Independent station commodity recommendation method and device, equipment and medium

Abstract

The invention relates to an independent station commodity recommendation method and device, equipment and a medium, and the method comprises the steps: responding to a commodity recommendation request, and obtaining the user data of a current user and the commodity data of commodities on shelf at a current independent station; inputting the user data and the commodity data into a complete space multi-task model, and generating a first user feature vector and a first commodity feature vector through a corresponding built-in coding network in a conversion index reasoning path in the model; a second user feature vector and a second commodity feature vector are generated through a corresponding amplification coding network in the path, and the amplification coding network independently carries out comparative learning training in advance; fusing each user feature vector and each commodity feature vector into a comprehensive feature vector through a prediction network in the path, and predicting conversion index data according to the comprehensive feature vector; and for a plurality of on-shelf commodities of the current independent station, screening part of on-shelf commodities based on the conversion index data and pushing the on-shelf commodities to the current user. The commodity recommendation accuracy of the independent station can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of commodity recommendation, in particular to a standalone station commodity recommendation method and device, equipment and medium. BACKGROUND

[0002] In the commodity recommendation scenario, the complete space multi-task model optimizes the click rate and the click conversion rate at the same time, extends the conversion estimation task that can only be trained on the click sample to the whole exposure space, thereby alleviating the sample selection bias, and enables the conversion branch to utilize the information of the non-click sample by sharing the embedding layer, and improves the estimation stability under the sparse label. The model organizes "whether to click" and "whether to purchase" into a cascaded multi-task structure from the perspective of the behavior chain, uses the supervision information of CTOR (click-to-conversion rate) and CTR (click rate) to train the network in the training stage, realizes the implicit learning of the CVR (conversion rate) task at the same time, and sorts the commodities by using these index data in the inference stage. Compared with the two-stage scheme of independently training the click rate model first and then training the conversion rate model based on the click sample, the complete space multi-task model avoids the bias accumulation caused by the distribution difference between the click space and the exposure space, and also saves the complex sample splicing process, and thus becomes the mainstream baseline of the commodity recommendation system in recent years.

[0003] However, when this model is applied to the standalone station, its performance is still significantly restricted by the data sparsity. The standalone station usually lacks the traffic scale of large platform e-commerce, the user behavior sequence is short, and the commodity exposure times are few, which leads to that the update frequency of a large number of items in the shared embedding layer is extremely low, and the parameters are difficult to fully converge. Under this condition, although the conversion branch can formally utilize all the exposure samples, the observable conversion positive examples are still concentrated in a few popular commodities, and the embedding representation of the long-tail commodities is still close to the initial random state; at the same time, the user-side embedding also lacks discrimination due to the scarcity of interaction records. The sharing mechanism becomes a kind of negative transfer in the sparse scenario: the gradient of the click branch dominates the embedding layer update, and the gradient of the conversion branch is submerged, so that the model's ability to capture potential purchase signals is limited. As a result, the complete space multi-task model gives a large variance and low confidence in the CVR estimation in the standalone station, and the top-ranked commodities are often concentrated in the head SKU (commodity single product) with existing transaction records, which cannot effectively tap the conversion potential of the long-tail inventory, and the novelty and coverage of the recommendation results are not ideal.

[0004] Therefore, how to enhance the feature representation capability under sparse data while maintaining the training advantages of the complete space multi-task model, ensure accurate prediction of the click and conversion related index data of the commodity, provide a more accurate ranking benchmark for commodity recommendation, and make the commodity recommendation more effective, has a positive significance for the standalone station commodity recommendation service. SUMMARY

[0005] The purpose of this application is to solve at least one of the above-mentioned problems by providing an independent website product recommendation method and corresponding apparatus, equipment, non-volatile readable storage medium, and computer program product.

[0006] According to one aspect of this application, a method for recommending products on an independent website is provided, comprising: Respond to the product recommendation request triggered by the current user on the current independent website, and obtain the user data of the current user and the product data of the products listed on the current independent website; The user data and the product data are input into a preset complete spatial multi-task model. The corresponding first user feature vector and first product feature vector are generated by the built-in encoding network in the conversion index inference path of the model. The corresponding amplified coding network within the conversion index inference path generates the corresponding second user feature vector and second product feature vector. The amplified coding network is independently trained in advance to acquire the corresponding feature representation ability. The prediction network in the conversion index inference path fuses the feature vectors of each user and each product into a comprehensive feature vector, and predicts the corresponding conversion index data based on the comprehensive feature vector. The prediction network is determined by fine-tuning training after the trained amplified coding network is added to the conversion index inference path. For the multiple products listed on the current independent website, some of the listed products are selected based on their corresponding conversion index data and pushed to the current user as recommended products.

[0007] According to another aspect of this application, an independent website product recommendation device is provided, comprising: The request and response module is set to respond to product recommendation requests triggered by the current user on the current independent website, and to obtain the user data of the current user and the product data of the products listed on the current independent website; The built-in representation module is configured to input the user data and the product data into a preset full-space multi-task model, and generate the corresponding first user feature vector and first product feature vector by the corresponding built-in encoding network in the conversion index inference path within the model. The augmented representation module is configured to generate a corresponding second user feature vector and a second product feature vector by the corresponding augmented coding network within the inference path of the conversion index. The augmented coding network is trained independently in advance by comparative learning to acquire the corresponding feature representation ability. The indicator prediction module is configured to use the prediction network in the conversion indicator inference path to fuse each user feature vector and each product feature vector into a comprehensive feature vector, and predict the corresponding conversion indicator data based on the comprehensive feature vector. The prediction network is determined by fine-tuning training after the trained amplified coding network is added to the conversion indicator inference path. The filtering and recommendation module is configured to select a portion of the listed products from the multiple products listed on the current independent website and push them to the current user as recommended products based on their corresponding conversion index data.

[0008] According to another aspect of this application, an independent website product recommendation device is provided, including a central processing unit and a memory, wherein the central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method described in this application.

[0009] According to another aspect of this application, a non-volatile readable storage medium is provided, which stores a computer program implemented according to the independent website product recommendation method in the form of computer-readable instructions, wherein the computer program, when invoked by a computer, performs the steps included in the method.

[0010] According to another aspect of this application, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the method.

[0011] Compared to traditional technologies, this application innovatively extends the traditional full-space multi-task model by introducing an augmented coding network trained through contrastive learning, significantly improving the model's feature representation capabilities. This network can reduce the similarity of features among similar products and amplify the differences in features among dissimilar products, generating higher-quality user and product feature vectors. This provides the prediction network with comprehensive and in-depth information, enabling more accurate prediction of product click and conversion-related metrics for ranking and filtering recommended products. Addressing the data sparsity problem of independent websites, this application effectively maintains sufficient model parameter convergence by adding an augmented coding network and conducting contrastive learning training, improving the embedding representation capability of long-tail products and user-side discriminative power, overcoming the data sparsity challenge, and improving recommendation performance. Furthermore, this application accurately predicts conversion metrics, filters products with high conversion potential, improves user satisfaction, conversion rate, and sales, while ensuring recommendation novelty and coverage, comprehensively enhancing the product recommendation experience on independent websites. Attached Figure Description

[0012] Figure 1 A schematic diagram of the network architecture of the optimized complete spatial multi-task model for this application; Figure 2 This is a flowchart illustrating one embodiment of the independent website product recommendation method of this application; Figure 3 This is a schematic diagram of the independent website product recommendation device of this application; Figure 4 This is a schematic diagram of the structure of an independent website product recommendation device used in this application. Detailed Implementation

[0013] An independent website refers to an online sales platform independently operated by a business or individual. It does not rely on large e-commerce platforms but sells goods directly to consumers through its own website. The product recommendation method for independent websites described in this application can be implemented as a computer program installed and running on the server of the independent website, providing an efficient product recommendation service to enhance user experience and improve sales conversion rates.

[0014] The method in this application can respond to a product recommendation request triggered by a user on an independent website, and obtain the user's current user data and the product data of the products currently listed on the independent website. User data may include, but is not limited to, specific behavioral data such as the user's browsing history, purchase records, and search habits, as well as user characteristic data such as gender and age. Product data may include information such as the product title, product summary text, product attributes, category, price, inventory status, and user reviews. This data is input into a pre-defined full-space multi-task model to generate corresponding user feature vectors and product feature vectors.

[0015] In the application scenario of this application, when users use the product recommendation service on an independent website, they can interact with the server deploying the product recommendation service through the website's interface. When a user visits the independent website, the website can automatically collect the user's real-time behavioral data and combine it with the user's long-term behavioral patterns to form the user data and product data required for the model, which are then analyzed through a complete spatial multi-task model. The built-in encoding network and the amplified encoding network in the conversion index inference path within the model work together to generate high-quality feature vectors, which are then fused into a comprehensive feature vector.

[0016] Based on the comprehensive feature vector, the prediction network in the conversion index inference path predicts the corresponding conversion index data, which reflects users' potential purchase intentions for different products. For the multiple products currently listed on the independent website, a selection of these products is selected as recommended products based on their corresponding conversion index data and pushed to the current user.

[0017] When users utilize the product recommendation service on an independent website, they can see recommended products on the homepage, product detail page, or search results page, depending on the specific context in which the product recommendation request was triggered. These recommended products are selected using the methods described in this application, based on the semantic relationship between the user's user data and the product data of the listed products in the independent website's product database, with the aim of increasing the user's willingness to purchase. Users can directly click on the recommended products to view or purchase them, thereby achieving a personalized shopping experience.

[0018] The complete spatial multitasking model used in this application is an optimized version of the traditional pre-trained complete spatial multitasking model (ESMM) with a determined network structure. For example... Figure 1 As shown, the optimized network structure includes two paths corresponding to the predicted click-through rate and the predicted conversion rate, namely the conversion metric inference path and the click metric inference path. The two paths and their connection relationship follow the inherent network structure of the full-space multi-task model. Both paths include their own encoding network and prediction network to complete the functions of feature representation and metric data prediction. The two paths share the encoding dictionary, and the outputs of their respective prediction networks are fused into the same output so as to provide supervised labels for training corresponding to the click-through rate (CTCVR).

[0019] Building upon this foundation, this application optimizes the entire model by introducing a pre-trained augmented coding network, trained through contrastive learning, into the conversion index inference path. It should be noted that, since the full-space multi-task model is also based on collaborative filtering principles and requires two outputs—user data and product data—each path's built-in coding network includes both user-side and product-side built-in coding networks. Similarly, the augmented coding network also includes augmented coding networks corresponding to both user-side and product-side data.

[0020] In the optimized network structure of the complete spatial multi-task model, the conversion metric inference path and the click metric inference path are responsible for handling the prediction tasks of conversion and click, respectively. The conversion metric inference path includes a built-in encoding network and an augmented encoding network. The augmented encoding network is pre-trained independently through contrastive learning to acquire the corresponding feature representation capabilities. Two built-in encoding networks are used to generate the first user feature vector and the first product feature vector, while two augmented encoding networks are used to generate the second user feature vector and the second product feature vector. These feature vectors are then fused into a comprehensive feature vector, providing the prediction network with more comprehensive and in-depth information, enabling the model to more accurately predict product click and conversion-related metric data. The prediction network is determined after fine-tuning training following the addition of the trained augmented encoding network to the conversion metric inference path, in order to optimize the model's prediction performance.

[0021] When the full-space multi-task model completes training and is put into online inference, any path in its network architecture can be flexibly selected. Since the training samples related to clicks are relatively abundant, while the training samples related to conversions are relatively sparse, this application mainly optimizes the inference path for conversion metrics. Therefore, the following description will focus on the inference path for conversion metrics.

[0022] Please see Figure 2 According to the product recommendation method for independent websites provided in this application, when applied to a server, in some embodiments, the method includes the following steps: Step S5100: Respond to the product recommendation request triggered by the current user on the current independent website, and obtain the user data of the current user and the product data of the products listed on the current independent website; When a user browses, searches, or clicks on products on the independent website, a product recommendation request is triggered. This request is submitted to the server of the current independent website, which then responds. After receiving the request, the server parses it, identifies the current user, and retrieves the user data as required by this application. Simultaneously, it retrieves the listed products and their data related to the request from the current independent website.

[0023] In one embodiment, the product recommendation request carries query text. The server can call a preset first recall channel based on the query text and recall multiple listed products and their product data that semantically match the query text from the product database of the current independent website according to the preset business logic of the channel.

[0024] In another embodiment, the product recommendation request specifies the target product that the user is currently browsing or clicking. The server can call a preset second recall channel based on the target product and recall co-purchased products or similar products and their product data from the product database of the current independent website according to the preset business logic of the channel. The co-purchased products and similar products are also the products listed on the current independent website.

[0025] User data and product data are important components of the original input specifications of the complete spatial multi-task model. In this application, the input specifications are continued to be used, that is, user data and product data are defined and provided according to the input specifications of the prototype of the complete spatial multi-task model. Based on its input specifications, those skilled in the art can flexibly determine the specific user data and product data.

[0026] User data can be categorized into user characteristic data and user behavior data. User behavior data includes, but is not limited to, specific behavioral data such as a user's browsing history, purchase records, and search habits. These behavioral data sets contain information corresponding to a user's specific access behavior to a particular product at a certain moment. Corresponding behavioral product sequences can be abstracted from this behavioral data as user behavior data to reflect user behavioral preferences. User characteristic data, on the other hand, mainly includes personalized data such as a user's gender, age, and interest tags.

[0027] Product data is useful data describing the characteristics of the listed products themselves. Its specific data composition is more flexible. Theoretically, it can use all or part of the product information of the listed products. This product information includes, but is not limited to, the title, summary text, attributes, category, price, inventory status, user reviews, and ranking information of the listed products. Specifically, it can be flexibly selected while adhering to the aforementioned input specifications.

[0028] In the practical scenario of an independent website, user and product data can be acquired through the website's backend system. When a user visits the website, the backend system can track the user's clickstream and page dwell time, while also collecting information proactively provided by the user, such as registration details. Product data is entered by the merchant when listing products, including detailed product information and attribute tags. The backend system then processes this collected user-side and product-side data into corresponding user and product data through appropriate feature engineering.

[0029] Step S5200: Input the user data and the product data into a preset complete spatial multi-task model, and generate the corresponding first user feature vector and first product feature vector by the corresponding built-in encoding network in the conversion index inference path of the model. As disclosed above, the network structure of the complete spatial multi-task model has been optimized in this application. This embodiment focuses on, but is not limited to, how the conversion index inference path is utilized, which is a key part of the model used to predict the conversion rate. In this path, user data and product data are first input into the corresponding built-in encoding networks on the user side and the product side, respectively.

[0030] Inherent Encoder Networks (IENs) are a standard component inherent in full-space multi-task models for learning user and item feature representations. On the user side, IENs receive user data and, through a series of neural network layers, transform this raw data into a fixed-length user feature vector. This vector captures user behavioral patterns and personalized features, providing rich user-side information for subsequent recommendation tasks.

[0031] Similarly, on the product side, a built-in encoding network processes product data, which is transformed into product feature vectors through a similar neural network structure. These vectors contain the key attributes and features of the products, helping the model understand the characteristics of the products and the potential interests of users.

[0032] The outputs of these two built-in encoding networks—the first user feature vector and the first product feature vector—provide the model with preliminary feature representations of users and products. These feature vectors form the basis for subsequent feature fusion and prediction tasks; they will be used to generate more advanced feature representations to improve the accuracy of conversion rate predictions.

[0033] Step S5300: Generate the corresponding second user feature vector and second product feature vector from the corresponding amplified coding network within the conversion index inference path. The amplified coding network has been independently trained through comparative learning in advance to acquire the corresponding feature representation ability. The augmented coding network is an additional introductory path for the transformation index of the traditional full-space multi-task model introduced in this application. It learns the corresponding feature representation capabilities through independent comparative learning training in advance, thereby enhancing the model's understanding of user and product features.

[0034] The introduction of the augmented coding network aims to improve the discriminative power of feature vectors through pre-training with contrastive learning. Contrastive learning is an unsupervised learning method that learns feature representations by constructing pairs of positive and negative samples, making similar samples closer together in the feature space and dissimilar samples further apart. In this application, the augmented coding network utilizes the rich feature representation capabilities obtained from contrastive learning training to further optimize the feature vectors of users and products.

[0035] In practice, the augmented coding network can include multiple neural network layers, such as convolutional layers, recurrent layers, or fully connected layers. These layers can be individual or in combination. For example, for a user-side augmented coding network, a convolutional layer can first extract local features from user behavior data, then a recurrent layer can capture the time-series features of user behavior, and finally a fully connected layer can integrate these features to generate a second user feature vector. Similarly, a product-side augmented coding network may use different network structures to process product data and generate a second product feature vector.

[0036] In one embodiment, each amplification coding network can be connected to a multilayer perceptron (MLP) after a Transformer encoder. In this way, the high-dimensional representation of key features is first captured by the encoder’s attention mechanism, and then projected through the MLP to obtain a low-dimensional representation that matches the dimension of the feature vector output by the built-in encoder, so as to facilitate the fusion of subsequent feature vectors.

[0037] During the training of the augmented coding network, a large amount of user and product data is used to construct anchor samples and corresponding positive and negative sample pairs. Anchor samples can be actual user data and product data of products that have been clicked and converted. Positive samples are obtained by augmenting the product data and / or user data based on the anchor samples, while maintaining the association between user behavior and actual access to the corresponding products. Negative samples, on the other hand, have no actual user access relationship with the product data. In this way, the user and product feature vectors learned by the augmented coding network can better reflect users' potential interest in products and the attractiveness of the products.

[0038] In practical applications, the training of augmented coding networks can employ various contrastive learning strategies, such as the InfoNCE loss function and the Triplet loss function. These strategies optimize feature representations by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs. For example, the InfoNCE loss function optimizes feature vectors by calculating the inner product of the feature vectors of positive sample pairs, dividing it by the temperature parameter, and then comparing it with the inner product of the feature vectors of other sample pairs.

[0039] Accordingly, the augmented coding network can generate high-quality second user feature vectors and second product feature vectors. These feature vectors are combined with the first feature vectors generated by the built-in coding network to provide the prediction network with more comprehensive and in-depth information, thereby improving the accuracy of the model in predicting conversion index data.

[0040] Step S5400: The prediction network in the conversion index inference path merges the feature vectors of each user and each product into a comprehensive feature vector, and predicts the corresponding conversion index data based on the comprehensive feature vector. The prediction network is determined by fine-tuning training after the trained amplified coding network is added to the conversion index inference path. The prediction network is an inherent module in the full-space multi-task model. This network includes fusion layers and fully connected layers, which are used to transform all feature vectors from the user side and the product side into the final conversion prediction result.

[0041] The fusion layer is responsible for integrating the first and second user feature vectors from the user side, and the first and second product feature vectors from the product side, into a single comprehensive feature vector. This can be achieved by concatenating or weighted summing of all feature vectors. The goal is to merge multi-dimensional information from users and products to form a feature representation that comprehensively represents a user's potential interest in a product.

[0042] Following the fusion layer, the fully connected layer further abstracts and processes the comprehensive feature vector through a series of neural network layers. A fully connected layer can contain one or more hidden layers, each consisting of multiple neurons connected to all neurons in the previous layer via weights. In the fully connected layer, the weight parameters learned by the model map the comprehensive feature vector to the predicted conversion metric, thereby outputting conversion-related result data, which theoretically corresponds to the conversion rate (CVR).

[0043] Based on the inherent implementation principle of the complete spatial multi-task model, the output of the prediction network in the conversion metric inference path is fused with the click-related result data output by the prediction network in the click metric inference path, ultimately outputting click-conversion related result data. The click-related result data theoretically corresponds to the click-through rate (CTR), while the final output click-conversion related result data corresponds to the click-through rate-conversion rate (CTCVR). Therefore, the final output is the model's probability estimate of a user's purchase behavior after clicking, i.e., the click-through rate-conversion rate (CTCVR). Those skilled in the art can flexibly use any one or more of the various metric data, namely CTR, CVR, and CTCVR. However, this embodiment mainly focuses on CVR and CTCVR. In particular, CTCVR can be obtained by dividing CTR, because the CVR obtained in this way is more accurate due to considering the influence of the click metric inference path branch.

[0044] Step S5500: For the multiple listed products on the current independent website, select some listed products as recommended products and push them to the current user based on their corresponding conversion index data.

[0045] To apply conversion rate (CVR) data to the actual product recommendation process, the conversion rate data of each listed product retrieved from the independent website's product database can be determined first, and then the products can be filtered and pushed based on this data.

[0046] When sorting listed products based on their conversion metrics, a threshold can be set; only products with conversion metrics exceeding this threshold will be selected as recommended products. Alternatively, a sorting mechanism can be used to rank products from highest to lowest conversion metrics, then selecting the top-ranked products as recommended products. These recommended products together form a recommendation list, which is then pushed to users to fulfill their product recommendation requests.

[0047] In some embodiments, other business factors, such as product inventory status, price sensitivity, and user's historical purchasing behavior, may also be considered when determining recommended products. For example, even if a product has high conversion rate data, it will not be selected as a recommended product if its inventory is insufficient. Similarly, for price-sensitive users, moderately priced products may be prioritized for recommendation.

[0048] Furthermore, to improve the diversity and coverage of recommendations, strategies can be introduced to balance the types of products in the recommendation list. For example, category diversity constraints can be set to ensure that the recommendation list includes products from different categories. Alternatively, a novelty metric can be introduced to prioritize recommending products that users have less exposure to or that are newly listed, thereby increasing the attractiveness of the recommendations.

[0049] After filtering the products, a final recommendation list is generated and pushed to the current user through the website's interface. Users can see these recommended products in various scenarios, such as the homepage, product detail page, or search results page, where the user triggered the recommendation request. The recommended products can be displayed in various formats, such as lists, grids, or carousels, depending on the website's design and the user's habits.

[0050] As can be seen from the above description of the embodiments, this application has many beneficial effects, including but not limited to: First, this application innovatively extends the traditional full-space multi-task model by introducing an augmented coding network trained through contrastive learning, significantly enhancing the model's feature representation capabilities. This augmented coding network, trained independently through contrastive learning, learns richer feature representations, reducing the feature similarity between similar products and increasing the feature similarity between dissimilar products. Therefore, when integrated into the conversion index inference path, it generates higher-quality second user feature vectors and second product feature vectors. These feature vectors, combined with the first user feature vector and second product feature vector generated by the built-in coding network, provide the prediction network with more comprehensive and in-depth information, enabling the model to more accurately predict product click and conversion-related index data. This optimizes the ranking criteria for product recommendations, making product recommendations more effective.

[0051] Secondly, this application specifically addresses the data sparsity problem in product recommendation scenarios on independent websites, proposing an effective solution. By adding an augmented encoding network to the full-space multi-task model and training it through contrastive learning, the model parameters can still achieve sufficient convergence even with short user behavior sequences and low product exposure frequency. This not only improves the model's embedding representation ability for long-tail products but also enhances the discriminative power of user-side embeddings, thereby effectively overcoming the challenges posed by data sparsity and improving the product recommendation performance on independent websites.

[0052] Furthermore, this application comprehensively enhances the product recommendation experience on independent websites. By accurately predicting conversion metrics, it can more effectively filter out products with high conversion potential and push these products to users. This not only improves user satisfaction but also increases the conversion rate and sales of independent websites. Simultaneously, it ensures the novelty and coverage of recommendation results, enabling users to discover more interesting and relevant products, thereby enhancing the overall shopping experience.

[0053] Based on any embodiment of the method in this application, before responding to the current user's product recommendation request on the current independent website, the method includes: Step S4100: Use user feature data and product sequence data from the training dataset to construct user data as training samples, and perform contrastive learning training on the corresponding first amplification coding network on the user side until convergence; Step S4200: Construct commodity data as training samples using commodity feature data from the training dataset, and perform contrastive learning training on the corresponding second amplification coding network on the commodity side until convergence; Step S4300: Connect each amplified coding network to the prediction network in the transformation index inference path of the pre-trained complete spatial multi-task model, grant the weight update permission of the prediction network, and freeze the weight update permission of the others. Step S4400: Construct user data and product data as training samples using the training dataset. Determine the supervision label of the training sample based on whether there is a factual click-through conversion relationship between the user to which the user data belongs and the product to which the product data belongs. Fine-tune the complete spatial multi-task model after connecting to each amplified coding network until convergence.

[0054] Based on any embodiment of the method in this application, user data as training samples is constructed using user feature data and product sequence data from the training dataset, and contrastive learning training is performed on the corresponding first amplification coding network on the user side until convergence, including: Step S4110: Construct the first anchor sample from the user feature data in the training dataset and the product sequence visited by the source users; User profile data can include basic user information such as age, gender, and interests, as well as behavioral data such as browsing history, search history, and purchase history. A list of products visited or interacted with by a user on an independent website reflects the user's interests and purchasing tendencies.

[0055] Anchor samples serve as baseline samples in contrastive learning, used to generate subsequent positive and negative samples so that the model learns the correct association between users and products. When constructing the first anchor sample, each user's feature data is associated with the corresponding sequence of products that user has actually visited, forming the corresponding anchor sample.

[0056] Step S4120: Perform data augmentation on the pervasive product sequence based on the first anchor point sample to obtain the first positive sample; Data augmentation of the pervasive product sequence in the first anchor sample can expand the first positive sample. Data augmentation can expand the training dataset, thereby improving the model's generalization ability.

[0057] In one embodiment, data augmentation can be achieved by perturbing the sequence of visited products. For example, the order of products in the sequence can be adjusted, or augmented samples can be generated by randomly replacing some products in the sequence with other similar products. Similar products can be determined based on product category, attributes, or user behavior data. For example, if a user's sequence of visited products includes sneakers, another pair of sneakers can be randomly selected to replace one sneaker in the sequence to generate a positive sample.

[0058] In another embodiment, data augmentation can be achieved by expanding the sequence of visited items. This can be done by adding other items that match the user's behavioral patterns. For example, if a user frequently purchases sports equipment, other types of sports equipment, such as sportswear or protein powder, can be added to the sequence to generate positive samples.

[0059] In another embodiment, data augmentation can be achieved by compressing the sequence of visited items. This can be done by removing some items from the sequence while preserving the user's primary interests and behavioral patterns. For example, if the sequence of items visited by a user is too long, some less relevant items can be removed, retaining the items the user is most likely to be interested in, to generate positive samples.

[0060] In all these embodiments, the goal of data augmentation is to generate positive samples that are semantically similar to the original anchor samples but differ in details. This helps the model learn richer and more diverse feature representations, thereby improving the model's understanding of user interests and product features.

[0061] Step S4130: Associate the source user of the user feature data in the first anchor point sample with the product sequence visited by other users to construct the first negative sample; Negative samples are crucial for the model to learn to distinguish the relationship between different users and products, and help improve the model's ability to identify user preferences. Various implementation methods can be used to construct the first negative sample.

[0062] In one embodiment, when constructing the first negative sample, the source user of the user feature data in the first anchor sample is associated with the product visit sequences of other users, but not with the product visit sequences of the source user itself. Specifically, other users different from those in the anchor sample can be selected from the training dataset, and then their product visit sequences are paired with the user feature data in the anchor sample to form the first negative sample. This pairing method ensures that there is no actual interaction between the user and the product in the negative sample, thus providing the model with correct negative feedback.

[0063] In another embodiment, a first negative sample can be constructed by randomly selecting certain items from the item sequence and replacing them with other randomly selected items. This replacement strategy can be completely random or based on certain rules, such as item category or price range. For example, if the item sequence in the anchor sample contains sports equipment, then items can be randomly selected from non-sports equipment categories for replacement to generate a negative sample.

[0064] In another embodiment, the first negative sample can be constructed by perturbing the product sequence. This perturbation could be changing the order of products in the sequence or inserting products unrelated to the user's historical behavior. For example, if a user's historical behavior shows an interest in electronics, then household items could be inserted into their visited product sequence to construct a negative sample.

[0065] In all these embodiments, the core objective of constructing the first negative sample is to generate a sample that is semantically dissimilar to the anchor sample but may be similar in form. This helps the model learn a more accurate representation of the user-item relationship, thereby improving the model's accuracy in predicting potential user interests.

[0066] Step S4140: Use the first anchor sample, the first positive sample, and the first negative sample as training samples to construct a batch training sample set, input it into the first amplified coding network to perform contrastive learning training and perform weight updates.

[0067] The first anchor point sample, the first positive sample, and the first negative sample are used as training samples. These samples are organized into a batch training sample set and used to input the first amplified coding network for comparative learning training. During this process, weight updates are implemented to improve the model's learning effect on the relationship between users and products in user data by optimizing the network weights.

[0068] In one embodiment, each anchor sample can be associated with one or more positive samples and one or more negative samples to form a batch of training samples for training. In the same batch of training samples, the anchor samples, positive samples and negative samples are pre-established with correlations so that comparative learning can be performed based on these correlations during training.

[0069] In another embodiment, the batch training sample set may include a mixture of anchor samples, positive samples, and negative samples that have established associations, in a certain proportion. For example, a batch may contain 60% anchor samples, 30% positive samples, and 10% negative samples. This mixing ratio can be adjusted according to the actual training effect to achieve the best learning effect. In this way, the model can simultaneously learn the similarity of positive samples and the differences of negative samples in each training batch, thereby improving the quality of feature representation.

[0070] In another embodiment, the batch training sample set can be constructed using a dynamic sampling method. Dynamic sampling means adjusting the ratio of positive to negative samples during training based on the model's performance in the current batch. For example, if the model performs poorly in distinguishing a particular type of product, the number of negative samples for that type of product can be increased to enhance the model's ability to identify such products.

[0071] By training the user-side augmented coding network through contrastive learning using multiple batches of training samples, the network can optimize its weights by minimizing the distance between anchor samples and positive samples while maximizing the distance between anchor samples and negative samples. This is typically achieved by calculating a loss function designed to reflect the model's performance in distinguishing between positive and negative samples. For example, triplet loss functions or InfoCE loss functions can be used, which encourage the model to bring anchor samples closer to positive samples while pushing them further away from negative samples.

[0072] Furthermore, weight updates can be implemented using the backpropagation algorithm, which adjusts network weights based on the gradient of the loss function to reduce prediction errors. During training, regularization techniques such as L2 regularization or Dropout may also be employed to prevent overfitting and improve the model's generalization ability.

[0073] The above embodiments, by utilizing a constructed batch training sample set to perform contrastive learning training on the user-side augmented coding network, can significantly improve the feature representation capability of user data. Under this training method, the model, through contrastive learning between anchor samples, positive samples, and negative samples, can more accurately capture the complex relationships between users and products in the user data. The introduction of positive samples enables the model to learn to identify and reinforce signals of user preferences, while negative samples help the model distinguish irrelevant or uninteresting products, thereby enhancing the model's discriminative ability. The augmented coding network trained in this way can generate more discriminative feature vectors. These feature vectors not only reflect the user's basic attributes and behavioral patterns but also reveal the user's potential interests and purchasing tendencies, thus providing more accurate and richer information for subsequent product recommendations.

[0074] Based on any embodiment of the method in this application, commodity data as training samples is constructed using commodity feature data from the training dataset, and contrastive learning training is performed on the corresponding second amplification coding network on the commodity side until convergence, including: Step S4210: Use the product data of any product in the training dataset as the second anchor sample; Product data for any item in the training dataset is selected as the second anchor sample. This product data is typically entered by merchants when listing the product and includes basic product information such as title, description, category, price, and inventory, but excludes information about other products. Anchor samples play a benchmark role in contrastive learning, serving as the basis for generating positive and negative samples, enabling the model to learn the correct relationships between products.

[0075] In one embodiment, the selection of the second anchor sample can be based on the popularity or sales volume of the product. For example, the product with the highest sales volume can be selected as the anchor sample because these products usually have richer user interaction data, thus providing a more reliable learning signal for the model.

[0076] In another embodiment, the selection of the second anchor sample can be based on the freshness of the product, i.e., the product that has recently been added or updated. This approach helps the model learn the latest market trends and user preferences, thereby improving the timeliness of recommendations.

[0077] In another embodiment, the second anchor sample can be selected from the training dataset through random sampling. This method ensures that the model is exposed to diverse product information, thereby improving the model's generalization ability.

[0078] Step S4220: Perform data augmentation on the product data based on the second anchor point sample, or obtain the product data of co-purchased products or similar products to which the product data belongs, as the second positive sample; Data augmentation can be achieved by perturbing the product data in the second anchor sample. For example, keywords in the product description can be adjusted, or synonyms can be added to the product title, thereby generating positive samples that are semantically similar to the anchor sample but differ in details. This perturbation should keep the core features of the product data unchanged, while introducing sufficient variation to simulate different expressions that users might encounter.

[0079] In another embodiment, data augmentation can be achieved by expanding product attributes. For example, if the product in the second anchor sample is athletic shoes, other attributes related to athletic shoes, such as different colors, sizes, or brands, can be added to generate positive samples. This approach helps the model learn the correlations between product attributes, thereby improving the accuracy of recommendations.

[0080] Furthermore, obtaining data on jointly purchased or similar products as positive samples is also an effective implementation method. Jointly purchased products typically refer to other products that users frequently buy together when purchasing a particular product, while similar products refer to products with similar functions or uses. By using this product data as positive samples, and using the information on jointly purchased and / or similar products of the products accessed by the user corresponding to the second anchor sample as corresponding positive samples, the model can learn the association patterns in user purchasing behavior, thereby improving the personalization of recommendations.

[0081] Step S4230: Take the product data of other products besides the arbitrary product as the second negative sample. The other products are not co-purchased products or similar products of the arbitrary product. The selection of negative samples directly affects the ability of the second amplified coding network to learn to distinguish product features. Therefore, the second negative sample should be semantically dissimilar to the second anchor sample, but may be similar in form. This helps the model learn accurate product representations, thereby improving the accuracy of predicting users' potential interests.

[0082] In one embodiment, constructing a second negative sample can be achieved by randomly selecting product data from the training dataset that is different from the anchor sample. This product data is not directly related to the products in the anchor sample in terms of category, attributes, or user interaction behavior, thus ensuring the validity of the negative sample. For example, if the anchor sample is athletic shoes, then products from completely different categories, such as kitchen utensils, can be selected as negative samples.

[0083] In another embodiment, a second negative sample can be constructed by randomly perturbing the product attributes. This perturbation can include changing attributes such as product category, price range, or brand to generate product data that is semantically dissimilar to the anchor sample. For example, if the product in the anchor sample is high-end athletic shoes, a negative sample can be constructed by lowering the price range or changing the brand, thereby simulating products that a user is unlikely to be interested in simultaneously.

[0084] In another embodiment, constructing the second negative sample can be achieved by considering the product's sales data. Products with low sales volume and few user reviews are selected as negative samples because these products differ significantly from the products in the anchor sample in terms of user preferences. For example, if the second anchor sample is a best-selling product, then products with poor sales volume can be selected as negative samples.

[0085] Furthermore, a second negative sample can be constructed by analyzing user behavior data. For example, items that users rarely or never purchase after buying the second anchor sample item can be selected as negative samples. This approach leverages users' actual purchasing behavior to enhance the model's understanding of user preferences.

[0086] It should be noted that the second negative sample cannot be the product information of the co-purchased goods or similar goods of the second anchor sample, so as to ensure that the amplification coding network can effectively perform comparative learning.

[0087] Step S4240: Use the second anchor sample, the second positive sample, and the second negative sample as training samples to construct a batch training sample set, input it into the second amplified coding network to perform contrastive learning training and perform weight updates.

[0088] The constructed second anchor sample, second positive sample, and second negative sample are integrated into a batch training sample set. These sample sets are then input into the second amplified coding network for contrastive learning training, and weights are updated during this process. The network weights are optimized to enhance the model's ability to represent the features of commodity data.

[0089] In one embodiment, each anchor sample can be associated with one or more positive samples and one or more negative samples to form a batch of training samples for training. In the same batch of training samples, the second anchor sample, the second positive sample, and the second negative sample are pre-established with an association so that comparative learning can be performed based on this association during training.

[0090] In another embodiment, the batch training sample set integration process may include mixing anchor samples, positive samples, and negative samples in a certain proportion. For example, a batch may contain 70% anchor samples, 20% positive samples, and 10% negative samples. Such a proportion helps the model learn the relationship between positive and negative samples in a balanced way in each training step, thereby improving the model's ability to identify and distinguish product features.

[0091] In another embodiment, the batch training sample set can be constructed using a dynamic sampling method. Dynamic sampling means adjusting the ratio of positive to negative samples based on the model's performance in the current training phase. For example, if the model performs poorly in identifying a certain type of product, the number of negative samples for that type of product can be increased to improve the model's ability to distinguish between them.

[0092] In all these embodiments, the goal of contrastive learning training is to optimize the weights of the augmented encoding network by minimizing the distance between anchor samples and positive samples while maximizing the distance between anchor samples and negative samples. This is typically achieved by computing a loss function designed to reflect the model's performance in distinguishing between positive and negative samples. For example, a triplet loss function or an InfoNCE loss function can be used, which encourages the model to bring anchor samples closer to positive samples while pushing them further away from negative samples.

[0093] Furthermore, weight updates are implemented through backpropagation, an algorithm that adjusts network weights based on the gradient of the loss function to reduce prediction errors. During training, regularization techniques, such as L2 regularization or Dropout, may also be employed to prevent overfitting and improve the model's generalization ability.

[0094] Through the above embodiments, this application ensures that the product-side amplification coding network effectively learns the rich features of product data during training, thereby providing high-quality feature representations for subsequent product recommendations. These feature representations not only capture the basic attributes of products but also reflect their attractiveness and market competitiveness, providing strong support for product recommendation services on independent websites.

[0095] Based on any embodiment of the method in this application, the complete spatial multi-task model after accessing each amplified coding network is fine-tuned and trained until convergence, including: Step S4310: Input multiple training samples consisting of user data and product data into the conversion index inference path and click index inference path of the complete spatial multi-task model in batches to obtain the click conversion index classification result output after the inference fusion of each inference path; After the complete spatial multi-task model is connected to the amplified coding network, it needs to be fine-tuned to adapt to the specific task requirements of this application. In this step, multiple training samples consisting of user data and product data are input into the conversion index inference path and click index inference path of the model in batches. Through these two inference paths, the fused output result, i.e., the click-conversion index classification result, is obtained.

[0096] The conversion metric inference path and the click metric inference path are core components of the full-space multi-task model, responsible for predicting conversions and clicks, respectively. By integrating augmented encoding networks on both the user and product sides, the conversion metric inference path can leverage richer feature representations to improve prediction accuracy.

[0097] Each batch of training samples input into the model undergoes processing through both conversion and click-through rate inference paths, ultimately generating a click-to-conversion rate (CTCVR) classification result. Since the two paths predict conversion rate and click-through rate separately and then merge to output the same result data, this classification result is actually the CTCVR, reflecting the model's probability estimate of a user's purchase behavior after clicking. It is a key output for model training and optimization. In this way, the entire model can be trained across the complete exposure space, thereby improving its performance and accuracy in independent website product recommendation scenarios.

[0098] Step S4320: Calculate the loss value of the classification result using the supervision label corresponding to each training sample, and perform backpropagation based on the loss value to correct the weight parameters of the prediction network in the conversion index inference path; The loss function used to determine the loss value consists of two parts: similarity loss and cross-loss. These two parts are weighted to determine the loss value corresponding to each batch, i.e., the overall loss value. The similarity loss value is calculated based on the similarity between the feature vector output by the built-in encoding network and the feature vector output by the augmented encoding network in the conversion metric inference path. This calculation of the similarity loss value helps the model learn the correct association between users and products, thereby improving prediction accuracy. The cross-loss value is determined based on the supervised labels of each training sample; it reflects the difference between the model's predicted click-to-conversion metric and the actual labels.

[0099] In this application, the integrated formula for the overall loss value can be expressed as follows: in, Represents the similarity loss value. The cross-loss value is calculated and determined based on the loss function originally applied to the full-space multi-task model, such as the cross-entropy function. and These are preset weights used to balance the importance of the two types of loss in the total loss.

[0100] In the actual calculation process, the similarity loss value for each training sample is first calculated, which can be achieved by calculating the cosine similarity or Euclidean distance between feature vectors. Then, the cross-loss value is calculated based on the supervision labels, which can be done using common cross-entropy loss functions or mean squared error functions. Finally, the similarity loss value and the cross-loss value are weighted and summed according to preset weights to obtain the overall loss value for each batch.

[0101] During training, the model continuously adjusts its weight parameters based on the overall loss value of each batch to minimize the overall loss. The backpropagation algorithm updates the weights in the network based on the gradient information of the loss function, thereby making the model's predictions gradually approach the true labels.

[0102] Step S4330: When the preset conditions are met, confirm that the model has converged and terminate the model training; otherwise, iteratively call other batches of training sample sets to continue training. The preset conditions include the loss value meeting the preset threshold condition or the number of iterations meeting the preset quantity condition.

[0103] The convergence test of model training determines when to terminate the training. This can be achieved by applying pre-defined conditions. Optionally, these conditions may include a pre-defined threshold for the loss value or a pre-defined number of iterations. When these conditions are met, the model can be considered converged, meaning its parameters have been sufficiently optimized, and further training is unlikely to lead to significant performance improvements.

[0104] In one embodiment, the preset threshold condition can be that the decrease in the loss value over multiple consecutive iterations is less than a specific decimal, such as 0.001 or less. This indicates that the model's predictive performance has stabilized, and the impact of parameter updates on the loss value becomes negligible, thus training can be stopped to avoid overfitting.

[0105] In another embodiment, the preset iteration count condition can be that the model has completed a predetermined number of iterations, such as 1000 or more. This ensures that the model has sufficient training time to learn patterns in the data, and even if the decrease in the loss value may not reach the threshold condition, reaching a certain number of iterations is still an indicator that the model is sufficiently trained.

[0106] If the model fails to converge after being judged by the preset conditions, the weights in step S4320 can be updated, and then the process returns to step S4310 to continue iteratively training the model using the next batch of training sample sets. If the model has converged, the training is terminated.

[0107] During implementation, these preset conditions can be determined by monitoring changes in the loss value during training and performance on the validation set. Once any condition is met, the training process can be terminated, the model parameters saved, and the model ready to be deployed to the production environment for actual product recommendation tasks.

[0108] In the above embodiments, the full-space multi-task model, after being connected to the augmented coding network and fine-tuned until convergence, significantly improves its feature representation capabilities, thus achieving technical advantages in several aspects. First, the augmented coding network enhances the model's ability to distinguish between user and product features through contrastive learning training, enabling the model to capture more subtle and richer feature information. Second, the fine-tuning training process allows the model to adapt to specific task requirements, optimizing the predictive capabilities of conversion rate inference paths and click rate inference paths, thereby improving the accuracy of click-through conversion rate (CTCVR) estimation. Furthermore, by fusing the feature vectors output by the augmented coding networks from the user and product sides, the model can output more accurate click-through conversion rate classification results, which directly reflect the probability of a user making a purchase after clicking. This contributes significantly to the performance and accuracy of product recommendation in independent website scenarios. When using the corresponding indicator data for product recommendation ranking, the product recommendation results become more reliable and effective.

[0109] Based on any embodiment of the method in this application, the loss value of the classification result is calculated using the supervision labels corresponding to each training sample, including: Step S4321: Calculate the similarity loss value between the feature vector output by the built-in encoding network in the transformation index inference path and the comprehensive feature vector after fusing the feature vector output by the amplified encoding network. Since the conversion index inference path of the complete spatial multi-task model introduces two amplified coding networks, a similarity loss value can be introduced to measure the impact of these two amplified coding networks when calculating the loss value. Specifically, the similarity loss value can be calculated between the feature vectors output by the built-in coding networks in the conversion index inference path and the combined feature vector. The feature vectors output by the two built-in coding networks include a first user feature vector and a first product feature vector, while the combined feature vector is obtained by fusing the second user feature vector and the second product feature vector output by the two amplified coding networks based on the aforementioned feature vectors. The similarity loss value aims to measure the degree of similarity between the fused feature vector and the original feature vector, thereby ensuring that the amplified coding networks do not introduce information that contradicts the original feature representation.

[0110] Specifically, the similarity loss value can be calculated in several ways. For example, cosine similarity can be used to measure the angle between two feature vectors; the closer the value is to 1, the more similar the two vectors are. Alternatively, Euclidean distance can be used to measure the straight-line distance between two vectors; the smaller the distance, the more similar the two vectors are. In this embodiment, the formula for calculating the similarity loss value can be expressed as: in, Indicates the first The comprehensive feature vector of each sample Indicates the first The original feature vectors of each sample, namely the first user feature vector and the first product feature vector. This refers to the batch size. The formula calculates the similarity between the integrated feature vector and the original feature vector. The similarity loss value is obtained by subtracting the similarity score from 1. The smaller the loss value, the better the feature fusion effect.

[0111] The calculation of this similarity loss helps the model maintain feature consistency during training, avoiding interference from the augmented coding network on the original feature representation. By minimizing the similarity loss, the model can learn more stable and reliable feature representations, thereby improving prediction accuracy and the model's generalization ability.

[0112] In practical applications, the calculation of similarity loss can be adjusted based on the specific feature vector representation and business requirements. For example, if the feature vectors are highly dimensional and sparse, a similarity metric more suitable for sparse data may be needed. Furthermore, the calculation of similarity loss can be combined with other types of losses (such as cross-loss) to achieve multi-objective optimization of model training. In this way, the model can not only learn the correct association between users and products but also maintain the consistency and stability of feature representations, thereby improving the performance of the entire recommendation system.

[0113] Step S4322: After weighting the similarity loss value according to the preset weight, add it to the cross loss value determined based on the supervision label, and use it as the loss value of the current batch.

[0114] For each training batch, an overall loss value is determined. The overall loss value is calculated according to the formula revealed earlier. That is, by associating preset weights, the similarity loss value and the cross loss value calculated by the full-space multi-task model according to its inherent principles are weighted and summed to obtain the overall loss value. Since the relevant formula has been revealed earlier, it will not be repeated here.

[0115] Through the above embodiments, the model continuously adjusts the weight parameters according to the overall loss value during training to minimize the overall loss value. Then, the backpropagation algorithm updates the weights in the network according to the gradient information of the loss function, so that the model's prediction results gradually approach the true labels. It can not only learn the correct association between users and products, but also maintain the consistency and stability of feature representation, thereby improving the performance of the entire recommendation system.

[0116] Based on any embodiment of the method in this application, the corresponding conversion index data is predicted based on the comprehensive feature vector, including: Step S5410: Obtain the click index data output by the click index inference path of the complete spatial multi-task model and the click conversion index data determined by fusing the prediction results of the conversion index inference path and the click index inference path; As revealed above, the complete space multi-task model determines click-to-conversion index data by fusing the prediction results of the conversion index inference path and the click index inference path. Therefore, the final classification result of the entire model is actually a click-to-conversion rate (CTCVR) that takes into account the prediction results of the click index inference path, i.e., the click-through rate (CTR), and the prediction results of the conversion index inference path, i.e., the conversion rate.

[0117] Specifically, the model's click metric inference path is responsible for predicting the click-through rate (CTR) of each item, i.e., the probability that a user will click on an item. This path uses user feature vectors and item feature vectors to predict the likelihood of a user clicking on each item. Click metric data reflects the degree of user interest in items and is one of the commonly used metrics in recommendation systems.

[0118] The conversion metric inference path is responsible for predicting the conversion rate of each product, i.e., the probability of a user making a purchase after clicking on it. This path also uses user feature vectors and product feature vectors to predict user purchasing behavior. Unlike the click metric inference path, this path includes feature vectors generated using an amplified coding network, in addition to the model's inherent feature vectors, ultimately merging them into a single comprehensive feature vector. Conversion metric data is an important indicator for evaluating product conversion potential and is significant for increasing sales and optimizing inventory management.

[0119] After the two paths infer their respective prediction results from their respective feature vectors through their respective prediction networks, the model then merges the prediction results of the two paths to determine the final click-to-conversion metric data, which serves as the click-to-conversion rate (CTCVR).

[0120] Step S5420: Divide the click conversion index data by the click index data to obtain the conversion index data.

[0121] In the theoretical framework of the complete space multi-task model, the click-to-conversion rate (CTR) is the product of the click-through rate (CTR) and the conversion rate (CVR). Therefore, by dividing the click-to-conversion metric data (CTCVR) by the click-through rate (CTR), the conversion metric data (CVR) can be obtained, serving as a representation of the conversion rate (CVR). The CTR data is determined by the prediction results of the complete space multi-task model, which fuses the conversion metric inference path and the click metric inference path, representing the probability that a user will make a purchase after clicking on a product. The click metric data is output by the click metric inference path, reflecting the probability that a user clicks on a product. Both the CTR and CTR are explicitly provided with corresponding supervision labels during the training phase, while the conversion rate is implicitly learned during training. Therefore, in this embodiment, it is more accurate to indirectly determine the conversion rate using the CTR and CTR. It should be noted that in another embodiment of this application, the online inference phase can also use the conversion metric inference path alone, omitting the click metric inference path and directly using the conversion rate determined by the conversion metric inference path instead.

[0122] Specifically, click-to-conversion (CPC) metrics can be represented by the model's prediction of the probability that a user will make a purchase after clicking, while click metrics represent the model's prediction of the probability that a user will click on a product. By dividing the CPC metrics by the click metrics, we obtain the conversion metrics, which are the percentage of users who actually make a purchase after clicking on a product. This ratio reflects the product's conversion rate and is an important indicator for evaluating product attractiveness and user purchase intent.

[0123] The above embodiments, during the online inference phase, achieve several technical advantages by indirectly calculating conversion rate (CVR), primarily in the following aspects: First, they leverage the predictive capabilities of a full-space multi-task model, determining the click-to-conversion rate (CTCVR) by fusing the prediction results of the conversion rate inference path and the click rate inference path. This fusion mechanism allows the model to simultaneously consider user click behavior and purchase behavior, thus providing a more comprehensive assessment of product attractiveness. Second, by indirectly obtaining the conversion rate by dividing the CTCVR data by the click rate data, the probability of a user actually making a purchase after clicking can be more accurately reflected. This helps independent websites more accurately assess the conversion potential of products and optimize inventory management and marketing strategies. Furthermore, it avoids the errors that may arise from directly predicting conversion rates because it utilizes the more readily available and stable click rate to assist in the calculation, thereby improving the stability and reliability of conversion rate estimation. Finally, it enhances the model's robustness in data-sparse environments, especially in environments like independent websites where user behavior data is relatively limited. By using the click rate as a mediating variable, it effectively mitigates the impact of data sparsity on the accuracy of conversion rate prediction, thereby enhancing the model's generalization ability and recommendation performance.

[0124] Please seeFigure 3 According to one aspect of this application, an independent website product recommendation device includes a request response module 5100, a built-in representation module 5200, an augmented representation module 5300, an indicator prediction module 5400, and a filtering and recommendation module 5500. The request response module 5100 is configured to respond to a product recommendation request triggered by a current user on the current independent website, and acquire user data of the current user and product data of products currently listed on the independent website. The built-in representation module 5200 is configured to input the user data and the product data into a preset complete spatial multi-task model, and generate corresponding first user feature vectors and first product feature vectors from the corresponding built-in encoding network in the conversion indicator inference path within the model. The augmented representation module 5300 is configured to generate corresponding first user feature vectors and first product feature vectors from the conversion... The corresponding amplified coding network within the indicator inference path generates a corresponding second user feature vector and a second product feature vector. The amplified coding network is trained independently beforehand through comparative learning to acquire the corresponding feature representation capabilities. The indicator prediction module 5400 is configured to use the prediction network in the conversion indicator inference path to fuse each user feature vector and each product feature vector into a comprehensive feature vector, and predict the corresponding conversion indicator data based on the comprehensive feature vector. The prediction network is determined after fine-tuning training following the addition of the trained amplified coding network to the conversion indicator inference path. The filtering and recommendation module 5500 is configured to select a portion of the listed products from the current independent website as recommended products and push them to the current user based on their corresponding conversion indicator data.

[0125] Based on any embodiment of the device in this application, prior to the request response module 5100, this device further includes: a user-side training module, configured to construct user data as training samples using user feature data and product sequence data in the training dataset, and perform contrastive learning training on the corresponding first amplified coding network on the user side until convergence; a product-side training module, configured to construct product data as training samples using product feature data in the training dataset, and perform contrastive learning training on the corresponding second amplified coding network on the product side until convergence; a network configuration module, configured to connect each amplified coding network to the prediction network in the conversion index inference path within the pre-trained complete spatial multi-task model, grant the weight update permission of the prediction network, and freeze the other weight update permissions; and a fine-tuning training module, configured to construct user data and product data as training samples using the training dataset, determine the supervision label of the training sample based on whether there is a factual click conversion relationship between the user to which the user data belongs and the product to which the product data belongs, and perform fine-tuning training on the complete spatial multi-task model after connecting each amplified coding network until convergence.

[0126] Based on any embodiment of the apparatus in this application, the user-side training module includes: a first anchor sample construction module, configured to construct a first anchor sample by combining user feature data in the training dataset with the pervasive product sequences of its source users; a first positive sample construction module, configured to perform data augmentation on the pervasive product sequences in the first anchor sample to obtain a first positive sample; a first negative sample construction module, configured to associate the source users of the user feature data in the first anchor sample with the pervasive product sequences of other users to construct a first negative sample; and a user-side batch training module, configured to construct a batch training sample set by using the first anchor sample, the first positive sample, and the first negative sample as training samples, inputting it into the first amplified coding network to perform contrastive learning training and weight updates.

[0127] Based on any embodiment of the device in this application, the product-side training module includes: a second anchor sample construction module, configured to use product data of any product in the training dataset as a second anchor sample; a second positive sample construction module, configured to perform data augmentation on the product data based on the second anchor sample, or to obtain product data of co-purchased products or similar products to which the product data belongs, as a second positive sample; a second negative sample construction module, configured to use product data of other products besides the arbitrary product as a second negative sample, wherein the other products are not co-purchased products or similar products of the arbitrary product; and a product-side batch training module, configured to use the second anchor sample, the second positive sample, and the second negative sample as training samples to construct a batch training sample set, input it into the second amplification coding network to perform contrastive learning training and perform weight updates.

[0128] Based on any embodiment of the device in this application, the fine-tuning training module includes: an inference prediction module, configured to input multiple training samples consisting of user data and product data into the conversion index inference path and click index inference path of the complete spatial multi-task model in batches, and obtain the click conversion index classification result output after inference fusion of each inference path; a back-update module, configured to calculate the loss value of the classification result using the supervision label corresponding to each training sample, perform back propagation based on the loss value, and correct the weight parameters of the prediction network in the conversion index inference path; and an iterative decision module, configured to confirm model convergence and terminate model training when a preset condition is met, otherwise iteratively call other batches of training sample sets to continue training, wherein the preset condition includes the loss value meeting a preset threshold condition or the number of iterations meeting a preset quantity condition.

[0129] Based on any embodiment of the device in this application, the reverse update module includes: a similarity loss calculation module, configured to calculate the similarity loss value between the feature vector output by the built-in encoding network in the transformation index inference path and the comprehensive feature vector after the feature vector is fused with the feature vector output by the amplified encoding network; and a comprehensive loss calculation module, configured to weight the similarity loss value according to a preset weight and add it to the cross-loss value determined based on the supervision label, as the loss value of the current batch.

[0130] Based on any embodiment of the device in this application, the indicator prediction module 5400 includes: a result data acquisition module, configured to acquire click indicator data output by the click indicator inference path of the complete spatial multi-task model and click conversion indicator data determined by fusing the prediction results of the conversion indicator inference path and the click indicator inference path; and a conversion indicator calculation module, configured to divide the click conversion indicator data by the click indicator data to obtain the conversion indicator data.

[0131] Another embodiment of this application also provides an independent website product recommendation device. For example... Figure 4 The diagram shows the internal structure of an independent website product recommendation device. This device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable, non-volatile storage medium stores an operating system, a database, and computer-readable instructions. The database can store information sequences, and when executed by the processor, these computer-readable instructions enable the processor to implement an independent website product recommendation method.

[0132] The processor of this independent website product recommendation device provides computing and control capabilities, supporting the operation of the entire device. The device's memory can store computer-readable instructions, which, when executed by the processor, cause the processor to perform the independent website product recommendation method of this application. The device's network interface is used for communication with a terminal.

[0133] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the independent website product recommendation device to which the present application is applied. The specific independent website product recommendation device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0134] In this embodiment, the processor is used to execute... Figure 3The specific functions of each module are described, and the memory stores the program code and various data required to execute the aforementioned modules or sub-modules. The network interface is used to enable data transmission between user terminals or the server. In this embodiment, the non-volatile readable storage medium stores the program code and data required to execute all modules in the independent website product recommendation device of this application. The server can call the server's program code and data to execute the functions of all modules.

[0135] This application also provides a non-volatile readable storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the independent website product recommendation method of any embodiment of this application.

[0136] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the method described in any embodiment of this application.

[0137] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a computer-readable storage medium such as a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM).

Claims

1. A method for recommending products on an independent website, characterized in that, include: Respond to the product recommendation request triggered by the current user on the current independent website, and obtain the user data of the current user and the product data of the products listed on the current independent website; The user data and the product data are input into a preset complete spatial multi-task model. The corresponding first user feature vector and first product feature vector are generated by the built-in encoding network in the conversion index inference path of the model. The corresponding amplified coding network within the conversion index inference path generates the corresponding second user feature vector and second product feature vector. The amplified coding network is independently trained in advance to acquire the corresponding feature representation ability. The prediction network in the conversion index inference path fuses the feature vectors of each user and each product into a comprehensive feature vector, and predicts the corresponding conversion index data based on the comprehensive feature vector. The prediction network is determined by fine-tuning training after the trained amplified coding network is added to the conversion index inference path. For the multiple products listed on the current independent website, some of the listed products are selected based on their corresponding conversion index data and pushed to the current user as recommended products.

2. The independent website product recommendation method according to claim 1, characterized in that, Before responding to the current user's product recommendation request on the current independent website, the following steps are included: User data, which is used as training samples, is constructed using user feature data and product sequence data from the training dataset. The corresponding first amplification coding network on the user side is then trained through contrastive learning until convergence. Product data is constructed using product feature data from the training dataset as training samples, and contrastive learning training is performed on the corresponding second amplification coding network on the product side until convergence. Each amplified coding network is connected to the prediction network in the transformation index inference path of the pre-trained full-space multi-task model. The weight update permission of the prediction network is granted, while the weight update permission of the other networks is frozen. User data and product data are constructed as training samples using a training dataset. The supervision label of the training sample is determined based on whether there is a factual click-to-conversion relationship between the user to which the user data belongs and the product to which the product data belongs. The complete spatial multi-task model connected to each amplified coding network is then fine-tuned and trained until convergence.

3. The independent website product recommendation method according to claim 2, characterized in that, User data, constructed as training samples, is generated using user feature data and product sequence data from the training dataset. Contrastive learning training is then performed on the corresponding first amplification coding network on the user side until convergence, including: The user feature data in the training dataset and the product sequence visited by the source users are used to construct the first anchor sample; Data augmentation is performed on the pervasive product sequence based on the first anchor point sample to obtain the first positive sample; The first negative sample is constructed by associating the source user of the user feature data in the first anchor sample with the product sequence visited by other users. The first anchor sample, the first positive sample, and the first negative sample are used as training samples to construct a batch training sample set, which is then input into the first amplified coding network to perform comparative learning training and weight updates.

4. The independent website product recommendation method according to claim 2, characterized in that, Product data is constructed using product feature data from the training dataset as training samples. Contrastive learning training is then performed on the corresponding second amplification coding network on the product side until convergence, including: Use the product data of any product in the training dataset as the second anchor sample; Data augmentation is performed on the product data based on the second anchor sample, or product data of co-purchased products or similar products of the product to which the product data belongs is obtained as the second positive sample; The product data of other products besides the aforementioned arbitrary product are used as the second negative sample. These other products are not co-purchased products or similar products of the aforementioned arbitrary product. The second anchor sample, the second positive sample, and the second negative sample are used as training samples to construct a batch training sample set, which is then input into the second amplified coding network to perform comparative learning training and weight updates.

5. The independent website product recommendation method according to claim 2, characterized in that, The complete spatial multi-task model, after being integrated with each amplified coding network, is fine-tuned and trained until convergence, including: Multiple training samples consisting of user data and product data are input into the conversion index inference path and click index inference path of the complete spatial multi-task model in batches to obtain the click conversion index classification result output after the inference fusion of each inference path. The loss value of the classification result is calculated using the supervision label corresponding to each training sample. Backpropagation is performed based on the loss value to correct the weight parameters of the prediction network in the inference path of the conversion index. When the preset conditions are met, the model is confirmed to have converged and the model training is terminated; otherwise, other batches of training sample sets are iteratively called to continue training. The preset conditions include the loss value meeting a preset threshold condition or the number of iterations meeting a preset quantity condition.

6. The independent website product recommendation method according to claim 5, characterized in that, The loss value of the classification result is calculated using the supervision labels corresponding to each training sample, including: The similarity loss value between the feature vector output by the built-in encoding network in the inference path of the conversion index and the comprehensive feature vector after fusing the feature vector output by the amplified encoding network; The similarity loss value is weighted according to a preset weight, and then added to the cross-loss value determined based on the supervision label to obtain the loss value for the current batch.

7. The independent website product recommendation method according to any one of claims 1 to 6, characterized in that, Based on the comprehensive feature vector, the corresponding conversion index data is predicted, including: Obtain the click index data output by the click index inference path of the complete spatial multi-task model and the click conversion index data determined by fusing the prediction results of the conversion index inference path and the click index inference path; Divide the click-to-conversion metric data by the click metric data to obtain the conversion metric data.

8. A product recommendation device for independent websites, characterized in that, include: The request and response module is set to respond to product recommendation requests triggered by the current user on the current independent website, and to obtain the user data of the current user and the product data of the products listed on the current independent website; The built-in representation module is configured to input the user data and the product data into a preset full-space multi-task model, and generate the corresponding first user feature vector and first product feature vector by the corresponding built-in encoding network in the conversion index inference path within the model. The augmented representation module is configured to generate a corresponding second user feature vector and a second product feature vector by the corresponding augmented coding network within the inference path of the conversion index. The augmented coding network is trained independently in advance by comparative learning to acquire the corresponding feature representation ability. The indicator prediction module is configured to use the prediction network in the conversion indicator inference path to fuse each user feature vector and each product feature vector into a comprehensive feature vector, and predict the corresponding conversion indicator data based on the comprehensive feature vector. The prediction network is determined by fine-tuning training after the trained amplified coding network is added to the conversion indicator inference path. The filtering and recommendation module is configured to select a portion of the listed products from the multiple products listed on the current independent website and push them to the current user as recommended products based on their corresponding conversion index data.

9. A product recommendation device for independent websites, comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A non-volatile readable storage medium, characterized in that, It stores a computer program in the form of computer-readable instructions, which, when invoked by a computer, performs the steps included in the method as described in any one of claims 1 to 7.

Images