Method, device, storage medium and program product for determining seasonal information

Abstract

Embodiments of the present application provide a kind of time information determination method, equipment, storage medium and program product.The method comprises: obtaining the object cluster to be analyzed, and the historical activity sequence corresponding to the object cluster, multiple objects to be analyzed are included in the object cluster, and multiple objects to be analyzed in the same object cluster at least correspond to the same leaf category;In the case where the object cluster is time object cluster with time attribute, based on the historical activity sequence, the predicted activity sequence corresponding to the object cluster is determined;Based on the predicted activity sequence, the time information corresponding to the object cluster is determined.In the embodiment, the object cluster is analyzed by the historical activity sequence of object cluster level, and the time attribute of each object in the cluster is represented by the time attribute of the cluster, so that the problem that the analysis result is not accurate when single object is analyzed and long tail time object is missed is solved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, storage medium and program product for determining seasonal information. Background Technology

[0002] Seasonal phenomena refer to the unique climate, produce, customs, or natural phenomena of a particular season or time of year, with particular emphasis on content related to time rhythms and seasonal changes. In today's highly mature era of instant retail and fresh food e-commerce, the competitive focus of merchandise operations has shifted from simply expanding the quantity of goods to deeply exploring the value of the "time dimension."

[0003] Currently, time-series analysis is often performed on a single product. This can easily lead to inaccurate seasonal analysis results when faced with long-tail products that lack historical data, have zero sales, or experience drastic random fluctuations. It can even result in a large number of long-tail seasonal products being missed. Summary of the Invention

[0004] This application provides a method, device, storage medium, and program product for determining seasonal information, which can solve the problems of inaccurate analysis results and missed detection of long-tail seasonal products when performing time-series analysis on a single product.

[0005] In a first aspect, embodiments of this application provide a method for determining seasonal information, the method comprising: Obtain the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. The object cluster includes multiple objects to be analyzed. The multiple objects to be analyzed in the same object cluster correspond to at least the same leaf category. The historical activity sequence includes activity information at multiple historical time points. In the case where the object cluster is a seasonal object cluster with seasonal attributes, a predicted activity sequence corresponding to the object cluster is determined based on the historical activity sequence. The predicted activity sequence includes the predicted activity at multiple future time points. Based on the predicted activity sequence, seasonal information corresponding to the object cluster is determined, and the seasonal information includes multiple seasonal phases.

[0006] Secondly, embodiments of this application provide a device for determining seasonal information, the device comprising: The first acquisition module is used to acquire the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. The object cluster includes multiple objects to be analyzed, and the multiple objects to be analyzed in the same object cluster correspond to at least the same leaf category. The historical activity sequence includes activity information at multiple historical time points.

[0007] The first determining module is used to determine a predicted activity sequence corresponding to the object cluster based on the historical activity sequence when the object cluster is a seasonal object cluster with seasonal attributes. The predicted activity sequence includes predicted activity at multiple future time points.

[0008] The second determining module is used to determine the timing information corresponding to the object cluster based on the predicted activity sequence, wherein the timing information includes multiple timing stages.

[0009] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the above-described method.

[0010] Fourthly, embodiments of this application provide a computer storage medium for storing a computer program that enables a computer to implement the above-described method when executed.

[0011] Fifthly, embodiments of this application provide a computer program product, the computer program product including a computer program, which, when executed by a processor, can implement the method described in the first aspect.

[0012] The seasonal information determination method provided in this embodiment obtains the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. It then performs seasonal analysis on the historical activity sequence at the cluster level to determine whether the object cluster has seasonal attributes. Finally, it predicts the future activity sequence of seasonal object clusters with seasonal attributes, ultimately determining the seasonal information corresponding to the seasonal object cluster. By using the seasonal attributes of the object cluster to characterize the seasonal attributes of each object within the cluster, it achieves dimensionality reduction and powerful noise reduction of the underlying data. Based on population statistical laws, it can effectively smooth abnormal fluctuations caused by individual objects. Even for long-tail products with extremely sparse historical data, it can rely on its high-quality cluster to obtain a robust time-series baseline, thereby fundamentally solving the problems of missed detection of long-tail seasonal products and unstable analysis results, and improving the accuracy of seasonal analysis results. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a method for determining seasonal information provided in an exemplary embodiment of this application; Figure 2 A flowchart illustrating a method for determining seasonal information provided in an exemplary embodiment of this application; Figure 3 A flowchart illustrating a variety information identification method provided for an exemplary embodiment of this application; Figure 4 A flowchart illustrating a method for determining seasonal information provided in an exemplary embodiment of this application; Figure 5 A flowchart illustrating a method for determining seasonal information provided in an exemplary embodiment of this application; Figure 6 An application diagram illustrating a method for determining seasonal information provided in an exemplary embodiment of this application; Figure 7 A schematic diagram of the structure of a device for determining seasonal information provided in an exemplary embodiment of this application; Figure 8 A schematic diagram of the structure of an electronic device provided for an exemplary embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] It should be noted that, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0016] To facilitate understanding of the method, apparatus, storage medium, and program product for determining seasonal information provided in the embodiments of this application, the relevant technologies are briefly described below: In the context of instant retail and fresh food e-commerce, accurate identification of the seasonality of products is crucial for supply chain scheduling and inventory turnover. Failure to accurately predict the seasonal window of a product will lead to errors in stock preparation and scheduling. However, determining seasonal information faces multiple challenges, specifically in three dimensions: data noise interference, subjective perception differences, and complex temporal interference.

[0017] On the one hand, the sparsity and high noise of single-product data make trend extraction difficult. E-commerce platforms have a vast number of products, and many long-tail products exhibit frequent listing and delisting, long periods of zero sales, or missing sales records in their historical Daily Active Customers (DAC) curves. Even for products with sales records, occasional promotional activities, artificial stockpiling, and inventory fluctuations can cause data oscillations. This random noise often masks the true seasonal patterns of products, rendering statistical analysis based on single-product DAC sequences ineffective and unable to accurately reflect their natural lifecycle trajectory.

[0018] On the other hand, the highly subjective nature of seasonal definitions and the lack of clear labels lead to ambiguous judgment criteria. The definition of seasonal attributes largely relies on human experience and perception. Different operational staff often have differing opinions on whether a product is "seasonal" and what level it belongs to (e.g., strongly seasonal or ordinary seasonal). Furthermore, there is a lack of a unified, standardized labeling system. This subjectivity makes relying solely on human experience or simple rules not only inefficient but also prone to conflict with objective sales data, making it difficult to establish reusable algorithmic benchmarks.

[0019] Furthermore, complex calendar changes and platform promotional activities interfere with the accuracy of seasonal cycle detection. The determination of seasonal cycles is influenced by multiple coupled external factors: on the one hand, traditional Chinese festivals are based on the lunar calendar, and their corresponding Gregorian calendar dates fluctuate annually, causing fixed Gregorian calendar time windows to fail to match peak sales periods for festive goods; on the other hand, frequent promotional activities on e-commerce platforms artificially create sales peaks, thus interfering with the detection of seasonal fluctuation patterns. Existing fixed threshold or static rule models struggle to adaptively isolate these interfering factors, leading to significant deviations in the division of cycle start and end points (such as the listing period and peak season).

[0020] To address the aforementioned issues, the relevant technical solutions all have limitations when dealing with complex scenarios: Short-term analysis solutions based on large language models suffer from both lag and insufficient foresight: these methods rely heavily on short-term sales snapshots, search popularity indices, and summaries of online articles. While they fully utilize the rich world knowledge of large language models to understand product semantics, they are essentially descriptions of phenomena that have already occurred. Lacking the ability to deeply analyze long-term historical sales trends, these solutions often only capture launch signals after a significant sales surge, failing to provide advance prediction and leading to a reactive supply chain response.

[0021] Traditional methods based on purely historical sales statistics are weak in their resistance to interference and have poor adaptability: methods that rely on historical activity sequences for periodic detection are highly susceptible to data noise contamination, and the sparsity of single-product data and abnormal peaks caused by promotions are often misjudged as seasonal features by traditional algorithms. Furthermore, traditional statistical models struggle to flexibly handle the interannual drift of lunar dates and cannot effectively distinguish between "artificial peaks during major promotions" and "natural seasonal peaks." When faced with the changing annual holidays and promotional rhythms, statistical methods based on fixed rules or rigid thresholds often fail, leading to distorted results in determining seasonal information.

[0022] In view of this, this application proposes a method for determining seasonal information to address the above-mentioned problems. By analyzing the historical activity sequence at the cluster level, the seasonal attributes of each object within the cluster are used to characterize the seasonal attributes of the object cluster. This overcomes the analytical difficulties caused by the sparsity of historical data for a single product, zero sales, and random noise interference. Even in scenarios where product sales records are discontinuous and fluctuate drastically, the true seasonal patterns of the product can still be stably identified. Furthermore, by predicting and analyzing the activity sequence of seasonal object clusters, the seasonal information corresponding to the object cluster is determined. This allows the acquisition of seasonal information to precede significant fluctuations in product sales, enabling downstream applications to formulate targeted inventory and scheduling strategies based on the seasonal differences of products. This effectively avoids losses due to expired or out-of-stock seasonal products caused by a lack of targeted planning.

[0023] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart illustrating a method for determining seasonal information provided in an embodiment of this application. The execution subject of this method is a device for determining seasonal information. This device can be implemented as software or a combination of software and hardware. When the device is implemented as hardware, it can be various electronic devices capable of determining seasonal information, including but not limited to e-commerce platforms, servers, etc. When the device is implemented as software, it can be installed in the aforementioned electronic devices. Figure 1 As shown, the method for determining this seasonal information may include: Step S101: Obtain the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. The object cluster includes multiple objects to be analyzed. Multiple objects to be analyzed in the same object cluster correspond to at least the same leaf category. The historical activity sequence includes activity information at multiple historical time points.

[0025] Step S102: When the object cluster is a seasonal object cluster with seasonal attributes, determine the predicted activity sequence corresponding to the object cluster based on the historical activity sequence. The predicted activity sequence includes the predicted activity of multiple future time points.

[0026] Step S103: Based on the predicted activity sequence, determine the timing information corresponding to the object cluster. The timing information includes multiple timing stages.

[0027] The specific implementation principles and effects of each of the above steps are explained in detail below: Step S101: Obtain the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. The object cluster includes multiple objects to be analyzed. Multiple objects to be analyzed in the same object cluster correspond to at least the same leaf category. The historical activity sequence includes activity information at multiple historical time points.

[0028] The method for determining seasonal information provided in this application can be applied to various technical scenarios that require predicting the patterns of temporal fluctuations. "Object" refers to an entity with observable temporal behavioral characteristics. For example, in e-commerce scenarios, the object can be a product or a product cluster, and seasonal information is used to guide supply chain inventory preparation and marketing scheduling; in content recommendation scenarios, the object can be video, text, or audio content, and seasonal information is used to identify seasonal hotspots or festival-related content to assist in content production and distribution; in advertising and marketing scenarios, the object can be advertising materials or marketing activities, and seasonal information is used to determine appropriate placement windows to avoid lagging behind user needs; in tourism planning scenarios, the object can be a tourist destination or travel supplies, and seasonal information is used to predict peak and off-peak passenger flows to guide ticketing and transportation capacity scheduling. Those skilled in the art will understand that any entity with observable temporal behavioral characteristics, whose behavioral patterns are driven by factors such as natural seasons, climate, or cultural festivals, can be considered an object of this application, and the method of this application can be used to determine seasonal information. This application does not exhaustively list all such entities.

[0029] In this embodiment, cluster-level analysis replaces single-item-level analysis, and the seasonal attribute of an object cluster represents the seasonal attribute of each object to be analyzed within the cluster. To ensure that multiple objects to be analyzed within the same object cluster have highly similar seasonal attributes, multiple objects to be analyzed within the same object cluster must correspond to at least the same leaf category. A leaf category refers to the last-level classification node in a classification system. For example, in e-commerce scenarios, "electric fan" and "down jacket" are leaf categories; in content recommendation scenarios, the content category nodes corresponding to "swimming instruction videos" and "Spring Festival blessing videos" can also be considered leaf categories. It is understandable that objects under the same leaf category usually exhibit similar temporal fluctuation patterns. For example, cooling products such as electric fans and air conditioners show obvious peak activity in summer and troughs in winter, while warming products such as down jackets and heaters show peak demand in winter. Therefore, leaf categories can be used as the basis for constructing object clusters, and all objects under the same leaf category can be regarded as one object cluster. For example, all brands and models of electric fans under the leaf category "electric fan" constitute one object cluster.

[0030] There are several ways to obtain the object clusters to be analyzed. If the system has pre-stored the object cluster partitioning results, these results can be read directly. Alternatively, clustering algorithms can be used to automatically mine groups of objects with similar seasonal waveforms from historical time-series data. To ensure basic alignment of temporal patterns within clusters, regardless of the method used, objects within the same cluster must share at least the same leaf category.

[0031] After obtaining the object cluster, it is necessary to acquire the corresponding historical activity sequence. The historical activity sequence includes activity information from multiple historical time points. This activity information can be quantified based on at least one of the following metrics: daily active customers, daily play count, daily search volume, or daily visits. The specific method can be flexibly selected according to the application scenario. This application does not limit the specific method for acquiring the historical activity sequence. As an optional implementation, the historical time window and time granularity can be determined first. To ensure the identification of stable annual cycles, the historical time window is typically at least two years, and the time granularity can be selected as daily or weekly depending on the application scenario. Then, the original behavior logs (containing at least object identifier, timestamp, and user identifier or behavior type) of each object within the object cluster are read from the database or log system within the specified time window. The activity index of each object at each time point is calculated according to the preset time granularity (e.g., Daily Active Customers (DAC) in an e-commerce scenario). Finally, the activity values ​​of all objects within the cluster at the same time point are aggregated (e.g., by taking the mean or median) to obtain a continuous time series representing the overall activity level of the object cluster. If there are missing time points in the series, the gaps can be retained for subsequent algorithm processing or filled using interpolation methods.

[0032] Step S102: When the object cluster is a seasonal object cluster with seasonal attributes, determine the predicted activity sequence corresponding to the object cluster based on the historical activity sequence. The predicted activity sequence includes the predicted activity of multiple future time points.

[0033] After obtaining the object cluster to be analyzed and the corresponding historical activity sequence, the historical activity sequence can be analyzed first to determine whether the object cluster has seasonal attributes.

[0034] Optionally, periodicity detection algorithms (such as Autocorrelation Function (ACF), Autoperiod (Hybrid FFT + ACF PeriodDetection, AUTOPERIOD), Exponential Moving Average (EMA), Robust Time-Frequency Mining for Multiple Periodicity Detection, etc.) can be used to analyze historical activity sequences and extract the principal period length from the sequence. Then, based on the extracted principal period length, it is determined whether the object cluster has seasonal attributes. Specifically, if the principal period length is greater than or equal to a preset period threshold, and the sequence corresponds to a peak amplitude or holiday pulse intensity, then the object cluster is determined to have seasonal attributes. If the principal period length is less than the preset period threshold, or the sequence does not correspond to a peak amplitude or holiday pulse intensity, then the object cluster is determined not to have seasonal attributes.

[0035] Optionally, the judgment results can be semantically calibrated using a large language model, leveraging its common-sense knowledge to correct for boundary ambiguities in periodicity detection. In scenarios where historical activity sequences are missing or data quality is insufficient to support periodicity detection, a large language model can be used alone to directly determine the seasonal attributes of object clusters based on its built-in natural knowledge and semantic understanding capabilities. For example, by inputting the leaf category name or related descriptive information of an object cluster into the large language model, the model can infer whether the corresponding object cluster possesses seasonal attributes based on natural seasonal patterns, climate characteristics, or common-sense knowledge of folk festivals.

[0036] However, in practical applications, the periodic characteristics of different objects are not completely consistent, and a single algorithm is difficult to take into account the seasonal patterns of various objects. Relying solely on the knowledge of large language models for judgment may also lead to deviations from real historical data. Therefore, a combination of multiple periodic detection algorithms and large language models can be used to jointly determine seasonal attributes.

[0037] Optionally, after obtaining the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster, the method in this embodiment may further include: processing the historical activity sequence using multiple different period detection algorithms to obtain multiple period detection results corresponding to the object cluster; processing the variety information and leaf category corresponding to the object cluster using a large language model to obtain an identification result of whether the object cluster has a seasonal attribute; and determining the object cluster as a seasonal object cluster if at least two of the multiple period detection results and identification results indicate that the object cluster has a seasonal attribute.

[0038] In practical implementation, various time-frequency domain periodicity detection algorithms, such as ACF, AUTOPERIOD, and EMA+RobustPeriod, can be used for analysis. Each algorithm has its advantages in periodicity identification: the autocorrelation function (ACF) is more suitable for detecting clearly repeating periodic structures in sequences; AUTOPERIOD is more robust to local noise; and the exponential moving average combined with the robust periodicity detection framework (EMA+RobustPeriod) can more accurately capture the main period in smoothed sequences. By introducing multiple algorithms in conjunction with a large language model for joint judgment, the bias caused by a single algorithm can be effectively reduced, improving the comprehensiveness and reliability of periodicity identification.

[0039] When determining whether an object cluster has a seasonal attribute, the determination can be made based on quantitative indicators such as the period length, peak amplitude, and holiday matching degree output by the period detection algorithm. Optionally, after obtaining multiple period detection results corresponding to the object cluster, the time periods with seasonal attributes included in the multiple period detection results can be determined; if the duration of the time period is within a preset duration range, the corresponding period detection result indicates that the object cluster has a seasonal attribute.

[0040] After analyzing historical activity sequences, the period detection algorithm can output the following results: the period length of the sequence (such as a 52-week annual cycle, a 26-week semi-annual cycle, etc.), the period strength or confidence score (used to quantify the significance of the period), the time location and amplitude of the peak, the location and amplitude of the trough, and whether pulse fluctuations related to specific holidays are detected (including pulse time, amplitude, and duration).

[0041] In one optional embodiment, when determining whether an object cluster has seasonal attributes, the annual cycle can be used as the primary detection target. Since the length of a natural year is approximately 52 weeks, considering the interannual drift of lunar and Gregorian calendar dates (e.g., the Spring Festival fluctuates between late January and mid-February), and the window adjustments caused by leap years and weekend distribution in different years, and considering the potential ± several weeks of calculation error in peak positioning by cycle detection algorithms (such as ACF, AUTOPERIOD, EMA+RobustPeriod, etc.), a main cycle length falling between 48 and 56 weeks can be considered a valid annual cycle. In other words, if a cycle detection result shows that the duration of a time period with seasonal attributes (i.e., the main cycle) is within the preset duration (48 to 56 weeks), then the corresponding cycle detection result indicates that the object cluster has seasonal attributes.

[0042] It should be noted that, to ensure the reliability of the annual cycle determination, at least two years of historical activity sequences can be used for cycle detection. The reason is that using only one year's data cannot distinguish between true annual cyclical patterns and accidental single fluctuations (such as a sales increase in a particular year due to abnormal weather or a one-off promotional event); while using data from at least two complete years, it is possible to observe whether the cycle appears stably in consecutive years, thereby effectively eliminating random disturbances and interference from abnormal events in a single year.

[0043] The input to the large language model can be the variety information and leaf category information corresponding to the object cluster. If the object cluster already has structured stored leaf category information or variety information, it can be read directly; otherwise, a multimodal recognition model can be used to analyze the title, descriptive text, or image of the object to output the corresponding leaf category and variety information. Then, the obtained leaf category and variety information can be used as input to construct natural language prompts, which are then input into the large language model. The large language model then combines its built-in world knowledge (including natural seasonal patterns, climate characteristics, and common knowledge of folk festivals) to perform reasoning and obtain the seasonal attribute analysis results.

[0044] Subsequently, a voting mechanism can be used to confirm the final seasonal attribute determination result for the analysis results of multiple periodic detection results and large language models. For example, if at least two of the multiple periodic detection results and recognition results indicate that the object cluster has a seasonal attribute, the object cluster is determined to be a seasonal object cluster.

[0045] If the judgment result shows that the object cluster does not have a seasonal attribute, the result can be output directly, indicating that the object to be analyzed does not have a seasonal attribute, and the judgment basis can be output (for example, the main cycle length is 60 weeks, which exceeds the range of 48-56 weeks); if the judgment result shows that the object cluster has a seasonal attribute, the activity prediction of the seasonal object cluster can continue.

[0046] In practical applications, it is common practice to predict the activity sequence for the coming year. Optionally, a time-series forecasting model can be used to determine the predicted activity sequence. Based on historical activity sequences, the predicted activity sequence corresponding to the object cluster is determined, including: inputting the historical activity sequence and holiday information into the time-series forecasting model for processing to obtain the seasonal time series and holiday distribution information corresponding to the object cluster. The seasonal time series is a Fourier series, which is used to represent the periodic change pattern of activity in the future time period, and the holiday distribution information follows a Gaussian distribution, which is used to represent the local fluctuations of activity around holidays in the future time period; the predicted activity sequence is determined based on the seasonal time series and holiday distribution information.

[0047] Time series forecasting models can include Prophet (Facebook's open-source additive model time series forecasting tool), STL (Seasonal-Trend Decomposition using LOESS), N-BEATS (Neural Basis Expansion Analysis for Time Series), ARIMA (Autoregressive Integrated Moving Average), LSTM (Long Short-Term Memory), etc. Holiday information can be user-defined based on application needs, such as lunar festivals like Spring Festival, Dragon Boat Festival, and Mid-Autumn Festival, as well as Gregorian calendar holidays like New Year's Day and Labor Day. It can also include e-commerce promotional days like Singles' Day and 618. By flexibly configuring the names and dates of holidays (for lunar festivals, dynamic calculation based on the corresponding Gregorian calendar dates each year) and their impact windows (such as the number of days before and after the holiday), time series forecasting models can more accurately model local fluctuations in activity during holidays.

[0048] After inputting historical activity sequences and holiday information into the time series forecasting model, the model can decompose the historical activity sequence into a trend term, a seasonal time series, and holiday distribution information. The trend term represents the long-term growth or decline trend of the object cluster (e.g., linear growth, saturation growth, or a stable trend); the seasonal time series is fitted with a Fourier series to represent the periodic changes in activity over future time periods (e.g., annual cycle, weekly cycle, and peaks and troughs); and the holiday distribution information follows a Gaussian distribution to represent the local fluctuations in activity around holidays over future time periods.

[0049] Subsequently, seasonal time series and holiday distribution information can be extracted. Based on the waveform morphology of the seasonal time series (such as peak amplitude, window length, symmetry, etc.) and the pulse intensity and holiday distribution information matching degree, a predicted activity sequence for the object cluster is generated based on the predicted specific time nodes. An optional implementation method is to cyclically copy the historical waveform of the seasonal time series to the next year, and simultaneously superimpose corresponding effect values ​​on the holiday distribution information according to the Gregorian calendar dates of holidays in the next year. Then, these two predicted values ​​are added together to obtain the complete predicted sequence for the next year.

[0050] Alternatively, other methods can be used to determine the predicted activity sequence corresponding to the object cluster. For example, based on the moving average method or weighted moving average method, the mean of several recent time points in the historical activity sequence can be used as the predicted value for the next time point, and the predicted activity sequence for multiple future time points can be generated recursively. Alternatively, based on the exponential smoothing method (such as the Holt-Winters model), historical data can be smoothed and trend and seasonal components can be captured for extrapolation. Alternatively, based on the Autoregressive Integrated Moving Average (ARIMA) model or its seasonal variant (Seasonal Autoregressive Integrated Moving Average (SARIMA), the autocorrelation and partial autocorrelation structure of the historical sequence can be modeled and predicted forward. Or based on the linear regression model, the time sequence number can be used as the independent variable and historical activity as the dependent variable, and a linear trend can be fitted for extrapolation and prediction.

[0051] Step S103: Based on the predicted activity sequence, determine the timing information corresponding to the object cluster. The timing information includes multiple timing stages.

[0052] In this embodiment, an object cluster can contain multiple seasonal cycles. For example, an air conditioner cluster may have both summer cooling peaks and winter heating peaks. The division of seasonal phases can be done independently for each peak. The complete process of an object cluster's activity rising and falling within a seasonal cycle can be divided into multiple seasonal phases. Taking e-commerce as an example, a complete seasonal cycle can typically be divided into four consecutive seasonal phases: the launch period (a warm-up phase where activity gradually rises from the daily baseline), the peak period (a peak phase where activity reaches the annual peak and remains at a high level), the decline period (a decline phase where activity begins to fall from the peak), and the end period (a closing phase where activity returns to below the daily baseline level). In some instances, the seasonal information also includes time information corresponding to each of the multiple seasonal phases. In this case, the "time information" corresponding to each seasonal phase includes the start and end dates of each phase. Optionally, the peak date of the peak period can also be recorded.

[0053] For example, suppose a product cluster is related to "mooncakes". Predictive activity sequence analysis shows that this cluster has a complete seasonal cycle from early August to late October each year. Specifically, the listing period is approximately August 1st to August 31st, the peak period is approximately September 1st to October 5th (peaking around September 29th), the decline period is approximately October 6th to October 20th, and the off-season is approximately October 21st to October 31st. The seasonal information for this cluster can include the time window corresponding to each complete seasonal cycle (August 1st to October 31st), and the start and end times of each seasonal phase (as described above).

[0054] After obtaining the predicted activity sequence, seasonal information can be determined in various ways. For example, it can be based on key change points: calculate the first-order difference of the predicted activity sequence to identify inflection points where activity changes from negative to positive and from positive to negative, and use local extrema detection to locate troughs, peaks, decline inflection points, and the end point where activity falls back below the baseline. Taking e-commerce as an example, when activity rises continuously and exceeds the daily baseline for the first time (such as the average of the past three months), it can be marked as the start point of the listing period; when the rate of increase changes from positive to negative, it is marked as the peak point; when the rate of decrease changes from negative to positive, it is marked as the decline inflection point; when activity falls back below the baseline and continues for a preset number of time points (such as three consecutive weeks), it is marked as the end point of the delisting period. In chronological order, the period from the start point of the listing period to the peak point is the listing period; the period extending several time steps before and after the peak point is the boom period; the period from the peak point to the decline inflection point is the decline period; and the period from the decline inflection point to the end point of the delisting period is the delisting period. The start and end dates of each stage are read directly from the time coordinates of the corresponding key points.

[0055] Alternatively, a preset threshold method can be used: trigger threshold conditions for each stage are pre-defined. For example, the listing period condition is that the weekly average activity level exceeds 1.2 times the annual average activity level for two consecutive weeks; the boom period condition is that the weekly average activity level exceeds 1.5 times the annual average activity level and the week-on-week growth rate is positive; the decline period condition is that the week-on-week growth rate turns negative and declines for two consecutive weeks; the delisting period condition is that the weekly average activity level falls below the annual average and there is no rebound for three consecutive weeks. Scanning the prediction sequence, the moment when the listing period condition is first met is recorded as the listing period start point, the moment when the boom period condition is first met is recorded as the boom period start point, the moment when the decline period condition is first met is recorded as the decline period start point, and the moment when the delisting period condition is first met is recorded as the delisting period start point. The interval between the start points of two adjacent stages corresponds to the time window of the previous stage (for example, the period from the listing period start point to the boom period start point is the listing period, the period from the boom period start point to the decline period start point is the boom period, the period from the decline period start point to the delisting period start point is the decline period, and the period after the delisting period start point is the delisting period). If the conditions for a certain stage are never met, then that stage is empty.

[0056] In addition, a template-based matching approach can be used: if the object cluster has waveform templates for standard seasonal periods in historical years, the templates can be matched with the current predicted sequence through dynamic time warping or correlation coefficient methods. The stage boundaries in the templates can be adapted to the current time axis through translation and scaling, thereby obtaining the time windows for each stage. Alternatively, a sequence labeling model can be used: a sequence labeling model is constructed, taking the predicted activity sequence as input, outputting stage labels (such as listing period, outbreak period, decline period, delisting period, and stable period) for each time point, and then extracting intervals with consecutive identical labels as the time windows for the corresponding seasonal stages.

[0057] It should be noted that the scheme in this application uses the seasonal attributes of object clusters to characterize the seasonal attributes of multiple objects to be analyzed within the cluster. In practical applications, the input of this scheme can be one or more objects to be analyzed, and the output result is the seasonal analysis result corresponding to these one or more objects to be analyzed, rather than directly outputting the cluster-level analysis result. If the object to be analyzed does not have seasonal attributes, the output result will mark the object as "non-seasonal" and output the corresponding judgment criteria, such as "the main cycle length does not fall within the range of 48 to 56 weeks" or "the peak amplitude of the seasonal item is less than 1.5 times the annual average and there are no significant holiday pulses". If the object to be analyzed has seasonal attributes, it will be marked as "seasonal" in the output results, and the judgment criteria and corresponding seasonal information will also be output: the judgment criteria can be "the main cycle length is 52 weeks, the peak amplitude of the seasonal item is 2.3 times the annual average, and the pulse amplitude before and after the Dragon Boat Festival exceeds 3 times the baseline"; the seasonal information can include multiple seasonal phases in the predicted future year and the time information corresponding to each seasonal phase, such as "listing period: May 1 to May 15; outbreak period: May 16 to June 10; decline period: June 11 to June 25; delisting period: after June 26".

[0058] By predicting the seasonal information of the objects to be analyzed, differentiated lifecycle management strategies can be configured for different objects. For example, resources can be allocated and initial exposure can be introduced during the listing period, resource investment can be increased and supply capacity can be guaranteed during the boom period, a gradual exit mechanism can be initiated during the decline period, and resource allocation can be stopped during the delisting period to prepare for the next cycle planning. Compared with relying on historical data from the same period, this can effectively reduce decision-making delays, synchronize actions at each stage with actual activity changes, and thus improve the timeliness of strategy response.

[0059] In practical applications, different seasonal objects possess different seasonal characteristics. For example, the market window for lychees is very short, lasting only a few weeks; the peak sales season for electric fans is concentrated in the summer and lasts for several months; and mooncakes experience a sales peak around the Mid-Autumn Festival, with the date varying annually according to the lunar calendar. Therefore, optionally, after determining the seasonal information corresponding to the object cluster, seasonal object clusters can be classified into seasonal layers based on their seasonal characteristics, and the seasonal layering results can be displayed in the output.

[0060] Optionally, the seasonal object cluster can be divided into three levels: strongly seasonal, seasonal, and festival-type. Strongly seasonal objects (such as lychees and mangoes in specific seasons) typically refer to objects that are not observable year-round but have concentrated activity windows; seasonal objects (such as electric fans in summer and down jackets in winter) typically refer to objects that are observable year-round but whose activity fluctuates significantly in specific seasons; festival-type objects (such as zongzi for the Dragon Boat Festival and Spring Festival couplets for the Spring Festival) typically refer to objects highly correlated with specific festivals or solar terms. For festival-type objects, the corresponding festival name also needs to be output.

[0061] In the specific implementation process, since the seasonal stratification task not only depends on the time series form, but also on object knowledge, consumption habits and festival semantic understanding, an optional implementation method can be to complete the seasonal stratification by using a few-sample discrimination method based on a large language model. By providing a small number of labeled examples (such as a few typical examples) to the language model, the model is guided to learn the discrimination criteria of seasonal objects at different levels, and makes a comprehensive judgment by combining the object's name, category information and the aforementioned seasonal cycle information. The number of labeled examples is not limited, usually a few examples are sufficient. These examples are concatenated in the prompt words as context to guide the model to make judgments.

[0062] To ensure the reliability of festival-type classification, a seasonal month constraint can be added to the output of the large language model: for objects predicted by the model as festival-type, the output festival or solar term must fall within the seasonal month range corresponding to the object's seasonal cycle; if the festival name does not match the seasonal month, it is considered that the constraint condition is not met, and it needs to be reclassified or reverted to another seasonal category level (strong seasonality or seasonal type). Through this constraint mechanism, the misjudgment caused by the large language model relying solely on knowledge association can be effectively reduced, ensuring the accuracy of the classification of festival-type seasonal objects.

[0063] Based on the seasonal level information in the output, users can formulate differentiated resource investment and time scheduling strategies for seasonal objects at different levels, thereby improving supply chain response efficiency, reducing the expiration losses of strongly seasonal objects and the stockout risk of festival objects, while avoiding resource misallocation of seasonal objects and achieving refined operation.

[0064] The method for determining seasonal information provided in this embodiment obtains the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. When the object cluster is a seasonal object cluster with seasonal attributes, it determines the predicted activity sequence corresponding to the object cluster based on the historical activity sequence, and then determines the seasonal information corresponding to the object cluster based on the predicted activity sequence. By analyzing objects at the cluster level and using the seasonal attributes of the object cluster to characterize the seasonal attributes of each object within the cluster, it can effectively smooth out abnormal fluctuations caused by individual objects, obtaining a more stable pattern of activity changes. Even for long-tail products with extremely sparse historical data, a robust time-series baseline can be obtained by attaching to the high-quality cluster to which they belong, thus obtaining accurate seasonal information. Furthermore, determining seasonal information based on prediction results rather than historical statistical data can effectively reduce lag in the decision-making process and improve the timeliness of response to activity changes.

[0065] Figure 2 This application provides a flowchart of a method for determining seasonal information according to an embodiment of the present application; based on the above embodiment, it also provides a method for obtaining object clusters, see attached diagram. Figure 2As shown, before obtaining the cluster of objects to be analyzed, an aggregation operation can be performed based on the aggregation parameters of multiple objects. This method can include the following steps: Step S201: Obtain multiple objects.

[0066] The method for acquiring multiple objects can be flexibly determined based on the application scenario and data source. For example, in an e-commerce scenario, product identifiers and their attribute information under a specified category can be read from a product database or data warehouse, including product ID, leaf category, standard variety, place of origin, brand, and specifications; alternatively, product lists uploaded by operators through a configuration interface can be received, or product data can be obtained from a third-party system via API. In a content recommendation scenario, identifiers and metadata of video, image, or audio objects can be read from a content management system, including content ID, leaf category, topic tags, publisher, and regional information. This application does not limit the specific means of acquiring multiple objects; any method that can obtain a set of objects with seasonal analysis requirements and their basic attribute information is applicable. After acquiring the objects, subsequent clustering, seasonal attribute identification, and prediction operations can be performed.

[0067] Step S202: Determine the aggregation parameters for performing aggregation operations on multiple objects based on the object types corresponding to the multiple objects.

[0068] Aggregation parameters refer to key features or attribute fields used to measure the similarity between objects and determine the grouping criteria when performing clustering operations on multiple objects. In this embodiment, the goal of clustering multiple objects is to make multiple objects within the same object cluster have similar seasonal patterns or temporal fluctuations, thereby enabling unified seasonal information identification and prediction at the cluster level. In reality, different objects have different factors that influence their seasonal attributes. For example, in content recommendation scenarios, the seasonal attributes of content under the category "swimming instruction videos" are mainly determined by the leaf category, with peak playback concentrated in summer; while the seasonal attributes of content such as "Spring Festival Gala replays" are mainly determined by the content theme, with peak playback strongly correlated with the Spring Festival date; similarly, the seasonal attributes of videos related to "cherry blossom festivals in a certain area" are jointly determined by the content theme and regional information, with different regions having different cherry blossom festival dates. Therefore, multiple objects can be classified according to the factors affecting their seasonal attributes, and aggregation parameters can be determined based on the corresponding object type for aggregation operations.

[0069] Optionally, clustering can be performed in the following manner, determining the aggregation parameters for the aggregation operation on multiple objects based on their respective object types, including: When the object type is a preset standard object, the leaf category of the object is determined as the aggregation parameter; When the object type is a first-class object, the leaf category and variety information of the object are determined as aggregation parameters. The first-class object is an object whose activity can be affected by the variety information. When the object type is the second type, the leaf category, variety information and origin information of the object are determined as aggregation parameters. The second type of object is the object whose activity can be affected by the variety information and origin information.

[0070] Preset standard objects refer to object types whose seasonal attributes are primarily influenced by leaf categories. For example, in e-commerce scenarios, for standardized industrial products such as electric fans and down jackets, their seasonal attributes are mainly affected by leaf categories (such as "electric fans" and "down jackets"). Different brands and models of products under the same leaf category have similar seasonal fluctuation patterns, so the leaf category is used as an aggregation parameter. In content recommendation scenarios, for content with clear themes such as "swimming instruction videos" and "skiing tutorials," their seasonal attributes are determined by leaf categories (such as "swimming" and "skiing"). Different video content under the same leaf category shows similar playback peaks in summer or winter, so the leaf category is used as an aggregation parameter. In tourism planning scenarios, for tourist destination types such as "island vacations" and "ski resorts," their seasonal attributes are determined by leaf categories (such as "island" and "ski resorts"). Different destinations under the same leaf category experience peak tourist flows in summer or winter, so the leaf category is used as an aggregation parameter.

[0071] The first type of object refers to those whose activity level can be affected by variety information. Activity level reflects the seasonality of an object, and the seasonality of the first type of object is not only related to its corresponding leaf category but also affected by variety information. For example, in e-commerce scenarios, for food and fresh produce, such as live shrimp, zongzi (sticky rice dumplings), and duck eggs, different varieties may have different market launch schedules and consumption patterns. For instance, different varieties of live shrimp (such as tiger prawns, crayfish, and mantis shrimp) each have different market launch windows; different filling varieties of zongzi (such as meat zongzi, red bean paste zongzi, and salted egg yolk zongzi) may have slightly different consumption peaks before and after the Dragon Boat Festival; and different processed varieties of duck eggs (such as salted duck eggs and preserved eggs) also have different festival-related patterns. The first type of object cannot effectively distinguish its seasonality based solely on leaf categories. Therefore, the leaf category and variety information of the first type of object can be used as aggregation parameters to achieve more accurate seasonal clustering.

[0072] The second category of objects refers to those whose seasonality can be influenced by both variety and origin information. For objects whose seasonal attributes cannot be effectively distinguished by leaf category and variety information alone, more granular aggregation parameters need to be introduced. Taking fruit objects as an example, such as lychee and mango, in addition to variety differences, origin has a significant impact on ripening and market availability: the market availability of the same variety in different origins (such as Hainan, Guangdong, and Guangxi) is successively delayed, forming a clear regional ladder; the ripening periods of different varieties (such as Fei Zi Xiao, Bai Tang Ying, and Gui Wei) can differ by several weeks. Therefore, relying solely on leaf category and variety information cannot accurately capture their seasonal attributes; it is necessary to combine leaf category, variety information, and origin information as aggregation parameters to construct an object cluster that can reflect the true seasonal attributes.

[0073] In practical applications, aggregation parameters can be flexibly adjusted according to different application scenarios. Taking content recommendation scenarios as an example: for content with obvious seasonality, such as "swimming instruction videos" and "skiing tutorials," leaf categories (e.g., "swimming," "skiing") can be used as aggregation parameters to group videos under the same category together, utilizing the high correlation between their peak playback times and the season for seasonal analysis; for festival-related content such as "Spring Festival Gala replays" and "Mid-Autumn Festival Gala highlights," leaf categories and content themes (variety information) can be combined as aggregation parameters, for example, grouping videos related to the "Spring Festival Gala" together to capture the strong correlation between their peak playback times and the festival date; for regional hot topics such as "live broadcast of a cherry blossom festival in a certain place" and "review of a marathon in a certain city," leaf categories, content themes, and regional information should be used together as aggregation parameters (e.g., "cherry blossom festival + Wuhan") to reflect the time window differences of the same event due to different regions. In addition, other aggregation features can be introduced according to application needs, such as content duration (short, medium, and long videos) or production party (different media organizations), to distinguish the seasonal performance differences of different types of content under the same event.

[0074] The above-mentioned object type classification can be based on preset rules using human experience, or it can be automatically determined by analyzing the clustering results of historical activity data, or dynamically adjusted by fusing rules and data. This application does not limit this approach; any method that can classify objects into multiple types based on the influencing factors of seasonal attributes is applicable.

[0075] After determining the classification rules for object types, a trained classification model (such as one based on decision trees, random forests, or lightweight neural networks) can be used to determine the type of the objects to be classified. Alternatively, a rule-based inference engine (e.g., matching a pre-defined keyword list based on leaf category names) or a combination of both can be employed. For objects with a defined object type, aggregation parameters can be determined based on the corresponding object type. This differentiated aggregation granularity strategy can effectively improve the accuracy of clustering, providing accurate input for subsequent seasonal analysis.

[0076] When the object type is a preset standard object, its leaf category can be obtained as an aggregation parameter by analyzing the object information. For example, when the object information is stored in a structured form in a database, the corresponding category field can be read directly; if only semi-structured or unstructured information such as the object's title and description text is available, a natural language processing model (such as a pre-trained language model) can be used for entity recognition and category inference to extract the leaf category; if the object image is included, an image recognition model or a multimodal model can be combined to assist in the judgment.

[0077] When the object type is a first-class object, it is necessary to obtain not only the leaf category information, but also the variety information.

[0078] Optionally, when the object type is a first type of object, the object context corresponding to each of the multiple objects is obtained, and the object context includes at least one of the following: title information, image, attribute information; based on the object context, the candidate varieties corresponding to each of the multiple objects are determined; based on the candidate varieties and the object context, the variety information corresponding to each of the multiple objects is determined.

[0079] In practical applications, the content included in the object context, such as title information, slice information, and attribute information, may contain variety information. For example, in e-commerce scenarios, the title of a product usually contains information about the variety, such as the explicit variety name "Feizixiao" in "Freshly Smile Lychees from Hainan". In content recommendation scenarios, the title or tags of a video may also contain variety information, such as "Spring Festival Gala" in "Spring Festival Gala Replay", which corresponds to the variety.

[0080] In one optional implementation, object context information can be identified using a multimodal model to determine variety information. Based on the object context, determining candidate varieties corresponding to multiple objects includes: inputting the object context into a pre-trained multimodal model to obtain predicted varieties output by the multimodal model; performing recall based on the predicted varieties to obtain multiple recalled varieties; and determining candidate varieties based on the predicted varieties and the multiple recalled varieties.

[0081] Multimodal models can extract cross-modal semantic features based on object contextual information (such as titles, descriptions, and images), outputting one or more predicted variety labels with confidence scores. For example, the model outputs a confidence score of 0.85 for "Feizixiao," 0.10 for "Baitangying," and 0.05 for "Guiwei." Then, these predicted varieties can be used as query conditions to perform a recall operation in a pre-defined variety knowledge base or product index. Recall methods can include: fuzzy matching based on string edit distance, extended matching based on a thesaurus, and similarity retrieval based on vector embedding. Through the recall operation, multiple recalled varieties with semantic similarity or association with the predicted variety are obtained. The predicted variety and the recalled varieties are concatenated to obtain a candidate variety set.

[0082] Another alternative implementation involves searching a pre-built variety dictionary based on contextual information to obtain a set of candidate varieties. The variety dictionary can be constructed based on human experience or historical data, and includes the standard variety names of various objects, their common aliases, misspellings, and abbreviations, specifically represented as a set of "standard word-synonym" mapping pairs. For example, for the leaf category "pear," the mapping pair "Snowflake Pear → [Snow Pear, Rock Sugar Pear, Rock Sugar Snow Pear]" can be stored; for "lychee," "Fei Zi Xiao → [Fei Zi Xiao Lychee, New Ball Honey Lychee]" can be stored. After obtaining the object context, the text information (such as titles and attributes) is first segmented and cleaned, and then the matched variety words are retrieved from the dictionary using string matching (such as maximum forward matching, AC automata) or regular expressions. For image information, OCR technology can be used to extract the text before dictionary matching. Through the above retrieval operations, one or more matched variety words are obtained, and these variety words are used as a set of candidate varieties. After obtaining the set of candidate varieties, the predicted variety is further determined from the candidate varieties. If the candidate variety set contains multiple variety terms, they can be sorted according to word frequency, position weight (e.g., words appearing at the beginning of the title have higher weight) or the variety priority preset by the dictionary, and the variety with the highest score can be selected as the predicted variety.

[0083] After obtaining the candidate variety set, the final variety information can be determined based on the candidate varieties and the object context.

[0084] Optionally, for each candidate variety, its matching confidence with the object context can be calculated. This can be achieved by comprehensively scoring dimensions such as the original confidence output from the multimodal model, the co-occurrence frequency with context keywords, and the consistency between the variety and leaf category, with the variety scoring the highest being output as the final variety information. Alternatively, preset variety priorities in the dictionary (e.g., priority for common varieties, priority for standard names) can be used as the primary basis, directly selecting the variety with the highest priority as the final variety information, i.e., directly using the predicted variety as the variety information.

[0085] It should be noted that the output variety information is a standard term, which refers to the variety name after standardized mapping, and is used to eliminate ambiguities caused by synonyms, aliases or typos, so as to ensure that the same variety has a consistent identifier among different objects. For example, "Feizixiao" can be mapped to the standard variety "Feizixiao litchi", "crayfish" is mapped to "Procambarus clarkii", and "cherry" is mapped to "Prunus avium". By outputting standard terms, a unified and reliable variety identifier can be provided for subsequent clustering and aggregation parameters, so as to avoid clustering deviation caused by inconsistent writing of variety names.

[0086] Optionally, a multimodal model and a variety dictionary can be used in combination, and a large language model is adopted to determine the final variety information. Determining the variety information corresponding to each of the multiple objects based on the candidate varieties and the object context includes: Determining a pre-constructed variety dictionary; using a large language model to process the candidate varieties, the object context and the variety dictionary to obtain the variety information corresponding to each of the multiple objects.

[0087] As Figure 3 shown in the schematic flowchart of a variety information identification method provided in the embodiments of the present application, the following combines Figure 3 to specifically describe the process of identifying variety information by using a multimodal model, a variety dictionary and a large language model in combination: First, the corresponding variety dictionary can be determined based on the application scenario, and retrieval is performed in the pre-constructed variety dictionary based on the object context to determine whether the object context directly contains content that matches the variety information recorded in the variety dictionary. All standard varieties and their aliases and variants in the pre-constructed variety dictionary can be presented in the form of a prefix tree (Trie), and the maximum forward matching algorithm can be used to scan the text of the object context (such as titles and attribute descriptions): starting from the starting position of the text, try to match the longest entry in the dictionary each time. If the matching is successful, the corresponding variety name is hit and its position is recorded; if no match is found, slide one character to continue the matching until the scanning is completed.

[0088] If the matching is successful, variety normalization processing is performed on the hit variety name to map it to a standard variety name (for example, mapping "Feizixiao" to "Feizixiao litchi"), and the standard term is output as the variety information.

[0089] If no variety term is matched, the multimodal model is used to perform semantic analysis on the object context to determine the predicted variety, and the predicted variety is used as a query condition to perform similarity retrieval in the pre-constructed variety table: methods such as cosine similarity calculation based on vector embedding, fuzzy matching based on edit distance, or direct screening based on confidence scores can be adopted to recall a plurality of candidate varieties that are semantically similar or associated with the predicted variety.

[0090] Next, the predicted variety and candidate varieties can be concatenated to construct a RAG Prompt (Retrieval Enhancement Prompt). This prompt can be designed as: "Given the following variety mapping relationships: {Standard word A: [Synonym 1, Synonym 2...], Standard word B: [...]}. Please convert the predicted variety word '{Predicted word}' into the most suitable standard word." This prompt is then input into a large language model. The large language model combines the object context, variety dictionary, and natural knowledge (such as the variety's origin association, growing season, common aliases, market perception, etc.) to perform semantic understanding and disambiguation on the candidate varieties, outputting normalized variety information. The large language model can make judgments based on various methods. For example, it can make judgments based on few-shot learning, providing several typical "candidate variety-context-standard variety" mapping examples in the prompt to guide the model to output the most matching standard word; alternatively, it can employ a self-consistent mechanism, sampling the same input multiple times and voting to select the output with the highest frequency to reduce the randomness of a single inference.

[0091] Optionally, the output of the large language model may include variety information and prompts, so as to perform dynamic updates of the variety dictionary based on the prompts.

[0092] After processing candidate varieties, object contexts, and variety dictionaries using a large language model, the method also includes: When the output of the large language model includes the first prompt information, the newly added synonyms are added to the variety dictionary based on the first prompt information. The first prompt information is used to identify the existence of newly added synonyms in the output. When the output of the large language model includes the second prompt information, the newly added standard words are identified as knowledge to be added to the database based on the second prompt information. If the knowledge to be added to the database passes the review, the newly added standard words are added to the variety dictionary. The second prompt information is used to identify the existence of newly added standard words in the output.

[0093] After the large language model analyzes the candidate varieties, object context, and variety dictionary, if it determines that the predicted variety is a synonym that is not included in the variety dictionary but corresponds to a standard word, it outputs the standard word and the first prompt information, and automatically adds the new synonym to the variety dictionary. The standard word represents the standard variety name corresponding to the predicted variety, and the synonym represents the alternative name, variant, etc. of the standard variety. If it determines that the predicted variety does not match any of the standard words in the variety dictionary, it outputs the predicted variety and the second prompt information, and then it is manually reviewed. After the review is approved, the predicted variety is added to the variety dictionary as a standard word.

[0094] Optionally, if the predicted variety is highly similar to a candidate variety after analysis (e.g., similarity greater than 98%), it can be considered that the predicted variety has been included as a synonym. The standard word corresponding to the predicted variety (i.e., the candidate variety) is output and the "synonym already exists" flag is output.

[0095] Optionally, if the large language model finds after analysis that the leaf category to which the input object currently belongs is inconsistent with the expected category corresponding to its actual variety, the variety information can be output as "None" and a "Classification Misclassified" label can be output to facilitate subsequent manual correction or automatic category migration.

[0096] Based on the above process, by combining variety dictionary and large language model with multimodal information to jointly judge the variety information of the object, the accuracy of recognition can be effectively improved. Furthermore, by continuously and dynamically updating the dictionary and enriching the entries and synonym mapping relationships of the variety dictionary, more comprehensive and accurate basic data support can be provided for subsequent variety recognition tasks.

[0097] In practical applications, there may be situations where relying solely on leaf category and variety information is insufficient to effectively distinguish the seasonal attributes of objects. In such cases, more granular aggregation parameters need to be introduced. For example, for fruit objects, different varieties of lychee (Feizixiao, Baitangying) have different ripening periods, and even the same variety of "Feizixiao" has a delayed market launch time depending on its origin (Hainan, Guangdong, Guangxi). For tea objects, the harvesting and market launch windows for spring and autumn teas differ for different origins of Pu'er tea (Yiwu, Banzhang, Jingmai). Therefore, origin information can be introduced as the basis for clustering. That is, when the object type is a second category object, the leaf category, variety information, and origin information are determined as aggregation parameters. Here, the second category object refers to objects whose activity can be affected by variety information and origin information.

[0098] Origin information can be obtained by analyzing and processing the object context. Optionally, when the object type is a second type of object, the method further includes: determining the object context of the second type of object, which includes title information, image and attribute information; and processing the object context using a pre-trained origin extraction model to determine the origin information of the object.

[0099] In practical applications, origin information can be obtained through origin extraction models. These models can employ techniques such as named entity recognition, multimodal feature fusion, or large language model generation to automatically identify and output origin information from the title, image, or attribute information of an object. For example, inputting the product title "Hainan Feizixiao Lychee" into a sequence labeling model based on Bidirectional Encoder Representations from Transformers (BERT) can extract "Hainan" as the origin. For objects containing only images, multimodal models or Optical Character Recognition (OCR) technology can be used to identify the origin text on the packaging. For products in e-commerce scenarios, image information can be the main product image or the product details page. In content recommendation or travel planning scenarios, image information can be video covers, content illustrations, or scenic spot promotional images. The object context can also include auxiliary information such as geographic tags, publication location, and timestamps.

[0100] By introducing finer-grained aggregation parameters (origin information), objects of the same variety but from different regions can be distinguished, making the clustering results more realistically reflect the temporal fluctuation characteristics of the objects and avoiding seasonal cycle shifts caused by origin differences from being mixed into the same cluster, thereby improving the accuracy of seasonal attribute identification.

[0101] Step S203: Aggregate the corresponding multiple objects based on the aggregation parameters to obtain multiple object clusters.

[0102] After determining the aggregation parameters, multiple objects can be grouped based on the values ​​of these parameters, grouping objects with the same aggregation parameter value into the same object cluster. For example, at the database level, a grouping operation (GROUP BY, i.e., the grouping clause in Structured Query Language) can be performed to group all products with the leaf category "lychee," the standard variety "Feizixiao," and the place of origin "Hainan" into one cluster; and those with the leaf category "lychee," the standard variety "Feizixiao," and the place of origin "Guangdong" into another cluster. If the aggregation parameters have many dimensions or a wide range of values, the aggregation parameter values ​​can be standardized and encoded first (e.g., concatenating the leaf category, variety, and place of origin into a string or hash value), and then hash grouping can be performed based on this encoding.

[0103] Alternatively, clustering algorithms (such as K-Means Clustering or Density-Based Spatial Clustering of Applications with Noise (DBSCAN)) can be used to group objects with similar aggregation parameter characteristics into different clusters. In this case, the aggregation parameters need to be converted into numerical feature vectors (e.g., one-hot encoding of categorical parameters). This application does not limit the specific implementation method of aggregation (such as database grouping, distributed computing frameworks, or clustering algorithms); any method that can divide objects into different clusters based on aggregation parameters is applicable.

[0104] Step S204: Obtain the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. The object cluster includes multiple objects to be analyzed. Multiple objects to be analyzed in the same object cluster correspond to at least the same leaf category. The historical activity sequence includes activity information at multiple historical time points.

[0105] Step S205: When the object cluster is a seasonal object cluster with seasonal attributes, determine the predicted activity sequence corresponding to the object cluster based on the historical activity sequence. The predicted activity sequence includes the predicted activity at multiple future time points.

[0106] Step S206: Based on the predicted activity sequence, determine the timing information corresponding to the object cluster. The timing information includes multiple timing stages.

[0107] The specific implementation of steps S204 to S206 is the same as that of steps S101 to S103 in the aforementioned embodiments, and will not be repeated here. For details, please refer to the relevant descriptions in the aforementioned embodiments.

[0108] It should be noted that a complete seasonal cycle needs to include the process of activity increasing and decreasing. Therefore, it is necessary to determine the seasonal information corresponding to each seasonal stage to ensure the integrity of the seasonal cycle. If the seasonal information corresponding to a certain seasonal stage cannot be identified under the preset conditions, the preset conditions for that stage can be adjusted individually until the stage can be identified or it is confirmed that the stage does not exist.

[0109] Optionally, after determining the predicted activity sequence corresponding to the object cluster, the method in this embodiment further includes: When the timing information corresponding to the object cluster cannot be determined based on the predicted activity sequence, an adjustable range is determined for adjusting the preset thresholds corresponding to the multiple timing stages in the timing information. Within the adjustable range, the preset thresholds corresponding to the multiple timing stages are reduced at least once to obtain multiple adjusted thresholds for determining the multiple timing stages, so as to determine the timing information corresponding to the object cluster based on the multiple adjusted thresholds.

[0110] In this application, "preset threshold" can be a broad concept, and its specific form depends on the method used. For example, in a key change point-based approach, the preset threshold can be a multiple threshold for determining "first exceeding the daily baseline," a duration threshold for determining "continuously exceeding a preset number of time points," or the change rate sign detection itself (which can be considered a zero threshold). In a preset threshold-based approach, the preset threshold is directly reflected in specific values ​​such as the set activity multiple, the positive or negative week-on-week growth rate, and the number of consecutive weeks. In a template matching-based approach, the preset threshold can be a similarity threshold between the template and the sequence or a path constraint threshold for dynamic time warping. In a sequence labeling model-based approach, the preset threshold can be a confidence threshold for the stage labels output by the model.

[0111] Therefore, the preset threshold in this application is not limited to numerical multiples, but can also include various forms such as duration, direction of change rate, similarity threshold, and confidence threshold. Any adjustable parameter used to control the tightness of seasonal phase division can be regarded as a preset threshold. The embodiments of this application do not limit the specific form of the preset threshold in the implementation method.

[0112] The threshold adjustment range can be flexibly set based on the historical data distribution of the object cluster or application experience. One approach is to use the lower quartile and upper quartile of a certain indicator (such as the ratio of peak to mean) in the historical activity sequence as the lower and upper limits of the adjustment, limiting the threshold adjustment range to between the two. This avoids misjudging non-seasonal fluctuations due to the threshold being too low below the lower quartile, or missing the detection of real seasonal signals due to the threshold being too high above the upper quartile. Alternatively, the adjustment boundary can be determined based on the confidence interval of historical data (such as the 90% confidence interval), or rules can be set according to tolerance, such as a single adjustment not exceeding 20% ​​of the preset threshold and a cumulative adjustment not exceeding 50% of the preset threshold.

[0113] In practical applications, the threshold can be adjusted multiple times within the adjustable range until the corresponding seasonal information can be determined. If the corresponding seasonal information cannot be determined at the boundary of the adjustable range, the seasonal attribute of the object cluster is determined to be unrecognizable, and the seasonal information is marked as "unrecognizable" or "no valid seasonal period" in the output result. At the same time, the reason for the failure can be output (such as insufficient fluctuation amplitude, periodic instability, etc.).

[0114] This threshold adjustment method can more flexibly adapt to the seasonal fluctuation characteristics of different object clusters, avoid missed detections or misjudgments caused by fixed thresholds, and thus improve the accuracy and robustness of seasonal phase division.

[0115] After the analysis based on the predicted activity sequence is completed, the predicted timing information corresponding to the object cluster can be output.

[0116] However, in practical applications, the predicted activity sequence may contain minor pulses caused by random events (such as short-term promotions or inventory fluctuations in e-commerce scenarios), as well as high-frequency fluctuations around the period or false peaks and troughs caused by data gap interpolation. This can lead to an excessive number of periods and the risk of over-segmentation. Therefore, optionally, after determining the seasonal information corresponding to the object cluster, the method further includes: The number of seasonal periods included in the statistical seasonal information is counted. If the number of seasonal periods is greater than or equal to a preset threshold, the preset thresholds corresponding to the multiple seasonal stages in a single seasonal period are increased to obtain multiple adjusted thresholds for determining multiple seasonal stages, so as to reduce the number of seasonal periods included in the seasonal information based on the multiple adjusted thresholds.

[0117] One possible implementation is to first smooth the predicted activity sequence using an exponential moving average (EMA), by giving higher weight to recent data (a relatively new segment of data on the timeline that is closer to the current prediction point) to filter out short-term random fluctuations and tiny pulses; then, on the smoothed sequence, a tightened threshold (for example, increasing the outbreak threshold from 1.5 times the annual average to 1.8 times, or increasing the minimum duration of peak detection from 1 week to 2 weeks) is used to re-detect the cycle and determine the seasonal phase.

[0118] By increasing the preset thresholds corresponding to each stage in a single cycle, adjacent peaks with an interval less than the preset threshold (such as two weeks) can be merged into a continuous stage, or weak cycles with amplitudes lower than the lower limit of the seasonal score (such as less than 1.2) can be eliminated, thereby achieving cycle merging, avoiding excessive segmentation, and ensuring that the output seasonal information is accurate and effective.

[0119] In this embodiment, multiple objects are acquired, and aggregation parameters for clustering are determined based on the object type corresponding to each object. Then, differentiated aggregation parameters are used to cluster the objects, resulting in multiple object clusters. This effectively ensures the accuracy of the clustering results, making multiple objects within the same object cluster highly consistent in terms of seasonal attributes. It avoids inconsistencies in seasonal attributes in the clustering results due to a single aggregation granularity failing to adapt to all objects, providing a more reliable input basis for subsequent seasonal analysis and thus improving the overall accuracy of the seasonal information determination method.

[0120] Figure 4 A flowchart illustrating another method for determining seasonal information provided as an exemplary embodiment of this application; based on any embodiment, refer to the appendix. Figure 4 As shown, when the seasonal information includes multiple seasonal phases and the time information corresponding to each of the multiple seasonal phases, the activity change information of the object cluster in the future time period can be determined first based on the predicted activity sequence, and then the seasonal information can be determined based on the activity change information. At this time, the seasonal information corresponding to the object cluster can be determined based on the predicted activity sequence, which may include: Step S401: Based on the predicted activity sequence, determine the activity change information of the object cluster in the future time period.

[0121] Activity change information may include at least one of the following: peak occurrence time, peak amplitude, trough occurrence time, trough amplitude, rising slope, falling slope, rising duration, falling duration, cycle length, cycle intensity, holiday pulse occurrence time, pulse amplitude, pulse duration, cycle length variation coefficient, and peak amplitude offset.

[0122] In one optional implementation, activity change information can be extracted using differencing and sliding windows: First, the difference or year-on-year change rate between adjacent time points in the predicted activity sequence is calculated to obtain a first-order differencing sequence to indicate the direction of rise and fall; then, local extrema are detected using a sliding window (window length, for example, 3 or 5 time points), recording the time and amplitude of peaks / troughs; and the average upward slope is calculated from the trough to the peak, and the average downward slope is calculated from the peak to the next trough, and the duration of rise / fall is statistically analyzed. For multi-year sequences, the peak time standard deviation and peak amplitude coefficient of variation are calculated by aligning the windows to the same period as indicators of cycle stability.

[0123] Another alternative implementation can extract activity change information based on a time-series decomposition model: the predicted activity sequence is input into a time-series decomposition model (e.g., STL or Prophet) to decompose it into trend terms, seasonal terms, and holiday terms. Local extreme points are located from the seasonal terms using second-order differencing or inflection point detection to obtain the time and amplitude of peaks / troughs, and the average slope and duration of rising / falling segments are calculated. Pulse information is extracted from the holiday terms (when the amplitude exceeds a preset baseline, the start and end times, peak amplitude, and width are recorded). For multi-period scenarios, the waveforms of the seasonal terms in each period are aligned, and the peak time offset variance and peak amplitude fluctuation range are calculated as stability indicators.

[0124] Optionally, activity change information can be determined based on the weekly activity value and the mean activity value. In this case, determining the activity change information of the object cluster in the future time period based on the predicted activity sequence can include: determining the activity values ​​of multiple weekly granularities included in the predicted activity sequence; determining the mean activity value corresponding to the object cluster based on the predicted activity sequence; and determining the activity change information based on the activity value and the mean activity value.

[0125] After obtaining the predicted activity sequence, activity values ​​at the daily or finer granularity can be aggregated by natural week (e.g., by taking the weekly average or sum) to obtain a weekly discrete activity sequence. Then, the arithmetic mean, weighted average, or median of this sequence can be calculated as the activity mean, which serves as the daily baseline for the object cluster. Based on the activity values ​​and the mean, change information can be determined in the following ways: calculating the ratio of activity to the mean at each time point (obtaining a relative activity sequence) or the difference (obtaining an offset sequence); calculating the rate of change at consecutive time points (e.g., week-on-week growth rate); or calculating deviation statistics (e.g., variance, standard deviation) through a sliding window. This application does not limit the specific calculation method; any method that can derive the change pattern based on a comparison of activity values ​​and the mean is applicable.

[0126] Step S402: Based on the activity change information, determine the multiple seasonal phases corresponding to the object cluster and the time information corresponding to each of the multiple seasonal phases.

[0127] One optional implementation involves first identifying the points of change in the direction of activity rise and fall using a first-order difference sequence, and then using local extremum detection to locate the trough (starting point of the listing period), peak (center of the outbreak), inflection point (starting point of the decline), and end point (end point of the delisting period). Taking an e-commerce scenario as an example: when activity rises continuously and exceeds the daily baseline (e.g., the average of the past three months) for the first time, it is marked as the starting point of the listing period; when the rate of increase changes from positive to negative, it is marked as the peak point; when the rate of decrease changes from negative to positive, it is marked as the inflection point of the decline period; when activity falls back below the baseline and persists for a preset number of time points (e.g., three consecutive weeks), it is marked as the end point of the delisting period. In chronological order, the period from the starting point of the listing period to the peak point is the listing period; the period extending several time steps before and after the peak point is the outbreak period; the period from the peak point to the inflection point of the decline period is the decline period; and the period from the inflection point of the decline period to the end point of the delisting period is the delisting period. The start and end dates of each stage are directly read from the time coordinates of the corresponding key points.

[0128] Another optional implementation method can pre-set trigger threshold conditions for each stage (referred to as preset thresholds). For example: the listing period condition is that the weekly average activity value exceeds 1.2 times the annual average for two consecutive weeks; the explosion period condition is that the weekly average activity value exceeds 1.5 times the annual average and the week-on-week growth rate is positive; the decline period condition is that the week-on-week growth rate turns from positive to negative and declines for two consecutive weeks; the delisting period condition is that the weekly average activity value falls below the annual average and there is no rebound for three consecutive weeks. Scan the prediction sequence and record the moment when the listing period condition is first met as the listing period start point, the moment when the explosion period condition is first met as the explosion period start point, the moment when the decline period condition is first met as the decline period start point, and the moment when the delisting period condition is first met as the delisting period start point. The interval between the start points of two adjacent stages corresponds to the time window of the previous stage (for example, from the listing period start point to the explosion period start point is the listing period, from the explosion period start point to the decline period start point is the explosion period, from the decline period start point to the delisting period start point is the decline period, and after the delisting period start point is the delisting period).

[0129] Step S403: Determine the season information based on multiple season stages and their corresponding time information.

[0130] After determining multiple seasonal phases and their corresponding time information, this information can be integrated to form complete seasonal information for the object cluster, such as the start and end dates, peak dates, and seasonal scores for each phase.

[0131] This application's embodiments determine seasonal information based on activity change information, directly utilizing the temporal fluctuation characteristics of the object cluster itself (such as peak, slope, duration, etc.) to automatically divide seasonal stages (such as listing period, outbreak period, decline period, and delisting period, etc.), avoiding reliance on fixed time windows or human experience. This ensures that the stage boundaries are synchronized with the actual activity evolution of the object cluster, eliminating division errors caused by fixed window misalignment or subjective bias, and improving the adaptability and accuracy of the seasonal information determination method.

[0132] Figure 5 A flowchart illustrating another method for determining seasonal information provided as an exemplary embodiment of this application; based on any embodiment, refer to the appendix. Figure 5 As shown, when determining the specific seasonal information of a seasonal object cluster based on the predicted activity sequence, a preset threshold for dividing the seasonal phase can be adaptively determined according to the activity change information corresponding to the seasonal object cluster, thereby improving the adaptability to different seasonal products. In this case, the method in this embodiment may further include: Step S501: Based on the predicted activity sequence, determine the activity change information of the object cluster in the future time period.

[0133] Step S502: Determine the preset thresholds corresponding to the multiple seasonal stages.

[0134] When determining the time period, a corresponding preset threshold can be set for each period as the basis for judging the current activity level. The preset threshold can be determined based on the historical data distribution of the object cluster or application experience.

[0135] Step S503: Adjust the preset threshold based on the activity change information to obtain adaptive thresholds corresponding to multiple seasonal stages.

[0136] In practical applications, different seasonal clusters of products exhibit varying seasonal cycle patterns. For example, the lychee cluster has a short and concentrated market window (only about 4 weeks), the electric fan cluster has a wide peak and a long duration (about 12 weeks), and the mooncake cluster's dates fluctuate annually due to the influence of lunar festivals (early September to early October in the Gregorian calendar). Using a uniform high threshold might miss some clusters (such as the gradual peak of electric fans), while using a low threshold might misjudge noise from other clusters as a surge (such as the random fluctuations of long-tail products). Therefore, using a uniform fixed threshold is difficult to adapt to the seasonal variation patterns of all product clusters.

[0137] Based on this, the embodiments of this application adopt an adaptive threshold strategy for different object clusters to improve the accuracy of seasonal identification.

[0138] Optionally, a seasonal score for adjusting a preset threshold can be determined based on the activity change information of the seasonal object cluster, and the preset threshold can be adjusted based on the seasonal score.

[0139] Adjusting a preset threshold based on activity change information to obtain adaptive thresholds for multiple seasonal stages can include: determining the activity peak corresponding to the object cluster based on activity change information; determining the seasonal score for adjusting the preset threshold based on the activity peak and the average activity; and adjusting the preset threshold based on the seasonal score to obtain adaptive thresholds for multiple seasonal stages.

[0140] The seasonal score is a comprehensive indicator used to quantify the intensity of seasonal fluctuations in an object cluster. It can be calculated based on the ratio or difference between the peak activity level and the mean activity level. The greater the difference between the peak activity level and the mean activity level, and the steeper the peak, the higher the seasonal score, indicating that the seasonal outbreak intensity of the object cluster is greater and the listing window is more concentrated. Conversely, if the seasonal score is low, it indicates that its seasonal attributes are not significant.

[0141] Optionally, adjusting the preset threshold based on the seasonal score to obtain adaptive thresholds corresponding to multiple seasonal stages may include: increasing the preset thresholds corresponding to multiple seasonal stages when the seasonal score is greater than the preset threshold, to obtain adaptive thresholds corresponding to multiple seasonal stages; and decreasing the preset thresholds corresponding to multiple seasonal stages when the seasonal score is less than the preset threshold, to obtain adaptive thresholds corresponding to multiple seasonal stages.

[0142] When adaptively adjusting a preset threshold based on activity level changes, the adjustment range can be determined by the difference or ratio between the current object cluster's seasonal score and the preset threshold. For example, when the ratio of the seasonal score to the preset threshold is greater than 1.2, the outbreak threshold is increased by 10%; when the ratio is less than 0.8, the outbreak threshold is decreased by 15%; and when the ratio is between 0.8 and 1.2, the preset threshold remains unchanged. Specific threshold adjustment rules can be determined through a preset mapping table lookup operation. This mapping table records the correspondence between seasonal score ranges and threshold adjustment coefficients. Dynamically adjusting the threshold based on the difference between the seasonal score and the preset threshold can avoid misjudgments or missed detections due to improper threshold settings.

[0143] Step S504: Based on the activity change information and the adaptive threshold, determine multiple seasonal phases corresponding to the object cluster and the time information corresponding to each of the multiple seasonal phases.

[0144] There are multiple ways to determine seasonal phases and time information based on activity change information and adaptive thresholds. For example, one approach is to detect the starting point of the listing period when activity first exceeds the adaptive threshold, the peak point where the rate of increase turns from positive to negative, the inflection point of the decline period where the rate of decrease turns from negative to positive, and the end point where activity falls below the threshold. These phases can then be divided into listing, peak, decline, and delisting periods in chronological order, with the start and end dates of each phase directly taken from the corresponding key points. Alternatively, adaptive threshold conditions can be set for each seasonal phase (e.g., the listing period threshold is 1.2 times the baseline for two consecutive weeks, the peak period threshold is 1.5 times the baseline with a positive month-on-month change, the decline period threshold is a negative month-on-month change with two consecutive weeks of decline, and the delisting period threshold is below the baseline for three consecutive weeks). The sequence is scanned, and the position where the condition is first met is taken as the starting point of that phase. The time window between adjacent starting points is the previous phase. This application does not limit the specific method used.

[0145] Step S505: Determine the season information based on multiple season phases and their corresponding time information.

[0146] In this embodiment, preset thresholds corresponding to multiple seasonal stages are adjusted based on activity change information. For seasonal objects with large fluctuations, the adaptive threshold automatically raises the trigger threshold for outbreaks, avoiding misjudging random noise as outbreaks due to excessively low thresholds. For seasonal objects with gentle fluctuations, the adaptive threshold appropriately lowers the threshold, preventing the detection of genuine seasonal signals due to excessively high thresholds. Simultaneously, when the activity of an object cluster shifts in different years (e.g., due to abnormal weather causing peaks to occur earlier or later), the adaptive threshold can automatically calibrate accordingly, reducing the workload of repeated manual parameter tuning. Through this mechanism, the problem of "high threshold missed detection, low threshold misjudgment" that easily occurs under fixed thresholds can be avoided, making the boundary division of seasonal stages more closely match the actual seasonal evolution trajectory of the object cluster, thereby improving the stability and accuracy of seasonal stage division. Furthermore, the adaptive threshold can also accommodate scenarios with multiple coexisting periods, further broadening the applicability of the method.

[0147] Figure 6 A schematic diagram illustrating the application of a method for determining seasonal information provided in this application is attached. Figure 6 As shown, taking an e-commerce scenario as an example, the seasonal information for "Sydney Pear" is determined. The following section combines this with the attached... Figure 6 The complete process for determining seasonal information is described below: First, obtain the product context information. The product context can include the product title, main image, or details page, as well as the attribute information filled in by the merchant. Multimodal models can be used to identify the product context in order to extract product category information. Multimodal recognition models can include CLIP (contrastive language-image pre-trained model), ALIGN (cross-modal model trained on large-scale noisy image-text pairs), ViLT (lightweight vision-language Transformer), Flamingo (visual language model), BLIP-2 (guided language image pre-trained model), GPT-4V (multimodal large language model supporting image understanding), Qwen-VL (common sense visual language model), and Alibaba MOON (e-commerce multimodal product understanding model), etc. Recognition based on multimodal information can obtain richer product information, providing an accurate data foundation for subsequent identification of product category, variety, or origin information.

[0148] Next, the product data is aggregated. First, the product type is determined based on the identified product category information. In this embodiment, the leaf category corresponding to "Snow Pear" is "Pear". The type to which "Snow Pear" belongs can be determined based on a pre-built classification rule base or classification decision model. After judgment, it is known that the product belongs to the second type of product. Its activity is affected by the leaf category, variety information, and place of origin information. Therefore, the variety information and place of origin information of the product are identified. When identifying the variety information of the product, it can first be searched in the pre-built variety dictionary based on the product text. If there is a clear matching variety description in the product text, it is directly mapped to the standard variety. Otherwise, a multimodal model is used to extract candidate varieties from the title, image, and attribute information (mainly the category). After obtaining the candidate varieties, the candidate varieties, product context, and variety dictionary are input into the big language model. The big language model combines the retrieved knowledge information to output a standardized result. The standard variety name corresponding to "Snow Pear" in the figure should be "Snowflake Pear". Synonyms or aliases discovered during the normalization process can be automatically added to the thesaurus; for newly emerging standard words that have not yet been included, corresponding prompts are output and the words are placed in a queue awaiting manual review to continuously improve the standard product system. The identification of product origin information can be obtained by processing the object's context using a pre-trained origin extraction model.

[0149] In terms of model implementation, a smaller-sized LLM (such as Qwen3-32B) can be selected as the base model, and post-training can be performed specifically for the origin extraction task. During training, a robust task baseline can be established first through supervised fine-tuning (SFT) to enable the model to master basic origin recognition capabilities. Subsequently, group relative policy optimization (GRPO) can be used to further focus on difficult samples, such as complex samples with multiple origin co-occurrences, origin abbreviations, ambiguous expressions, or unclear visual information, to improve the model's discrimination ability in complex scenarios. At the same time, the concept of curriculum learning is introduced during training. Simple samples are used first to help the model quickly establish a stable pattern, and then more difficult samples are gradually added to enhance the model's recognition performance in complex boundary situations.

[0150] After determining the leaf category, variety information, and origin information of a product, these three fields can be used as aggregation parameters to obtain the corresponding product clusters. For example, all products with the leaf category "pear," the variety information "snow pear," and the origin information "Hebei" can be clustered into one product cluster. The clustering algorithm can be K-means clustering, density-based noisy spatial clustering, or hierarchical clustering, etc., and this application does not limit the specific clustering. Based on a differentiated aggregation granularity strategy, the appropriate aggregation granularity, such as leaf category, leaf category-variety, or leaf category-variety-origin, can be adaptively selected for different types of objects, so that objects within the same object cluster maintain a high degree of consistency in seasonal attributes, thereby improving the clustering quality and the accuracy of seasonal information identification.

[0151] Next, seasonality classification is performed, specifically determining whether a product cluster possesses seasonal attributes. This can be done by obtaining the historical activity sequence corresponding to the product cluster and performing joint classification based on multiple periodicity detection algorithms and a large language model to determine whether the product cluster has seasonal attributes. Periodicity detection algorithms can include various time-frequency domain periodicity detection algorithms such as ACF, AUTOPERIOD, and EMA+RobustPeriod. By introducing multiple algorithms for parallel judgment, the bias introduced by a single method can be reduced.

[0152] If a product cluster does not have a seasonal attribute, the large language model can output the product as an "non-seasonal" identifier and output the reason for the judgment. If the product cluster is determined to be a seasonal product cluster with a seasonal attribute, the subsequent seasonal cycle mining and seasonal layering judgment steps can be performed on the product cluster.

[0153] Seasonal cycle mining is the process of obtaining seasonal information for product clusters. Historical activity sequences and custom holiday information can be input into a time series prediction model, which then outputs a predicted activity sequence. Time series prediction models can be Prophet, STL, N-BEATS, ARIMA, LSTM, etc. These models can decompose historical activity sequences to obtain seasonal time series and holiday distribution information, and then infer the predicted activity sequence based on this information.

[0154] Next, the predicted activity sequence can be analyzed to determine the seasonal information corresponding to the product clusters. This seasonal information includes multiple seasonal phases and the time information corresponding to each phase. For example, taking "Snowflake Pear" as an example, its seasonal information for the next year can include: the number of seasonal cycles (e.g., one main cycle), and within that cycle: the listing period (August 1st to August 31st), the peak period (September 1st to October 5th), the decline period (October 6th to October 20th), and the off-season period (October 21st to October 31st). The start and end dates of each phase, as well as the peak occurrence time of the peak period (e.g., September 29th), are recorded. Each seasonal phase is determined based on an adaptive dynamic threshold. Dividing seasonal phases based on an adaptive threshold strategy can solve problems such as the inability of fixed threshold schemes to adapt to fluctuation differences between different product clusters and the interference of lunar calendar drift in cycle positioning.

[0155] Next, the seasonal stratification step is executed. Product clusters can be categorized into strongly seasonal products, seasonal products, and festival products based on their seasonal characteristics. Seasonal stratification can be achieved using a few-sample discrimination method based on a large language model. This stratification is then combined with the product variety and category information obtained in the aforementioned seasonal variety stratification step, as well as the seasonal information obtained in the seasonal cycle mining step, for comprehensive judgment. Specifically, the festival dates for festival products correspond to the months of the seasonal cycle in the seasonal information. This seasonal stratification method, which uses a two-way verification between large language model semantic discrimination and data mining results, can solve the problems of illusion bias and inconsistencies between stratification results and actual cycle data that can easily arise from using only a large language model for stratification. Through seasonal type stratification guidance, resource allocation at each stage can be better matched with actual activity changes, improving overall operational efficiency.

[0156] Then, the large language model can output the seasonal information corresponding to the seasonal products. The output results can include seasonal level identifiers (such as strong seasonal products, seasonal products or festival products) and product prediction cycle information (such as the start and end dates of the listing period, peak period, decline period and off-season period). If it is a festival product, the corresponding festival name can also be output.

[0157] Appendix Figure 6The output could be: the seasonality of this product (pear / snowflake pear) is seasonal, with different varieties from different producing areas having concentrated and slightly staggered market periods. Specific seasonal information is as follows: Hebei Crown Pear's market period is from early to late July, with a peak period from early to late August, and a peak in mid-August; Northeast Golden Jade Pear's market period is from mid-August to early September, with a peak period from mid-September to early October, and a peak in late September; Snowflake Pear (taking Hebei as an example)'s market period is late August, with a peak period from mid-September to early October. The decline and end of the market for each variety are concentrated in mid-to-late October.

[0158] The technical solution provided in this application embodiment can achieve the following effects: (1) Overcoming the problem of "single product analysis being limited by data sparsity and high noise": By performing aggregation analysis on the basis of object clusters, the seasonal attributes of each product within the cluster are used to characterize the seasonal attributes of the cluster, thus achieving dimensionality reduction and powerful noise reduction of the underlying data. The statistical regularity of the group smooths out the individual fluctuations of a single product caused by frequent listing and delisting, stockouts, or random purchases. Even for long-tail objects with extremely sparse historical data, a robust time-series baseline can be obtained, thereby solving the problems of missed detection of long-tail seasonal objects and unstable analysis results. At the same time, based on the differentiated clustering strategy, the clustered product clusters have consistency in seasonal attributes, providing an accurate data foundation for seasonal analysis based on the cluster level.

[0159] (2) Overcoming the problem of “lack of noise stripping ability and difficulty in distinguishing between marketing activities and seasonal fluctuations”: By using a cross-voting mechanism based on multiple cycle detection algorithms and large language models, the objective transaction cycle pattern is aligned with the subjective seasonal semantic cognition. Only when multiple discrimination results show that the product to be analyzed is confirmed as a seasonal product, effectively filtering out the interference of false sales peaks caused by “top influencers selling goods” or “platform shopping festival promotions”.

[0160] (3) Overcoming the problems of "post-event hype, insufficient foresight, and inability to adapt to complex calendars and product category differences": By using a time-series prediction model, the historical activity sequence is accurately decoupled into trend items, seasonal items, and holiday items. Combined with customized holiday information, a predicted activity sequence can be generated to predict the listing window and trend before the sales of goods explode, overcoming the lag, while adapting to the lunar calendar drift and the differentiated seasonal patterns of different product categories. By using adaptive thresholds, the model adapts to the drastically different amplitude characteristics of different product categories (mooncakes with explosive growth vs. home appliances with gradual fluctuations), breaking the positioning bias caused by a single fixed threshold. It not only provides e-commerce platforms with an accurate "seasonal marketing calendar", but also provides downstream supply chain with an "advanced prediction window" for advance preparation, warehousing and logistics resource allocation, and front-end marketing preheating, realizing a leap from lagging hype to pre-event prediction.

[0161] (4) Regarding the problem of "lack of refined seasonal stratification and inability to distinguish festival / seasonal relationships": By analyzing the seasonal characteristics of goods, goods can be divided into strong seasonal goods, seasonal goods, or festival goods, and the corresponding festival names (if any) can be output. This allows downstream applications to formulate differentiated strategies based on different seasonal levels, avoiding the risk of expiration and loss of strong seasonal goods and stockouts of festival goods. At the same time, by requiring that the festival or solar term names output by the large language model must fall within the actual peak sales month, the illusion of a mismatch between festival and sales time that may be generated by the large model is effectively improved, ensuring the logical consistency of the stratification results.

[0162] It should be noted that the above process is merely an exemplary implementation of this application and is not intended to limit the scope of protection of this application. Those skilled in the art can make appropriate adjustments or equivalent substitutions to the order of steps, parameter settings, algorithm selection, etc., without departing from the principles of this application, and such adjusted solutions should still fall within the scope of protection of this application.

[0163] Figure 7 A schematic diagram of a device for determining seasonal information provided in an exemplary embodiment of this application; see attached diagram. Figure 7 As shown, this embodiment provides a device for determining seasonal information, which is used to perform the above-described... Figure 1 The data processing method shown may include a device for determining the seasonal information, which may include: The first acquisition module 11 is used to acquire the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. The object cluster includes multiple objects to be analyzed. The multiple objects to be analyzed in the same object cluster correspond to at least the same leaf category. The historical activity sequence includes activity information at multiple historical time points.

[0164] The first determining module 12 is used to determine the predicted activity sequence corresponding to the object cluster based on the historical activity sequence when the object cluster is a seasonal object cluster with seasonal attributes. The predicted activity sequence includes the predicted activity of multiple future time points.

[0165] The second determining module 13 is used to determine the timing information corresponding to the object cluster based on the predicted activity sequence. The timing information includes multiple timing stages.

[0166] In some instances, before acquiring the cluster of objects to be analyzed, the first acquisition module 11 and the first determination module 12 in this embodiment are used to perform the following steps: The first acquisition module 11 is used to acquire multiple objects; The first determining module 12 is used to determine the aggregation parameters for performing aggregation operations on multiple objects based on the object types corresponding to the multiple objects respectively; and to aggregate the corresponding multiple objects based on the aggregation parameters to obtain multiple object clusters.

[0167] In some instances, when the first determining module 12 determines the aggregation parameters for performing aggregation operations on multiple objects based on the object types corresponding to the multiple objects respectively, the first determining module 12 is used to: determine the leaf category of the object as the aggregation parameter when the object type is a preset standard object; determine the leaf category and variety information of the object as the aggregation parameters when the object type is a first type of object, wherein the first type of object is an object whose activity can be affected by the variety information; and determine the leaf category, variety information and origin information of the object as the aggregation parameters when the object type is a second type of object, wherein the second type of object is an object whose activity can be affected by the variety information and origin information.

[0168] In some instances, when the object type is a first type of object, the first acquisition module 11 and the first determination module 12 in this embodiment are used to perform the following steps: The first acquisition module 11 is used to acquire the object context corresponding to multiple objects respectively. The object context includes at least one of the following: title information, image, and attribute information. The first determining module 12 is used to determine candidate varieties corresponding to multiple objects based on the object context; and to determine variety information corresponding to multiple objects based on the candidate varieties and the object context.

[0169] In some instances, when the first determining module 12 determines the variety information corresponding to multiple objects based on the candidate varieties and the object context, the first determining module 12 is used to perform: determining a pre-constructed variety dictionary; and using a large language model to process the candidate varieties, the object context, and the variety dictionary to obtain the variety information corresponding to multiple objects.

[0170] In some instances, after processing the candidate varieties, object context, and variety dictionary using a large language model, the first determining module 12 in this embodiment is used to perform the following: when the output of the large language model includes a first prompt, add the new synonym to the variety dictionary based on the first prompt, wherein the first prompt is used to identify the existence of a new synonym in the output; when the output of the large language model includes a second prompt, determine the new standard word as knowledge to be added to the database based on the second prompt, and add the new standard word to the variety dictionary if the knowledge to be added to the database passes the review, wherein the second prompt is used to identify the existence of a new standard word in the output.

[0171] In some instances, after obtaining the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster, the first determining module 12 in this embodiment is used to perform: processing the historical activity sequence using a variety of different period detection algorithms to obtain multiple period detection results corresponding to the object cluster; processing the variety information and leaf category corresponding to the object cluster using a large language model to obtain the identification result of whether the object cluster has seasonal attributes; and determining the object cluster as a seasonal object cluster if at least two of the multiple period detection results and identification results indicate that the object cluster has seasonal attributes.

[0172] In some instances, after obtaining multiple periodic detection results corresponding to an object cluster, the first determining module 12 in this embodiment is used to perform: determining the time periods with seasonal attributes included in the multiple periodic detection results; and determining that the corresponding periodic detection result identifies the object cluster with seasonal attributes if the duration of the time period is within a preset duration range.

[0173] In some instances, when the first determining module 12 determines the predicted activity sequence corresponding to the object cluster based on the historical activity sequence, the first determining module 12 is used to perform the following: inputting the historical activity sequence and holiday information into the time series prediction model for processing to obtain the seasonal time series and holiday distribution information corresponding to the object cluster, wherein the seasonal time series is a Fourier series, which is used to represent the periodic change pattern of activity in the future time period, and the holiday distribution information follows a Gaussian distribution, which is used to represent the local fluctuation of activity around the holidays in the future time period; and determining the predicted activity sequence based on the seasonal time series and holiday distribution information.

[0174] In some instances, the timing information further includes: time information corresponding to the plurality of timing stages respectively; when the second determining module 13 determines the timing information corresponding to the object cluster based on the predicted activity sequence, the second determining module 13 is used to perform: determining the activity change information of the object cluster in a future time period based on the predicted activity sequence; determining the plurality of timing stages corresponding to the object cluster and the time information corresponding to the plurality of timing stages respectively based on the activity change information; and determining the timing information based on the plurality of timing stages and the corresponding time information.

[0175] In some instances, when the second determining module 13 determines the activity change information of an object cluster within a future time period based on the predicted activity sequence, the second determining module 13 is used to perform: determining the activity values ​​of multiple weekly granularities included in the predicted activity sequence; determining the average activity value corresponding to the object cluster based on the predicted activity sequence; and determining the activity change information based on the activity value and the average activity value.

[0176] In some instances, when the second determining module 13 determines multiple seasonal periods corresponding to an object cluster and the time information corresponding to each of the multiple seasonal periods based on activity change information, the second determining module 13 is used to perform: determining preset thresholds corresponding to each of the multiple seasonal periods; adjusting the preset thresholds based on activity change information to obtain adaptive thresholds corresponding to each of the multiple seasonal periods; and determining multiple seasonal periods corresponding to an object cluster and the time information corresponding to each of the multiple seasonal periods based on activity change information and adaptive thresholds.

[0177] In some instances, when the second determining module 13 adjusts the preset threshold based on the activity change information to obtain adaptive thresholds corresponding to multiple seasonal stages, the second determining module 13 is used to perform: determining the activity peak value corresponding to the object cluster based on the activity change information; determining the seasonal score for adjusting the preset threshold based on the activity peak value and the average activity value; and adjusting the preset threshold based on the seasonal score to obtain adaptive thresholds corresponding to multiple seasonal stages.

[0178] In some instances, when the second determining module 13 adjusts the preset threshold based on the seasonal score to obtain adaptive thresholds corresponding to multiple seasonal periods, the second determining module 13 is used to perform the following: when the seasonal score is greater than the preset threshold, increase the preset thresholds corresponding to the multiple seasonal periods to obtain adaptive thresholds corresponding to the multiple seasonal periods; when the seasonal score is less than the preset threshold, decrease the preset thresholds corresponding to the multiple seasonal periods to obtain adaptive thresholds corresponding to the multiple seasonal periods.

[0179] In some instances, after determining the predicted activity sequence corresponding to the object cluster, the second determining module 13 in this embodiment is used to perform: counting the number of time periods included in the time period information; when the number of time periods is greater than or equal to a preset threshold, increasing the preset threshold used to determine the multiple time periods in a single time period, obtaining multiple adjusted thresholds used to determine multiple time periods, so as to reduce the number of time periods included in the time period information based on the multiple adjusted thresholds.

[0180] The device for determining the seasonal information in this embodiment can also perform the aforementioned embodiments. Figures 1-6 The description of the embodiments shown is for reference only, and will not be elaborated upon here.

[0181] like Figure 8As shown, this embodiment provides an electronic device, which can refer to a complete set of hardware and software infrastructure supporting software operation, data processing, and task execution. It can provide computing resources, storage, communication, and a runtime environment for applications (such as large model inference, scientific simulations, and web services). This electronic device is used to perform the above-described tasks. Figure 1 The method for determining the seasonal information shown may, in practice, include a memory 24 and a processor 25.

[0182] Memory 24 is used to store computer programs and can be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0183] The processor 25, coupled to the memory 24, is used to execute a computer program in the memory 24 for: acquiring an object cluster to be analyzed and a historical activity sequence corresponding to the object cluster, wherein the object cluster includes multiple objects to be analyzed, and the multiple objects to be analyzed in the same object cluster correspond to at least the same leaf category, and the historical activity sequence includes activity information at multiple historical time points; in the case that the object cluster is a seasonal object cluster with seasonal attributes, determining a predicted activity sequence corresponding to the object cluster based on the historical activity sequence, the predicted activity sequence including predicted activity at multiple future time points; and determining seasonal information corresponding to the object cluster based on the predicted activity sequence, the seasonal information including multiple seasonal stages.

[0184] Furthermore, such as Figure 8 As shown, the electronic device also includes other components such as a communication component 26, a display 27, a power supply component 28, and an audio component 29. Figure 8 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 8 The components shown. Additionally... Figure 8 The components within the center frame are optional, not mandatory, and their specific requirements depend on the product form of the work node. In this embodiment, the work node can be a terminal device such as a desktop computer, laptop computer, smartphone, or Internet of Things (IoT) device, or a server-side device such as a conventional server, cloud server, or server array. If the work node in this embodiment is implemented as a terminal device such as a desktop computer, laptop computer, or smartphone, it may include... Figure 8 The components within the center frame; if the working node in this embodiment is implemented as a server-side device such as a conventional server, cloud server, or server array, it may not include... Figure 8 The component within the center frame.

[0185] The aforementioned memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0186] The aforementioned communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as 2G, 3G, 4G / LTE, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0187] The aforementioned display includes a screen, which may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen can be implemented as a touchscreen to receive input signals from the user. The Touch Panel includes one or more touch sensors to sense touches, swipes, and gestures on the Touch Panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0188] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0189] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0190] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium.

[0191] Accordingly, this application also provides a computer program product, which includes a computer program or instructions. When the computer program or instructions are executed by a processor, the processor is able to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. In addition, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, so that the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device can be implemented as a means to implement the corresponding functions in the above method embodiments.

[0192] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0193] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for determining seasonal information, characterized in that, include: Obtain the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster. The object cluster includes multiple objects to be analyzed. The multiple objects to be analyzed in the same object cluster correspond to at least the same leaf category. The historical activity sequence includes activity information at multiple historical time points. In the case where the object cluster is a seasonal object cluster with seasonal attributes, a predicted activity sequence corresponding to the object cluster is determined based on the historical activity sequence. The predicted activity sequence includes the predicted activity at multiple future time points. Based on the predicted activity sequence, seasonal information corresponding to the object cluster is determined, and the seasonal information includes multiple seasonal phases.

2. The method according to claim 1, characterized in that, Before obtaining the cluster of objects to be analyzed, the method further includes: Retrieve multiple objects; Based on the object types corresponding to the multiple objects, determine the aggregation parameters used to perform aggregation operations on the multiple objects; Multiple objects are aggregated based on the aggregation parameters to obtain multiple object clusters.

3. The method according to claim 2, characterized in that, Based on the object types corresponding to the multiple objects, aggregation parameters for performing aggregation operations on the multiple objects are determined, including: When the object type is a preset standard object, the leaf category of the object is determined as the aggregation parameter; When the object type is a first type of object, the leaf category and variety information of the object are determined as the aggregation parameters, wherein the first type of object is an object whose activity can be affected by the variety information; When the object type is the second type of object, the leaf category, variety information and origin information of the object are determined as the aggregation parameters, wherein the second type of object is an object whose activity can be affected by the variety information and origin information.

4. The method according to claim 3, characterized in that, When the object type is a first-class object, the method further includes: Obtain the object context corresponding to each of the plurality of objects, wherein the object context includes at least one of the following: title information, image, and attribute information; Based on the object context, candidate varieties corresponding to the plurality of objects are determined respectively; Based on the candidate varieties and the object context, the variety information corresponding to each of the multiple objects is determined.

5. The method according to claim 4, characterized in that, Based on the candidate varieties and the object context, variety information corresponding to multiple objects is determined, including: Determine the pre-constructed variety dictionary; The candidate varieties, the object context, and the variety dictionary are processed using a large language model to obtain variety information corresponding to multiple objects.

6. The method according to claim 5, characterized in that, After processing the candidate varieties, the object context, and the variety dictionary using a large language model, the method further includes: When the output of the large language model includes a first prompt, the newly added synonym is added to the variety dictionary based on the first prompt, wherein the first prompt is used to identify the existence of a newly added synonym in the output; When the output of the large language model includes the second prompt information, the newly added standard words are identified as knowledge to be added to the database based on the second prompt information. If the knowledge to be added to the database passes the review, the newly added standard words are added to the variety dictionary. The second prompt information is used to identify the existence of newly added standard words in the output.

7. The method according to any one of claims 1-6, characterized in that, After obtaining the object cluster to be analyzed and the historical activity sequence corresponding to the object cluster, the method further includes: The historical activity sequence is processed using various periodicity detection algorithms to obtain multiple periodicity detection results corresponding to the object cluster; The variety information and leaf category corresponding to the object cluster are processed using a large language model to obtain the identification result of whether the object cluster has seasonal attributes; If at least two of the multiple periodic detection results and the identification results indicate that the object cluster has a seasonal attribute, the object cluster is determined to be the seasonal object cluster.

8. The method according to claim 7, characterized in that, After obtaining multiple periodic detection results corresponding to the object cluster, the method further includes: Determine the time periods with seasonal attributes included in the results of multiple periodic detections; If the duration of the time period is within a preset duration range, the corresponding period detection result indicates that the object cluster has a seasonal attribute.

9. The method according to any one of claims 1-6, characterized in that, Based on the historical activity sequence, determine the predicted activity sequence corresponding to the object cluster, including: The historical activity sequence and holiday information are input into a time series prediction model for processing to obtain the seasonal time series and holiday distribution information corresponding to the object cluster. The seasonal time series is a Fourier series, which is used to represent the periodic change pattern of activity in the future time period. The holiday distribution information follows a Gaussian distribution, which is used to represent the local fluctuation of activity around the holidays in the future time period. Based on the seasonal time series and holiday distribution information, the predicted activity series is determined.

10. The method according to any one of claims 1-6, characterized in that, The timing information further includes: time information corresponding to each of the plurality of timing stages; and determining the timing information corresponding to the object cluster based on the predicted activity sequence, including: Based on the predicted activity sequence, determine the activity change information of the object cluster in the future time period; Based on the activity change information, multiple seasonal phases corresponding to the object cluster and the time information corresponding to each of the multiple seasonal phases are determined. The seasonal information is determined based on the multiple seasonal stages and their corresponding time information.

11. The method according to claim 10, characterized in that, Based on the predicted activity sequence, determine the activity change information of the object cluster in future time periods, including: Determine the activity values ​​of multiple weekly granularities included in the predicted activity sequence; Based on the predicted activity sequence, determine the mean activity value corresponding to the object cluster; The activity change information is determined based on the activity value and the average activity value.

12. The method according to claim 10, characterized in that, Based on the activity change information, multiple seasonal phases corresponding to the object cluster and the time information corresponding to each of the multiple seasonal phases are determined, including: Determine the preset thresholds corresponding to multiple seasonal phases; The preset threshold is adjusted based on the activity change information to obtain adaptive thresholds corresponding to multiple seasonal stages. Based on the activity change information and the adaptive threshold, multiple seasonal phases corresponding to the object cluster and the time information corresponding to each of the multiple seasonal phases are determined.

13. The method according to claim 12, characterized in that, Based on the activity change information, the preset threshold is adjusted to obtain adaptive thresholds corresponding to multiple seasonal stages, including: Based on the activity change information, determine the activity peak value corresponding to the object cluster; Based on the peak activity level and the average activity level, a seasonal score is determined for adjusting the preset threshold. The preset threshold is adjusted based on the seasonal score to obtain adaptive thresholds corresponding to multiple seasonal stages.

14. The method according to claim 13, characterized in that, Based on the seasonal score, the preset threshold is adjusted to obtain adaptive thresholds corresponding to multiple seasonal stages, including: When the seasonal score is greater than a preset threshold, the preset thresholds corresponding to the multiple seasonal stages are increased to obtain the adaptive thresholds corresponding to the multiple seasonal stages. If the seasonal score is less than a preset threshold, the preset thresholds corresponding to the multiple seasonal stages are reduced to obtain the adaptive thresholds corresponding to the multiple seasonal stages.

15. The method according to any one of claims 1-6, characterized in that, After determining the predicted activity sequence corresponding to the object cluster, the method further includes: In the case where the seasonal information corresponding to the object cluster cannot be determined based on the predicted activity sequence, an adjustable range is determined for adjusting the preset thresholds corresponding to multiple seasonal stages in the seasonal information. Within the adjustable range, the preset thresholds corresponding to the multiple seasonal stages are reduced at least once to obtain multiple adjusted thresholds for determining the multiple seasonal stages, so as to determine the seasonal information corresponding to the object cluster based on the multiple adjusted thresholds.

16. The method according to any one of claims 1-6, characterized in that, After determining the timing information corresponding to the object cluster, the method further includes: The number of seasonal periods included in the aforementioned seasonal information is counted; When the number of the seasons is greater than or equal to a preset threshold, the preset thresholds for determining the multiple seasons in a single season are increased to obtain multiple adjusted thresholds for determining multiple seasons, so as to reduce the number of seasons included in the season information based on the multiple adjusted thresholds.

17. An electronic device, characterized in that, include: A memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions, when executed by the processor, implement the method of any one of claims 1-16.

18. A computer storage medium, characterized in that, Used to store a computer program that, when executed by a computer, implements the method of any one of claims 1-16.

Images