E-commerce live broadcast trend analysis system based on live broadcast data

Abstract

The invention relates to the technical field of e-commerce live broadcast, and particularly discloses a live broadcast data-based e-commerce live broadcast trend analysis system, which is used for extracting live broadcast commodity sales data and historical live broadcast sales data of a target commodity to obtain a live broadcast sales data time sequence set and a plurality of historical live broadcast sales data time sequence sets, performing time sequence feature analysis on the live broadcast sales data time sequence set and each historical live broadcast sales data time sequence set by using an AI-based data analysis and processing algorithm; therefore, the commodity sales data speculation value of the next time interval of the target commodity is intelligently pre-determined according to the selection range ratio dynamic semantic query expression between the live sales data time sequence characteristics and the historical live sales data time sequence characteristics. According to the method, nonlinear and changeable market dynamics can be better captured, so that the model not only pays attention to a short-term trend, but also can understand a long-term mode, and the understanding depth of a sales volume change rule is improved.

Description

Technical Field

[0001] The present application relates to the field of e-commerce live broadcast technology, and more specifically, to an e-commerce live broadcast trend analysis system based on live broadcast data. Background Art

[0002] With the rapid development of Internet technology, the e-commerce live broadcast industry has experienced explosive growth and has become an important force driving the consumer market. This emerging field has not only spawned a large number of e-commerce platforms, but also generated a huge amount of live broadcast data. However, the high unpredictability of audience size and purchasing decisions has brought great challenges to sales forecasting, which makes it difficult for live broadcast e-commerce companies to avoid the risk of overstocking or understocking when preparing goods in advance.

[0003] In this regard, patent CN116957647A discloses an e-commerce live streaming system based on classification prediction, which collects commodity sales information before the current time interval through a first statistical device, and summarizes historical data for the same period through a second statistical device. The prediction and analysis device uses a deep neural network to analyze these data and predict the sales volume of commodities in the current interval. The stocking processing mechanism prepares inventory according to the forecast, and the transfer request mechanism automatically issues a replenishment request when the inventory is insufficient to ensure that the sales forecast needs are met, thereby optimizing the inventory management based on historical sales patterns and improving the operational efficiency of live streaming e-commerce.

[0004] The patent is about training deep neural networks to predict products. However, traditional deep neural methods may have difficulty in fully mining and utilizing the complex features in time series data, especially those nonlinear and changeable market dynamics, which limits the model's understanding of the law of sales volume changes. In addition, traditional neural networks cannot well capture the semantic information hidden behind sales data, such as the impact of promotional activities, holiday effects, etc. on sales, thus affecting the accuracy of predictions.

[0005] Therefore, an optimized e-commerce live streaming trend analysis solution based on live streaming data is desired. Summary of the invention

[0006] This application provides an e-commerce live broadcast trend analysis system based on live broadcast data, which can better capture nonlinear and changeable market dynamics, so that the model not only focuses on short-term trends, but also understands long-term patterns, thereby improving the depth of understanding of the law of sales volume changes. At the same time, the concepts of "selection range ratio" and "dynamic semantic query representation" are introduced, which can intelligently consider the impact of events such as promotions and holidays on sales, and quickly optimize search strategies to provide more accurate and timely sales forecasts.

[0007] According to one aspect of the present application, there is provided an e-commerce live broadcast trend analysis system based on live broadcast data, comprising: a data collection module, used to collect and count live broadcast commodity sales data of each commodity in multiple time intervals within a scheduled live broadcast time period of a target e-commerce live broadcast platform; a historical data collection module, used to collect historical live broadcast sales data of each commodity in multiple time intervals that are the same as the scheduled live broadcast time period of the target e-commerce live broadcast platform in the past multiple days; a prediction module, used to obtain trend prediction results of each commodity based on the live broadcast commodity sales data and historical live broadcast sales data of each commodity; a replenishment module, used to issue a replenishment request in response to the trend prediction result of the target commodity exceeding the inventory of the commodity in the next time interval; characterized in that the prediction module comprises: a data acquisition unit, used to obtain the trend prediction results of each commodity based on the live broadcast commodity sales data and historical live broadcast sales data of each commodity. The method comprises the following steps: extracting the live commodity sales data and the historical live sales data of the target commodity to obtain a live sales data time series set and multiple historical live sales data time series sets; a data encoding unit, which is used to perform time series encoding on the live sales data time series set and the multiple historical live sales data time series sets respectively to obtain a set of live sales data time series features and historical live sales data time series features; a data query response unit, which is used to perform a feature dynamic semantic query response with a selection range ratio anchoring on the set of live sales data time series features and the set of historical live sales data time series features to obtain a live sales time series semantic query response feature, and based on the live sales time series semantic query response feature, obtain the trend prediction result, which is used to represent the commodity sales data inference value of the next time interval of the target commodity object.

[0008] The present application provides an e-commerce live broadcast trend analysis system based on live broadcast data, which extracts the live broadcast commodity sales data and historical live broadcast sales data of the target commodity to obtain a live broadcast sales data time series set and multiple historical live broadcast sales data time series sets, and uses AI-based data analysis and processing algorithms to perform time series feature analysis on the live broadcast sales data time series set and each historical live broadcast sales data time series set, respectively, so as to intelligently predict the commodity sales data inference value of the next time interval of the target commodity according to the selection range ratio dynamic semantic query representation between the live broadcast sales data time series features and each historical live broadcast sales data time series features. The present application can better capture the nonlinear and changeable market dynamics, which enables the model to not only focus on short-term trends, but also understand long-term patterns, thereby improving the depth of understanding of the law of sales volume changes. At the same time, the concepts of "selection range ratio" and "dynamic semantic query representation" are introduced, which can intelligently consider the impact of events such as promotional activities and holidays on sales, and quickly optimize search strategies to provide more accurate and timely sales forecasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application are briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the present application, and are not intended to limit the present application.

[0010] Figure 1 This is a schematic block diagram of an e-commerce live broadcast trend analysis system based on live broadcast data according to an embodiment of the present application.

[0011] Figure 2 This is a schematic block diagram of a prediction module in an e-commerce live broadcast trend analysis system based on live broadcast data in an embodiment of the present application.

[0012] Figure 3 This is a schematic diagram of the data flow of the prediction module in the e-commerce live broadcast trend analysis system based on live broadcast data in an embodiment of the present application.

[0013] Figure 4 This is a schematic block diagram of a data query response unit in an e-commerce live broadcast trend analysis system based on live broadcast data in an embodiment of the present application.

[0014] Figure 5 This is a schematic block diagram of a current-historical sales data time series query response unit in an e-commerce live broadcast trend analysis system based on live broadcast data in an embodiment of the present application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work also fall within the scope of protection of the present application.

[0016] Based on patent CN116957647A, this application proposes an e-commerce live broadcast trend analysis system based on live broadcast data, such as Figure 1 As shown, the e-commerce live broadcast trend analysis system based on live broadcast data includes: a data collection module 10, which is used to collect and count the live broadcast commodity sales data of each commodity in multiple time intervals within the live broadcast scheduled time period of the target e-commerce live broadcast platform; a historical data collection module 20, which is used to collect the historical live broadcast sales data of each commodity in multiple time intervals that are the same as the live broadcast scheduled time period of the target e-commerce live broadcast platform in the past multiple days; a prediction module 30, which is used to obtain the trend prediction results of each commodity based on the live broadcast commodity sales data and historical live broadcast sales data of each commodity; a replenishment module 40, which is used to respond to the trend prediction result of the target commodity exceeding the inventory of the commodity in the next time interval, and issue a replenishment request. In particular, the sales data is sales volume.

[0017] Exemplarily, in the data collection module 10, the live broadcast commodity sales data of each commodity in multiple time intervals within the scheduled live broadcast time period of the target e-commerce live broadcast platform are collected and counted. It should be understood that by collecting and counting the live broadcast commodity sales data of each commodity in multiple time intervals within the scheduled live broadcast time period of the target e-commerce live broadcast platform, the sales patterns and consumer behavior characteristics in different time periods can be captured, which is crucial for understanding market dynamics and predicting future sales trends. Specifically, this module will record the sales of all commodities in each time interval, including but not limited to the specific sales quantity of each category of commodities. Such data accumulation not only helps to identify short-term fluctuations, but also reveals long-term trend changes, providing rich training materials for the model.

[0018] In one embodiment, the data collection module will monitor and record the sales of all products on the platform in real time according to preset time periods (such as every hour, half an hour, etc.) during the live broadcast event. At the end of each predetermined time interval, the module will automatically summarize the transaction information that occurred during this period and organize it into a structured sales report. This process involves parsing each successful order, extracting the types of goods involved and their corresponding sales volumes, and then organizing this information in time series to form a complete set of product sales information containing multi-dimensional information. In addition, in order to ensure the quality and integrity of the data, the data collection module will also take measures to prevent data loss or repeated calculations, such as setting up a reasonable caching mechanism and regularly checking the consistency and accuracy of the data.

[0019] Exemplarily, in the historical data collection module 20, the historical live broadcast sales data of each commodity on the target e-commerce live broadcast platform in multiple time intervals that are the same as the scheduled time period of the live broadcast in the past days are collected. It should be understood that historical data provides an important reference point, which can help the system identify which factors may have affected past sales, such as promotional activities, holiday effects, etc., and can reveal long-term trend changes. By learning from these historical data, the deep neural network model can more accurately understand market dynamics, thereby providing more reliable sales forecasts for current and future time intervals. In addition, historical data can also help the system distinguish between random fluctuations and stable trends, thereby improving the accuracy of predictions.

[0020] In one embodiment, the historical data collection module works by comprehensively recording and counting the historical live broadcast sales of each product on the target e-commerce live broadcast platform within multiple identical time periods in the past (for example, if the data currently being analyzed is from 10 a.m. to 12 p.m. on Mondays, the data from 10 a.m. to 12 p.m. on all Mondays in the previous weeks or even months will be reviewed). This means that for each time interval of interest, the module will trace back past sales performance in similar time periods, including sales volumes and other relevant indicators for different product categories. In this way, the system can obtain a series of historical samples that are comparable to the current time interval, which not only contain direct sales figures, but also imply various influencing factors under the market environment at the time.

[0021] Exemplarily, in the prediction module 30, based on the live broadcast commodity sales data and historical live broadcast sales data of each commodity, the trend prediction results of each commodity are obtained. It should be understood that by integrating the commodity sales data during the live broadcast and the historical sales data of the same time period in the past, the prediction module can capture the dynamic changes of the market and provide more accurate and timely commodity sales trend forecasts, thereby helping live e-commerce to formulate reliable and effective stocking strategies for future time intervals. The core of the prediction module is to use a deep neural network model to process and analyze these data. It not only focuses on the actual sales of each commodity during the current live broadcast, but also combines the sales patterns in similar time periods in historical data. This approach allows the system to identify the difference between short-term fluctuations and long-term trends, and intelligently consider the impact of special events such as promotional activities and holidays on sales. By learning a large amount of data, the model can better understand the nonlinear and changeable market dynamics, improve the depth of understanding of the law of sales volume changes, and thus provide a solid foundation for future sales forecasts.

[0022] Exemplarily, in the replenishment module 40, in response to the trend prediction result of the target commodity exceeding the inventory of the commodity in the next time interval, a replenishment request is issued. In particular, the sales data is the sales volume. It should be understood that the replenishment module is used to issue a replenishment request in response to the trend prediction result of the target commodity exceeding the inventory of the commodity in the next time interval. In this way, the system can take timely actions to ensure sufficient supply of goods when it is predicted that future demand exceeds the existing inventory, avoid missing sales opportunities due to out-of-stock, and also prevent excessive stockpiling of inventory from wasting resources.

[0023] In one embodiment, after the prediction module obtains trend prediction results based on the live commodity sales data and historical live sales data of each commodity, if it is found that the expected sales volume of a target commodity in the next time interval will exceed the current inventory, the replenishment module will start the corresponding mechanism to handle this situation. It will automatically evaluate the gap between the existing inventory level and the predicted demand, and issue a replenishment request accordingly. This request not only contains the name of the commodity that needs to be replenished, but also carries the difference between the number of commodities reserved for a certain category of commodities in the current time interval set by the live e-commerce and the predicted sales quantity. Such information is very important for supply chain management and inventory control because it can clearly indicate how much specific commodities need to be replenished to meet the expected demand.

[0024] Specifically, once the replenishment demand is determined, the module can immediately notify the relevant parties (such as suppliers or internal logistics departments) and provide detailed replenishment information, including but not limited to the type of goods, the required quantity, and the degree of urgency. In addition, the replenishment module can also be combined with other business processes, such as automated order generation and tracking functions, to ensure that the entire replenishment process is efficient and transparent. At the same time, in order to adapt to different operating scenarios, the replenishment module can also be customized according to the specific needs of the enterprise, such as setting different warning thresholds or defining priority rules to better balance cost-effectiveness and service levels.

[0025] Accordingly, in the prediction module, the technical concept of the present application is to extract the live commodity sales data and historical live sales data of the target commodity to obtain a live sales data time series set and multiple historical live sales data time series sets, and use AI-based data analysis and processing algorithms to perform time series feature analysis on the live sales data time series set and each historical live sales data time series set, respectively, so as to intelligently predict the commodity sales data inference value of the next time interval of the target commodity according to the selection range ratio dynamic semantic query representation between the live sales data time series features and the selection range ratio between the time series features of each historical live sales data. The present application can better capture the nonlinear and changeable market dynamics, which enables the model to not only focus on short-term trends, but also understand long-term patterns, thereby improving the depth of understanding of the law of sales volume changes. At the same time, the concepts of "selection range ratio" and "dynamic semantic query representation" are introduced, which can intelligently consider the impact of events such as promotional activities and holidays on sales, and quickly optimize search strategies to provide more accurate and timely sales forecasts.

[0026] In one embodiment, Figure 2 and Figure 3As shown, the prediction module 30 includes: a data acquisition unit 31, which is used to extract the live commodity sales data and historical live sales data of the target commodity to obtain a live sales data time series set and multiple historical live sales data time series sets; a data encoding unit 32, which is used to perform time series encoding on the live sales data time series set and the multiple historical live sales data time series sets respectively to obtain a set of live sales data time series features and historical live sales data time series features; a data query response unit 33, which is used to perform a feature dynamic semantic query response with a selection range ratio anchoring on the live sales data time series features and the set of historical live sales data time series features to obtain a live sales time series semantic query response feature, and based on the live sales time series semantic query response feature, obtain the trend prediction result, and the trend prediction result is used to represent the commodity sales data inference value of the next time interval of the target commodity object.

[0027] Exemplarily, in the data acquisition unit 31, the live commodity sales data and historical live sales data of the target commodity are extracted to obtain a live sales data time series set and multiple historical live sales data time series sets. It should be understood that the live sales data time series set reflects the sales of the commodity during the current live broadcast, while the multiple historical live sales data time series sets provide historical references for sales patterns within a specific time period. Combining these two types of data, the model can not only identify short-term fluctuations, but also reveal long-term trend changes, thereby improving the depth of understanding of the laws of sales volume changes. In addition, the introduction of historical data can also help the system intelligently consider the impact of special events such as promotions and holidays on sales, quickly optimize search strategies, and provide more accurate and timely sales forecasts.

[0028] In one embodiment, the data acquisition unit is used to extract the live sales data of the target product in multiple time intervals within the live scheduled time period from the data collection module to obtain a live sales data time series set, and extract the live sales data time series set of the target product in multiple time intervals within the live scheduled time period in the past days from the historical data collection module to obtain the multiple historical live sales data time series sets. That is, using the data collection module, the sales of all commodities in each scheduled time interval are monitored and recorded in real time during the live broadcast process, including the specific sales quantity of each category of commodities. For historical data, the historical data collection module is responsible for tracing the sales performance in the same time period in the past, ensuring that a sufficiently long time span is covered to obtain rich samples. Once enough data is collected, the next step is to organize them in chronological order to form an ordered data sequence. At the same time, in order to cope with the challenges brought by the wide variety of commodities, the system needs to accurately extract the live sales data of the target commodity in multiple time intervals within the live scheduled time period from the data collection module to form a live sales data time series set. At the same time, the system also needs to extract the live broadcast sales data of the same product in multiple time intervals within the same live broadcast scheduled time period in the past few days from the historical data collection module, and construct multiple historical live broadcast sales data time series sets.

[0029] Exemplarily, in the data encoding unit 32, the live sales data time series set and the multiple historical live sales data time series sets are respectively time-series encoded to obtain a set of live sales data time series features and historical live sales data time series features. It should be understood that, considering that the live sales data time series set and each historical live sales data time series set have obvious time series characteristics, their sales data are arranged in chronological order over time, and contain information such as the periodicity of sales peaks and troughs, and growth or decline trends in different time periods. And considering that the 1D-FCN network has unique advantages in processing sequence data, it can automatically capture deep and representative features from the live sales data time series set and multiple historical live sales data time series sets. Based on this, in the technical solution of the present application, the live sales data time series set and the multiple historical live sales data time series sets are respectively passed through a live sales time series encoder based on a 1D-FCN network to utilize the convolution kernel to slide along the time axis to perform convolution operations on the data at different time points, so as to mine the key features such as correlation and periodicity hidden in the sales data in different time intervals in the time series dimension, and obtain a set of live sales data time series feature vectors and historical live sales data time series feature vectors.

[0030] In one embodiment, the data encoding unit is used to: respectively pass the live sales data timing set and the multiple historical live sales data timing sets through a live sales timing encoder to obtain a live sales data timing feature vector as the live sales data timing feature and a set of historical live sales data timing feature vectors as the set of historical live sales data timing features.

[0031] In one embodiment, the live sales timing encoder is a live sales timing encoder based on a 1D-FCN network.

[0032] Exemplarily, in the data query response unit 33, a dynamic semantic query response with a selection range ratio anchoring feature is performed on the set of the live sales data timing features and the historical live sales data timing features to obtain the live sales timing semantic query response feature, and based on the live sales timing semantic query response feature, the trend prediction result is obtained, and the trend prediction result is used to represent the commodity sales data inference value of the next time interval of the target commodity object. It should be understood that considering that the live sales data timing features and the historical live sales data timing features have an inherent semantic association in different time intervals. For example, the sales growth trend of a certain product before a specific festival in history may have a corresponding relationship with the sales change trend when approaching a similar festival in the current live broadcast, reflecting the semantic connection of consumer purchasing habits and other aspects. In order to efficiently query matching and deeply understand the semantic information in these large-scale feature vector libraries, the present application introduces a dynamic semantic query response method with a selection range ratio anchoring feature to perform a semantic query response on the set of the live sales data timing feature vector and the historical live sales data timing feature vector to obtain a live sales timing semantic query response encoding vector. That is, this method can identify the feature vector that best represents the query intent to accurately locate those features that are important for understanding the sales volume change pattern and judging sales trends. It overcomes the limitation that traditional methods may blindly process all feature vectors and have difficulty focusing on key information. It enables subsequent analysis to revolve around truly valuable features, which helps to gain in-depth insights into sales-related situations, thereby characterizing the semantic representation that best represents the query features (time series features of live sales data).

[0033] In one embodiment, Figure 4As shown, the data query response unit 33 includes: a current-historical sales data time series query response unit 331, which is used to select a range-ratio anchored feature dynamic semantic query response for the set of the live sales data time series feature vector and the historical live sales data time series feature vector to obtain a live sales time series semantic query response encoding vector as the live sales time series semantic query response feature; a commodity sales data prediction unit 332, which is used to obtain the trend prediction result based on the live sales time series semantic query response encoding vector, and the trend prediction result is used to represent the commodity sales data inference value of the next time interval of the target commodity object.

[0034] In one embodiment, Figure 5 As shown, the current-historical sales data timing query response unit 331 includes: a commodity sales dynamic search anchor subunit 3311, which is used to calculate the live-historical sales data dynamic search anchor vector based on the live-historical sales data timing feature vector and the set of historical live-historical sales data timing feature vectors; a historical live-historical sales data selection range ratio determination subunit 3312, which is used to determine the historical live-historical sales data selection range ratio based on the feature distribution characteristics of the live-historical sales data dynamic search anchor vector; a commodity sales dynamic query subunit 3313, which is used to perform semantic similarity dynamic encoding on each feature in the live-historical sales data timing feature vector and the historical live-historical sales data selection range ratio to obtain the live-historical sales timing semantic query response encoding vector.

[0035] In one embodiment, the commodity sales dynamic search anchoring subunit 3311 is used to calculate the internal relationship score value of the live sales data time series feature vector relative to each historical live sales data time series feature vector in the set of historical live sales data time series feature vectors to obtain a set of live-historical sales data internal relationship score values. Specifically, the process can be expressed by the formula: ;in, is the set of time series feature vectors of the historical live sales data, and is the first, second, and third in the set of time series feature vectors of the historical live sales data. and The time series feature vector of historical live sales data, is the weight matrix of historical live sales data, is the time series feature vector of live sales data, is the live sales data weight matrix, is the hyperbolic tangent function, is the weight coefficient vector, It is the first in the set of scores of the internal relationship between live broadcast and historical sales data. The score value of the internal relationship between live broadcast and historical sales data.

[0036] The historical live sales data time series feature vector corresponding to the largest live-historical sales data internal relationship score value in the set of the live-historical sales data internal relationship score values ​​is used as the live-historical sales data dynamic search anchor vector. Specifically, the process can be expressed by the formula: ;in, is the vector at the location of the maximum value, It is the live broadcast-historical sales data dynamic search anchor vector.

[0037] Specifically, the internal relationship score values ​​of the live sales data time series feature vector relative to each historical live sales data time series feature vector in the set of historical live sales data time series feature vectors are calculated. This process aims to quantify the similarity and potential connection between the current live sales model and the sales model in a similar time period in the past. In this way, the system can capture information that is critical to understanding market dynamics. When the system calculates these internal relationship scores, a set of live-historical sales data internal relationship score values ​​is formed. This set is not just a series of numbers, it actually represents the multi-dimensional correlation between the current live sales situation and the historical sales model. A high score means that the two are highly similar or closely related in some aspects, which may be due to factors such as the same time period, similar promotional strategies, or the same consumer behavior. Low scores may indicate large differences, but even so, these scores still help to reveal the changing patterns and the impact of special events in different time periods.

[0038] Next, the system will select the historical live sales data time series feature vector corresponding to the maximum value from this score value set as the live-historical sales data dynamic search anchor vector. Selecting the feature vector corresponding to the maximum score value is not simply finding a highest point, but comprehensively considering the global distribution characteristics to ensure that the selected anchor point is representative. The selection of this anchor vector is crucial because it not only reflects the core characteristics of the current live sales situation, but also serves as the basis for further optimization. By selecting the most representative historical sales patterns as anchor points, the system can better understand and predict future sales trends. In addition, selecting such an anchor vector also helps to focus on those feature vectors that have a high correlation with the current query, thereby narrowing the search space and improving efficiency.

[0039] Exemplarily, in the historical live sales data selection range ratio determination subunit 3312, the historical live sales data selection range ratio is determined based on the characteristic distribution characteristics of the live-historical sales data dynamic search anchor vector. In one embodiment, the vector of the starting position of the historical live sales data selection range ratio is the live-historical sales data dynamic search anchor vector, and each historical live sales data time series feature vector in the historical live sales data selection range ratio is defined as a historical live sales data dynamic search optimization library feature vector. Specifically, the process can be expressed by the formula: ;in, and They are The variance and mean of is the modulation parameter, To obtain The maximum value in Indicates rounding up. is the size of the selection range ratio of historical live sales data, is a set of feature vectors of the dynamic search optimization library of historical live sales data in the selection range of historical live sales data, and Select the range ratio for the historical live sales data and A dynamic search optimization library feature vector of historical live sales data.

[0040] Specifically, when the system selects the live broadcast-historical sales data dynamic search anchor vector, it has actually found the most representative historical sales model as a reference point. This anchor vector reflects the core characteristics of the current live broadcast sales situation and is selected after a comprehensive evaluation of the global distribution characteristics. Next, the system will determine a reasonable selection range ratio based on the feature distribution characteristics of this anchor vector. The selection range ratio is defined to focus on those feature vectors that have a high correlation with the query, while excluding irrelevant parts, so as to narrow the search space and improve efficiency. In this process, the system considers the feature distribution around the anchor vector, such as the feature vectors of historical sales data in its neighboring area. These feature vectors are defined as dynamic search optimization library feature vectors, which together constitute a local subset of historical data. Each feature vector in this subset has a certain similarity or correlation with the anchor vector, so it can better represent the potential market behavior in the current live broadcast sales situation. The starting position of the selection range ratio is occupied by the anchor vector, and the other feature vectors in the range are spread around this core, forming a tightly connected data cluster. The process of determining the selection range ratio is not just a static operation. It combines statistical analysis or machine learning models to ensure that the selected optimized library feature vector can retain the original information to the greatest extent and reduce noise interference. For example, in an e-commerce live broadcast environment, certain time periods may show abnormally high sales due to specific promotions or holidays. If these outliers are included in the selection range, the prediction results may deviate from reality. Therefore, the system will intelligently adjust the selection range by analyzing the feature distribution characteristics, so that the final selected data subset can reflect normal sales trends and appropriately consider the impact of special events.

[0041] In one embodiment, the commodity sales dynamic query subunit 3313 is used to: input the live sales data time series feature vector and each historical live sales data dynamic search optimization library feature vector in the historical live sales data selection range ratio into a dynamic semantic search encoder to obtain the live sales time series semantic query response encoding vector. Specifically, the process can be expressed by the formula: ;in, is the length of the vector, for and The semantic similarity between It is the encoding vector of the live sales temporal semantic query response.

[0042] Specifically, after the system has determined the time series feature vectors of live sales data and the feature vectors of the dynamic search optimization library of historical live sales data screened out based on the selection range ratio, it has actually identified and focused on a set of data that best represents the current market conditions and its potential influencing factors. In order to further explore the deep patterns in these data and transform them into useful prediction bases, the system inputs them together into the dynamic semantic search encoder. The dynamic semantic search encoder itself is a complex neural network architecture, and its design goal is an embedding space that can effectively express semantic information. In this space, samples that were originally scattered in a high-dimensional feature space are reorganized into a compact representation with clear semantic meaning. The final output live sales time series semantic query response encoding vector should have a high degree of abstract generalization ability and be able to perform well in various downstream tasks.

[0043] Exemplarily, in the commodity sales data prediction unit 332, the trend prediction result is obtained based on the live sales time series semantic query response coding vector, and the trend prediction result is used to represent the commodity sales data speculation value of the next time interval of the target commodity object. In one embodiment, the commodity sales data prediction unit is used to: pass the live sales time series semantic query response coding vector through a decoder-based e-commerce live trend predictor to obtain the trend prediction result. That is, the live sales data time series feature vector and the set of historical live sales data time series feature vectors are used to select the live sales time series semantic query response coding vector obtained by the dynamic semantic query for decoding processing to intelligently predict the commodity sales data speculation value of the next time interval of the target commodity.

[0044] In one embodiment, the decoder-based e-commerce live streaming trend predictor uses a long short-term memory network (LSTM). LSTM is a recurrent neural network (RNN) variant that is particularly suitable for processing and predicting time series data. It can effectively capture long-term dependencies and has good adaptability to nonlinear and changeable market dynamics in e-commerce live streaming. Specifically, the whole process starts with receiving the encoding vector of the response to the live sales time series semantic query. This encoding vector is the result of processing by the dynamic semantic search encoder, and contains complex association information between the current live sales situation and the historical sales pattern. Next, these encoding vectors are input into the decoder based on the LSTM architecture. The LSTM unit contains three components: the forget gate, the input gate, and the output gate, which work together to control the flow of information and selectively remember or forget the past state. During the decoding process, the LSTM will gradually parse the input encoding vector and use its internal memory cells to track important time series features. With the advancement of each time step, the LSTM not only considers the input at the current moment, but also combines the information at all previous time points to ensure a more comprehensive understanding of the overall trend. In order to generate the final trend prediction results, the LSTM decoder will output the estimated sales data value of the target product in the next time interval at the last time step.

[0045] Here, when the live sales data timing feature vector and the set of historical live sales data timing feature vectors respectively represent the local timing correlation features of the live sales data timing set and multiple historical live sales data timing sets in a single sample domain and a full sample domain, when performing feature dynamic semantic search encoding based on selection range ratio anchoring, the complexity of the selection range ratio anchoring mechanism of the full sample domain relative to the single sample domain will lead to insufficient long-distance dynamic search coding representation of the live sales timing semantic query response encoding vector, thereby reducing the expression effect of the live sales timing semantic query response encoding vector and affecting the accuracy of the trend prediction results obtained by the decoder-based e-commerce live trend predictor.

[0046] Therefore, preferably, in one example, when the live sales time series semantic query response coding vector is passed through a decoder-based e-commerce live trend predictor to obtain a trend prediction result, the live sales time series semantic query response coding vector is optimized, and the optimization step includes: arranging the eigenvalues ​​of the live sales time series semantic query response coding vector in ascending order as a live sales time series semantic query response sequence coding vector; The eigenvalue and Covariance between eigenvalues ​​Covariance is less than or equal to the distance difference hyperparameter , calculate the The eigenvalues ​​are similar to the The weighted sum between the eigenvalues ​​is the optimized Eigenvalues: ;in, and respectively represent the first Eigenvalues ​​and Eigenvalues, represents the distance difference hyperparameter, is the feature mean of the sequential encoding vector of the live sales time series semantic query response, and and is the weight hyperparameter, Represents the optimized Eigenvalue.

[0047] Calculate the hyperbolic sine function value of the sum of squares of all eigenvalues ​​of the encoding vector of the live sales time series semantic query response: ;in, represents the length of the encoding vector of the live sales time series semantic query response, represents the hyperbolic sine function, Represents the value of the hyperbolic sine function.

[0048] And calculate its exponential value with the natural constant as the base and then divide it by the square of the length of the live sales time series semantic query response encoding vector to obtain the live sales time series semantic query response potential manifold value ,in, represents a natural constant, Represents the latent manifold value of the live sales temporal semantic query response.

[0049] In response to the live sales time series semantic query response sequence encoding vector The eigenvalue and The absolute value of the covariance between the eigenvalues ​​is greater than the distance difference hyperparameter , multiply the live sales time series semantic query response potential manifold value by the first After the eigenvalue, calculate the product with the first The weighted reduction between the eigenvalues ​​is the optimized : ;in, and represents the weight hyperparameter, Represents the optimized Eigenvalue.

[0050] On the basis of keeping the first eigenvalue of the sequential encoding vector of the live sales time series semantic query response unchanged, the optimized Feature values ​​are used to obtain optimized live sales temporal semantic query response encoding vectors.

[0051] That is, under the condition of preset semantic space trajectory distribution of the semantic feature set of the live sales time series semantic query response coding vector, the phenomenon of cross-domain semantic positioning efficiency attenuation is mainly due to the long-range action distance beyond the local association window. By adopting a nonlinear manifold potential unit representation method based on covariance tensor integration, the multimodal association topological structure implicit in its parameter space is analyzed, and by constructing a nonlinear manifold potential unit model with scale adaptability, the adaptive association representation mechanism between the parameters of the real-time interactive scene time series association mapping matrix is ​​re-established. Therefore, through this reconstruction strategy of the physical state evolution trajectory, the continuous state transition modeling under the long-range association of parameters is effectively realized, which significantly improves the semantic field reconstruction efficiency. At the same time, the original encoding representation characteristics of the live sales time series semantic query response coding vector are maintained, and the accuracy of the trend prediction results obtained by the decoder-based e-commerce live trend predictor is improved.

[0052] In summary, according to the embodiment of the present application, an e-commerce live broadcast trend analysis system based on live broadcast data is explained, which extracts the live broadcast commodity sales data and historical live broadcast sales data of the target commodity to obtain a live broadcast sales data time series set and multiple historical live broadcast sales data time series sets, and uses AI-based data analysis and processing algorithms to perform time series feature analysis on the live broadcast sales data time series set and each historical live broadcast sales data time series set, respectively, so as to intelligently predict the commodity sales data inference value of the next time interval of the target commodity according to the selection range ratio dynamic semantic query representation between the live broadcast sales data time series feature and each historical live broadcast sales data time series feature. The present application can better capture the nonlinear and changeable market dynamics, which enables the model to not only focus on short-term trends, but also understand long-term patterns, thereby improving the depth of understanding of the law of sales volume changes. At the same time, the concepts of "selection range ratio" and "dynamic semantic query representation" are introduced, which can intelligently consider the impact of events such as promotional activities and holidays on sales, and quickly optimize search strategies to provide more accurate and timely sales forecasts.

[0053] An embodiment of the present application further provides a computer program product, which includes a computer program code. When the computer program code runs on a computer, the computer implements the methods in the above embodiments of the present application.

[0054] An embodiment of the present application further provides a computer-readable storage medium, which stores computer instructions. When the computer instructions are executed on a computer, the computer implements the methods in the above embodiments of the present application.

[0055] An embodiment of the present application also provides a chip, including a circuit, for executing the methods in the above embodiments of the present application.

[0056] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0057] In the description of the embodiments of the present application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in this article is a kind of association relationship that describes associated objects, indicating that three relationships can exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In this application, "at least one" means one or more, and "multiple" means two or more. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0058] The prefixes such as "first" and "second" used in the embodiments of the present application are only used to distinguish different description objects, and have no limiting effect on the position, order, priority, quantity or content of the described objects. The use of prefixes such as ordinal numbers used to distinguish description objects in the embodiments of the present application does not constitute a limitation on the described objects. For the statement of the described objects, please refer to the description in the context of the claims or embodiments, and the use of such prefixes should not constitute an unnecessary limitation.

[0059] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0060] In the various embodiments of the present application, unless otherwise specified or logically conflicting, the terms and / or descriptions between the various embodiments are consistent and can be referenced to each other, and the technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationships.

[0061] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0062] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0063] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An e-commerce live broadcast trend analysis system based on live broadcast data, comprising: A data collection module is used to collect and count the live broadcast commodity sales data of each commodity in multiple time intervals within the scheduled live broadcast time period of the target e-commerce live broadcast platform; a historical data collection module is used to collect the historical live broadcast sales data of each commodity in multiple time intervals that are the same as the scheduled live broadcast time period on the target e-commerce live broadcast platform in the past multiple days; a prediction module is used to obtain the trend prediction results of each commodity based on the live broadcast commodity sales data and historical live broadcast sales data of each commodity; a replenishment module is used to issue a replenishment request in response to the trend prediction result of the target commodity exceeding the inventory of the commodity in the next time interval; the prediction module comprises: a data acquisition unit, which is used to extract the live broadcast commodity sales data and historical live broadcast sales data of the target commodity Historical live sales data to obtain a live sales data time series set and multiple historical live sales data time series sets; a data encoding unit, used to perform time series encoding on the live sales data time series set and the multiple historical live sales data time series sets respectively to obtain live sales data time series features and a set of historical live sales data time series features; a data query response unit, used to perform a feature dynamic semantic query response with a selection range ratio anchored on the live sales data time series features and the set of historical live sales data time series features to obtain live sales time series semantic query response features, and based on the live sales time series semantic query response features, obtain the trend prediction result, and the trend prediction result is used to represent the commodity sales data inference value of the next time interval of the target commodity object.

2. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 1 is characterized in that: The data acquisition unit is used to: extract the live sales data of the target product in multiple time intervals within the live broadcast scheduled time period from the data collection module to obtain a live sales data time series set, and extract the live sales data time series sets of the target product in multiple time intervals within the live broadcast scheduled time period in the past multiple days from the historical data collection module to obtain the multiple historical live sales data time series sets.

3. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 2 is characterized in that: The data encoding unit is used to: respectively pass the live sales data timing set and the multiple historical live sales data timing sets through a live sales timing encoder to obtain a live sales data timing feature vector as the live sales data timing feature and a set of historical live sales data timing feature vectors as the set of historical live sales data timing features.

4. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 3 is characterized in that: The live sales timing encoder is a live sales timing encoder based on the 1D-FCN network.

5. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 4 is characterized in that: The data query response unit includes: a current-historical sales data time series query response unit, which is used to select a range-ratio anchored feature dynamic semantic query response for the set of the live sales data time series feature vector and the historical live sales data time series feature vector to obtain a live sales time series semantic query response encoding vector as the live sales time series semantic query response feature; a commodity sales data prediction unit, which is used to obtain the trend prediction result based on the live sales time series semantic query response encoding vector, and the trend prediction result is used to represent the commodity sales data inference value of the next time interval of the target commodity object.

6. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 5 is characterized in that: The current-historical sales data timing query response unit includes: a commodity sales dynamic search anchor subunit, which is used to calculate the live-historical sales data dynamic search anchor vector based on the live sales data timing feature vector and the set of historical live sales data timing feature vectors; a historical live sales data selection range ratio determination subunit, which is used to determine the historical live sales data selection range ratio based on the feature distribution characteristics of the live-historical sales data dynamic search anchor vector; a commodity sales dynamic query subunit, which is used to perform semantic similarity dynamic encoding on each feature in the live sales data timing feature vector and the historical live sales data selection range ratio to obtain the live sales timing semantic query response encoding vector.

7. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 6 is characterized in that: The commodity sales dynamic search anchoring subunit is used to: calculate the internal relationship score value of the live sales data timing feature vector relative to each historical live sales data timing feature vector in the set of historical live sales data timing feature vectors to obtain a set of live-historical sales data internal relationship score values; and use the historical live sales data timing feature vector corresponding to the largest live-historical sales data internal relationship score value in the set of live-historical sales data internal relationship score values ​​as the live-historical sales data dynamic search anchoring vector.

8. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 7 is characterized in that: The vector of the starting position of the historical live sales data selection range ratio is the live-historical sales data dynamic search anchor vector, and each historical live sales data time series feature vector in the historical live sales data selection range ratio is defined as a historical live sales data dynamic search optimization library feature vector.

9. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 8, characterized in that: The commodity sales dynamic query subunit is used to: input the live sales data time series feature vector and each historical live sales data dynamic search optimization library feature vector in the historical live sales data selection range ratio into a dynamic semantic search encoder to obtain the live sales time series semantic query response encoding vector.

10. The e-commerce live broadcast trend analysis system based on live broadcast data according to claim 9, characterized in that: The commodity sales data prediction unit is used to: pass the live sales time series semantic query response encoding vector through a decoder-based e-commerce live trend predictor to obtain the trend prediction result.

Images