral stage state layer is constructed in the trade behavior state space. At the same time point, according to the preset mapping rules, multiple basic indicator states are combined and mapped to a single behavioral stage state corresponding to the time point, forming a behavioral stage state sequence that changes over time. Based on the state sequence of behavioral stages, a third evolutionary pattern state layer is constructed in the state space of trade behavior. The duration and state switching frequency of each behavioral stage state are statistically analyzed within a sliding time window. The distribution of behavioral stage states within each time window is classified into preset evolutionary pattern states, generating an evolutionary pattern state sequence that changes over time. Based on the state sequence of basic indicators, the state sequence of behavioral stages, and the state sequence of evolution patterns, the three-layer state corresponding to each time point of the same commodity-market combination in the same data source within the analysis time window is connected sequentially in chronological order to generate the trade behavior evolution path.

5. The trade product selection recommendation method based on multi-source data analysis according to claim 1, characterized in that, The obtained reliable reconstructed multi-source trade time series data includes: For the same commodity-market combination, trade behavior evolution paths are obtained from various data sources, and each trade behavior evolution path is associated with the corresponding data source identifier to form a set of trade behavior evolution paths of the commodity-market combination under different data sources. Based on the set of trade behavior evolution paths, the analysis time window is divided into several continuous time segments. Within each time segment, the state sequence of the trade behavior evolution path corresponding to each data source is statistically analyzed at three levels: basic indicator state, behavior stage state, and evolution mode state. The state sequences of different data sources within the same time segment are compared, and time segments with consistent state sequences and time segments with inconsistent state sequences are marked. Based on the consistency marking results of each data source in each time segment, the number of consistent time segments and inconsistent time segments of the data source within the analysis time window are statistically analyzed. According to the ratio between the number of consistent time segments and the total number of time segments, a corresponding data credibility weight is generated for each data source. Based on data credibility weights, a weighted calculation is performed on the multi-source time series data corresponding to the same commodity-market combination at each time point. The values ​​of the same indicator from different data sources are weighted and superimposed according to the data credibility weights to obtain the weighted fusion indicator value at the time point. The weighted fusion index values ​​obtained at each time point within the analysis time window are arranged in chronological order to generate reliable reconstructed multi-source trade time series data for the corresponding commodity-market combination.

6. The trade product selection recommendation method based on multi-source data analysis according to claim 1, characterized in that, The calculation of the response shift of the corresponding indicator relative to the time of the market event for each data source is specifically as follows: Based on the reliably reconstructed multi-source trade time series data, a response analysis time window is set for each commodity-market combination. Within the response analysis time window, an indicator sequence related to changes in the market environment is selected for the identification of market events. Within the response analysis time window, the magnitude and frequency of changes in the indicator series related to exchange rate changes, tariff or regulatory policy changes, promotional activities and holidays are detected. When the magnitude or frequency of change at a time point meets the preset event judgment conditions, the time point is marked as the market event occurrence time of the corresponding commodity-market combination, and the time point is associated with the corresponding event type. For each commodity-market combination, within a preset time range before and after the occurrence of the market event, extract indicator sequences of transaction data, market behavior data, and supply chain fulfillment data corresponding to each data source from the reliably reconstructed multi-source trade time series data, and determine the start time, peak time, or inflection point time of change relative to the baseline level for each indicator sequence. The time difference between the determined start time, peak time, or inflection point time of the change and the corresponding market event occurrence time is calculated to obtain the response time difference of each indicator in each data source relative to the market event occurrence time. The response time difference is recorded as the response displacement of the indicator relative to the market event. Statistical processing is performed on the response displacements of each indicator in the same data source to generate summary response displacement information of the data source under the corresponding commodity-market combination.

7. The trade product selection recommendation method based on multi-source data analysis according to claim 1, characterized in that, The event-driven aligned multi-source trade time series data includes: Obtain reliable reconstructed multi-source trade time series data, as well as aggregated response displacement information generated for each data source for the same commodity-market combination, and read the time displacement corresponding to the data source in the aggregated response displacement information of each data source; For the same commodity-market combination, within a unified analysis time window, based on the reliable reconstructed multi-source trade time series data, the indicator sequences of transaction data, market behavior data, and supply chain fulfillment data of each data source are selected respectively. The time displacement corresponding to the data source is applied to the indicator sequence, and the time axis is shifted for each time point in the indicator sequence according to the time displacement to obtain the trade time series data within the data source after time shifting. After completing the time axis shifting process within each data source, for the same commodity-market combination, time point alignment processing is performed on all time-shifted trade time series data within the data source on a unified time coordinate. Values ​​from different data sources that correspond to the same time point and the same indicator are merged into a unified time index, forming a multi-source aligned indicator sequence on a unified time axis. For time gaps that occur during time axis translation and time point alignment, interpolation is used to fill the gaps based on the effective index values ​​of adjacent time points to mark the missing data. The multi-source aligned indicator sequences formed on a unified time axis will be aggregated according to the commodity-market combination to generate multi-source trade time series data that is event-driven aligned with the commodity-market combination.

8. The trade product selection recommendation method based on multi-source data analysis according to claim 1, characterized in that, The step of generating trade product recommendation results based on the trade opportunity assessment results includes: Based on multi-source trade time series data aligned by event-driven processes, preset indicators are extracted according to commodity-market combinations. The values ​​of each indicator for the same commodity-market combination on a unified time axis are arranged in order to construct a multivariate trade time series input set corresponding to the commodity-market combination. The multivariate trade time series input set is preprocessed by performing missing value imputation, outlier correction and numerical normalization. The preprocessed series is then segmented according to the historical observation window and the prediction window to generate the input series of the improved TimesFM model. An improved TimesFM model is used to predict the input sequence. This improved TimesFM model, from bottom to top, includes a basic time series encoding layer, a cross-market correlation attention layer, a multi-scale seasonal adapter layer, and an opportunity score output layer, wherein: The basic time series coding layer has the same structure as the original TimesFM coding layer and is used to perform initial feature encoding on the input sequence; The cross-market association attention layer is located after the basic time series encoding layer. It is used to perform cross-series attention operations on the same product sequences from different markets within the same batch. The cross-market association attention layer is connected to the upstream encoding layer through a residual connection. The multi-scale seasonal adapter layer is located after the cross-market association attention layer and is used to adjust the seasonal pattern of the encoding results at three time scales: weekly, monthly and quarterly. The multi-scale seasonal adapter layer is connected to the cross-market association attention layer through layer normalization. The opportunity score output layer is located after the multi-scale seasonal adapter layer and is used to generate a multi-step forecast sequence of the target indicator and output a trade opportunity score for the commodity-market combination. The training of the improved TimesFM model includes a pre-training phase and a fine-tuning phase. The pre-training phase uses historical aligned data divided by commodity-market combination to perform self-supervised learning, and the fine-tuning phase uses labeled target metrics of commodity-market combination to perform supervised learning and update the TimesFM model parameters. The multi-step forecast sequence is weighted and aggregated in the time dimension and adjusted in combination with risk parameters to generate the corresponding commodity-market combination trade opportunity assessment results; The trade opportunity assessment results of each commodity-market combination are used as the ranking criteria to rank and screen all candidate commodity-market combinations, and output the recommended trade products.

Images