Information pushing method and device, model training method and device, equipment and storage medium

Abstract

The embodiment of the invention provides an information pushing method and device, a model training method and device, equipment and a storage medium. The method comprises the following steps: acquiring commodity multi-source data of a commodity; carrying out feature extraction on the commodity multi-source data of the commodity to obtain a total feature vector of the commodity; a shared encoder based on a multi-task model performs shared encoding processing on the full-quantity feature vector of the commodity to obtain a shared representation vector of the commodity; wherein the shared representation vector represents context semantic information of the commodity; processing the shared representation vector of the commodity based on a multi-task head of the multi-task model to obtain prediction information of the commodity in a future time period; wherein the prediction information comprises at least one of the following contents: pricing suggestion, comprehensive transaction probability and demand change trend; and carrying out information push processing on the prediction information of the commodity. According to the method, the transaction efficiency and the user satisfaction of the second-hand commodity transaction platform can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an information push method, a model training method, a device, an equipment, and a storage medium. Background Technology

[0002] In the current secondhand goods trading market, sellers price their goods, and when buyers search for goods, the recommendation system faces multiple pain points in recommending goods (including items to buyers): When buyers are present, existing recommendation systems rely solely on keyword matching or basic collaborative filtering algorithms, failing to combine key attributes such as the reasonableness of the product's pricing and its condition for in-depth matching. This results in a discrepancy between the recommended results and the user's actual needs (such as "high cost-performance ratio" or "urgent sale").

[0003] In existing technologies, product pricing is based on historical average transaction prices; product recommendations are made through keyword matching or basic collaborative filtering algorithms.

[0004] However, in the above methods, sellers rely too much on historical average transaction prices when setting prices, causing product pricing to become out of touch with actual market demand; at the same time, relying solely on keyword matching or collaborative filtering results in a single matching dimension and ignores the reasonableness of prices, leading to a disconnect between recommendation results and user needs. Summary of the Invention

[0005] This application provides information push methods, model training methods, devices, equipment, and storage media, which improve the scientific nature of dynamic commodity pricing, the accuracy of supply and demand matching, and the foresight of market forecasting, thereby improving the transaction efficiency and user satisfaction of second-hand commodity trading platforms.

[0006] Firstly, this application provides an information push method, including:

[0007] Obtain multi-source data of the product; and extract features from the multi-source data of the product to obtain the full feature vector of the product;

[0008] A shared encoder based on a multi-task model performs shared encoding processing on the full feature vector of the product to obtain a shared representation vector of the product; wherein, the shared representation vector represents the contextual semantic information of the product.

[0009] Based on the multi-task head of the multi-task model, the shared representation vector of the product is processed to obtain the prediction information of the product in the future time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend;

[0010] The predicted information for the product is processed for information push.

[0011] In one possible implementation, the step of extracting features from the multi-source data of the product to obtain the full feature vector of the product includes:

[0012] Feature extraction is performed on the internal data of the product from the multi-source data to obtain the internal feature vector of the product; wherein, the internal data includes static attributes of the product, dynamic data of the product, transaction data, and user data;

[0013] Feature extraction is performed on the external data in the multi-source data of the product to obtain the external feature vector of the product; wherein, the external data includes platform data, public opinion data and market data;

[0014] The internal and external feature vectors of the product are concatenated to obtain the full feature vector of the product.

[0015] In one possible implementation, the internal feature vector includes structured feature vectors and unstructured feature vectors; the structured feature vectors include numerical features, category features, and time series features; the unstructured feature vectors include text semantic vectors and user intent vectors.

[0016] And / or, the external feature vector includes key ratio features, public opinion features, and quadruples.

[0017] In one possible implementation, the multi-task head based on the multi-task model processes the shared representation vector of the goods to obtain prediction information of the goods within a future time period, including:

[0018] Based on the pricing regression head of the multi-task head, the shared representation vector is processed to obtain pricing suggestions in the prediction information of the product;

[0019] And / or, based on the supply and demand matching head of the multi-task head, process the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the prediction information of the product;

[0020] And / or, based on the trend prediction head of the multi-task head, the shared representation vector is processed to obtain the demand change trend in the prediction information of the commodity.

[0021] In one possible implementation, the supply and demand matching head based on the multi-task head processes the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the product's prediction information, including:

[0022] Based on the supply and demand matching head, the similarity between the shared representation vector and the user intent vector in the full feature vector is determined;

[0023] A reasonableness score is obtained by calculating the pricing recommendations in the forecast information of the product and the current price;

[0024] Based on the fully connected network of the supply and demand matching head, the similarity, the rationality score, and the seller's positive review rate in the full feature vector are fused to obtain the comprehensive transaction probability in the product's prediction information.

[0025] In one possible implementation, the trend prediction head based on the multi-task head processes the shared representation vector to obtain the demand change trend in the prediction information of the commodity, including:

[0026] The historical sequence of the product is encoded to obtain a temporal context; wherein, the historical sequence includes the historical price and transaction volume of the product category within a historical time period;

[0027] Based on the trend prediction head, the temporal context and the shared representation vector are fused and predicted to obtain the demand change trend in the predicted information of the commodity.

[0028] In one possible implementation, the information push processing of the predicted information for the goods includes:

[0029] If it is determined that the change value corresponding to the demand change trend represented by the predicted information is greater than a preset threshold, then a push message for the product is generated based on the predicted information of the product; wherein, the push message includes the demand change trend and a jump link;

[0030] The push message is processed asynchronously.

[0031] Secondly, this application provides a model training method for information push applications, including:

[0032] Obtain multi-source data of at least one product to be trained; and extract features from the multi-source data of the product to be trained to obtain the full feature vector of the product to be trained.

[0033] The shared encoder based on the initial model performs shared encoding processing on the full feature vector of the product to be trained to obtain the shared representation vector of the product to be trained; wherein, the shared representation vector represents the contextual semantic information of the product to be trained.

[0034] Based on the multi-task head of the initial model, the shared representation vector of the product to be trained is processed to obtain the prediction information of the product to be trained within a preset time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend;

[0035] Based on the prediction information of the product to be trained, the initial model is trained to obtain a multi-task model; wherein, the multi-task model is used to process the multi-source data of the product as described in the first aspect to obtain the prediction information of the product.

[0036] In one possible implementation, the step of extracting features from the multi-source data of the product to be trained to obtain the full feature vector of the product to be trained includes:

[0037] Feature extraction is performed on the internal data in the multi-source data of the product to obtain the internal feature vector of the product to be trained; wherein, the internal data includes static attributes of the product, dynamic data of the product, transaction data, and user data;

[0038] Feature extraction is performed on the external data in the multi-source data of the product to obtain the external feature vector of the product to be trained; wherein, the external data includes platform data, public opinion data and market data;

[0039] The internal and external feature vectors of the product to be trained are concatenated to obtain the full feature vector of the product to be trained.

[0040] In one possible implementation, the internal feature vector includes structured feature vectors and unstructured feature vectors; the structured feature vectors include numerical features, category features, and time series features; the unstructured feature vectors include text semantic vectors and user intent vectors.

[0041] And / or, the external feature vector includes key ratio features, public opinion features, and quadruples.

[0042] In one possible implementation, the multi-task head based on the initial model processes the shared representation vector of the product to be trained to obtain prediction information of the product to be trained within a preset time period, including:

[0043] Based on the pricing regression head of the multi-task head, the shared representation vector is processed to obtain pricing suggestions in the prediction information of the product to be trained;

[0044] And / or, based on the supply and demand matching head of the multi-task head, process the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the prediction information of the product to be trained;

[0045] And / or, based on the trend prediction head of the multi-task head, the shared representation vector is processed to obtain the demand change trend in the prediction information of the product to be trained.

[0046] In one possible implementation, the supply and demand matching head based on the multi-task head processes the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the prediction information of the product to be trained, including:

[0047] Based on the supply and demand matching head, the similarity between the shared representation vector and the user intent vector in the full feature vector is determined;

[0048] A reasonableness score is obtained by calculating the pricing suggestions and current prices in the prediction information of the product to be trained.

[0049] Based on the fully connected network of the supply and demand matching head, the similarity, the rationality score and the seller's positive review rate in the full feature vector are fused to obtain the comprehensive transaction probability in the prediction information of the product to be trained.

[0050] In one possible implementation, the trend prediction head based on the multi-task head processes the shared representation vector to obtain the demand change trend in the prediction information of the product to be trained, including:

[0051] The historical sequence of the product to be trained is encoded to obtain the temporal context; wherein, the historical sequence includes the historical price and transaction volume of the product category in the historical time period;

[0052] Based on the trend prediction head, the temporal context and the shared representation vector are fused and predicted to obtain the demand change trend in the prediction information of the product to be trained.

[0053] In one possible implementation, training the initial model based on the prediction information of the product to be trained to obtain a multi-task model includes:

[0054] The pricing loss function is determined based on the actual pricing results of the product to be trained and the pricing suggestions in the prediction information;

[0055] The matching loss function is determined based on the actual overall transaction probability of the product to be trained and the overall transaction probability in the predicted information.

[0056] The trend loss function is determined based on the actual demand change trend of the product to be trained and the demand change trend in the predicted information;

[0057] The total loss function is determined based on the pricing loss function, the matching loss function, and the trend loss function;

[0058] The initial model is trained based on the total loss function to obtain the multi-task model.

[0059] Thirdly, this application provides an information push device, comprising:

[0060] The extraction module is used to acquire multi-source data of the product; and to extract features from the multi-source data of the product to obtain the full feature vector of the product.

[0061] The encoding module is used to perform shared encoding processing on the full feature vector of the product based on a multi-task model to obtain the shared representation vector of the product; wherein the shared representation vector represents the contextual semantic information of the product.

[0062] The prediction module is used to process the shared representation vector of the product based on the multi-task head of the multi-task model to obtain prediction information of the product in the future time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend;

[0063] The push module is used to push information about the predicted information of the product.

[0064] Fourthly, this application provides a model training device for information push, comprising:

[0065] An extraction module is used to acquire multi-source data of at least one product to be trained; and to extract features from the multi-source data of the product to be trained to obtain the full feature vector of the product to be trained.

[0066] The encoding module is used to perform shared encoding processing on the full feature vector of the product to be trained based on the shared encoder of the initial model, so as to obtain the shared representation vector of the product to be trained; wherein, the shared representation vector represents the contextual semantic information of the product to be trained.

[0067] The prediction module is used to process the shared representation vector of the product to be trained based on the multi-task head of the initial model to obtain prediction information of the product to be trained within a preset time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend;

[0068] The training module is used to train the initial model based on the prediction information of the product to be trained, thereby obtaining a multi-task model; wherein the multi-task model is used to process the multi-source data of the product as described in the third aspect to obtain the prediction information of the product.

[0069] Fifthly, this application provides an electronic device, comprising: a memory and a processor; the memory storing computer-executable instructions; the processor executing the computer-executable instructions stored in the memory, causing the processor to perform the methods of the first aspect and / or the second aspect above.

[0070] In a sixth aspect, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods of the first and / or second aspects described above.

[0071] In a seventh aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the methods of the first and / or second aspects described above.

[0072] The information push method, model training method, apparatus, device, and storage medium provided in this application extract features from multi-source data of a product to obtain a full feature vector of the product. A shared encoder of a multi-task model performs shared encoding processing on this full feature vector to obtain a shared representation vector representing the contextual semantic information of the product. The multi-task head of the multi-task model processes this shared representation vector to predict the product's forecast information for a future time period, including at least one of the following: pricing suggestions, overall transaction probability, and demand change trends. This forecast information is then pushed to users. Furthermore, by extracting the contextual semantic information from the multi-source data of the product through the shared encoder and processing this contextual semantic information based on the multi-task head, the collaborative optimization of pricing, recommendation, and prediction tasks is achieved. This improves the scientific nature of dynamic product pricing, the accuracy of supply and demand matching, and the forward-looking nature of market forecasting, thereby enhancing the transaction efficiency and user satisfaction of the second-hand goods trading platform. Attached Figure Description

[0073] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0074] Figure 1 A schematic diagram of a digital operation architecture provided for this application;

[0075] Figure 2 A flowchart illustrating an information push method provided in an embodiment of this application;

[0076] Figure 3 A flowchart illustrating another information push method provided in an embodiment of this application;

[0077] Figure 4 A schematic diagram of an application layer workflow provided in an embodiment of this application;

[0078] Figure 5 A flowchart illustrating a model training method for information push applications provided in this application embodiment;

[0079] Figure 6 A flowchart illustrating another model training method for information push provided in an embodiment of this application;

[0080] Figure 7 This is a schematic diagram of the structure of an information push device provided in an embodiment of this application;

[0081] Figure 8 A schematic diagram of the structure of a model training device for information push provided in an embodiment of this application;

[0082] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0083] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0084] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0085] It should be noted that 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, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, they do not violate public order and good morals, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0086] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on user rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0087] Figure 1 A schematic diagram of a digital operation architecture provided for this application, such as Figure 1 As shown, the system of this application is deployed in the digital operation architecture of a second-hand goods trading platform. Its architecture adopts a layered microservice design, mainly including a data acquisition layer, a data processing and model service layer, and an application service layer. Specifically, the data acquisition layer accesses the database; the data processing layer deploys feature engineering; the model service layer deploys a shared encoder and multi-task heads for multi-task models to achieve joint training and inference; and the application service layer interacts with users through push notifications, recommendation lists, and pricing suggestion interfaces.

[0088] Based on the above scenarios, it can be seen that sellers rely too heavily on historical average transaction prices when setting prices, causing product pricing to become out of touch with actual market demand. At the same time, relying solely on keyword matching or collaborative filtering fails to combine the reasonableness of product pricing for in-depth matching, resulting in recommendation results that are out of touch with user needs.

[0089] The information push method provided in this application extracts contextual semantic information from multi-source product data through a shared encoder, and processes this contextual semantic information based on a multi-task head to achieve collaborative optimization of pricing, recommendation and prediction tasks.

[0090] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0091] Figure 2 This is a flowchart illustrating an information push method provided in an embodiment of this application, as shown below. Figure 2 As shown, the method includes:

[0092] 201. Obtain multi-source data of the product; and extract features from the multi-source data of the product to obtain the full feature vector of the product.

[0093] For example, the execution subject in this embodiment can be an electronic device, hereinafter referred to as the device. The device, based on the application layer, responds to task triggering instructions and performs multi-source data acquisition, that is, obtains multi-source data for each product from a multi-source database, including product static attributes, dynamic transaction data, user behavior data, and external market data. Based on feature engineering, the multi-source data of the products is fused and aligned to obtain fused data, and features are extracted from this fused data to obtain a unified high-dimensional full-feature vector for each product, i.e., a full-feature vector, which serves as the data input for subsequent processing.

[0094] For example, a distributed task scheduling framework can be used to configure a scheduled task to update at 2 AM every day, or a list of all products currently on sale can be read from the database in a paginated manner. For each product, the service calls the feature engineering pipeline to map structured data, unstructured text, and time-series information from multiple sources to a unified feature vector space. The resulting full feature vector provides accurate input to the multi-task model, improving the accuracy of subsequent pricing, matching, and prediction.

[0095] 202. A shared encoder based on a multi-task model performs shared encoding on the full feature vector of a product to obtain a shared representation vector of the product; wherein, the shared representation vector represents the contextual semantic information of the product.

[0096] For example, the device invokes a preset multi-task model, which deploys a shared encoder consisting of a multi-layer Transformer encoder. Its self-attention mechanism efficiently captures complex interactions and dependencies between features. For instance, the combination of the brand / model feature "iXX 16" and the storage feature "128G" is far more valuable than their individual summations. Based on this shared encoder, the full feature vector of the product is shared-encoded to obtain a shared representation vector. This shared representation vector is a high-level semantic representation vector that deeply integrates all information and is context-sensitive, representing the product's contextual semantic information. Through the shared encoder, the model can deeply integrate structured, unstructured, and time-series data, learning a richer and more accurate representation of product value.

[0097] 203. Based on the multi-task head of the multi-task model, the shared representation vector of the product is processed to obtain the prediction information of the product in the future time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend.

[0098] For example, the multi-task model also deploys a multi-task head. Based on the content of the current task trigger instruction, the current trigger scenario is determined, and based on the trigger scenario, it is decided which one or more target task heads in the multi-task head set to invoke. Based on the target task head, the shared representation vector of the product is processed to predict the product's forecast information for the future time period, including at least one of the following: pricing recommendations, overall transaction probability, and demand change trends.

[0099] For example, when a seller publishes / modifies a product, the shared representation vector of the product is processed based on the corresponding target task header to predict the product's pricing suggestions for a future time period; or, when a scheduled task (such as 2 a.m.) is triggered, the shared representation vector of the product is processed based on the corresponding target task header to predict the product's demand trend for a future time period.

[0100] 204. Process and push information on commodity forecasts.

[0101] For example, the obtained product prediction information can be packaged into a push message and pushed directly to the target user's (seller's or buyer's) device.

[0102] This includes recording whether sellers / buyers view the forecast information or adopt the suggestions in the forecast information, and using this recorded data as feedback signals to flow back to the data layer for subsequent evaluation of the effectiveness of the early warning strategy and optimization of the performance of the multi-task model.

[0103] This embodiment provides an information push method that extracts contextual semantic information from multi-source product data using a shared encoder and processes this contextual semantic information based on a multi-task head. This enables collaborative optimization of pricing, recommendation, and prediction tasks, improving the scientific nature of dynamic product pricing, the accuracy of supply and demand matching, and the foresight of market prediction. Consequently, it enhances the transaction efficiency and user satisfaction of the second-hand goods trading platform.

[0104] Figure 3 A flowchart illustrating another information push method provided in this application embodiment is shown below. Figure 3 As shown, the method includes:

[0105] 301. Obtain multi-source data for the product.

[0106] For example, this step can be referred to as step 201, which will not be repeated here.

[0107] 302. Extract features from the internal data of the multi-source data of the product to obtain the internal feature vector of the product; wherein, the internal data includes the static attributes of the product, the dynamic data of the product, the transaction data, and the user data.

[0108] For example, the multi-source data of a product includes internal data, which includes static product attributes, dynamic product data, transaction data, and user data. Based on refined feature engineering, the device extracts features from the product's internal data to obtain multiple feature data sets, and then fuses and concatenates these feature data sets to obtain the product's internal feature vector, representing the characteristics of the product's entity and internal operational data.

[0109] The static attributes of the products are obtained through forms filled out by the sellers, including structured data such as brand, model, memory, color, purchase time, condition (e.g., 95% new), repair history, and parts list; the dynamic data of the products includes time-series data such as product release time, seller's pricing history (each modification record), daily pageviews, inquiries, and favorites; the transaction data includes the final transaction price, transaction cycle, and buyer and seller IDs; and the user data includes buyer search terms, click sequence, dwell time, order records, seller's historical positive review rate, return rate, and response speed.

[0110] The internal feature vectors include structured feature vectors and unstructured feature vectors; structured feature vectors include numerical features, categorical features, and time series features; unstructured feature vectors include text semantic vectors and user intent vectors.

[0111] Specifically, based on refined feature engineering, numerical features of continuous values ​​such as price and size are extracted and standardized using tools. High-cardinality category features such as brand and model are extracted to learn semantic relationships between categories, and then their low-dimensional embedding vectors are learned. Each category is mapped to a D-dimensional dense vector. For the historical price and transaction volume sequence of a certain product category, a lightweight one-dimensional convolutional neural network (CNN) is used for feature extraction. The input of this network is a sequence window of the past T days, and the output is a fixed-length trend representation vector. This vector encodes recent trends, volatility, periodicity and other pattern information, compressing the dynamically changing time series information into semantically rich time series features. In turn, multiple features of the internal structured data are obtained, namely the structured feature vector. A pre-trained model (Bidirectional Encoder Representations from Transformers, or BERT for short) is used to encode product description text (such as "the screen has slight scratches" or "the packaging box is still there") to generate text semantic vectors. BERT is also used to encode aggregated buyer search terms and historical browsing behavior to generate user intent vectors, which accurately capture users' immediate needs and long-term preferences, providing a foundation for intelligent matching. Furthermore, through the vectorization of product text descriptions and user intent, multiple features of the internal unstructured data are obtained, namely unstructured feature vectors.

[0112] 303. Extract features from external data in multi-source commodity data to obtain the external feature vector of the commodity; where external data includes platform data, public opinion data and market data.

[0113] For example, the multi-source data of a product also includes external data, including platform data, public opinion data, and market data. Based on refined feature engineering, features are extracted from the external data in the multi-source data of the product to obtain multiple external feature data. These multiple external feature data are then fused and concatenated to obtain the external feature vector of the product, which represents the characteristics of the product's external market data.

[0114] The external feature vector includes key ratio features, public opinion features, and quadruples. Specifically, it involves associating and calculating platform data and new phone market data, such as associating "brand + model" with second-hand goods and calculating key ratio features, such as second-hand price / new phone price, (new phone current price - second-hand price) / new phone current price (measuring discount), etc.; it involves quantifying the sentiment and popularity of public opinion data to obtain public opinion features, such as using sentiment analysis models to score captured social media texts and calculating the average sentiment score and negative public opinion ratio of a specific model within a time window; and it involves extracting structured information from macro reports in market data, such as obtaining the plain text and publication time of industry reports through Optical Character Recognition (OCR) technology, and constructing (indicator, value, direction of change, time) quadruples using named entity recognition and relation extraction. For example, it involves extracting (smartphone shipment growth rate, +6.5%, year-on-year growth, 2025-Q4) from "Q4 smartphone shipments increased by 6.5%".

[0115] 304. Concatenate the internal and external feature vectors of the product to obtain the full feature vector of the product.

[0116] For example, based on vector concatenation technology, the internal and external feature vectors of a product are concatenated into a unified, high-dimensional full feature vector of the product, thus obtaining the full feature vector of the product to ensure that the product value can be fully captured in subsequent steps. For example, the matching degree feature (1-dimensional) and the semantic representation vector (512-dimensional) are concatenated into a 513-dimensional vector.

[0117] By collecting static product attributes (such as brand and model), dynamic transaction data (such as page views and transaction prices), and external market data (such as new product prices and public opinion sentiment scores), a unified representation is generated through standardization, embedding, and feature extraction. This provides reliable data input for subsequent processing, enabling the generated dynamic price range to scientifically reflect the actual value of the product and reduce unsold inventory or losses caused by pricing deviations. Furthermore, it achieves precise optimization of supply and demand matching, thereby significantly improving the conversion rate of the recommendation system.

[0118] 305. A shared encoder based on a multi-task model performs shared encoding on the full feature vector of a product to obtain a shared representation vector of the product; wherein the shared representation vector represents the contextual semantic information of the product.

[0119] For example, this step can be referred to as step 202, which will not be repeated here.

[0120] After step 305, steps 306, 307, and / or 308 may be performed.

[0121] 306. A pricing regression head based on a multi-task head processes the shared representation vector to obtain pricing suggestions from the product prediction information.

[0122] For example, after step 305, when the current triggering scenario is a seller publishing / modifying a product, the shared representation vector of the product can be input into the pricing regression head of the multi-task head. The structure of the pricing regression head is: FC(ReLU)->FC(ReLU)->Linear(2). Based on the pricing regression head, the shared representation vector is processed to obtain the pricing suggestion in the prediction information of the product; for example, two scalar values ​​[price_low, price_high] are output to represent the suggested price range.

[0123] 307. A supply and demand matching head based on a multi-task head processes the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the product prediction information.

[0124] For example, after step 305, when the current triggering scenario is a buyer searching for / browsing products, the shared representation vector of each product in all product lists and the user intent vector from the full feature vector can be input into the supply and demand matching head of the multi-task head. Based on this supply and demand matching head, the shared representation vector and the user intent vector from the full feature vector are fused to obtain the comprehensive transaction probability in the product's prediction information, including the comprehensive transaction probability representing "person-product matching degree + price competitiveness + trust degree".

[0125] In one possible implementation, step 307 includes: based on the supply and demand matching head, the similarity Sim between the shared representation vector H and the user intent vector U (e.g., the user intent vector after BERT encoding of the user's current search term) in the full feature vector can be calculated using a cosine similarity calculation formula, for example, Sim=Cosine Similarity(H,U). And based on a preset calculation rule, a reasonableness score is calculated for the pricing suggestions and current price in the product's prediction information, for example, reasonableness score = -|(seller's current listed price - (price_low + price_high) / 2)| / ((price_low + price_high) / 2), where price_low and price_high are the suggested minimum price and suggested maximum price in the pricing suggestions, respectively. The feature concatenation is performed on the feature Sim (corresponding to similarity), the feature Rationality_Score (corresponding to reasonableness score), and the feature Seller_Rating (corresponding to seller positive review rate in the full feature vector). The concatenated feature is then input into the fully connected network of the supply-demand matching head for final scoring. For example, the overall transaction probability Matching_Score in the product prediction information is predicted by calculating the following formula: Matching_Score = Sigmoid(FC(concat(Sim, Rationality_Score, Seller_Rating))). By sharing the similarity calculation between the representation vector and the user intent vector, precise optimization of supply-demand matching is achieved.

[0126] 308. A trend prediction head based on a multi-task head processes the shared representation vector to obtain the demand change trend in the prediction information of commodities.

[0127] For example, after step 305, when the current triggering scenario is a timed task (such as 2 AM), the shared representation vector of each product can be input into the trend prediction head of the multi-task head. Based on this trend prediction head, feature processing is performed on the shared representation vector to obtain the demand change trend in the product's prediction information.

[0128] For example, by calling the trend forecasting model in the deployed trend forecasting header and passing in the shared representation vector of the product, the trend forecasting model outputs a prediction result, including the field {"item_id":"12345","predicted_demand_change":0.15,"confidence": 0.88}, indicating that the demand for this product is predicted to increase by 15% in the future, and the confidence level of the returned result is 88%.

[0129] By processing the shared representation vector through multiple task heads (pricing head, matching head, and prediction head), dynamic pricing, supply and demand matching, and trend prediction tasks are completed respectively. This significantly improves the scientific nature of dynamic pricing (avoiding unsold goods or losses caused by subjective pricing), the accuracy of supply and demand matching (prioritizing the recommendation of high-quality and low-priced goods), and the foresight of market prediction (proactively pushing price adjustment warnings), thereby improving the transaction efficiency and user satisfaction of the second-hand goods trading platform.

[0130] In one possible implementation, step 308 includes: obtaining the historical sequence of the commodity, including the historical price and trading volume of the commodity category over a historical period, and performing feature encoding processing on the historical sequence of the commodity based on an encoder to obtain the time-series context (such as volatility, periodicity). For example, the price sequence of the past 7 days is encoded into a 128-dimensional trend representation vector corresponding to the time-series context. Through the trend prediction head, the time-series context is connected with the shared representation vector to form a more comprehensive input feature to enhance the model's ability to perceive market dynamics. For example, by using ΔDemand=Linear(FC(concat(H, C_seq))), the trend representation vector C_seq (128 dimensions) and the shared representation vector H (512 dimensions) are concatenated into a 640-dimensional vector, and feature processing is performed on this input feature to predict the demand change trend in the commodity's forecast information.

[0131] By introducing historical trend representation vectors, the trend forecasting head can more accurately capture market dynamics (such as the price drop of older models due to the release of new models). For example, it can proactively push alerts to sellers such as "Demand is predicted to drop by 20%, and price reductions are recommended," thereby reducing the risk of unsold inventory.

[0132] 309. If the change value corresponding to the demand change trend represented by the forecast information is greater than a preset threshold, then a push message for the product is generated based on the product's forecast information; wherein, the push message includes the demand change trend and a jump link.

[0133] For example, after steps 306, 307, and 308, when a scenario triggering demand change trend prediction is encountered, the change value corresponding to the demand change trend in the prediction information is extracted. A minimum change threshold is set, and the change value is compared with the minimum change threshold. If the change value is greater than the minimum change threshold, a jump link for the product is generated. The demand change trend in the product's prediction information and the jump link are then packaged together to generate a corresponding push message. For example, the most suitable copywriting template is selected from a message template library, and the demand change trend in the product's prediction information and the jump link are assembled to obtain the corresponding push message.

[0134] 310. Perform asynchronous push processing on push messages.

[0135] For example, an asynchronous message queuing mechanism is used to process each product's push message asynchronously before pushing it to the target device. This allows users to directly modify the current product's price based on demand trends indicated in the push message after clicking the link. The push can be implemented through multiple channels, such as app push notifications or in-app information pushes.

[0136] For example, Figure 4 A schematic diagram of an application layer workflow is provided for an embodiment of this application, such as... Figure 4 As shown, a scheduled task is configured via a timed trigger, set to update daily at 2 AM. Based on the price alert service, it iterates through all available products. For each product, it calls the trend prediction model in the trend forecasting section to perform model prediction processing on the latest characteristics of the product and its category, obtaining the model's predicted demand change trend. It determines whether the change corresponding to the model's predicted demand change trend exceeds a threshold. If it does, a personalized alert message is generated based on the model's predicted demand change trend and sent to a message queue. Leveraging the asynchronous and peak-shaving mechanisms of push notification tools (such as Push Worker), the message is pushed to sellers via App push notifications, in-app messages, SMS, and other push channels. The message includes a link to publish the product. Following the suggestion in the message, the seller clicks the link and operates on the product management page to modify the current product's price. Simultaneously, the system records whether the seller views and adopts the suggestion, using this as feedback signals fed back to the data layer for subsequent evaluation of the alert strategy's effectiveness and model optimization, forming a feedback loop. By predicting changes in demand and proactively sending out alerts, sellers are empowered to adjust their strategies in a timely manner, seize the best sales opportunities, and thus improve the transaction rate and user satisfaction of the second-hand goods trading platform.

[0137] In this embodiment, based on the above embodiments, an end-to-end solution from data input to final application (pricing suggestions, recommendation lists, and warning notifications) is provided, which reduces the possible deviations and inconsistencies between multiple independent systems and results in better overall performance.

[0138] Figure 5 This application provides a flowchart illustrating a model training method for information push applications, as shown in the embodiments below. Figure 5 As shown, the method includes:

[0139] 401. Obtain multi-source data of at least one product to be trained; and extract features from the multi-source data of the product to be trained to obtain the full feature vector of the product to be trained.

[0140] For example, the execution subject of this embodiment may be an electronic device, hereinafter referred to as the device. The device collects training data, that is, it obtains multi-source data of each product to be trained from the training database, including product static attributes, dynamic transaction data, user behavior data, and external market data. Based on feature engineering, the multi-source data of the product to be trained is fused and aligned to obtain fused data, and features are extracted from the fused data to obtain a unified high-dimensional full-feature vector of each product to be trained, that is, the full-feature vector, which is used as the data input for subsequent processing.

[0141] For example, for each product to be trained, the service invokes the feature engineering pipeline to map structured data, unstructured text, and time-series information from multiple sources to a unified feature vector space. The resulting full feature vector provides accurate input to the initial model, improving the accuracy of subsequent models in pricing, matching, and prediction.

[0142] 402. Based on the initial model, the shared encoder performs shared encoding on the full feature vector of the product to be trained to obtain the shared representation vector of the product to be trained; wherein, the shared representation vector represents the contextual semantic information of the product to be trained.

[0143] For example, the device invokes a preset initial model, which deploys a shared encoder consisting of a multi-layer Transformer encoder. Its self-attention mechanism efficiently captures complex interactions and dependencies between features. Based on this shared encoder, the full feature vectors of the product to be trained are shared-encoded to obtain a shared representation vector for the product. This shared representation vector is a context-sensitive, high-level semantic representation vector that deeply integrates all information to represent the contextual semantic information of the product. Through the shared encoder, the model can deeply integrate structured, unstructured, and time-series data to learn richer and more accurate product values.

[0144] 403. Based on the multi-task head of the initial model, the shared representation vector of the product to be trained is processed to obtain the prediction information of the product to be trained within a preset time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend.

[0145] For example, the initial model also deploys multiple task heads, determining which one or more target task heads to invoke based on a preset trigger scenario. Based on the target task head, the shared representation vector of the product to be trained is processed to predict information about the product to be trained within a future time period, including at least one of the following: pricing recommendations, overall transaction probability, and demand change trends. For instance, based on the target task head, the shared representation vector of the product to be trained is processed to predict pricing recommendations for the product to be trained within a preset time period; or, when a scheduled task (e.g., 2 AM) is triggered, based on the corresponding target task head, the shared representation vector of the product to be trained is processed to predict the demand change trends for the product to be trained within a preset time period.

[0146] 404. Based on the prediction information of the products to be trained, the initial model is trained to obtain a multi-task model; wherein, the multi-task model is used to process multi-source data of products such as information push methods to obtain prediction information of products.

[0147] For example, based on the predicted and actual information of the product to be trained, the model parameters of the initial model are optimized to obtain a multi-task model. This multi-task model is then used to perform multi-source data prediction processing on the product in the information push method to obtain the predicted information for that product. The obtained predicted information can be packaged into a push message and directly pushed to the target user's (seller's or buyer's) device.

[0148] This embodiment provides a model training method for information push. By extracting contextual semantic information from multi-source product data through a shared encoder, and processing this contextual semantic information based on a multi-task head, the method achieves collaborative optimization of pricing, recommendation, and prediction tasks. This is used to train an initial model to obtain a multi-task model, thereby achieving accurate pricing, recommendation, and trend prediction. This improves the scientific nature of dynamic product pricing, the accuracy of supply and demand matching, and the foresight of market prediction, thus enhancing the transaction efficiency and user satisfaction of the second-hand goods trading platform.

[0149] Figure 6 A flowchart illustrating another model training method for information push provided in this application embodiment is shown below. Figure 6 As shown, the method includes:

[0150] 501. Obtain multi-source data for at least one product to be trained.

[0151] For example, this step can be referred to as step 401, which will not be repeated here.

[0152] 502. Extract features from the internal data of the multi-source data of the product to obtain the internal feature vector of the product to be trained; wherein, the internal data includes the static attributes of the product, the dynamic data of the product, the transaction data, and the user data.

[0153] For example, the multi-source data of each product to be trained includes internal data, which includes product static attributes, product dynamic data, transaction data, and user data. Based on refined feature engineering, the device extracts features from the internal data of each product to be trained, obtaining multiple feature data, and then fuses and splices these multiple feature data to obtain the internal feature vector of each product to be trained, thereby representing the characteristics of the entity and internal operational data of the product to be trained.

[0154] The static attributes of the products are obtained through forms filled out by the sellers, including structured data such as brand, model, memory, color, purchase time, condition (e.g., 95% new), repair history, and parts list; the dynamic data of the products includes time-series data such as product release time, seller's pricing history (each modification record), daily pageviews, inquiries, and favorites; the transaction data includes the final transaction price, transaction cycle, and buyer and seller IDs; and the user data includes buyer search terms, click sequence, dwell time, order records, seller's historical positive review rate, return rate, and response speed.

[0155] The internal feature vectors include structured feature vectors and unstructured feature vectors; structured feature vectors include numerical features, categorical features, and time series features; unstructured feature vectors include text semantic vectors and user intent vectors.

[0156] Specifically, based on refined feature engineering, numerical features of continuous values ​​such as price and size are extracted and standardized using tools. High-cardinality category features such as brand and model are extracted to learn semantic relationships between categories, and then their low-dimensional embedding vectors are learned. Each category is mapped to a D-dimensional dense vector. For the historical price and transaction volume sequence of a certain product category, a lightweight one-dimensional convolutional neural network (CNN) is used for feature extraction. The input of this network is a sequence window of the past T days, and the output is a fixed-length trend representation vector. This vector encodes recent trends, volatility, periodicity and other pattern information, compressing the dynamically changing time series information into semantically rich time series features. In turn, multiple features of the internal structured data are obtained, namely the structured feature vector. Using the pre-trained BERT model, product description text (such as "slight scratches on the screen" or "the packaging box is still there") is encoded to generate text semantic vectors. BERT is also used to encode aggregated buyer search terms and historical browsing behavior to generate user intent vectors, accurately capturing users' immediate needs and long-term preferences, providing a foundation for intelligent matching. Furthermore, through the vectorization of product text descriptions and user intent, multiple features of internal unstructured data are obtained, namely unstructured feature vectors.

[0157] 503. Extract features from external data in multi-source product data to obtain the external feature vector of the product to be trained; where external data includes platform data, public opinion data and market data.

[0158] For example, the multi-source data of each product to be trained also includes external data, including platform data, public opinion data, and market data. Based on refined feature engineering, features are extracted from the external data in the multi-source data of each product to be trained to obtain multiple external feature data. These multiple external feature data are then fused and concatenated to obtain the external feature vector of each product to be trained, which represents the characteristics of the external market data of each product to be trained.

[0159] The external feature vector includes key ratio features, public opinion features, and quadruples. Specifically, it correlates and calculates platform data and new product market data, such as associating "brand + model" with second-hand goods and calculating key ratio features; it quantifies the sentiment and popularity of public opinion data to obtain public opinion features, such as calculating the average sentiment score and negative public opinion ratio of a specific model within a time window; and it extracts structured information from macro reports in market data, obtains the plain text and publication time of industry reports, and uses named entity recognition and relation extraction to construct quadruples (indicator, value, direction of change, time).

[0160] 504. Concatenate the internal and external feature vectors of the product to be trained to obtain the full feature vector of the product to be trained.

[0161] For example, based on vector concatenation technology, the internal and external feature vectors of each product to be trained are concatenated into a unified, high-dimensional full feature vector of the product, thus obtaining the full feature vector of each product to be trained, to ensure that the value of each product to be trained can be fully captured in subsequent training. For example, the matching degree feature (1-dimensional) and the semantic representation vector (512-dimensional) are concatenated into a 513-dimensional vector.

[0162] By collecting static product attributes (such as brand and model), dynamic transaction data (such as page views and transaction prices), and external market data (such as new product prices and public opinion sentiment scores), a unified representation is generated through standardization, embedding, and feature extraction. This provides reliable data input for subsequent multi-task models, improving their performance and enabling the dynamic price range generated by the multi-task models to scientifically reflect the actual value of products, reducing unsold inventory or losses caused by pricing deviations. It also enables the multi-task models to achieve precise optimization of supply and demand matching.

[0163] 505. Based on the initial model, the shared encoder performs shared encoding on the full feature vector of the product to be trained, and obtains the shared representation vector of the product to be trained; wherein, the shared representation vector represents the contextual semantic information of the product to be trained.

[0164] For example, this step can be referred to as step 402, which will not be repeated here.

[0165] After step 505, steps 506, 507, and / or 508 may be performed.

[0166] 506. A pricing regression head based on a multi-task head processes the shared representation vector to obtain pricing suggestions from the prediction information of the product to be trained.

[0167] For example, after step 505, when the trigger scenario is set to the time when a seller publishes / modifies a product, the shared representation vector of each product to be trained can be input into the pricing regression head of the multi-task head. The structure of the pricing regression head is: FC(ReLU)->FC(ReLU)->Linear(2). Based on the pricing regression head, the shared representation vector of each product to be trained is processed to obtain the pricing suggestion in the prediction information of each product to be trained.

[0168] 507. A supply and demand matching head based on a multi-task head processes the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the prediction information of the product to be trained.

[0169] For example, after step 505, when setting the trigger scenario to a buyer searching for / browsing products, the shared representation vector and the user intent vector from the full feature vector of each product to be trained can be input into the supply and demand matching head of the multi-task head. Based on this supply and demand matching head, the shared representation vector and the user intent vector from the full feature vector of each product to be trained are fused to obtain the comprehensive transaction probability in the prediction information of each product to be trained, including the comprehensive transaction probability representing "person-product matching degree + price competitiveness + trust degree".

[0170] In one possible implementation, step 507 includes: based on the supply and demand matching head, the similarity Sim between the shared representation vector H of each product to be trained and the user intent vector U (e.g., the user intent vector after BERT encoding of the user's current search term) can be calculated using a cosine similarity calculation formula, for example, Sim=Cosine Similarity(H,U). And based on a preset calculation rule, a reasonableness score is calculated for the pricing suggestions and current price in the prediction information of each product to be trained, for example, reasonableness score = -|(seller's current listed price - (price_low + price_high) / 2)| / ((price_low + price_high) / 2), where price_low and price_high are the suggested minimum price and suggested maximum price in the pricing suggestions, respectively. For each product to be trained, the features Sim (similarity score), Rationality_Score (reasonability score), and Seller_Rating (seller rating in the full feature vector) are concatenated. The concatenated features are then input into the fully connected network of the supply and demand matching head for final scoring. For example, the comprehensive transaction probability Matching_Score of each product to be trained is predicted by calculating the following formula: Matching_Score=Sigmoid(FC(concat(Sim, Rationality_Score, Seller_Rating))).

[0171] By calculating the similarity between the shared representation vector and the user intent vector, precise optimization of supply and demand matching is achieved. For example, when a user searches for "high-performance tablet at a reasonable price," the subsequent multi-task model can combine semantic information such as "screen intact" and "complete accessories" in the product description to prioritize recommending products with high matching degrees, thereby significantly improving the conversion rate of the recommendation system.

[0172] 508. A trend prediction head based on a multi-task head processes the shared representation vector to obtain the demand change trend in the prediction information of the product to be trained.

[0173] For example, after step 505, when the trigger scenario is set to a timed task (e.g., 2 AM), the shared representation vector of each product to be trained can be input into the trend prediction head of the multi-task head. Based on this trend prediction head, feature processing is performed on the shared representation vector of each product to be trained to obtain the demand change trend in the prediction information of each product to be trained.

[0174] For example, the trend prediction model in the deployed trend prediction header is invoked, and the shared representation vector of each product to be trained is passed in to obtain the demand change trend in the prediction information, including: {"item_id":"12345","predicted_demand_change":0.16,"confidence": 0.87}, which indicates that the demand for each product to be trained will increase by 16% in the future, and the confidence level of the returned result is 87%.

[0175] The multi-task learning framework enables the pricing, matching, and prediction tasks to mutually reinforce each other; accurate pricing prediction provides "reasonableness" features for matching; and matching and transaction data, in turn, provide high-quality training signals for the pricing and prediction models, forming a positive feedback loop.

[0176] In one possible implementation, step 508 includes: acquiring the historical sequence of each product to be trained, including the historical price and trading volume of the category to which each product belongs in the historical time period, and performing feature encoding processing on the historical sequence of each product to be trained based on the encoder to obtain the time-series context (such as volatility, periodicity). For example, the price sequence of the past 7 days is encoded into a 128-dimensional trend representation vector corresponding to the time-series context. Through the trend prediction head, the time-series context of each product to be trained is connected with the shared representation vector. For example, by using ΔDemand=Linear(FC(concat(H, C_seq))), the trend representation vector C_seq (128 dimensions) and the shared representation vector H (512 dimensions) are concatenated into a 640-dimensional vector to form a more comprehensive input feature, thereby enhancing the model's ability to perceive market dynamics. The input feature is then processed to predict the demand change trend in the prediction information of each product to be trained.

[0177] By introducing historical trend representation vectors, the trend forecasting head can more accurately capture market dynamics (such as the price drop of older models due to the release of new models). For example, it can proactively push alerts to sellers such as "Demand is predicted to drop by 20%, and price reductions are recommended," thereby reducing the risk of unsold inventory.

[0178] 509. Determine the pricing loss function based on the actual pricing results of the product to be trained and the pricing suggestions in the prediction information.

[0179] For example, the device obtains the actual pricing result of each product to be trained, and calculates the pricing loss function based on the actual pricing result and the pricing suggestion in the prediction information for each product to be trained; for example, L_price=Huber Loss(price_low, y1_true)+HuberLoss(price_high, y2_true), where price_low is the suggested minimum price in the prediction information, price_high is the suggested maximum price in the prediction information; y1_true is the minimum price in the actual pricing result; and y2_true is the maximum price in the actual pricing result.

[0180] 510. Determine the matching loss function based on the actual overall transaction probability of the product to be trained and the overall transaction probability in the predicted information.

[0181] For example, the device obtains the actual overall transaction probability of each product to be trained, and calculates the actual overall transaction probability and the overall transaction probability in the prediction information for each product to be trained based on the matching loss function calculation formula to obtain the matching loss function.

[0182] 511. Determine the trend loss function based on the actual demand change trend of the product to be trained and the demand change trend in the forecast information.

[0183] For example, the device acquires the actual demand change trend of each product to be trained, and calculates the trend loss function based on the trend loss function calculation formula, such as the mean squared error loss calculation formula, for the actual demand change trend and the demand change trend in the predicted information of each product to be trained.

[0184] 512. Determine the total loss function based on the pricing loss function, matching loss function, and trend loss function.

[0185] For example, the obtained pricing loss function L_price, matching loss function L_matching, and trend loss function L_trend are weighted and combined to obtain the total loss function L_total; for example, the total loss function is: L_total=α*L_price+β*L_matching+γ*L_trend, and the optimal weights α, β, and γ are determined by hyperparameter search.

[0186] 513. Train the initial model based on the total loss function to obtain the multi-task model.

[0187] For example, based on the total loss function, the parameters of the initial model are optimized by an optimizer to reduce the total loss function, resulting in a multi-task model for information push processing.

[0188] Based on the joint training mechanism, by weighted summation of the loss functions of multiple tasks, the shared encoder learns general features that are beneficial to all tasks. For example, the output of the pricing model provides "reasonableness" features for the matching model, and the transaction data of the matching model feeds back into the training signal of the initial model, thereby improving the model performance.

[0189] In this embodiment, based on the above embodiments, the multi-task joint training mechanism enables pricing, matching, and prediction tasks to promote each other. For example, the transaction data of the matching model feeds back into the model training, forming a positive feedback loop and improving the overall model performance and robustness.

[0190] Figure 7 This is a schematic diagram of the structure of an information push device provided in an embodiment of this application, as shown below. Figure 7 As shown, the device includes:

[0191] The extraction module 601 is used to acquire multi-source data of the product; and to extract features from the multi-source data of the product to obtain the full feature vector of the product.

[0192] The encoding module 602 is used for a shared encoder based on a multi-task model to perform shared encoding processing on the full feature vector of the product to obtain the shared representation vector of the product; wherein, the shared representation vector represents the contextual semantic information of the product;

[0193] The prediction module 603 is used to process the shared representation vector of the product based on the multi-task head of the multi-task model to obtain the prediction information of the product in the future time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend;

[0194] The push module 604 is used to push information on the prediction information of goods.

[0195] In one possible implementation, the extraction module 601 is used to: extract features from the internal data of the multi-source data of the product to obtain the internal feature vector of the product; wherein the internal data includes static attributes of the product, dynamic data of the product, transaction data and user data; extract features from the external data of the multi-source data of the product to obtain the external feature vector of the product; wherein the external data includes platform data, public opinion data and market data; and concatenate the internal feature vector and the external feature vector of the product to obtain the full feature vector of the product.

[0196] In one possible implementation, the internal feature vector includes structured feature vectors and unstructured feature vectors; the structured feature vectors include numerical features, categorical features, and time series features; the unstructured feature vectors include text semantic vectors and user intent vectors; and / or, the external feature vectors include key ratio features, public opinion features, and quadruples.

[0197] In one possible implementation, the prediction module 603 is specifically used for: processing the shared representation vector based on a pricing regression head with a multi-task head to obtain pricing suggestions in the product's prediction information; and / or processing the user intent vector in the shared representation vector and the full feature vector based on a supply-demand matching head with a multi-task head to obtain the comprehensive transaction probability in the product's prediction information; and / or processing the shared representation vector based on a trend prediction head with a multi-task head to obtain the demand change trend in the product's prediction information.

[0198] In one possible implementation, the prediction module 603 is specifically used to: determine the similarity between the shared representation vector and the user intent vector in the full feature vector based on the supply and demand matching head; calculate the pricing suggestion and current price in the product prediction information to obtain a reasonableness score; and fuse the similarity, reasonableness score and seller positive review rate in the full feature vector based on the fully connected network of the supply and demand matching head to obtain the comprehensive transaction probability in the product prediction information.

[0199] In one possible implementation, the prediction module 603 is specifically used to: encode the historical sequence of the commodity to obtain a temporal context; wherein the historical sequence includes the historical price and transaction volume of the commodity category in a historical time period; and based on the trend prediction head, perform fusion prediction on the temporal context and the shared representation vector to obtain the demand change trend in the prediction information of the commodity.

[0200] In one possible implementation, the push module 604 is specifically used to: if it is determined that the change value corresponding to the demand change trend represented by the prediction information is greater than a preset threshold, then generate a push message for the product based on the product's prediction information; wherein the push message includes the demand change trend and a jump link; and perform asynchronous push processing on the push message.

[0201] The apparatus in this embodiment can execute the technical solutions in the above method. Its specific implementation process and technical principles are the same, and will not be repeated here.

[0202] Figure 8 This application provides a schematic diagram of the structure of a model training device for information push, as shown in the embodiments of this application. Figure 8 As shown, the device includes:

[0203] The extraction module 701 is used to acquire multi-source data of at least one product to be trained; and to extract features from the multi-source data of the product to be trained to obtain the full feature vector of the product to be trained.

[0204] The encoding module 702 is used to perform shared encoding processing on the full feature vector of the product to be trained based on the initial model, so as to obtain the shared representation vector of the product to be trained; wherein, the shared representation vector represents the contextual semantic information of the product to be trained.

[0205] The prediction module 703 is used to process the shared representation vector of the product to be trained based on the multi-task head of the initial model to obtain the prediction information of the product to be trained within a preset time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend.

[0206] The training module 704 is used to train the initial model based on the prediction information of the product to be trained, so as to obtain a multi-task model. The multi-task model is used to process multi-source data of products such as those in information push devices to obtain prediction information of the products.

[0207] In one possible implementation, the extraction module 701 is specifically used for: extracting features from the internal data in the multi-source data of the product to obtain the internal feature vector of the product to be trained; wherein the internal data includes static attributes of the product, dynamic data of the product, transaction data, and user data; extracting features from the external data in the multi-source data of the product to obtain the external feature vector of the product to be trained; wherein the external data includes platform data, public opinion data, and market data; and concatenating the internal feature vector and the external feature vector of the product to be trained to obtain the full feature vector of the product to be trained.

[0208] In one possible implementation, the internal feature vector includes structured feature vectors and unstructured feature vectors; the structured feature vectors include numerical features, categorical features, and time series features; the unstructured feature vectors include text semantic vectors and user intent vectors; and / or, the external feature vectors include key ratio features, public opinion features, and quadruples.

[0209] In one possible implementation, the prediction module 703 is specifically used for: processing the shared representation vector based on a pricing regression head with a multi-task head to obtain pricing suggestions in the prediction information of the product to be trained; and / or processing the user intent vector in the shared representation vector and the full feature vector based on a supply-demand matching head with a multi-task head to obtain the comprehensive transaction probability in the prediction information of the product to be trained; and / or processing the shared representation vector based on a trend prediction head with a multi-task head to obtain the demand change trend in the prediction information of the product to be trained.

[0210] In one possible implementation, the prediction module 703 is specifically used for: determining the similarity between the shared representation vector and the user intent vector in the full feature vector based on the supply and demand matching head; calculating the pricing suggestion and current price in the prediction information of the product to be trained to obtain a reasonableness score; and fusing the similarity, reasonableness score and seller positive review rate in the full feature vector based on the fully connected network of the supply and demand matching head to obtain the comprehensive transaction probability in the prediction information of the product to be trained.

[0211] In one possible implementation, the prediction module 703 is specifically used to: encode the historical sequence of the product to be trained to obtain a temporal context; wherein the historical sequence includes the historical price and transaction volume of the product category in the historical time period; and based on the trend prediction head, perform fusion prediction on the temporal context and the shared representation vector to obtain the demand change trend in the prediction information of the product to be trained.

[0212] In one possible implementation, the training module 704 is specifically used to: determine a pricing loss function based on the actual pricing results of the product to be trained and the pricing suggestions in the prediction information; determine a matching loss function based on the actual comprehensive transaction probability of the product to be trained and the comprehensive transaction probability in the prediction information; determine a trend loss function based on the actual demand change trend of the product to be trained and the demand change trend in the prediction information; determine a total loss function based on the pricing loss function, the matching loss function, and the trend loss function; and train the initial model based on the total loss function to obtain a multi-task model.

[0213] The apparatus in this embodiment can execute the technical solutions in the above method. Its specific implementation process and technical principles are the same, and will not be repeated here.

[0214] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 9 As shown, the electronic device includes a memory 801 and a processor 802; the memory 801 is used to store executable instructions of the processor 802. The processor 802 is configured to execute the method provided in the above embodiments. The electronic device also includes a receiver 803 and a transmitter 804. The receiver 803 is used to receive instructions and data sent by other devices, and the transmitter 804 is used to send instructions and data to external devices. The specific implementation process of the processor can be found in the above method embodiments, and its implementation principle and technical effects are similar; therefore, it will not be repeated here.

[0215] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed on a computer, cause the computer to perform the technical solutions described above.

[0216] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. The readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in a device.

[0217] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when the at least one processor executes the computer program, it can implement the technical solutions in the above embodiments.

[0218] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as magnetic disks or optical disks.

[0219] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. An information push method, characterized in that, include: Acquire multi-source data of the product; and extract features from the multi-source data of the product to obtain the full feature vector of the product; A shared encoder based on a multi-task model performs shared encoding processing on the full feature vector of the product to obtain a shared representation vector of the product; wherein, the shared representation vector represents the contextual semantic information of the product. Based on the multi-task head of the multi-task model, the shared representation vector of the product is processed to obtain the prediction information of the product in the future time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend; The predicted information for the product is pushed to the system.

2. The method according to claim 1, characterized in that, The step of extracting features from the multi-source data of the product to obtain the full feature vector of the product includes: Feature extraction is performed on the internal data of the product from the multi-source data to obtain the internal feature vector of the product; wherein, the internal data includes static attributes of the product, dynamic data of the product, transaction data, and user data; Feature extraction is performed on the external data in the multi-source data of the product to obtain the external feature vector of the product; wherein, the external data includes platform data, public opinion data and market data; The internal and external feature vectors of the product are concatenated to obtain the full feature vector of the product.

3. The method according to claim 2, characterized in that, The internal feature vectors include structured feature vectors and unstructured feature vectors; the structured feature vectors include numerical features, category features, and time series features; the unstructured feature vectors include text semantic vectors and user intent vectors. And / or, the external feature vector includes key ratio features, public opinion features, and quadruples.

4. The method according to claim 1, characterized in that, The multi-task head based on the multi-task model processes the shared representation vector of the product to obtain prediction information of the product within a future time period, including: Based on the pricing regression head of the multi-task head, the shared representation vector is processed to obtain pricing suggestions in the prediction information of the product; And / or, based on the supply and demand matching head of the multi-task head, process the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the prediction information of the product; And / or, based on the trend prediction head of the multi-task head, the shared representation vector is processed to obtain the demand change trend in the prediction information of the commodity.

5. The method according to claim 4, characterized in that, The supply and demand matching head based on the multi-task head processes the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the product's prediction information, including: Based on the supply and demand matching head, the similarity between the shared representation vector and the user intent vector in the full feature vector is determined; A reasonableness score is obtained by calculating the pricing recommendations in the forecast information of the product and the current price; Based on the fully connected network of the supply and demand matching head, the similarity, the rationality score, and the seller's positive review rate in the full feature vector are fused to obtain the comprehensive transaction probability in the product's prediction information.

6. The method according to claim 4, characterized in that, The trend prediction head based on the multi-task head processes the shared representation vector to obtain the demand change trend in the predicted information of the commodity, including: The historical sequence of the product is encoded to obtain a temporal context; wherein, the historical sequence includes the historical price and transaction volume of the product category within a historical time period; Based on the trend prediction head, the temporal context and the shared representation vector are fused and predicted to obtain the demand change trend in the predicted information of the commodity.

7. The method according to any one of claims 1-6, characterized in that, The information push processing of the predicted information of the product includes: If it is determined that the change value corresponding to the demand change trend represented by the predicted information is greater than a preset threshold, then a push message for the product is generated based on the predicted information of the product; wherein, the push message includes the demand change trend and a jump link; The push message is processed asynchronously.

8. A model training method for information push applications, characterized in that, include: Obtain multi-source data of at least one product to be trained; and extract features from the multi-source data of the product to be trained to obtain the full feature vector of the product to be trained. The shared encoder based on the initial model performs shared encoding processing on the full feature vector of the product to be trained to obtain the shared representation vector of the product to be trained; wherein, the shared representation vector represents the contextual semantic information of the product to be trained. Based on the multi-task head of the initial model, the shared representation vector of the product to be trained is processed to obtain the prediction information of the product to be trained within a preset time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend; Based on the prediction information of the product to be trained, the initial model is trained to obtain a multi-task model; wherein, the multi-task model is used to process the multi-source data of the product as described in any one of claims 1-7 to obtain the prediction information of the product.

9. The method according to claim 8, characterized in that, The step of extracting features from the multi-source data of the product to be trained, to obtain the full feature vector of the product to be trained, includes: Feature extraction is performed on the internal data in the multi-source data of the product to obtain the internal feature vector of the product to be trained; wherein, the internal data includes static attributes of the product, dynamic data of the product, transaction data, and user data; Feature extraction is performed on the external data in the multi-source data of the product to obtain the external feature vector of the product to be trained; wherein, the external data includes platform data, public opinion data and market data; The internal and external feature vectors of the product to be trained are concatenated to obtain the full feature vector of the product to be trained.

10. The method according to claim 9, characterized in that, The internal feature vectors include structured feature vectors and unstructured feature vectors; the structured feature vectors include numerical features, category features, and time series features; the unstructured feature vectors include text semantic vectors and user intent vectors. And / or, the external feature vector includes key ratio features, public opinion features, and quadruples.

11. The method according to claim 8, characterized in that, The multi-task head based on the initial model processes the shared representation vector of the product to be trained to obtain prediction information of the product to be trained within a preset time period, including: Based on the pricing regression head of the multi-task head, the shared representation vector is processed to obtain pricing suggestions in the prediction information of the product to be trained; And / or, based on the supply and demand matching head of the multi-task head, process the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the prediction information of the product to be trained; And / or, based on the trend prediction head of the multi-task head, the shared representation vector is processed to obtain the demand change trend in the prediction information of the product to be trained.

12. The method according to claim 11, characterized in that, The supply-demand matching head based on the multi-task head processes the user intent vector in the shared representation vector and the full feature vector to obtain the comprehensive transaction probability in the prediction information of the product to be trained, including: Based on the supply and demand matching head, the similarity between the shared representation vector and the user intent vector in the full feature vector is determined; A reasonableness score is obtained by calculating the pricing suggestions and current prices in the prediction information of the product to be trained. Based on the fully connected network of the supply and demand matching head, the similarity, the rationality score and the seller's positive review rate in the full feature vector are fused to obtain the comprehensive transaction probability in the prediction information of the product to be trained.

13. The method according to claim 11, characterized in that, The trend prediction head based on the multi-task head processes the shared representation vector to obtain the demand change trend in the prediction information of the product to be trained, including: The historical sequence of the product to be trained is encoded to obtain the temporal context; wherein, the historical sequence includes the historical price and transaction volume of the product category in the historical time period; Based on the trend prediction head, the temporal context and the shared representation vector are fused and predicted to obtain the demand change trend in the prediction information of the product to be trained.

14. The method according to any one of claims 8-13, characterized in that, The step of training the initial model based on the prediction information of the product to be trained to obtain a multi-task model includes: The pricing loss function is determined based on the actual pricing results of the product to be trained and the pricing suggestions in the prediction information; The matching loss function is determined based on the actual overall transaction probability of the product to be trained and the overall transaction probability in the predicted information. The trend loss function is determined based on the actual demand change trend of the product to be trained and the demand change trend in the predicted information; The total loss function is determined based on the pricing loss function, the matching loss function, and the trend loss function; The initial model is trained based on the total loss function to obtain the multi-task model.

15. An information push device, characterized in that, include: The extraction module is used to acquire multi-source data of the product; and to extract features from the multi-source data of the product to obtain the full feature vector of the product. The encoding module is used to perform shared encoding processing on the full feature vector of the product based on a multi-task model to obtain the shared representation vector of the product; wherein the shared representation vector represents the contextual semantic information of the product; The prediction module is used to process the shared representation vector of the product based on the multi-task head of the multi-task model to obtain prediction information of the product in the future time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend; The push module is used to push information about the predicted information of the product.

16. A model training device for information push, characterized in that, include: An extraction module is used to acquire multi-source data of at least one product to be trained; and to extract features from the multi-source data of the product to be trained to obtain the full feature vector of the product to be trained. The encoding module is used to perform shared encoding processing on the full feature vector of the product to be trained based on the shared encoder of the initial model, so as to obtain the shared representation vector of the product to be trained; wherein the shared representation vector represents the contextual semantic information of the product to be trained. The prediction module is used to process the shared representation vector of the product to be trained based on the multi-task head of the initial model to obtain prediction information of the product to be trained within a preset time period; wherein, the prediction information includes at least one of the following: pricing suggestions, comprehensive transaction probability and demand change trend; The training module is used to train the initial model based on the prediction information of the product to be trained, thereby obtaining a multi-task model; wherein the multi-task model is used to process the multi-source data of the product as described in claim 15 to obtain the prediction information of the product.

17. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-14.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-14.

19. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-14.

Images