A data recall method, device, equipment, medium and program product

Abstract

Embodiments of the present application provide a data recall method, device, equipment, medium and program product, which are used to improve the accuracy of data recall. It can be applied to the fields of artificial intelligence, data recall recommendation, etc. It includes: obtaining object behavior data and commodity feature data, the object behavior data including advertisement domain behavior data and natural flow behavior data of a target object, wherein the advertisement domain behavior data is historical behavior data and real-time behavior data of the target object on advertising information, and the natural flow behavior data is historical behavior data and real-time behavior data of the target object on actively accessed information; performing feature extraction on the object behavior data and the commodity feature data respectively to obtain object behavior feature representation and commodity feature representation; performing commodity recall based on the object behavior feature representation and the commodity feature representation to obtain a recall commodity set; and determining a recall advertisement set based on the recall commodity set.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a data retrieval method, apparatus, device, medium, and program product. Background Technology

[0002] With the rapid development of internet technology and the widespread adoption of mobile devices, the e-commerce industry is booming. Merchants widely advertise through internet advertising platforms to promote their products and increase sales. However, the core problem that advertising recommendation systems need to solve is how to accurately recommend ads to the target audience on the platform, thereby achieving efficient conversions.

[0003] Current ad recall models typically employ a dual-tower structure, modeling object features and ad features separately. Embedding vectors are generated from the platform operator tower and the ad tower, respectively. During online inference, the dot product of these two vectors is used to calculate similarity, thereby recalling high-scoring ads. However, this approach relies heavily on behavioral data from the ad domain (such as click and conversion records of platform operators), which prevents existing recall systems from fully understanding the interests and conversion tendencies of platform operators, thus limiting the effectiveness of ad recall.

[0004] Therefore, there is an urgent need to provide a solution that can improve data recall. Summary of the Invention

[0005] This application provides a data retrieval method, apparatus, device, medium, and program product to improve the accuracy of data retrieval.

[0006] In view of this, this application provides a data recall method, comprising: acquiring object behavior data and product feature data, wherein the object behavior data includes advertising domain behavior data and natural traffic behavior data of the target object, wherein the advertising domain behavior data is the target object's historical behavior data and real-time behavior data on advertising information, and the natural traffic behavior data is the target object's historical behavior data and real-time behavior data on active access information; extracting features from the object behavior data and the product feature data respectively to obtain object behavior feature representation and product feature representation; recalling products based on the object behavior feature representation and the product feature representation to obtain a recalled product set; and determining a recalled advertising set based on the recalled product set.

[0007] Another aspect of this application provides a data retrieval device, including: an acquisition module for acquiring object behavior data and product feature data, wherein the object behavior data includes advertising domain behavior data and natural traffic behavior data of the target object, wherein the advertising domain behavior data is the historical behavior data and real-time behavior data of the target object towards advertising information, and the natural traffic behavior data is the historical behavior data and real-time behavior data of the target object towards actively accessing information;

[0008] The processing module is used to extract features from the object behavior data and the product feature data respectively to obtain object behavior feature representation and product feature representation; to perform product recall based on the object behavior feature representation and the product feature representation to obtain a recall product set; and to determine a recall advertisement set based on the recall product set.

[0009] In one possible design, in another implementation of another aspect of the embodiments of this application, the processing module is used to call the feature processing layer of the recall model to perform feature embedding on the object behavior data and the product feature data to obtain object behavior feature vector and product feature vector, wherein the recall model includes a feature processing layer, a first feature extraction network and a second feature extraction network.

[0010] The first feature extraction network of the recall model is invoked to extract features from the object's behavioral feature vector to obtain a representation of the object's behavioral features.

[0011] The second feature extraction network of the recall model is invoked to extract features from the product feature vector to obtain the product feature representation.

[0012] In one possible design, in another implementation of another aspect of the embodiments of this application, the processing module is used to perform a dot product between the object behavior feature representation and the product feature representation to obtain the estimated conversion rate of each product.

[0013] Based on the estimated conversion rate, products are recalled in a top-N manner to obtain the recalled product set.

[0014] In one possible design, in another implementation of another aspect of the embodiments of this application, the acquisition module is used to acquire an initial recall model and training samples. The initial recall model includes an initial feature processing layer, an initial first feature extraction network, an initial second feature extraction network, an initial third feature extraction network, and an output layer. The training samples include training object data, training product data, and conversion tags. The training object data includes advertising domain behavior data and organic traffic behavior data of the training objects. The training product data includes advertising domain product data and organic traffic product data. The conversion tags are used to indicate whether the product has been successfully converted.

[0015] The processing module is used to call the initial feature processing layer to embed features into the training object data and the training product data to obtain training object behavior feature vectors and training product feature vectors. The training product feature vectors include advertising domain product feature vectors and organic traffic product feature vectors. It then calls the initial first feature extraction network to extract features from the training object behavior feature vectors to obtain a training object feature representation. Next, it calls the initial second feature extraction network to extract features from the advertising domain product feature vectors to obtain an advertising domain product feature representation. Finally, it calls the initial third feature extraction network to extract features from the organic traffic product feature vectors to obtain an organic traffic product feature representation. The module then performs a dot product between the training object feature representation and the advertising domain product feature representation to obtain a first predicted conversion rate for the advertising domain product. It also performs a dot product between the training object feature representation and the organic traffic product feature representation to obtain a second predicted conversion rate for the organic traffic product. Based on the first predicted conversion rate and the conversion tag, it calculates a first loss value and a second loss value. Finally, it trains the initial recall model based on the first loss value and the second loss value to obtain the recall model.

[0016] In one possible design, in another implementation of another aspect of the embodiments of this application, the initial second feature extraction network and the initial third feature extraction network have the same network structure; the processing module is used to transfer the network parameters of the initial third feature extraction network to the initial second feature extraction network by means of exponential moving average.

[0017] In one possible design, in another implementation of another aspect of the embodiments of this application, the acquisition module is used to acquire object behavior data and the product conversion data corresponding to the object behavior data;

[0018] When the product conversion data indicates that the first product has been successfully converted, construct a positive sample using the object's behavior data and the first product, and set the conversion label to "product conversion successful";

[0019] When the conversion data of the product indicates that the first product has not been successfully converted, a negative sample is constructed from the object's behavior data and the first product, and the conversion label is set to "product not successfully converted".

[0020] In one possible design, in another implementation of another aspect of the embodiments of this application, the acquisition module is used to acquire the category label of the first product, the category label being used to indicate different factors for the conversion failure of the first product;

[0021] When the category label indicates that the first product is a product that the object has viewed or clicked, a first-class secondary conversion label is added to the negative sample. The first-class secondary conversion label is used to indicate that the conversion rate of the first product is lower than a first threshold.

[0022] When the category label indicates that the first product is an unconverted product under the same sales object, a second-level conversion label is added to the negative sample. The second-level conversion label is used to indicate that the conversion rate of the first product is not lower than the first threshold and is lower than the second threshold.

[0023] When the category label indicates that the first product is an undiscovered, unconverted product, a third-class secondary conversion label is added to the negative sample. The third-class secondary conversion label is used to indicate that the conversion rate of the first product is not lower than the second threshold.

[0024] This application also provides a computer device, including: a memory, a processor, and a bus system;

[0025] The memory is used to store programs;

[0026] The processor is used to execute programs in memory, and the processor is used to execute the methods mentioned above according to the instructions in the program code;

[0027] Bus systems are used to connect memory and processor to enable communication between them.

[0028] Another aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0029] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the above aspects.

[0030] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: introducing behavioral data from natural traffic provides more dimensional behavioral information, enabling a full exploration of the interests and preferences of the platform's users, and enriching the feature information; at the same time, aggregating data through the product dimension alleviates the sparsity of advertising domain conversion data, thereby improving the richness and reliability of advertising recall. Attached Figure Description

[0031] Figure 1 A schematic diagram of an architecture for existing ad recall;

[0032] Figure 2 This is a schematic diagram of an application system architecture for data retrieval in an embodiment of this application;

[0033] Figure 3 This is a schematic diagram of a data retrieval process in an embodiment of this application;

[0034] Figure 4 This is a schematic diagram of the recall model in the embodiments of this application;

[0035] Figure 5 This is a schematic diagram of a training process for the recall model in an embodiment of this application;

[0036] Figure 6 This is a schematic diagram of one embodiment of data recall in this application.

[0037] Figure 7 This is a schematic diagram of one embodiment of the data retrieval device in this application.

[0038] Figure 8 This is a schematic diagram of one embodiment of the server in this application;

[0039] Figure 9 This is a schematic diagram of one embodiment of the terminal device in this application. Detailed Implementation

[0040] This application provides a data retrieval method, apparatus, device, medium, and program product to improve the accuracy of data retrieval.

[0041] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0042] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0043] With the rapid development of internet technology and the widespread adoption of mobile devices, the e-commerce industry is booming. Merchants widely advertise through internet advertising platforms to promote their products and increase sales. However, the core problem that advertising recommendation systems need to solve is how to accurately recommend ads to the target audience on the platform, thereby achieving efficient conversions.

[0044] Current recall models typically use a two-tower structure, the structure of which can be as follows: Figure 1 The proposed method models object features and ad features separately. Embedding vectors are generated using the platform operation object tower and ad tower respectively. During online inference, the dot product of these two vectors is used to calculate similarity, thereby recalling high-scoring ads. However, this method primarily relies on behavioral data from the ad domain (such as click and conversion records of platform operation objects on ads). This prevents existing recall systems from fully exploring the interests and conversion tendencies of platform operation objects, limiting the effectiveness of ad recall. Therefore, there is an urgent need to provide a solution that can improve data recall effectiveness.

[0045] To address the aforementioned technical issues, this application provides the following technical solution: Acquiring object behavior data and product feature data, wherein the object behavior data includes advertising domain behavior data and organic traffic behavior data of the target object, whereby the advertising domain behavior data comprises the target object's historical and real-time behavior data regarding advertising information, and the organic traffic behavior data comprises the target object's historical and real-time behavior data regarding active access information; extracting features from the object behavior data and the product feature data respectively to obtain object behavior feature representations and product feature representations; performing product recall based on the object behavior feature representations and product feature representations to obtain a recalled product set; and determining a recalled advertising set based on the recalled product set. This introduction of organic traffic behavior data provides more multi-dimensional behavioral information, enabling a full exploration of the interests and preferences of platform users and enriching feature information; simultaneously, aggregating data through the product dimension alleviates the sparsity of advertising domain conversion data, thereby improving the richness and reliability of advertising recall.

[0046] For ease of understanding, some terms used in this application are explained below.

[0047] Advertising domain data refers to behavioral data of platform users acquired through paid advertising campaigns. This data originates from ad clicks, impressions, etc., and enters the platform or application through the advertiser's placement strategy. The sources of advertising domain data include the following: 1. Search advertising. Such as bidding rankings on platforms like XX. 2. Display advertising. Such as feed ads, banner ads, video ads, etc. 3. Social media advertising. Such as ads published on various social media platforms. 4. Programmatic advertising. Ads placed through real-time bidding platforms or demand-side platforms. 5. Affiliate advertising. Traffic generated through advertising networks.

[0048] Based on the sources of the aforementioned advertising domain data, its characteristics can be summarized as follows: 1. Passivity: The behavior of platform users is usually triggered by advertising and may not reflect genuine interest. 2. Paid Acquisition: Advertising domain traffic requires the platform to pay advertising fees to acquire users. 3. High Conversion Orientation: Advertising traffic is typically aimed at direct conversion (such as purchase, registration, download, etc.). 4. Short-Term Volatility: Advertising traffic may decrease rapidly due to the suspension of promotional activities.

[0049] Organic traffic data refers to data generated when users actively access a platform or application through non-paid channels. This type of traffic typically stems from users' interest in content, products, or services, without the direct impact of advertising. Sources of organic traffic data usually include: 1. Direct access data: Users access the platform by directly entering a URL, clicking on a bookmark, or launching the application. 2. Organic search engine data: Users enter the website through organic search engine rankings, rather than clicking on paid ads. 3. External links: Users access the platform through organic links from other websites, social media, or content recommendation systems. 4. Content attraction: Users actively access content (such as blogs, articles, videos, etc.) or participate in community interactions on the platform. 5. User retention: For example, returning users due to prior knowledge or trust in the platform.

[0050] Based on the above source descriptions, common characteristics of organic traffic data can be summarized as follows: 1. Proactive: The behavior of users on the platform is spontaneous and proactive, and their access motivation is closer to their genuine interests. 2. No Paid Promotion: Organic traffic data does not rely on paid advertising. 3. High-Quality Users: Compared to users acquired through paid advertising, organic traffic users are likely to have greater stickiness and loyalty. 4. Content- or Brand-Driven: Organic traffic is usually driven by high-quality content, product appeal, or brand influence.

[0051] Organic traffic data typically includes the following records: 1. Platform user behavior data, such as visit duration, clicks, page views, and bounce rate. 2. Source data, such as visit source (search engine, direct access, external links) and device information (mobile, PC). 3. Basic platform user attributes, such as the platform user's location, age, and gender. 4. Conversion data, i.e., the platform user's conversion behavior towards products (such as registration, purchase, download).

[0052] As described above, organic traffic data can more accurately reflect the interests and behavioral patterns of platform users, thus allowing for the optimization of data retrieval algorithms through the analysis of organic traffic data.

[0053] The technical solutions of this application and their effects are described below through several exemplary embodiments. It should be noted that the following embodiments can be referenced, borrowed from, or combined with each other. Identical terms, similar features, and similar implementation steps in different embodiments will not be repeated.

[0054] The data retrieval method provided in this application embodiment can be applied to, for example, Figure 2 The system shown includes a terminal 100, a server 200, a database 300, and a network 400. The terminal 100, server 200, and database 300 are connected via the network 400. The database 300 is used for product data and historical behavior data of the objects. Figure 2 The number of terminals, servers, and databases in the system shown is merely an example. For instance, there may be multiple terminals, servers, and databases. This application does not limit the number of terminals, servers, and databases.

[0055] The aforementioned system is used in the fields of artificial intelligence and data recall recommendation, and this application embodiment does not limit its application to these areas. During the ad recall process, server 200 can obtain object behavior data of the target object and global product feature data; then, it calls the recall model to perform feature processing on the object behavior data and product feature data to obtain corresponding object behavior feature representations and product feature representations; finally, it generates a product recall set based on the object behavior feature representations and product feature representations; then, it determines the ad recall set based on the product recall set; and finally, it outputs the ad recall set. It should be noted that the object behavior data and product feature data are authorized by the platform operation object or fully authorized by all parties.

[0056] In this process, terminal 100 communicates with server 200 via a network. Database 300 can be integrated onto server 200 or located in the cloud or on another server. The data retrieval process can involve interaction between terminal 100 and server 200. For example, a platform user generates a short video request or other information request through terminal 100, which then sends this request to server 200. Server 200 acquires the target object's behavior data and global product feature data. It then uses a retrieval model to process the object behavior data and product feature data to obtain corresponding object behavior feature representations and product feature representations. Finally, it creates a product retrieval set based on these representations, determines an advertising retrieval set, and finally feeds the advertising retrieval set back to terminal 100.

[0057] Terminal 100 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, smart voice interaction device, smart home appliance, or in-vehicle terminal, but is not limited to these. Server 200 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery network (CDN), big data, and artificial intelligence platforms.

[0058] In short, a database can be viewed as an electronic filing cabinet—a place to store electronic files. Platform operators can perform operations such as adding, querying, updating, and deleting data in the files. A "database" is a collection of data stored together in a certain way, shared by multiple platform operators, with minimal redundancy, and independent of application programs. A Database Management System (DBMS) is a computer software system designed to manage databases, generally possessing basic functions such as storage, retrieval, security, and backup. DBMSs can be classified according to the database model they support, such as relational or Extensible Markup Language (XML); or according to the type of computer they support, such as server clusters or mobile phones; or according to the query language used, such as Structured Query Language (SQL) or XQuery; or according to performance priorities, such as maximum scale or maximum operating speed; or other classification methods. Regardless of the classification method used, some DBMSs can cross categories, for example, simultaneously supporting multiple query languages.

[0059] It is understood that in the specific implementation of this application, data related to the object's behavior data, advertising domain product data, and natural traffic product data are involved. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the platform operator is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0060] Based on the above description, the following is an example: Figure 3 The system flow shown and Figure 4 The recall model architecture shown illustrates the solution provided in this application.

[0061] like Figure 3 As shown, the specific process can be as follows:

[0062] The system acquires ad domain data and organic traffic data. Then, based on the ad domain data, it constructs positive and negative samples according to whether a product is converted. Similarly, it constructs positive and negative samples based on the organic traffic data. Simultaneously, it performs feature processing on the organic traffic data to obtain corresponding object behavior features. Then, it uses an ad domain product conversion rate prediction network to learn features from the positive and negative samples constructed from the ad domain data to recall the corresponding products. Similarly, it uses an organic traffic product conversion rate prediction network to learn features from the positive and negative samples constructed from the organic traffic data to recall the corresponding products. In online application, for the top-N products recalled by the ad domain product conversion rate prediction network, each product recalls M ads, resulting in a total of M*N ads being recalled.

[0063] Based on the above description and this Figure 4 As shown in the recall model, a three-tower structure was used during training. Specifically, the object behavior data corresponds to the first feature extraction network, the advertising domain product data corresponds to the second feature extraction network, and the organic traffic product data corresponds to the third feature extraction network; moreover, the object behavior data, the advertising domain product data, and the organic traffic product data can share a feature processing layer. The first and second feature extraction networks constitute the advertising domain product conversion rate prediction network, and the first and third feature extraction networks constitute the organic traffic product conversion rate prediction network.

[0064] Based on the above Figure 4 The architecture of the recall model shown is illustrated, and the training process of this recall model can be as follows: Figure 5 As shown, it can be specifically described as follows:

[0065] 501. Obtain the initial recall model and training samples. The initial recall model includes an initial feature processing layer, an initial first feature extraction network, an initial second feature extraction network, an initial third feature extraction network, and an output layer. The training samples include training object data, training product data, and conversion tags. The training object data includes advertising domain behavior data and organic traffic behavior data of the training objects. The training product data includes advertising domain product data and organic traffic product data. The conversion tags are used to indicate whether the product has been successfully converted.

[0066] In this embodiment, the initial recall model can be Figure 4 The architecture shown can be such that the initial first feature extraction network, the initial second feature extraction network, and the initial third feature extraction network have the same network structure or different network structures. No specific limitation is made here, as long as feature extraction can be achieved.

[0067] It should be understood that when the initial second feature extraction network and the initial third feature extraction network have the same network structure, the initial third feature extraction network can use the EMA method to transfer its network parameters to the initial second feature extraction network during training. This allows the second feature extraction network to learn more information during online inference, thereby improving the richness and reliability of data recall. Specifically, during the training of the recall model, the organic traffic product conversion rate prediction network will act as a student network, transferring its network parameters to the advertising domain product conversion rate prediction network using an exponential moving average method. This transfers the information learned by the organic traffic product conversion rate prediction network to the advertising domain product conversion rate prediction network. To enhance the transfer effect, self-supervised learning can be introduced to perform network parameter transfer between the organic traffic and advertising domains. This self-supervised learning can be cross-domain contrastive learning. That is, using contrastive learning methods, the conversion rates of products are compared and learned based on object behavior in both the organic traffic and advertising domains, generating more consistent product representations. For example, methods such as SimCLR or BYOL can be used to make the product embedding representations generated from the natural traffic behavior data and the advertising domain behavior data of the same object more similar.

[0068] When constructing the training sample, the training sample includes training object data, product data, and conversion tags.

[0069] The training data refers to the historical behavioral data of platform operators (i.e., platform users) towards other objects (such as advertisements and products), used to describe the interactive activities of platform operators within the platform or application. This training data includes advertising domain behavioral data, organic traffic behavioral data, and contextual information about the behavioral data. For example, organic traffic behavioral data includes behaviors such as browsing products, clicking on products, adding to cart, favorited products, and placing orders. Advertising domain behavioral data includes behaviors such as clicking on ads, skipping ads, and conversions after clicking (such as placing orders or making purchases). The contextual information of this behavioral data includes the time the behavior occurred, the device used (e.g., mobile phone or computer), the duration of the behavior, and the product category in which the behavior occurred.

[0070] This product data describes the product's static and dynamic characteristics, representing its attribute information. Static characteristics may include product identifier, product category, brand, price, sales volume, and ratings. Dynamic characteristics may include the product's real-time status (such as inventory and discount information) and its popularity (such as recent pageviews and clicks). It may also include advertising-related characteristics, such as whether the product participates in advertising promotions, advertising bids, historical ad click-through rates, and ad conversion rates.

[0071] This conversion tag represents the conversion result of a platform user towards a specific product or advertisement, and is used to supervise the training of the model. In this embodiment, a positive training sample means that the platform user has performed a clear conversion behavior (such as placing an order or making a purchase); a negative training sample means that the platform user has not performed a conversion behavior (such as only browsing or clicking but not purchasing). When constructing this conversion tag based on the above scheme, the final behavior of the platform user can be mapped to the tag. For example, if the platform user purchases a product, the conversion tag of the training sample consisting of that product can be set to 1; if the platform user does not purchase a product, the conversion tag of the training sample consisting of that product can be set to 0.

[0072] The following process can be used to construct the training samples: extract historical behavioral data of platform users from the organic traffic domain and the advertising domain; organize the interaction sequences of platform users in chronological order to generate behavioral features; simultaneously acquire the static and dynamic features of the products involved in the behavioral data; associate the product features with the user behavior to generate complete training samples; and generate positive and negative samples based on the final behavioral result of the platform user (such as whether a purchase was made), where a label of 1 represents a conversion and a label of 0 represents no conversion. This integration of user behavior data, product features, and conversion labels can more comprehensively reflect the interests and conversion tendencies of platform users. Furthermore, by constructing conversion labels, the recall model can simultaneously learn the conversion and non-conversion behavior patterns of platform users, improving recall effectiveness. On the other hand, the combination of behavioral data and product data with embedded representations enables the recall model to efficiently learn the matching relationship between platform users and products.

[0073] To provide more product information, secondary conversion tags can be created for negative samples based on their product conversion rates. For example, the category tag of the first product can be obtained, indicating different factors contributing to its conversion failure. If the category tag indicates that the first product has been viewed or clicked, a first-class secondary conversion tag is added to the negative sample, indicating that its conversion rate is below a first threshold. If the category tag indicates that the first product is an unconverted product within the same sales target, a second-class secondary conversion tag is added, indicating that its conversion rate is not lower than the first threshold and is lower than the second threshold. If the category tag indicates that the first product is an undiscovered unconverted product, a third-class secondary conversion tag is added, indicating that its conversion rate is not lower than the second threshold.

[0074] 502. Call the initial feature processing layer to perform feature processing on the training object data and the training product data to obtain the training object behavior feature vector and the training product feature vector. The training product feature vector includes the advertising domain product feature vector and the natural traffic product feature vector.

[0075] In this embodiment, the training samples constructed in step 501 are input into the initial feature processing layer of the initial recall model, so that the initial feature processing layer performs feature processing on the training object data and the training product data in the training samples, and outputs the training object behavior feature vector and the training product feature vector.

[0076] It should be understood that the training object data and the training product data can have different data types. For example, discrete data, continuous data, and sequence data. Discrete data can include object identifiers, product identifiers, advertisement identifiers, product categories, brands, etc. Continuous data can include product prices, the age of the platform user, product sales volume, advertisement click-through rates, conversion rates, etc. Sequence data can be the historical behavior sequence of the platform user (products viewed, advertisements clicked, etc.).

[0077] Based on the different data types mentioned above, the initial feature processing layer can employ different processing methods when performing feature processing. For example, discrete data can be mapped to identifiers before embedding. This involves mapping discrete data (such as object identifiers or product identifiers) to integer indices, and then using the embedding layer to map these indices to a low-dimensional vector space. Continuous data can be normalized and then directly used as input to the recall model. For instance, continuous data (such as prices or sales volume) typically have different numerical ranges and need to be normalized to [0,1] or standardized to a mean of 0 and a variance of 1; then, the normalized continuous data can be directly used as input to the recall model.

[0078] 503. Call the initial first feature extraction network to extract features from the behavioral feature vector of the training object to obtain the feature representation of the training object.

[0079] After obtaining the input data of the recall model, the initial first feature extraction network is called to extract features from the behavioral feature vector of the training object to obtain the feature representation of the training object.

[0080] In this embodiment, the initial first feature extraction network can be a deep neural network (DNN), or other network structures, such as convolutional neural networks (CNN), recurrent neural networks (RNN), and Transformers. No specific limitation is made here.

[0081] The following uses a DNN as an example to illustrate the feature extraction of the behavioral feature vector of the training object: The DNN network extracts features layer by layer using multiple fully connected layers; at the same time, each fully connected layer extracts the non-linear relationship of the features through a non-linear activation function; finally, the feature representation of the training object is output through the output layer.

[0082] 504. Call the initial second feature extraction network to extract features from the feature vector of the product in the advertising domain to obtain the feature representation of the product in the advertising domain.

[0083] After obtaining the input data of the recall model, the initial second feature extraction network is called to extract features from the feature vector of the product in the advertising domain, so as to obtain the feature representation of the product in the advertising domain.

[0084] In this embodiment, the initial second feature extraction network can be a deep neural network (DNN), or other network structures, such as convolutional neural networks (CNN), recurrent neural networks (RNN), and Transformers. No specific limitation is made here.

[0085] The following uses a DNN as an example to illustrate the feature extraction of the product feature vector in the advertising domain: The DNN network extracts features layer by layer using multiple fully connected layers; at the same time, each fully connected layer extracts the nonlinear relationship of the features through a nonlinear activation function; finally, the output layer outputs the feature representation of the product in the advertising domain.

[0086] 505. Call the initial third feature extraction network to extract features from the feature vector of the natural traffic product to obtain the feature representation of the natural traffic product.

[0087] After obtaining the input data of the recall model, the initial third feature extraction network is called to extract features from the feature vector of the organic traffic product to obtain the feature representation of the organic traffic product.

[0088] In this embodiment, the initial third feature extraction network can be a deep neural network (DNN), or other network structures, such as a convolutional neural network (CNN), a recurrent neural network (RNN), and a Transformer. No specific limitation is made here.

[0089] The following uses a DNN as an example to illustrate the feature extraction of the product feature vector of natural traffic: The DNN network extracts features layer by layer using multiple fully connected layers; at the same time, each fully connected layer extracts the non-linear relationship of the features through a non-linear activation function; finally, the output layer outputs the feature representation of the training object.

[0090] It should be understood that there is no time limit between steps 503 and 505 above, and they can be executed simultaneously; no specific time limit is set here.

[0091] 506. Perform a dot product between the feature representation of the training object and the feature representation of the product in the advertising domain to obtain the first predicted conversion rate of the product in the advertising domain.

[0092] After obtaining the feature representations of the training object, the product features of the advertising domain, and the product features of the organic traffic, the feature representation of the training object is multiplied by the product features of the advertising domain and the product features of the organic traffic, respectively. The result of the multiplication of the feature representations of the training object and the product features of the advertising domain is defined as the first predicted conversion rate of the product in the advertising domain.

[0093] The result of the dot product calculation is the magnitude of the projections of the two embedded vectors in the vector space. The larger the value, the closer the directions of the two vectors are. Therefore, in the data retrieval scenario of this embodiment, the platform operation object and the product embedded vector are set in a unified vector space, and the dot product can well reflect the interest relevance of the platform operation object to the product.

[0094] 507. Perform a dot product between the feature representation of the training object and the feature representation of the natural traffic product to obtain the second predicted conversion rate of the natural traffic product.

[0095] After obtaining the feature representations of the training object, the advertising domain product, and the organic traffic product, the feature representation of the training object is multiplied by the feature representations of the advertising domain product and the organic traffic product, respectively. The result of this multiplication is defined as the second predicted conversion rate of the organic traffic product.

[0096] The result of the dot product calculation is the magnitude of the projections of the two embedded vectors in the vector space. The larger the value, the closer the directions of the two vectors are. Therefore, in the data retrieval scenario of this embodiment, the platform operation object and the product embedded vector are set in a unified vector space, and the dot product can well reflect the interest relevance of the platform operation object to the product.

[0097] 508. Calculate a first loss value based on the first predicted conversion rate and the conversion tag, and calculate a second loss value based on the second predicted conversion rate and the conversion tag.

[0098] In this embodiment, the loss value is calculated using the conversion tag, the first predicted conversion rate, and the second predicted conversion rate, respectively, to obtain the first loss value and the second loss value.

[0099] When calculating the first loss value and the second loss value, the cross-entropy loss function or other loss functions can be used. There are no specific restrictions here, as long as the loss value can be calculated.

[0100] 509. Train the initial recall model based on the first loss value and the second loss value to obtain the recall model.

[0101] After calculating the first loss value and the second loss value, the initial first feature network and the initial second feature network are trained using the first loss value, and the initial first feature network and the initial third feature network are trained using the second loss value, so as to update the network parameters and output the recall model.

[0102] It should be understood that during training, if the initial second feature network and the initial third feature network have the same network structure, the network parameters of the initial third feature network can be transferred to the initial second feature network using the corresponding transfer method. The specific transfer method is not limited here.

[0103] Based on the above introduction, the data retrieval method in this application will be described below. Please refer to [link / reference]. Figure 6 One embodiment of the data retrieval method in this application includes:

[0104] 601. Obtain object behavior data and product feature data. The object behavior data includes the target object's advertising domain behavior data and natural traffic behavior data. The advertising domain behavior data is the target object's historical and real-time behavior data on advertising information, and the natural traffic behavior data is the target object's historical and real-time behavior data on actively accessing information.

[0105] In this embodiment, the object behavior data refers to the historical and real-time behavior data of the platform operation object (i.e., the platform operation object) towards other objects (such as advertisements and products), which is used to describe the interactive activities of the platform operation object on the platform or application. This object behavior data includes advertising domain behavior data, organic traffic behavior data, and contextual information of the behavior data. For example, the organic traffic behavior data includes behaviors such as browsing products, clicking on products, adding products to the cart, favorited products, and placing orders. The advertising domain behavior data includes behaviors such as clicking on advertisements, skipping advertisements, and conversions after clicking (such as placing orders or making purchases). The contextual information of the behavior data includes the time the behavior occurred, the device used (e.g., mobile phone or computer), the duration of the behavior, and the product category in which the behavior occurred.

[0106] This product feature data describes the product's static and dynamic characteristics, representing its attribute information. Static features can include product identifier, product category, brand, price, sales volume, and ratings. Dynamic features can include the product's real-time status (e.g., inventory, discount information) and its popularity (e.g., recent pageviews, clicks). It can also include advertising-related features, such as whether the product participates in advertising promotions, advertising bids, historical ad click-through rates, and ad conversion rates.

[0107] It should be understood that during online inference, the product feature data should be the feature data of all products in the entire product pool.

[0108] 602. Perform feature extraction on the object behavior data and the product feature data respectively to obtain object behavior feature representation and product feature representation.

[0109] In this embodiment, when the recall model is deployed to online inference, it may only include the advertising domain product conversion rate prediction network. Therefore, during online inference, the first feature extraction network can be invoked to extract features from the object's behavioral data to obtain the object's behavioral feature representation; simultaneously, the second feature extraction network can be invoked to extract features from the product's feature data to obtain the product's feature representation.

[0110] 603. Based on the object's behavioral characteristics and the product's characteristics, recall products to obtain a set of recalled products.

[0111] In this embodiment, after obtaining the object behavior feature representation and the product feature representation, the prediction network layer performs a dot product on the object behavior feature representation and the product feature representation to obtain the recall probability of the product; then, the products are sorted from high to low according to the recall probability, and the top-N products are recalled to obtain the recall product set.

[0112] 604. Determine the recall advertising set based on the recall product set.

[0113] In this embodiment, there is an association between the product and the advertisement. Therefore, after obtaining the recalled product, M advertisements corresponding to each product can be recalled according to the association between the product and the advertisement to obtain the recalled advertisement set.

[0114] It should be understood that when recalling advertisements for each product, the association between the product and the advertisement can be identified by the degree of association; then, they are sorted in descending order of association degree, and the top M advertisements are used as the recall advertisements for the target product. That is, in this embodiment, the final recall advertisement set will include M*N advertisements.

[0115] In this embodiment, products are used as a medium bridging organic traffic and advertising domains. The platform learns the conversion rates of products in organic traffic and migrates them to the conversion rates of products in advertising. By recalling the corresponding product ads first and then the ads, data sparsity can be alleviated. That is, for a product, if there are ads 1, 2, and 3, where ads 1 and 2 both have product conversions, but ad 3 has no product conversions, the solution provided in this embodiment can learn that the product has two conversions, namely ads 1 and 2. In other words, compared with the prior art, after learning the conversion rate of ads, this application can also directly determine the conversion rate of different ads for the same product through product classification.

[0116] In this embodiment, after the server obtains the recall ad set, it can also output the M*N ads to the ad ranking system for ad recommendation.

[0117] The data retrieval device in this application is described in detail below. Please refer to [link / reference]. Figure 7 , Figure 7 This is a schematic diagram of one embodiment of the data retrieval device in this application. The data retrieval device 20 includes:

[0118] The acquisition module 201 is used to acquire object behavior data and product feature data. The object behavior data includes the target object's advertising domain behavior data and natural traffic behavior data. The advertising domain behavior data is the target object's historical behavior data and real-time behavior data on advertising information, and the natural traffic behavior data is the target object's historical behavior data and real-time behavior data on active access information.

[0119] The processing module 202 is used to extract features from the object behavior data and the product feature data respectively to obtain object behavior feature representation and product feature representation; to perform product recall based on the object behavior feature representation and the product feature representation to obtain a recall product set; and to determine a recall advertisement set based on the recall product set.

[0120] This application provides a data retrieval device. By employing this device, behavioral data from organic traffic is introduced, providing more multi-dimensional behavioral information and enabling a full exploration of the interests and preferences of platform users, thus enriching the feature information. Simultaneously, by aggregating data through product dimensions, the sparsity of advertising domain conversion data is alleviated, thereby improving the richness and reliability of advertising retrieval.

[0121] Optionally, in the above Figure 7 Based on the corresponding embodiments, in another embodiment of the data retrieval device 20 provided in this application,

[0122] The processing module 202 is used to call the feature processing layer of the recall model to embed features into the object behavior data and the product feature data to obtain object behavior feature vectors and product feature vectors. The recall model includes a feature processing layer, a first feature extraction network and a second feature extraction network.

[0123] The first feature extraction network of the recall model is invoked to extract features from the object's behavioral feature vector to obtain a representation of the object's behavioral features.

[0124] The second feature extraction network of the recall model is invoked to extract features from the product feature vector to obtain the product feature representation.

[0125] This application provides a data retrieval device. Using this device, the feature extraction networks of the platform operation object tower and the advertising tower of the pre-trained retrieval model are used to extract features from the corresponding data to obtain corresponding feature representations. The inclusion of natural traffic behavior data in the object behavior data provides more dimensional behavioral information, enabling a full exploration of the interests and preferences of the platform operation objects, enriching the feature information, and thus improving the richness and reliability of advertising retrieval.

[0126] Optionally, in the above Figure 7 Based on the corresponding embodiments, in another embodiment of the data recall device 20 provided in this application, the processing module 202 is used to perform a dot product between the object behavior feature representation and the product feature representation to obtain the estimated conversion rate of each product;

[0127] Based on the estimated conversion rate, products are recalled in a top-N manner to obtain the recalled product set.

[0128] This application provides a data retrieval device. Using this device, dot product is a simple and efficient calculation method used to measure the similarity between the platform operation object embedding vector and the advertisement embedding vector, resulting in high computational efficiency and faster retrieval speed. Simultaneously, using a Top-N filtering mechanism for product retrieval can significantly reduce the number of candidate products that need to be passed to the ranking stage, thereby saving computational resources.

[0129] Optionally, in the above Figure 7 Based on the corresponding embodiments, in another embodiment of the data retrieval device 20 provided in this application,

[0130] The acquisition module 201 is used to acquire an initial recall model and training samples. The initial recall model includes an initial feature processing layer, an initial first feature extraction network, an initial second feature extraction network, an initial third feature extraction network, and an output layer. The training samples include training object data, training product data, and conversion tags. The training object data includes advertising domain behavior data and organic traffic behavior data of the training objects. The training product data includes advertising domain product data and organic traffic product data. The conversion tags are used to indicate whether the product has been successfully converted.

[0131] The processing module 202 is used to call the initial feature processing layer to embed features into the training object data and the training product data to obtain training object behavior feature vectors and training product feature vectors, wherein the training product feature vectors include advertising domain product feature vectors and organic traffic product feature vectors; call the initial first feature extraction network to extract features from the training object behavior feature vectors to obtain training object feature representations; call the initial second feature extraction network to extract features from the advertising domain product feature vectors to obtain advertising domain product feature representations; call the initial third feature extraction network to extract features from the organic traffic product feature vectors to obtain organic traffic product feature representations; perform a dot product between the training object feature representations and the advertising domain product feature representations to obtain a first predicted conversion rate for advertising domain products; perform a dot product between the training object feature representations and the organic traffic product feature representations to obtain a second predicted conversion rate for organic traffic products; calculate a first loss value based on the first predicted conversion rate and the conversion label; calculate a second loss value based on the second predicted conversion rate and the conversion label; and train the initial recall model based on the first loss value and the second loss value to obtain the recall model.

[0132] This application provides a data retrieval device. By employing this device, behavioral data from natural traffic is introduced, providing more multi-dimensional behavioral and product information. This allows for a fuller understanding of the interests and preferences of platform users and a richer learning of product characteristics, thus enriching the feature information. Simultaneously, by aggregating data through product dimensions, the sparsity of advertising domain conversion data is alleviated, thereby improving the richness and reliability of advertising retrieval.

[0133] Optionally, in the above Figure 7 Based on the corresponding embodiments, in another embodiment of the data retrieval device 20 provided in this application,

[0134] The initial second feature extraction network and the initial third feature extraction network have the same network structure; the processing module 202 is used to transfer the network parameters of the initial third feature extraction network to the initial second feature extraction network by means of exponential moving average.

[0135] This application provides a data retrieval device. Using this device, the same feature network is used for both the feature extraction network for the advertising domain and the feature extraction network for organic traffic, and parameter transfer is performed using EMA. This effectively alleviates the negative transfer problem caused by distribution differences, while improving sensitivity to the latest training data, thus enhancing the model's stability and robustness.

[0136] Optionally, in the above Figure 7 Based on the corresponding embodiments, in another embodiment of the data retrieval device 20 provided in this application, the acquisition module 201 is used to acquire object behavior data and product conversion data corresponding to the object behavior data;

[0137] When the product conversion data indicates that the first product has been successfully converted, construct a positive sample using the object's behavior data and the first product, and set the conversion label to "product conversion successful";

[0138] When the conversion data of the product indicates that the first product has not been successfully converted, a negative sample is constructed from the object's behavior data and the first product, and the conversion label is set to "product not successfully converted".

[0139] This application provides a data retrieval device. Using this device, object behavior data, product data, and conversion tags are used to construct training positive and negative samples. This allows the model to learn more product and object behavior feature information, thereby improving the richness and reliability of ad retrieval.

[0140] Optionally, in the above Figure 7Based on the corresponding embodiments, in another embodiment of the data recall device 20 provided in this application, the acquisition module 201 is used to acquire the classification label of the first product, which is used to indicate different factors of the conversion failure of the first product;

[0141] When the category label indicates that the first product is a product that the object has viewed or clicked, a first-class secondary conversion label is added to the negative sample. The first-class secondary conversion label is used to indicate that the conversion rate of the first product is lower than a first threshold.

[0142] When the category label indicates that the first product is an unconverted product under the same sales object, a second-level conversion label is added to the negative sample. The second-level conversion label is used to indicate that the conversion rate of the first product is not lower than the first threshold and is lower than the second threshold.

[0143] When the category label indicates that the first product is an undiscovered, unconverted product, a third-class secondary conversion label is added to the negative sample. The third-class secondary conversion label is used to indicate that the conversion rate of the first product is not lower than the second threshold.

[0144] This application provides a data recall device. Using this device, constructing multi-level negative samples allows the model to learn more product information, thereby broadening the richness of the recalled prediction samples.

[0145] The data retrieval device provided in this application can be a server; please refer to [link / reference]. Figure 8 , Figure 8 This is a schematic diagram of a server structure provided in an embodiment of this application. The server 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 322 (e.g., one or more processors) and memory 332, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 342 or data 344. The memory 332 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the server. Furthermore, the CPU 322 may be configured to communicate with the storage media 330 and execute the series of instruction operations stored in the storage media 330 on the server 300.

[0146] Server 300 may also include one or more power supplies 326, one or more wired or wireless network interfaces 350, one or more input / output interfaces 358, and / or one or more operating systems 341, such as Windows Server.TM Mac OS X TM Unix TM Linux TM FreeBSD TM etc.

[0147] The steps performed by the server in the above embodiments can be based on this Figure 8 The server structure shown.

[0148] The data retrieval device provided in this application can be used in terminal devices. Please refer to [link / reference]. Figure 9 For ease of explanation, only the parts relevant to the embodiments of this application are shown. For specific technical details not disclosed, please refer to the method section of the embodiments of this application. In the embodiments of this application, a smartphone is used as an example for illustration:

[0149] Figure 9 This is a block diagram illustrating a portion of the structure of a smartphone related to the terminal device provided in the embodiments of this application. (Reference) Figure 9 The smartphone includes components such as a radio frequency (RF) circuit 410, a memory 420, an input unit 430, a display unit 440, a sensor 450, an audio circuit 460, a wireless fidelity (WiFi) module 470, a processor 480, and a power supply 490. Those skilled in the art will understand that... Figure 9 The smartphone structure shown does not constitute a limitation on smartphones and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0150] The following is combined Figure 9 A detailed introduction to the various components of a smartphone:

[0151] RF circuit 410 can be used for receiving and transmitting signals during information transmission or calls. Specifically, it receives downlink information from the base station and processes it with processor 480; additionally, it transmits uplink data to the base station. Typically, RF circuit 410 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low-noise amplifier (LNA), a duplexer, etc. Furthermore, RF circuit 410 can also communicate wirelessly with networks and other devices. The aforementioned wireless communication can use any communication standard or protocol, including but not limited to Global System for Mobile Communication (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), email, Short Message Service (SMS), etc.

[0152] The memory 420 can be used to store software programs and modules. The processor 480 executes various functions and data processing of the smartphone by running the software programs and modules stored in the memory 420. The memory 420 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the smartphone (such as audio data, phonebook, etc.). In addition, the memory 420 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0153] The input unit 430 can be used to receive input numerical or character information, and to generate key signal inputs related to the platform operation object settings and function control of the smartphone. Specifically, the input unit 430 may include a touch panel 431 and other input devices 432. The touch panel 431, also known as a touch screen, can collect touch operations on or near the platform operation object (such as operations performed by the platform operation object using a finger, stylus, or any suitable object or accessory on or near the touch panel 431), and drive the corresponding connection device according to a pre-set program. Optionally, the touch panel 431 may include two parts: a touch detection device and a touch controller. The touch detection device detects the touch position of the platform operation object and detects the signal generated by the touch operation, and transmits the signal to the touch controller; the touch controller receives touch information from the touch detection device, converts it into touch point coordinates, sends it to the processor 480, and can receive and execute commands sent by the processor 480. In addition, the touch panel 431 can be implemented using various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the touch panel 431, the input unit 430 may also include other input devices 432. Specifically, other input devices 432 may include, but are not limited to, one or more of the following: physical keyboard, function keys (such as volume control buttons, power buttons, etc.), trackball, mouse, joystick, etc.

[0154] The display unit 440 can be used to display information input by the platform operator or information provided to the platform operator, as well as various menus of the smartphone. The display unit 440 may include a display panel 441, optionally configured as a liquid crystal display (LCD), organic light-emitting diode (OLED), or similar form. Further, a touch panel 431 may cover the display panel 441. When the touch panel 431 detects a touch operation on or near it, it transmits the information to the processor 480 to determine the type of touch event. Subsequently, the processor 480 provides corresponding visual output on the display panel 441 according to the type of touch event. Although in Figure 9 In this embodiment, the touch panel 431 and the display panel 441 are two separate components to realize the input and output functions of the smartphone. However, in some embodiments, the touch panel 431 and the display panel 441 can be integrated to realize the input and output functions of the smartphone.

[0155] The smartphone may also include at least one sensor 450, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor, wherein the ambient light sensor can adjust the brightness of the display panel 441 according to the ambient light level, and the proximity sensor can turn off the display panel 441 and / or the backlight when the smartphone is moved to the ear. As a type of motion sensor, an accelerometer sensor can detect the magnitude of acceleration in various directions (generally three axes), and can detect the magnitude and direction of gravity when stationary. It can be used for applications that recognize the smartphone's posture (such as landscape / portrait switching, related games, magnetometer posture calibration), vibration recognition-related functions (such as pedometer, tapping), etc. Other sensors that may be configured in the smartphone, such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, will not be described in detail here.

[0156] Audio circuit 460, speaker 461, and microphone 462 provide an audio interface between the platform and the smartphone. Audio circuit 460 converts received audio data into electrical signals and transmits them to speaker 461, where speaker 461 converts them into sound signals for output. On the other hand, microphone 462 converts collected sound signals into electrical signals, which are received by audio circuit 460, converted into audio data, and then processed by processor 480 before being transmitted via RF circuit 410 to, for example, another smartphone, or the audio data can be output to memory 420 for further processing.

[0157] WiFi is a short-range wireless transmission technology. Smartphones, through their WiFi modules (470), can help platform users send and receive emails, browse web pages, and access streaming media, providing them with wireless broadband internet access. Although Figure 9 WiFi module 470 is shown, but it is understood that it is not an essential component of a smartphone and can be omitted as needed without changing the nature of the invention.

[0158] The processor 480 is the control center of the smartphone, connecting various parts of the smartphone through various interfaces and lines. It performs various functions and processes data by running or executing software programs and / or modules stored in the memory 420, and by calling data stored in the memory 420, thereby providing overall monitoring of the smartphone. Optionally, the processor 480 may include one or more processing units; optionally, the processor 480 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, platform operation object interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the aforementioned modem processor may also not be integrated into the processor 480.

[0159] The smartphone also includes a power supply 490 (such as a battery) that supplies power to various components. Optionally, the power supply can be logically connected to the processor 480 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system.

[0160] Although not shown, smartphones may also include a camera, Bluetooth module, etc., which will not be described in detail here.

[0161] The steps performed in the above embodiments can be based on this Figure 9 The terminal device structure is shown.

[0162] This application also provides a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the methods described in the foregoing embodiments.

[0163] This application also provides a computer program product including a program, which, when run on a computer, causes the computer to perform the methods described in the foregoing embodiments.

[0164] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0165] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0166] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0167] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0168] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0169] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A data retrieval method, characterized in that, include: Acquire object behavior data and product feature data. The object behavior data includes advertising domain behavior data and natural traffic behavior data of the target object. The advertising domain behavior data is the historical and real-time behavior data of the target object towards advertising information, and the natural traffic behavior data is the historical and real-time behavior data of the target object towards actively accessing information. Feature extraction is performed on the object behavior data and the product feature data respectively to obtain object behavior feature representation and product feature representation; Based on the object behavior feature representation and the product feature representation, product recall is performed to obtain a recall product set; The recall advertising set is determined based on the recall product set.

2. The method according to claim 1, characterized in that, The step of extracting features from the object behavior data and the product feature data respectively to obtain object behavior feature representations and product feature representations includes: The feature processing layer of the recall model is invoked to embed features into the object behavior data and the product feature data to obtain object behavior feature vectors and product feature vectors. The recall model includes a feature processing layer, a first feature extraction network, and a second feature extraction network. The first feature extraction network of the recall model is invoked to extract features from the object behavior feature vector to obtain the object behavior feature representation; The second feature extraction network of the recall model is invoked to extract features from the product feature vector to obtain the product feature representation.

3. The method according to claim 1, characterized in that, Based on the object behavior feature representation and the product feature representation, product recall is performed to obtain a recall product set, including: The object behavior feature representation and the product feature representation are multiplied by a dot to obtain the estimated conversion rate of each product; Based on the estimated conversion rate, products are recalled in a top-N manner to obtain the recalled product set.

4. The method according to any one of claims 2 to 3, characterized in that, The method further includes: An initial recall model and training samples are obtained. The initial recall model includes an initial feature processing layer, an initial first feature extraction network, an initial second feature extraction network, an initial third feature extraction network, and an output layer. The training samples include training object data, training product data, and conversion tags. The training object data includes advertising domain behavior data and organic traffic behavior data of the training objects. The training product data includes advertising domain product data and organic traffic product data. The conversion tags are used to indicate whether the product has been successfully converted. The initial feature processing layer is invoked to perform feature processing on the training object data and the training product data to obtain the training object behavior feature vector and the training product feature vector. The training product feature vector includes the advertising domain product feature vector and the natural traffic product feature vector. The initial first feature extraction network is invoked to extract features from the behavioral feature vector of the training object, so as to obtain the feature representation of the training object; The initial second feature extraction network is invoked to extract features from the feature vector of the advertising domain product to obtain the feature representation of the advertising domain product; The initial third feature extraction network is invoked to extract features from the natural traffic product feature vector to obtain a natural traffic product feature representation; The feature representation of the training object is multiplied by the feature representation of the advertising domain product to obtain the first predicted conversion rate of the advertising domain product; The feature representation of the training object is multiplied by the feature representation of the organic traffic product to obtain the second predicted conversion rate of the organic traffic product; A first loss value is calculated based on the first predicted conversion rate and the conversion label, and a second loss value is calculated based on the second predicted conversion rate and the conversion label; The initial recall model is trained based on the first loss value and the second loss value to obtain the recall model.

5. The method according to claim 4, wherein the initial second feature extraction network and the initial third feature extraction network have the same network structure; the method further comprises: The network parameters of the initial third feature extraction network are transferred to the initial second feature extraction network using an exponential moving average method.

6. The method according to claim 4, characterized in that, The acquisition of training samples includes: Obtain object behavior data and the corresponding product conversion data; When the product conversion data indicates that the first product has been successfully converted, a positive sample is constructed by combining the object behavior data and the first product, and the conversion tag is set to "product conversion successful". When the product conversion data indicates that the first product has not been successfully converted, a negative sample is constructed by combining the object behavior data and the first product, and the conversion label is set to "product not successfully converted".

7. The method according to claim 6, characterized in that, The step of constructing a negative sample from the object behavior data and the first product includes: Obtain the category tags of the first product, which are used to indicate different factors that cause the conversion failure of the first product; When the category label indicates that the first product is a product that the object has viewed or clicked, a first type of secondary conversion label is added to the negative sample. The first type of secondary conversion label is used to indicate that the conversion rate of the first product is lower than a first threshold. When the classification label indicates that the first product is an unconverted product under the same sales object, a second type of secondary conversion label is added to the negative sample. The second type of secondary conversion label is used to indicate that the conversion rate of the first product is not lower than the first threshold and is lower than the second threshold. When the classification label indicates that the first product is an undiscovered, unconverted product, a third-level secondary conversion label is added to the negative sample. The third-level secondary conversion label is used to indicate that the conversion rate of the first product is not lower than the second threshold.

8. A data retrieval device, characterized in that, include: The acquisition module is used to acquire object behavior data and product feature data. The object behavior data includes advertising domain behavior data and natural traffic behavior data of the target object. The advertising domain behavior data is the historical and real-time behavior data of the target object towards advertising information, and the natural traffic behavior data is the historical and real-time behavior data of the target object towards actively accessing information. The processing module is used to extract features from the object behavior data and the product feature data respectively to obtain object behavior feature representation and product feature representation; to perform product recall based on the object behavior feature representation and the product feature representation to obtain a recall product set; and to determine a recall advertisement set based on the recall product set.

9. A computer device, characterized in that, include: Memory, processor, and bus system; The memory is used to store programs; The processor is configured to execute a program in the memory, and the processor is configured to execute the method of any one of claims 1 to 7 according to instructions in the program code; The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

10. A computer-readable storage medium comprising instructions, when executed on a computer, causing the computer to perform the method as claimed in any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that, The computer program is executed by a processor using the method as described in any one of claims 1 to 7.

Images