Commodity recall method and device and electronic equipment
By generating a hierarchical navigable small-world HNSW graph and a target product recall model, the problem of inaccurate product recall was solved, and the relevance of recalled products to user interests and product detail pages was realized, thereby improving the user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING WODONG TIANJUN INFORMATION TECH CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-08
AI Technical Summary
Existing product recall methods on product detail pages have issues such as recalled products not meeting user expectations and being irrelevant to the products on the product detail page, resulting in a poor user experience.
By obtaining the representation vectors of the products to be recalled, a hierarchical navigable small-world (HNSW) graph is generated, and the target products to be recalled are determined from it using the target product recall model. Fine ranking is then performed in conjunction with a deep interest network model.
This improved the accuracy and reliability of product recalls, enhanced the user experience, and ensured that recalled products aligned with user interests and were relevant to the main product on the product details page.
Smart Images
Figure CN121998675A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of information processing technology, and in particular to a product recall method, apparatus and electronic device. Background Technology
[0002] Currently, product recall scenarios for product detail pages often include recall sources such as same-product recall, similar-product recall, and multi-interest network with dynamic (MIND) recall, and these sources are often combined. However, products recalled based on these sources are not always accurate. For example, recalled products may not meet user expectations, or they may be unrelated to the products on the product detail page, resulting in a poor user experience. Therefore, improving the accuracy of product recall to enhance the user experience has become an urgent problem to be solved. Summary of the Invention
[0003] This disclosure discloses a product recall method, apparatus, electronic device, storage medium, and computer program product.
[0004] The first aspect of this disclosure proposes a product recall method, which includes: acquiring N products to be recalled, and determining a representation vector for each product to be recalled based on a pre-trained target product recall model, where N is a natural number greater than or equal to 1; layering the representation vectors of the N products to be recalled to generate a hierarchical navigable small-world (HNSW) graph, wherein the HNSW graph includes multiple layers of navigable small-world (NSW) graphs arranged vertically; and determining target products to be recalled from the products to be recalled based on the HNSW graph and the target product recall model.
[0005] In this embodiment, N products to be recalled are obtained, and a representation vector for each product to be recalled is determined based on a pre-trained target product recall model, where N is a natural number greater than or equal to 1. The representation vectors of the N products to be recalled are layered to generate a hierarchical navigable small-world (HNSW) graph. The HNSW graph includes multiple layers of navigable small-world (NNSW) graphs arranged vertically. Based on the HNSW graph and the target product recall model, target products to be recalled are determined from the products to be recalled. This disclosure, based on the HNSW graph and the target product recall model, can determine target products to be recalled from the products to be recalled. The target products not only match user interests but are also related to the main products on the product details page, improving the accuracy and reliability of determining target products to be recalled and enhancing the user experience.
[0006] A second aspect of this disclosure provides a product recall device, comprising: a first acquisition module, configured to acquire N products to be recalled and determine a representation vector for each product to be recalled based on a pre-trained target product recall model, where N is a natural number greater than or equal to 1; a second acquisition module, configured to layer the representation vectors of the N products to be recalled to generate a layered navigable small-world (HNSW) graph, wherein the HNSW graph includes multiple layers of navigable small-world (NSW) graphs arranged vertically; and a determination module, configured to determine a target product to be recalled from the products to be recalled based on the HNSW graph and the target product recall model.
[0007] A third aspect of this disclosure provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a product recall method as described in the first aspect of the present disclosure.
[0008] A fourth aspect of this disclosure provides a computer-readable storage medium storing computer instructions for causing the computer to perform a product recall method as described in the first aspect above.
[0009] A fifth aspect of this disclosure provides a computer program product including a computer program that, when executed by a processor, implements the product recall method of the first aspect of this disclosure.
[0010] Additional aspects and advantages of this disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this disclosure. Attached Figure Description
[0011] The above and / or additional aspects and advantages of this disclosure will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 A schematic flowchart of a product recall method provided in an embodiment of this disclosure; Figure 2 A schematic flowchart of a product recall method provided in another embodiment of this disclosure; Figure 3 A schematic flowchart of a product recall method provided in another embodiment of this disclosure; Figure 4 A schematic flowchart of a product recall method provided in another embodiment of this disclosure; Figure 5 This is a schematic diagram of the structure of a product recall model provided in an embodiment of the present disclosure; Figure 6This is a schematic diagram of the structure of a product recall device provided in an embodiment of the present disclosure; Figure 7 A block diagram of an electronic device provided according to an embodiment of this disclosure. Detailed Implementation
[0012] Embodiments of this disclosure are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting this disclosure.
[0013] The following description, with reference to the accompanying drawings, outlines a product recall method, apparatus, electronic device, and storage medium according to embodiments of the present disclosure.
[0014] Figure 1 This is a schematic flowchart of a product recall method provided in an embodiment of the present disclosure.
[0015] like Figure 1 As shown, the method includes the following steps: S101, obtain N products to be recalled, and determine the representation vector of each product to be recalled based on the pre-trained target product recall model, where N is a natural number greater than or equal to 1.
[0016] It should be noted that this disclosure can pre-obtain a training dataset, which includes user tags, user historical click sequences, main product features and candidate product features corresponding to the product details page. Based on the training dataset, the initial product recall model is trained to obtain a trained target product recall model.
[0017] The target product recall model includes a sequentially connected embedding layer, fusion layer, attention layer, feature cross layer, and prediction layer.
[0018] In this embodiment of the disclosure, after obtaining N products to be recalled, the representation vector of each product to be recalled can be determined based on a pre-trained target product recall model.
[0019] S102, the representation vectors of N items to be recalled are layered to generate a hierarchical navigable small-world HNSW graph, wherein the HNSW graph includes multiple layers of navigable small-world NSW graphs arranged vertically.
[0020] Hierarchical Navigable Small World (HNSW) is a graph-based data structure. The HNSW graph divides nodes into different levels and greedily traverses elements from the upper level until a local minimum is reached. Then it switches to the next level and uses the local minimum in the upper level as the new element to start traversing again until the lowest level has been traversed.
[0021] It should be noted that after obtaining the representation vector of each product to be recalled, the representation vector of each product to be recalled can be used as a node in the HNSW graph.
[0022] In this embodiment of the disclosure, after obtaining the representation vectors of N products to be recalled, the representation vectors of the N products to be recalled can be layered to generate a hierarchical navigable small world (HNSW) graph, wherein the HNSW graph includes one or more representation vectors.
[0023] It should be noted that after inserting the representation vectors of N products to be recalled as nodes into the graph, a multi-layer NSW graph arranged from top to bottom can be formed, that is, the HNSW graph includes a multi-layer NSW graph arranged from top to bottom.
[0024] Optionally, after obtaining the representation vectors of N products to be recalled, the number of layers of the representation vectors can be determined according to the number of representation vectors. For example, if the number of layers of the representation vectors is 4, then there are 4 NSW diagrams.
[0025] In this embodiment of the disclosure, the representation vector can be used as the node to be inserted, the target level into which the node to be inserted can fall is obtained, the node to be inserted is inserted at each of the current level and the upper level of the target level, and the neighboring nodes of the node to be inserted are determined in each NSW graph. The neighboring nodes are connected to the node to be inserted until the HNSW graph includes all representation vectors to generate the HNSW graph.
[0026] S103, Based on the HNSW diagram and the target product recall model, determine the target recall products from the products to be recalled.
[0027] In this embodiment of the disclosure, after obtaining the HNSW diagram and the target product recall model, the target recall product can be determined from the products to be recalled based on the HNSW diagram and the target product recall model.
[0028] Among them, the target recall product recall model can be used to determine the predicted click-through rate and post-click conversion rate for each product to be recalled.
[0029] Optionally, starting from the topmost NSW graph in the HNSW graph, the search can proceed through each NSW graph layer in a top-down order until the bottommost NSW graph is reached. Based on the predicted click-through rate and / or the predicted post-click conversion rate, the target representation vector in the bottommost NSW graph that meets the set conditions is selected, and the product to be recalled corresponding to the target representation vector is taken as the target recall product.
[0030] Optionally, after obtaining the target recall products, the target recall products can be fine-ranked, for example, by using a fine-ranking model such as the Deep Interest Network (DIN) model to sort the target recall products and recommend them to users.
[0031] In this embodiment, N products to be recalled are obtained, and a representation vector for each product to be recalled is determined based on a pre-trained target product recall model, where N is a natural number greater than or equal to 1. The representation vectors of the N products to be recalled are layered to generate a hierarchical navigable small-world (HNSW) graph. The HNSW graph includes multiple layers of navigable small-world (NNSW) graphs arranged vertically. Based on the HNSW graph and the target product recall model, target products to be recalled are determined from the products to be recalled. This disclosure, based on the HNSW graph and the target product recall model, can determine target products to be recalled from the products to be recalled. The target products not only match user interests but are also related to the main products on the product details page, improving the accuracy and reliability of determining target products to be recalled and enhancing the user experience.
[0032] Figure 2 This is a flowchart illustrating a product recall method provided in one embodiment of the present disclosure. Based on the above embodiment, it further incorporates... Figure 2 The process of determining target recall products from products to be recalled based on HNSW and the target product recall model is explained, including the following steps: S201, starting from layer h of the HNSW graph, determine the model input vector corresponding to the current layer i and input it into the target product recall model to obtain the target prediction value corresponding to the current layer i, wherein the target prediction value is at least one of the click-through rate prediction value and the post-click conversion rate value.
[0033] In this embodiment of the disclosure, for the top layer h, the representation vector in the NSW graph of the top layer h is determined as the input vector in the target product recall model; for non-top layers, the second target representation vector adjacent to the first target representation vector and the first target representation vector are determined as the input vectors in the target product recall model.
[0034] It should be noted that when obtaining adjacent representation vectors, the representation vector closest to the representation vector of the current layer entering the HNSW graph is obtained, where the distance can be the L2 distance.
[0035] S202, select the representation vector corresponding to the target predicted value that meets the set conditions as the first target representation vector, and enter the i-1 layer of the current layer. Repeat the above process until the bottom layer of the HNSW diagram, select the representation vector corresponding to the target predicted value that meets the set conditions as the target representation vector, and take the product to be recalled corresponding to the target representation vector as the target recall product.
[0036] In this embodiment of the disclosure, for each layer of the HNSW graph, the target prediction values output by each layer of the NSW graph are sorted, and the representation vectors corresponding to the top K target prediction values are used as representation vectors that satisfy the set conditions, where K is a positive integer.
[0037] In this embodiment, the first target representation vector enters the i-1 layer of the current layer. In the i-1 layer of the current layer, the second target representation vector adjacent to the first target representation vector and the first target representation vector are input to the target product recall model. The target predicted value of the product to be recalled corresponding to the first target representation vector and the second target representation vector is output. The target predicted values corresponding to the i-1 layer of the current layer are sorted, and the representation vectors corresponding to the top K target predicted values are taken as the first target representation vectors that meet the set conditions and enter the i-2 layer of the current layer. The above steps are repeated until the bottom layer of the HNSW graph is reached. The representation vector corresponding to the target predicted value that meets the set conditions is selected as the target representation vector, and the product to be recalled corresponding to the target representation vector is taken as the target recall product.
[0038] In this embodiment of the disclosure, target recall products are determined from the products to be recalled based on the HNSW graph and the target product recall model. Based on the HNSW graph and the target product recall model, the disclosure can determine target recall products from the products to be recalled. The target recall products not only match user interests but are also related to the main products on the product details page, which improves the accuracy and reliability of determining target recall products and increases the efficiency of determining target recall products. By introducing the HNSW graph, the accuracy and timeliness of target recall products are guaranteed, and the user experience is improved.
[0039] The training process of the target product recall model proposed in this disclosure will be explained below.
[0040] Figure 3 This is a flowchart illustrating a product recall method provided in one embodiment of the present disclosure. Based on the above embodiment, it further incorporates... Figure 3 The training process of the target product recall model is explained, including the following steps: S301, Obtain the training dataset, which includes user tags, user historical click sequences, main product features and candidate product features corresponding to the product details page.
[0041] It should be noted that this disclosure does not limit the specific method for obtaining the training dataset; the appropriate method can be selected based on the actual situation. Optionally, the user's historical click sequence and the main product features corresponding to the product details page can be obtained within a preset time window to determine candidate product features. The user tags, user historical click sequence, main product features corresponding to the product details page, and candidate product features can be used as training datasets.
[0042] For example, a preset time window can be set to 3 months, 2 months, etc., and 1,000 clicks within the preset time window can be selected as the user's historical click sequence.
[0043] User tags, or user profiles, can include tags such as a user's gender, age, and location.
[0044] The user's historical click sequence is the user's historical click behavior arranged in chronological order. The user's historical click sequence may include features such as the product code (Identity document, or ID) corresponding to the user's click behavior, the product category code, and timestamps.
[0045] Among them, the main product features corresponding to the product details page (hereinafter referred to as the product details page) are the product features of the current product details page. The main product features include the product code of the user's click behavior, the product category code, and the timestamp, etc.
[0046] The candidate product features include positive and negative samples corresponding to the user's current product details page. Positive samples are the products that will be clicked next in the recommendation section (positive samples). Negative samples can be randomly selected globally, for example, 20 products can be randomly selected from all products clicked by the user as negative samples. Negative samples can also be randomly selected from the same category, for example, 20 products of the same category as the positive samples can be randomly selected from all products clicked by the user as negative samples.
[0047] In this embodiment of the disclosure, the true value of the click-through rate and the true value of the post-click conversion rate corresponding to the positive sample are 1, and the true value of the click-through rate and the true value of the post-click conversion rate corresponding to the negative sample are 0.
[0048] S302, Based on the training dataset, train the initial product recall model to obtain the trained target product recall model, wherein the target product recall model is used to determine the click-through rate and post-click conversion rate of the product.
[0049] The click-through rate (CTR) of a product is the product's click-through rate.
[0050] The Click Through & Conversion Rate (CTCVR) of a product is the probability that a click on the product will result in a conversion.
[0051] In this embodiment of the disclosure, after obtaining the training dataset, the initial product recall model can be trained based on the training dataset to obtain the trained target product recall model.
[0052] Optionally, the training dataset can be input into the product recall model to obtain the predicted click-through rate (CTR) and post-click conversion rate (PCC) values of candidate products. Based on the predicted CTR and actual CTR values, as well as the predicted PCC and actual PCC values, the product recall model can be adjusted, and the adjusted product recall model can be trained until the training termination condition is met to obtain the target product recall model.
[0053] For example, the loss function of the product recall model can be obtained based on the predicted and actual click-through rates (CTRs), as well as the predicted and actual post-click conversion rates (PCCs). The model parameters of the product recall model can be updated based on the loss function, and the next training data can be used to continue training the large model with adjusted model parameters until the model training termination condition is met to obtain the target product recall model.
[0054] It should be noted that this disclosure does not impose any restrictions on the setting of the training termination conditions for the product recall model, and the training termination conditions can be set according to the actual situation.
[0055] Optionally, the model training termination condition can be set to the loss function value being less than a preset loss threshold; alternatively, the model training termination condition can also be set to the number of times the model parameters of the product recall model are adjusted reaching a preset number threshold.
[0056] In this embodiment, a training dataset is obtained, including user tags, user historical click sequences, main product features corresponding to the product details page, and candidate product features. Based on the training dataset, an initial product recall model is trained to obtain a trained target product recall model. The target product recall model is used to determine the click-through rate (CTR) and post-click conversion rate (PCC) of a product. By inputting user tags, user historical click sequences, main product features corresponding to the product details page, and candidate product features into the product recall model for training, this disclosure integrates user-side behavioral information, main product features corresponding to the product details page, and candidate product features. This improves the accuracy and reliability of obtaining the CTR and PCC values of a product, laying the foundation for more accurate determination of the target recall product in the future.
[0057] Figure 4 This is a flowchart illustrating a product recall method provided in one embodiment of the present disclosure. Based on the above embodiment, it further incorporates... Figure 4 The process of training a product recall model based on a training dataset to obtain a trained target product recall model includes the following steps: S401. Input the training dataset into the product recall model to obtain the predicted click-through rate and post-click conversion rate of candidate products.
[0058] In this embodiment, a training dataset can be input into a product recall model. Embedding vectors for various data types in the training dataset are obtained, and these embedding vectors are fused to obtain multi-dimensional representation vectors. These multi-dimensional representation vectors include a first representation vector corresponding to user tags, a second representation vector corresponding to main product features, a third representation vector corresponding to candidate product features, and a higher-order representation vector corresponding to the user's historical click sequence. Based on the second and third representation vectors, a first and second aggregation vector of the higher-order representation vectors are obtained. Feature cross-validation is performed on the second and third representation vectors to obtain a first feature cross-validation vector and a second feature cross-validation vector. Based on the second and third representation vectors, the first and second aggregation vectors, and the first and second feature cross-validation vectors, the predicted click-through rate (CTR) and post-click conversion rate (PCC) of the candidate products are output.
[0059] In this embodiment, the embedding vectors of user tags, user historical click sequences, main product features, and candidate product features are obtained. The embedding vectors of user tags, user historical click sequences, main product features, and candidate product features are fused to obtain their respective representation vectors. The embedding vectors of user tags, main product features, and candidate product features are fused to obtain their respective representation vectors. The embedding vector of the user historical click sequence is processed by Transformer and attention mechanisms to obtain the high-order representation vector corresponding to the user historical click sequence.
[0060] S402, Based on the predicted and actual click-through rates (CTRs), the predicted and actual post-click conversion rates (PCCs), the product recall model is adjusted, and the adjusted product recall model is trained until the training termination condition is met to obtain the target product recall model.
[0061] In this embodiment of the disclosure, after obtaining the predicted click-through rate (CTR) and the predicted post-click conversion rate (PCC), a first loss function of the product recall model can be determined based on the predicted CTR and the actual CTR. A second loss function of the product recall model can also be determined based on the predicted post-click conversion rate and the actual PCC. The product recall model is then adjusted based on the first and second loss functions until the training termination condition is met to obtain the target product recall model.
[0062] It should be noted that this disclosure does not limit the types of the first loss function and the second loss function, and they can be selected according to the actual situation.
[0063] Optionally, the first loss function and the second loss function can be noise contrastive estimation (NCE) loss functions.
[0064] It should be noted that this disclosure does not impose any restrictions on the setting of the training termination conditions for the product recall model, and the training termination conditions can be set according to the actual situation.
[0065] Optionally, the model training termination condition can be set to the loss function value being less than a preset loss threshold; alternatively, the model training termination condition can also be set to the number of times the model parameters of the product recall model are adjusted reaching a preset number threshold.
[0066] Optionally, real online recommendation data can be used to evaluate the effectiveness of the target product recall model, with metrics such as hit rate (HR). Multiple model evaluation results are accumulated, and the average metric of the model is calculated. When the model is used online, the model file can be updated to the online model library on a regular basis when the model metric is at or above the average level. If the model evaluation metric is below the average level, it is not updated. The training and evaluation cycle of the offline model is weekly, that is, the model is trained and evaluated every week.
[0067] The following explains the specific process of inputting the training dataset into the product recall model to obtain the predicted click-through rate and post-click conversion rate of candidate products.
[0068] For example, such as Figure 5 As shown, the product recall model includes a sequentially connected embedding layer, fusion layer, attention layer, feature cross layer, and prediction layer.
[0069] For the embedding layer of the product recall model: user tags, user historical click sequences, main product features and candidate product features can be input into the embedding layer, and the embedding layer outputs the first embedding vector of user tags, the second embedding vector of user historical click sequences, the third embedding vector of main product features and the fourth embedding vector of candidate product features.
[0070] For the fusion layer of the product recall model: the first, second, third, and fourth embedding vectors can be input into the fusion layer. The user network in the fusion layer fuses the first embedding vector to obtain the first representation vector, and the item network in the fusion layer fuses the third and fourth embedding vectors to obtain the second representation vector. and the third representation vector The higher-order representation vector of the second embedding vector is obtained by the Transformer network (sequence higher-order representation learning network) and attention mechanism in the fusion layer. .
[0071] In this embodiment of the disclosure, for higher-order representation vectors It can obtain the weight values of the attention mechanism and the weight matrix of the Transformer layer, and based on the weight values, weight matrix, and second embedding vector, obtain the higher-order representation vector of the second embedding vector. .
[0072] Optionally, a higher-order representation of click behavior in the user's historical click sequence is learned based on a Transformer network and an attention mechanism (Self-attention), with the attention mechanism network learning attention weights. Then, through weighted sum pooling, the higher-order vectors of each click in the user's historical click sequence are obtained, which are the higher-order representation vectors of the second embedding vector. .
[0073] For example, the higher-order representation vector of the second embedding vector can be obtained using the following formula:
[0074] in, The higher-order representation vector of the second embedding vector. For the second embedding vector, Here, n is the weight matrix, and n is the number of second embedding vectors. This is the weight value.
[0075] Optionally, for weight values The dimension of the second embedding vector can be obtained, wherein, in this embodiment of the disclosure, the dimension of the second embedding vector is 32, that is... The similarity between two second embedding vectors is obtained based on the dimension of the second embedding vector. The weights are then normalized using the Softmax normalization function to obtain the weight values. .
[0076] Alternatively, the weight values can be obtained using the following formula. :
[0077]
[0078] in, For weight values, For any two second embedding vectors, the similarity... For the second embedding vector, , For the weight matrix, Let n be the dimension of the second embedding vector and n be the number of second embedding vectors.
[0079] For the attention layer of the product recall model: the second representation vector can be... Third representation vector and higher-order representation vectors The input is fed into the attention layer to obtain the second representation vector. and higher-order representation vectors First attention value Third representation vector and higher-order representation vectors Second attention value And obtain the first aggregated vector of the higher-order representation vector. Second aggregation vector .
[0080] Alternatively, the second representation vector can be obtained using the following formula. and higher-order representation vectors The first attention value after passing through the Attention Unit, also known as a Multilayer Perceptron (MLP). Third representation vector and higher-order representation vectors The second attention value after the Attention Unit :
[0081]
[0082] in, For the first attention value, For the first The higher-order representation vector of each click For the first The higher-order representation vector of each click For the second representation vector, For the second attention value, This is the third representation vector.
[0083] In this embodiment of the disclosure, after obtaining the first attention value Second attention value Then, through weighted summation and pooling operations, we can obtain the overall representation vector of the user's historical click behavior sequence, which is the first aggregated vector of the higher-order representation vectors. Second aggregation vector .
[0084] Alternatively, the first aggregate vector of the higher-order representation vector can be obtained using the following formula. Second aggregation vector :
[0085]
[0086] in, For the first aggregation vector, For the second aggregation vector, For the first attention value, For each higher-order representation vector, For the second representation vector, For the second attention value, For the third representation vector, This indicates the attention operation.
[0087] For the feature cross-layer of the product recall model: the second representation vector can be... and the third representation vector The input is fed into the feature cross layer, which then processes the second representation vector. and the third representation vector Perform cross product and subtraction operations to obtain the first feature cross vector. Second feature cross vector .
[0088] It should be noted that the second representation vector and the third representation vector Simultaneously, the input is fed into the feature cross layer, which can refine the second representation vector. and the third representation vector By performing cross-product and subtraction operations, the product recall model learns the interaction information between the main product and candidate products, namely, the difference information and the consistency information, to ensure that the candidate products and the main product are strongly correlated, thus obtaining the first feature cross vector. Second feature cross vector .
[0089] For the prediction layer of the product recall model, the prediction layer performs a cross-vector analysis on the first feature. Second feature cross vector Second representation vector Third representation vector First aggregation vector Second aggregation vector The vectors are concatenated and then input into the MLP. Finally, the MLP outputs the predicted click-through rate and the predicted conversion rate after clicking.
[0090] It should be noted that, depending on the actual situation, one can choose whether to concatenate the first representation vector corresponding to the user tag to enhance the expressive power of the product recall model.
[0091] For example, the first feature cross vector can be... Second feature cross vector Second representation vector Third representation vector First aggregation vector Second aggregation vector The vector is concatenated with the first representation vector, and the concatenated vector is then input into the MLP. Finally, the MLP outputs the click-through rate prediction value and the post-click conversion rate prediction value.
[0092] In summary, in this embodiment of the present disclosure, the product recall model includes a sequentially connected embedding layer, fusion layer, attention layer, feature cross layer, and prediction layer. By introducing a Transformer into the fusion layer, it is possible to learn the user's click behavior representation at a higher order. The attention layer learns the association between candidate products, main products, and the user's historical click behavior sequence. The feature cross layer learns the feature cross between the main product and candidate products, thereby ensuring information interaction between the user's click sequence, main product, and candidate products, and improving the accuracy of obtaining the predicted click-through rate and post-click conversion rate of products.
[0093] To implement the product recall method of the first aspect embodiment described above, this disclosure proposes a product recall device. Figure 6 This is a schematic diagram of the structure of a product recall device according to an embodiment of this disclosure. Figure 6 As shown, the product recall device 600 includes: The first acquisition module 610 is used to acquire N products to be recalled and determine the representation vector of each product to be recalled based on the pre-trained target product recall model, where N is a natural number greater than or equal to 1. The second acquisition module 620 is used to hierarchically divide the representation vectors of the N products to be recalled to generate a hierarchical navigable small world (HNSW) graph, wherein the HNSW graph includes a multi-layered navigable small world (NSW) graph arranged vertically. The determination module 630 is used to determine the target recall product from the products to be recalled based on the HNSW diagram and the target product recall model.
[0094] In one embodiment of this disclosure, the second acquisition module 620 is further configured to: use the representation vector as a node to be inserted, acquire the target level at which the node to be inserted can fall into the HNSW graph, insert the node to be inserted into the current level of the node to be inserted and each level above the target level, determine the neighboring nodes of the node to be inserted in each layer of the NSW graph, and connect the neighboring nodes to the node to be inserted until the HNSW graph includes all representation vectors to generate the HNSW graph.
[0095] In one embodiment of this disclosure, the determining module 630 is further configured to: starting from the top layer h of the HNSW graph, determine the model input vector corresponding to the current layer i and input it into the target product recall model to obtain the target predicted value corresponding to the current layer i, wherein the target predicted value is at least one of the click-through rate (CTR) prediction value and the post-click conversion rate (PCC) value; select the representation vector corresponding to the target predicted value that meets the set conditions as the first target representation vector, and enter the i-1 layer of the current layer, repeating the above process until at the bottom layer of the HNSW graph, select the representation vector corresponding to the target predicted value that meets the set conditions as the target representation vector, and use the product to be recalled corresponding to the target representation vector as the target recalled product. In one embodiment of this disclosure, the target predicted value is at least one of the CTR prediction value and the PCC value.
[0096] In one embodiment of this disclosure, the determining module 630 is further configured to: for the top layer h, determine the representation vector in the NSW graph of the top layer h as the input vector in the target product recall model; for non-top layers, determine the second target representation vector adjacent to the first target representation vector and the first target representation vector as the input vector in the target product recall model.
[0097] In one embodiment of this disclosure, the determining module 630 is further configured to: sort the target prediction values output by each NSW graph layer for each layer of the HNSW graph, and use the representation vectors corresponding to the top K target prediction values as the representation vectors that satisfy the set conditions, wherein K is a positive integer.
[0098] In one embodiment of this disclosure, the training process of the target product recall model includes: acquiring a training dataset, wherein the training dataset includes user tags, user historical click sequences, main product features and candidate product features corresponding to the product details page; and training an initial product recall model based on the training dataset to obtain a trained target product recall model, wherein the target product recall model is used to determine the click-through rate (CTR) and post-click conversion rate (PCC) of the product.
[0099] In one embodiment of this disclosure, the apparatus 600 is further configured to: input the training dataset into the product recall model to obtain the predicted click-through rate (CTR) and post-click conversion rate (PCC) of candidate products; adjust the product recall model based on the predicted CTR and actual CTR, as well as the predicted PCC and actual PCC, and continue training the adjusted product recall model until the training termination condition is met to obtain the target product recall model.
[0100] In one embodiment of this disclosure, the apparatus 600 is further configured to: determine a first loss function of the product recall model based on the predicted click-through rate (CTR) value and the actual CTR value; and determine a second loss function of the product recall model based on the predicted post-click conversion rate (PCC) value and the actual PCC conversion rate (PCC) value; and adjust the product recall model based on the first loss function and the second loss function until the training termination condition is met to obtain the target product recall model.
[0101] In one embodiment of this disclosure, the apparatus 600 is further configured to: input the training dataset into a product recall model; obtain embedding vectors for various types of data in the training dataset, and fuse the embedding vectors for various types of data to obtain multi-dimensional representation vectors, wherein the multi-dimensional representation vectors include a first representation vector corresponding to user tags, a second representation vector corresponding to main product features, a third representation vector corresponding to candidate product features, and a higher-order representation vector corresponding to user historical click sequences; obtain a first aggregation vector and a second aggregation vector of the higher-order representation vectors based on the second representation vector and the third representation vector; perform feature cross-validation on the second representation vector and the third representation vector to obtain a first feature cross-validation vector and a second feature cross-validation vector; and output the predicted click-through rate and predicted post-click conversion rate of the candidate products based on the second representation vector and the third representation vector, the first aggregation vector and the second aggregation vector, and the first feature cross-validation vector and the second feature cross-validation vector.
[0102] In one embodiment of this disclosure, the apparatus 600 is further configured to: acquire the embedding vectors of the user tag, the user historical click sequence, the main product feature, and the candidate product feature; fuse the embedding vectors of the user tag, the user historical click sequence, the main product feature, and the candidate product feature to obtain their respective representation vectors; fuse the embedding vectors of the user tag, the main product feature, and the candidate product feature to obtain their respective representation vectors; and perform Transformer and attention mechanism processing on the embedding vector of the user historical click sequence to obtain the higher-order representation vector corresponding to the user historical click sequence.
[0103] In one embodiment of this disclosure, the apparatus 600 is further configured to: obtain attention values between the second representation vector and the third representation vector and the higher-order representation vector, respectively, based on the second representation vector, the third representation vector, and the higher-order representation vector; obtain the first aggregate vector based on the attention value corresponding to the second representation vector and the higher-order vector; and obtain the second aggregate vector based on the attention value corresponding to the third representation vector and the higher-order representation vector.
[0104] In this embodiment, N products to be recalled are obtained, and a representation vector for each product to be recalled is determined based on a pre-trained target product recall model, where N is a natural number greater than or equal to 1. The representation vectors of the N products to be recalled are layered to generate a hierarchical navigable small-world (HNSW) graph. The HNSW graph includes multiple layers of navigable small-world (NNSW) graphs arranged vertically. Based on the HNSW graph and the target product recall model, target products to be recalled are determined from the products to be recalled. This disclosure, based on the HNSW graph and the target product recall model, can determine target products to be recalled from the products to be recalled. The target products not only match user interests but are also related to the main products on the product details page, improving the accuracy and reliability of determining target products to be recalled and enhancing the user experience.
[0105] It should be noted that the above explanation of the product recall method embodiment of the first aspect also applies to the product recall device of the present disclosure embodiment, and the specific process will not be repeated here.
[0106] like Figure 7 The diagram shown is a block diagram of a product recall method or an electronic device for a product recall method according to an embodiment of this disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as intelligent voice interaction devices, personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.
[0107] like Figure 7 As shown, the electronic device includes one or more processors 701, a memory 702, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components are interconnected via different buses and can be mounted on a common motherboard or otherwise as required. The processor 701 can process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple storage devices, if desired. Similarly, multiple electronic devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 7 Take the 701 processor as an example.
[0108] The memory 702 is the non-transitory computer-readable storage medium provided in this disclosure. The memory stores instructions executable by at least one processor to cause the at least one processor to perform the product recall method provided in the first aspect embodiment or the second aspect embodiment of this disclosure. The non-transitory computer-readable storage medium of this disclosure stores computer instructions for causing a computer to perform the product recall method provided in the first aspect embodiment of this disclosure.
[0109] The memory 702, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the product recall method provided in the first aspect embodiment of this disclosure. The processor 701 executes various functional applications and data processing of the server by running the non-transitory software programs, instructions, and modules stored in the memory 702, thereby implementing the product recall method in the above method embodiment.
[0110] Memory 702 may include a program storage area and a data storage area. The program storage area may store an operating system and applications required for at least one function; the data storage area may store data created by the use of the electronic device according to the product recall method. Furthermore, memory 702 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory 802 may optionally include memory remotely located relative to processor 801, and these remote memories may be connected to the product recall method or the electronic device of the product recall method via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
[0111] The product recall method or electronic device for the product recall method may further include: an input device 703 and an output device 704. The processor 701, memory 702, input device 703, and output device 704 can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.
[0112] Input device 703 can receive input digital or character information, as well as key signal inputs related to user settings and function control of electronic devices used in product recall methods, such as touchscreens, keypads, mice, trackpads, touchpads, joysticks, one or more mouse buttons, trackballs, joysticks, etc. Output device 804 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The display device may include, but is not limited to, liquid crystal displays (LCDs), light-emitting diode (LED) displays, and plasma displays. In some embodiments, the display device may be a touchscreen.
[0113] To implement the above embodiments, this disclosure also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the product recall method as proposed in the first aspect of the present disclosure.
[0114] To implement the above embodiments, this disclosure proposes a computer program product, including a computer program that, when executed by a processor, implements the product recall method provided in the first aspect of the present disclosure.
[0115] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.
[0116] These computational programs (also referred to as programs, software, software applications, or code) include machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms “machine-readable medium” and “computer-readable medium” refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device (PLD)) used to provide machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term “machine-readable signal” refers to any signal used to provide machine instructions and / or data to a programmable processor.
[0117] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).
[0118] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.
[0119] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server, or simply "VPS") services, such as high management difficulty and weak business scalability.
[0120] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.
[0121] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise expressly and specifically defined.
[0122] Although embodiments of the present disclosure have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present disclosure. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present disclosure.
Claims
1. A product recall method, characterized in that, The method includes: Obtain N products to be recalled, and determine the representation vector of each product to be recalled based on the pre-trained target product recall model, where N is a natural number greater than or equal to 1; The representation vectors of the N items to be recalled are hierarchically layered to generate a hierarchical navigable small-world HNSW graph, wherein the HNSW graph includes multiple layers of navigable small-world NSW graphs arranged vertically. Based on the HNSW diagram and the target product recall model, target recall products are determined from the products to be recalled.
2. The method according to claim 1, characterized in that, The step of hierarchically layering the representation vectors of the N items to be recalled to generate a hierarchical navigable small-world (HNSW) graph includes: Using the representation vector as the node to be inserted, the target level into which the node to be inserted can fall is obtained. The node to be inserted is inserted into the current level of the node to be inserted and the level above the target level. The neighboring nodes of the node to be inserted are determined in each NSW graph and connected to the node to be inserted until the HNSW graph includes all representation vectors to generate the HNSW graph.
3. The method according to claim 2, characterized in that, The step of determining the target recall product from the products to be recalled based on the HNSW diagram and the target product recall model includes: Starting from the top layer h of the HNSW graph, determine the model input vector corresponding to the current layer i and input it into the target product recall model to obtain the target prediction value corresponding to the current layer i, wherein the target prediction value is at least one of the click-through rate prediction value and the post-click conversion rate value; Select the representation vector corresponding to the target predicted value that meets the set conditions as the first target representation vector, and enter the i-1 layer of the current layer. Repeat the above process until the bottom layer of the HNSW diagram is reached. Select the representation vector corresponding to the target predicted value that meets the set conditions as the target representation vector, and take the product to be recalled corresponding to the target representation vector as the target recall product.
4. The method according to claim 3, characterized in that, Determining the input vector of the current layer i into the target product recall model includes: For the top-level h layer, the representation vector in the NSW graph of the top-level h layer is determined to be the input vector in the target product recall model; For non-top-level targets, the second target representation vector adjacent to the first target representation vector and the first target representation vector are determined as input vectors in the target product recall model.
5. The method according to any one of claims 3-4, characterized in that, The step of selecting the representation vector corresponding to the target predicted value that meets the set conditions also includes: For each layer of the HNSW graph, the target prediction values output by each layer of the NSW graph are sorted, and the representation vectors corresponding to the top K target prediction values are used as the representation vectors that satisfy the set conditions, where K is a positive integer.
6. The method according to claim 1, characterized in that, The training process of the target product recall model includes: Obtain a training dataset, wherein the training dataset includes user tags, user historical click sequences, main product features and candidate product features corresponding to the product details page; Based on the training dataset, the initial product recall model is trained to obtain a trained target product recall model, wherein the target product recall model is used to determine the click-through rate (CTR) and post-click conversion rate (PCC) of the product.
7. The method according to claim 1, characterized in that, The step of training the product recall model based on the training dataset to obtain a trained target product recall model includes: The training dataset is input into the product recall model to obtain the predicted click-through rate and post-click conversion rate of candidate products. Based on the predicted and actual click-through rates (CTRs), the predicted and actual post-click conversion rates (PCCs), the product recall model is adjusted, and the adjusted product recall model is trained until the training termination condition is met to obtain the target product recall model.
8. The method according to claim 7, characterized in that, The step of training the initial product recall model based on the training dataset to obtain a trained target product recall model includes: Based on the predicted click-through rate (CTR) and the actual CTR, a first loss function for the product recall model is determined, and based on the predicted post-click conversion rate (PCC) and the actual PCC conversion rate (PCC), a second loss function for the product recall model is determined. The product recall model is adjusted based on the first loss function and the second loss function until the training termination condition is met to obtain the target product recall model.
9. The method according to claim 8, characterized in that, The step of inputting the training dataset into the product recall model to obtain the predicted click-through rate (CTR) and post-click conversion rate (PCC) of the candidate products includes: Input the training dataset into the product recall model; The embedding vectors of various types of data in the training dataset are obtained and fused to obtain a multi-dimensional representation vector. The multi-dimensional representation vector includes a first representation vector corresponding to the user tag, a second representation vector corresponding to the main product feature, a third representation vector corresponding to the candidate product feature, and a higher-order representation vector corresponding to the user's historical click sequence. Based on the second representation vector and the third representation vector, obtain the first aggregate vector and the second aggregate vector of the higher-order representation vector; The second representation vector and the third representation vector are subjected to feature cross-interaction to obtain a first feature cross-vector and a second feature cross-vector. Based on the second and third representation vectors, the first and second aggregation vectors, and the first and second feature cross vectors, the predicted click-through rate and post-click conversion rate of the candidate product are output.
10. The method according to claim 8, characterized in that, The process involves obtaining embedding vectors for various types of data in the training dataset and fusing these embedding vectors to obtain a multi-dimensional representation vector, including: Obtain the embedding vectors of the user tags, the user's historical click sequence, the main product features, and the candidate product features; The embedding vectors of the user tags, the user's historical click sequence, the main product features, and the candidate product features are fused to obtain their respective representation vectors; The embedding vectors of the user tags, main product features, and candidate product features are fused to obtain the representation vectors of the user tags, main product features, and candidate product features respectively. The embedding vector of the user's historical click sequence is processed using Transformer and attention mechanisms to obtain the high-order representation vector corresponding to the user's historical click sequence.
11. The method according to claim 10, characterized in that, The step of obtaining the first aggregated vector and the second aggregated vector of the higher-order vector based on the second representation vector and the third representation vector includes: Based on the second representation vector, the third representation vector, and the higher-order representation vector, obtain the attention values of the second representation vector and the third representation vector with respect to the higher-order representation vector; The first aggregation vector is obtained based on the attention value corresponding to the second representation vector and the higher-order vector; The second aggregation vector is obtained based on the attention value corresponding to the third representation vector and the higher-order representation vector.
12. A product recall device, characterized in that, The device includes: The first acquisition module is used to acquire N products to be recalled and determine the representation vector of each product to be recalled based on the pre-trained target product recall model, where N is a natural number greater than or equal to 1. The second acquisition module is used to hierarchically divide the representation vectors of the N products to be recalled to generate a hierarchical navigable small world (HNSW) graph, wherein the HNSW graph includes a multi-layered navigable small world (NSW) graph arranged vertically. The determination module is used to determine the target recall product from the products to be recalled based on the HNSW diagram and the target product recall model.
13. An electronic device, characterized in that, Including memory and processor; The memory and the processor communicate with each other through an internal connection path. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory. When the processor executes the instructions stored in the memory, it causes the processor to perform the method of any one of claims 1-11.
14. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when run on a computer, implements the method according to any one of claims 1-11.
15. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-11.