A method and system for material handling

Abstract

Embodiments of the present application provide a material processing method and system, comprising obtaining recalled material, inputting the recalled material into a trained coarse classification model; extracting user features and material features from the recalled material according to a pre-determined input feature limitation category; obtaining output results of a first pre-determined number of improved double-tower models based on the user features and the material features, as preliminary screening coarse classification material; obtaining output results of a second pre-determined number of deep structure models based on the preliminary screening coarse classification material, as coarse classified material, wherein the second pre-determined number is less than the first pre-determined number. In the technical solution, the double-tower structure model and the deep structure model are organically combined, not only meeting the time-consuming requirement of the recommendation system for the coarse classification module, but also improving the scoring and sorting ability of the coarse classification module, solving the pain point that the coarse classification model cannot use the deep structure model due to time consumption, and greatly improving the business indicators.

Description

Technical Field

[0001] This invention relates to the field of network technology, and more specifically to a material handling method and system. Background Technology

[0002] Information flow recommendation systems typically consist of a cascaded structure of recall, coarse ranking, fine ranking, and re-ranking. Coarse ranking, as an intermediate module bridging recall and fine ranking, is a crucial part of the recommendation system, playing a pivotal role. It is responsible for sorting the recall results and selecting a shortlist of thousands of candidates that meet the requirements of fine ranking for further scoring and sorting. Optimizing the coarse ranking architecture can significantly improve the overall business performance. Due to limitations in computing power and system time requirements, the existing coarse ranking module has lower complexity than the fine ranking module in terms of sample selection, feature selection, and online scoring and ranking. Regarding sample selection, to avoid additional overhead, coarse ranking uses the same samples as fine ranking, selecting exposure and clicks as positive samples and exposure without clicks as negative samples to train the click-through rate model. For feature selection, a subset of features is usually selected based on human experience; this subset is the same as the features used in fine ranking. In online scoring and ranking, fine-grained ranking typically involves concatenating user-side and material-side features and inputting them into a deep structured neural network for training and online ranking and scoring. After being concatenated, user and material features can fully interact within a fully connected neural network, generating a wealth of important information that aids the model's learning. Figure 3 As shown. Unlike fine-ranking, coarse-ranking is usually based on a dual-tower structure model for training and online ranking and scoring, such as... Figure 4 As shown, user-side features and material-side features are fed into different deep neural networks for calculation.

[0003] During training, user-side features and material-side features are processed through an embedding layer to generate corresponding vectorized results, which are then fed into their respective neural network layers. The user tower and material tower output corresponding user output vectors and material output vectors, respectively. The outputs are obtained through dot product or cosine similarity, and the final loss function is calculated for model training. During online scoring and ranking, the user side constructs current user features in real time and inputs them into the corresponding neural network to obtain the user output vector. The material side constructs features offline in advance and inputs them into the corresponding neural network to obtain the material output vector, which is stored in the Faiss database. Based on the user-side output vector, the top N most similar material-side features are found in the Faiss database, and then the final ranking score is obtained based on click or cosine similarity (to maintain consistency with training).

[0004] In the process of realizing this invention, the inventors discovered at least the following problems in the prior art:

[0005] Current recommendation systems typically use a relatively simple dual-tower structure model for the coarse-ranking stage, training and scoring the user and material sides separately. This prevents user and material features from interacting within the neural network. Compared to deep structure models that directly concatenate user and material features into the same neural network, this results in a lower offline evaluation metric (AUC) and worse online real-world performance. Therefore, improving the coarse-ranking model to enhance its scoring and ranking capabilities, thereby improving its real-world performance, is a problem that needs to be addressed. Summary of the Invention

[0006] This invention provides a material processing method and system to solve the problem of poor real business performance indicators caused by inherent problems in the coarse-ranking model in existing recommendation systems.

[0007] To achieve the above objectives, in one aspect, embodiments of the present invention provide a material processing method, comprising: acquiring recalled materials obtained through a recall model; inputting the recalled materials into a trained coarse-ranking model; wherein the trained coarse-ranking model includes an improved dual-tower model and a trained deep structure model, and the trained coarse-ranking model is obtained by training an initial coarse-ranking model; extracting user features and material features of the recalled materials from the recalled materials according to a predetermined input feature category; inputting the user features of the recalled materials into the user tower of the improved dual-tower model to obtain a user output vector; and inputting the material features of the recalled materials into the material tower of the improved dual-tower model. The process involves obtaining a material output vector; based on the user output vector and the material output vector, obtaining a first preset number of output results from the improved dual-tower model as the initial screening and coarse-sorting material; extracting user features and material features of the initial screening and coarse-sorting material from the initial screening and coarse-sorting material using the deep structure model; and performing a splicing process on the user features and material features of the initial screening and coarse-sorting material using the deep structure model; based on the splicing process, obtaining a second preset number of output results from the deep structure model as the coarse-sorted material, where the second preset number is less than the first preset number; and using the coarse-sorted material as input to the fine-sorting model to obtain the final recommendation result.

[0008] On the other hand, embodiments of the present invention provide a material handling system, including: a recalled material acquisition module, configured to acquire recalled materials obtained through a recall model, and input the recalled materials into a trained coarse-ranking model; wherein, the trained coarse-ranking model includes an improved dual-tower model and a trained deep structure model, the trained coarse-ranking model being obtained by training an initial coarse-ranking model; the improved dual-tower model is configured to extract user features and material features of the recalled materials from the recalled materials based on predetermined input features to define categories; input the user features of the recalled materials into the user tower of the improved dual-tower model to obtain a user output vector; and input the material features of the recalled materials into the improved dual-tower model. In the material tower of the model, a material output vector is obtained; based on the user output vector and the material output vector, a first preset number of the output results of the improved dual-tower model are obtained as the initial screening and coarse discharge material; a second input feature extraction module is used; the deep structure model is used to extract the user features and material features of the initial screening and coarse discharge material from the initial screening and coarse discharge material; and to perform splicing processing on the user features and material features of the initial screening and coarse discharge material; based on the splicing processing result, a second preset number of the output results of the deep structure model are obtained as the coarsely discharged material, the second preset number being less than the first preset number; the coarsely discharged material is used as input to the fine discharge model to obtain the final recommendation result.

[0009] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned material handling method.

[0010] Furthermore, embodiments of the present invention also provide a computer device, comprising: one or more processors; a storage device for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the aforementioned material handling method.

[0011] The above technical solution has the following beneficial effects:

[0012] In this technical solution, a new coarse-ranking model is formed by organically combining the dual-tower structure model and the deep structure model. This divides the scoring and ranking in the coarse-ranking stage into two phases: first, the dual-tower structure model is used to initially score and rank a large number of materials, and then the highest-scoring portion is selected from the output of the initial scoring and ranking. Next, the deep structure model is used to perform a second coarse-ranking and ranking of these selected portions. Finally, the highest-scoring portion of the output is given to the fine-ranking model. This approach enables the coarse-ranking model to rank a large number of materials and the deep structure model to rank a small number of materials. This not only meets the time requirements of the recommendation system for the coarse-ranking module but also improves the scoring and ranking capabilities of the coarse-ranking module. It solves the pain point that the coarse-ranking model cannot use the deep structure model due to time constraints, and greatly improves business metrics.

[0013] In addition, this technical solution also has the following characteristics:

[0014] 1. In existing technologies, input feature categories are typically selected based on human experience for model training. However, due to the inherent errors in human experience, model training may not reach its optimal state, thus impacting business performance metrics. This application adds a SENET layer to the existing dual-tower model. When training the dual-tower model with the SENET layer, the input feature categories can be determined. This allows the improved coarse-ranking model to achieve or even surpass models trained using human-experience-based feature selection with fewer features. Furthermore, the training and online scoring / ranking times are reduced by nearly half.

[0015] 2. In existing technologies, the same samples are used in both the coarse-ranking and fine-ranking stages, resulting in poor ranking performance. This new technology avoids using the same real exposure samples from user exposure and click behavior as the fine-ranking model. Instead, it trains the model using a sample stream composed of real exposure samples from user exposure and click behavior and negative samples randomly sampled from the material library, coarse-ranking scores, and fine-ranking and re-ranking scores. This improves the model's scoring and ranking capabilities. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a material handling method according to an embodiment of the present invention;

[0018] Figure 2This is a structural framework diagram of a material handling system according to an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the architecture of deep structured models in existing technologies;

[0020] Figure 4 This is a schematic diagram of the structure of a twin-tower model in the existing technology;

[0021] Figure 5 This is a schematic diagram of a dual-tower model containing a SENET layer in an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram illustrating the composition of the sample stream in an embodiment of the present invention;

[0023] Figure 7 This is a schematic diagram of the scoring and sorting of the coarse-ranking model in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, an embodiment of the present invention provides a material handling method, including:

[0026] S101. Obtain the recalled materials obtained through the recall model, and input the recalled materials into the trained coarse-ranking model; wherein, the trained coarse-ranking model includes an improved dual-tower model and a trained deep structure model, and the trained coarse-ranking model is obtained by training the initial coarse-ranking model;

[0027] S102. Based on the predetermined input features, define the categories and extract the user features and material features of the recalled materials from the recalled materials.

[0028] S103. Input the user characteristics of the recalled materials into the user tower of the improved dual-tower model to obtain the user output vector;

[0029] S104. Input the material characteristics of the recalled materials into the material tower of the improved dual-tower model to obtain the material output vector;

[0030] S105. Based on the user output vector and the material output vector, obtain a first preset number of the output results of the improved dual-tower model as the initial screening and coarse discharge material;

[0031] S106. Extract the user characteristics and material characteristics of the primary screening material from the primary screening material using the deep structural model.

[0032] S107. The user characteristics and material characteristics of the primary screening and coarse discharge material are spliced ​​together using the deep structural model.

[0033] S108. Based on the result of the splicing process, a second preset number of the output results of the deep structure model are obtained as coarsely sorted materials. The second preset number is less than the first preset number. The coarsely sorted materials are used as input to the fine sorting model to obtain the final recommendation result. In general, the output results of the fine sorting model need to be sorted by the rearrangement model to obtain the final recommendation result.

[0034] To address the aforementioned technical issues, this technical solution combines the dual-tower model with the deep structure-based model used in the fine-ranking stage of existing technologies. The new coarse-ranking model is a combination of the dual-tower model and the deep structure-based model. During operation, the dual-tower model is first used for "mass selection," and then a subset is selected as input for the deep structure-based model for "fine-tuning." This combines the speed of the dual-tower model with the high accuracy of the deep structure-based model, thereby improving the business metrics in the coarse-ranking stage. To reduce time consumption, the deep structure-based model used in this coarse-ranking stage selects a subset of input features from a defined category. For example, while existing fine-ranking models typically use over 200 features, in the coarse-ranking stage of this technical solution, only the 50 most important features are selected based on the defined categories.

[0035] Furthermore, the initial recommendation system includes the recall model, the initial coarse-ranking model, and the fine-ranking model, wherein the initial coarse-ranking model adopts a dual-tower model structure;

[0036] The method for obtaining the trained coarse-ranking model from the initial coarse-ranking model includes:

[0037] S010. Add a SENET layer to the dual-tower model to obtain the initial dual-tower model for training.

[0038] S020. Collect positive samples and negative samples and concatenate them into a sample stream. The positive samples are material samples in the final recommendation results output by the initial recommendation system and which the user has already followed. The negative samples are material samples outside the final recommendation results output by the initial recommendation system and which the user has already followed.

[0039] S030. Use the sample stream to train the initial training dual-tower model and the initial deep structure model. After training, remove the SeNet layer from the trained dual-tower model to obtain the improved dual-tower model. Use the improved dual-tower model and the trained deep structure model as the trained coarse-ranked model.

[0040] Existing coarse-ranking architectures suffer from inaccurate feature selection. Due to computational limitations, and lacking a robust feature selection scheme, feature selection is typically based on manual experience. However, this manual experience introduces errors, leading to some important features being excluded while less important ones are included in training. This prevents the model from reaching its optimal performance, impacting business efficiency metrics. To address this issue, this technical solution, in addition to adding a deep structured model in the coarse-ranking stage, improves the existing dual-tower model. Specifically, a SENet structure is added between the vector layer and the neural network layer in the dual-tower model. This SENet structure helps determine the input feature categories during model training, identifying which categories of features for each material need to be extracted and input into the model for online application.

[0041] Furthermore, existing technologies have another problem: in terms of sample selection, using the same samples as the fine-ranking model, since coarse-ranking is located between recall and fine-ranking, it directly scores and ranks the N recalled outputs, and the K highest-scoring outputs are given to fine-ranking for further scoring and ranking. However, the number of N recalled outputs is usually four times or more than the number of the K highest-scoring outputs in coarse-ranking. Directly using the same samples as fine-ranking leads to sample selection bias because the samples trained for fine-ranking are based on the K highest-scoring samples in coarse-ranking combined with user feedback. Using only these samples for training means that the model does not learn about the samples in the N recalled outputs that are not in the K highest-scoring outputs in coarse-ranking. However, the coarse-ranking model also needs to score and rank these samples online, so it cannot rank these samples well. Therefore, this technical solution also incorporates the output results of the recall, coarse ranking, and fine ranking stages to expand the range of sample selection (mainly negative samples). Then, positive and negative samples are concatenated into a sample stream, which is used as input for model training. This solves the sample bias problem and improves the online real business performance indicators.

[0042] Furthermore, in step S030:

[0043] The SENET layer is used to determine input feature definition categories, which are used to determine the categories of user features and material features extracted from the recalled materials.

[0044] Furthermore, step S020 specifically includes:

[0045] S021. Collect the first negative sample from the output of the recall model of the initial recommendation system;

[0046] S022. Collect a second negative sample from the output of the initial coarse-ranking model of the initial recommendation system;

[0047] S023. Collect a third negative sample from the output of the fine-ranking model of the initial recommendation system;

[0048] S024. Select a fourth negative sample from the final recommendation results output by the initial recommendation system that the user has not paid attention to;

[0049] S025. Select positive samples from the final recommendation results output by the initial recommendation system that the user has already followed;

[0050] S026. The first negative sample, the second negative sample, the third negative sample, the fourth negative sample, and the positive sample are concatenated into a sample stream.

[0051] The sample stream is composed of real exposure samples (based on user exposure and click behavior used for training the fine-ranking model) and negative sample streams randomly sampled from the material library, coarse-ranking scores, and fine-ranking and re-ranking scores. The coarse-ranking model trained based on this sample stream can be trained not only using the user exposure and click samples from the fine-ranking model but also using negative samples randomly sampled from the material library, coarse-ranking scores, and fine-ranking and re-ranking scores, thus resolving the sample selection bias problem.

[0052] like Figure 2 As shown, an embodiment of the present invention provides a material handling system, including:

[0053] The recalled material acquisition module 21 is used to acquire the recalled materials obtained through the recall model and input the recalled materials into the trained coarse-ranking model; wherein, the trained coarse-ranking model includes an improved dual-tower model and a trained deep structure model, and the trained coarse-ranking model is obtained by training the initial coarse-ranking model;

[0054] The improved dual-tower model 22 is used to extract user features and material features of recalled materials from the recalled materials based on predetermined input features that define categories; input the user features of the recalled materials into the user tower of the improved dual-tower model to obtain a user output vector; input the material features of the recalled materials into the material tower of the improved dual-tower model to obtain a material output vector; and based on the user output vector and the material output vector, obtain a first predetermined number of output results of the improved dual-tower model as preliminary screening and coarse discharge materials.

[0055] The deep structure model 23 is used to extract user features and material features of the primary screening and coarse-sorted materials from the primary screening and coarse-sorted materials; and to perform splicing processing on the user features and material features of the primary screening and coarse-sorted materials; based on the splicing processing result, a second preset number of output results of the deep structure model are obtained as coarse-sorted materials, the second preset number being less than the first preset number; the coarse-sorted materials are used as input to the fine-sorting model to obtain the final recommendation result.

[0056] Furthermore, the initial recommendation system includes the recall model, the initial coarse-ranking model, and the fine-ranking model, wherein the initial coarse-ranking model adopts a dual-tower model structure;

[0057] The material handling system also includes a coarse-scale model training module 20, used for:

[0058] A SENET layer is added to the dual-tower model to obtain the initial training dual-tower model. Positive and negative samples are collected and concatenated to form a sample stream. The positive samples are material samples from the final recommendation results output by the initial recommendation system and followed by the user, and the negative samples are material samples from the final recommendation results output by the initial recommendation system and not followed by the user. The sample stream is used to train the initial training dual-tower model and the initial deep structure model. After training, the SENET layer in the trained training dual-tower model is removed to obtain the improved dual-tower model. The improved dual-tower model and the trained deep structure model are used as the trained coarse-ranking model.

[0059] Furthermore, the coarse-ranking model training module 20 also includes a sample stream splicing submodule, used for:

[0060] A first negative sample is collected from the output of the recall model of the initial recommendation system; a second negative sample is collected from the output of the initial coarse-ranking model of the initial recommendation system; a third negative sample is collected from the output of the fine-ranking model of the initial recommendation system; a fourth negative sample is selected from the final recommendation results output by the initial recommendation system that the user has not followed; a positive sample is selected from the final recommendation results output by the initial recommendation system that the user has followed; the first negative sample, the second negative sample, the third negative sample, the fourth negative sample, and the positive sample are concatenated into a sample stream.

[0061] The technical solutions of the present invention will be described in detail below with reference to a specific application example:

[0062] This embodiment can be applied to information feed recommendation scenarios on social media platforms such as Weibo. This embodiment improves upon the problems of feature selection, sample selection, and online scoring and ranking in the coarse-ranking architecture of recommendation systems from three aspects, thereby optimizing the entire coarse-ranking architecture.

[0063] Regarding feature selection: To address the difficulty of feature selection, an effective feature selection scheme was constructed. Specifically, firstly, the current coarse-ranking model structure was modified, such as... Figure 5 As shown, a SENet structure is added between the feature embedding layer and the fully connected deep neural network layer. The principle is as follows: SENet learns an S-vector during each training iteration. Each dimension of the vector corresponds to a weight for each feature, calculated based on the contribution of each feature to the model's final loss. Features with greater contributions have higher weights. The formula for calculating feature importance weights is as follows:

[0064] S=σ(W[E1,...,Em]+b)

[0065] Ei is the vector for each feature, m represents the number of features, S is an M-dimensional vector where each dimension represents the weight of the corresponding feature, W is a kxM-dimensional vector, and b is an M-dimensional vector. W and b are parameters to be learned during model training. Ultimately, the vector for each feature is the product of the weight at the corresponding position in S and Ei, which is then input into the subsequent deep neural network for training. Next, a new model structure is used to train based on all features, recording the weight of each feature output for each training session. The importance of each feature is obtained by summing its weights and dividing by the total number of training sessions. Finally, based on the importance of each feature, the N features with the highest weights are selected as the final features (i.e., input features constrained to categories), and a new coarse-ranking model is trained based on these features for online scoring and ranking.

[0066] Regarding sample selection: To address the sample selection bias problem, a sample stream was constructed, such as... Figure 6 As shown, the sample stream is composed of real exposure samples consisting of user exposure and click behavior used for training the fine-ranking model, and a negative sample stream randomly sampled from a library of 2 million materials, 6,000 coarse-ranking scores, and 1,000 fine-ranking and re-ranking scores. The coarse-ranking model trained based on this sample stream can not only use the user exposure and click samples trained on the fine-ranking model as in existing technologies, but also use the negative samples randomly sampled from the library of 2 million materials, 6,000 coarse-ranking scores, and 1,000 fine-ranking and re-ranking scores, thus solving the sample selection bias problem.

[0067] Taking the click-through rate (CTR) model as an example, the training process for a model based on the added SeNet feature selection module and sample stream samples is as follows:

[0068] 1. Determine the training samples. Positive samples are those of users who actually expose and click (or are followed by users). Negative samples are those of users who expose but do not click (or are not followed by users), 10 samples randomly sampled from the 6000 coarse ranking scores, and 8 samples randomly sampled from the 1000 fine ranking scores.

[0069] 2. In each training iteration, 512 samples (including both positive and negative samples) are fed into a dual-tower model containing a SENet layer for batch training. Each sample inputs user features and material features into the user tower and material tower of the model, respectively. The user tower and material tower have the same structure. First, they pass through the embedding layer to obtain user feature embeddings and material feature embeddings, and then they are input into their respective SENet layers. SENet calculates the weight of each feature embedding, and then multiplies it by each feature embedding to obtain the output of the SENet module. Important features will have higher weights, which strengthens the features that are beneficial to model training.

[0070] 3. The outputs of the SENet modules on the user side and the material side are fed into their respective neural network layers to obtain output vectors of the same dimension. The inner product of the user output vector and the material output vector is calculated. The loss is calculated based on the positive and negative sample labels and their weights. The gradient is solved by backpropagation based on the loss, and the parameters in the model network are updated based on the gradient. During training, AUC is calculated. Initially, AUC will continuously increase. When the AUC value tends to stabilize at a certain point in time, the model training converges and can be used for online service. During this process, the N features with the highest weights are selected as input features to limit the categories.

[0071] For online scoring and ranking, a deep structured model was added to the existing dual-tower structured model. Due to time constraints, directly upgrading to a deep structured model would significantly increase processing time. Therefore, two optimizations were made: First, based on feature selection, the 50 most important features were chosen from the 200+ features currently used in the fine-ranking model as input features to limit categories. This reduces the number of features by nearly 150 compared to the fine-ranking model, resulting in the coarse-ranking model taking half the time compared to the fine-ranking model when training with the same deep structured model, while also significantly improving model performance. Second, a "coarse selection followed by fine screening" approach was proposed. In the Weibo scenario, the coarse-ranking module needs to score 6000 retrieved items and select the top 1000 for further scoring and ranking in the fine-ranking module. Directly using a deep structured model for scoring would significantly increase processing time, failing to meet the low-time requirement of a recommendation system. The improvement involves first scoring 6000 items using a dual-tower structure model, selecting the top 3000, and then using a deep structured model to score these 3000 items again, selecting the top 1000 for fine-grained ranking. This not only meets the requirement of low time consumption for recommendation systems but also achieves significantly better results than directly using a dual-tower structure model to score 6000 items. Figure 7 As shown, scoring 6000 records directly using the dual-tower structure model takes 12ms, while scoring 6000 records directly using the deep structure model takes 60ms, which is much longer than the scoring time of the dual-tower structure model. However, when scoring using the above method, the time is only 32ms, which meets the current time requirements.

[0072] When scoring and ranking online, in the coarse ranking stage, each material will be analyzed according to the previously determined input features to obtain the corresponding user features and material features, and then the corresponding user features and material features will be input into the trained dual-tower model.

[0073] The disclosed embodiments have been described above to enable any person skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments without departing from the spirit and scope of this disclosure. Therefore, this disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in this application.

[0074] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A material handling method, characterized in that, include: Obtain the recalled materials obtained through the recall model, and input the recalled materials into the trained coarse-ranking model; wherein, the trained coarse-ranking model includes an improved dual-tower model and a trained deep structure model, and the trained coarse-ranking model is obtained by training the initial coarse-ranking model; Based on the predetermined input features that define the categories, extract the user features and material features of the recalled materials from the recalled materials. The user characteristics of the recalled materials are input into the user tower of the improved dual-tower model to obtain the user output vector; The material characteristics of the recalled materials are input into the material tower of the improved dual-tower model to obtain the material output vector; Based on the user output vector and the material output vector, a first preset number of the output results of the improved dual-tower model are obtained as the initial screening and coarse discharge material; The user characteristics and material characteristics of the primary screening coarse discharge material are extracted from the primary screening coarse discharge material using the deep structural model; and, The user characteristics and material characteristics of the primary screening and coarse discharge materials are spliced ​​together. Based on the splicing process, a second preset number of the output results of the deep structure model are obtained as coarsely sorted materials. The second preset number is less than the first preset number. The coarsely sorted materials are used to input the fine sorting model to obtain the final recommended result. The deep structure model is used to perform a second coarse sorting and ranking of the initially screened coarsely sorted materials.

2. The material handling method as described in claim 1, characterized in that, The initial recommendation system includes the recall model, the initial coarse-ranking model, and the fine-ranking model, wherein the initial coarse-ranking model adopts a dual-tower model structure; The method for obtaining the trained coarse-ranking model from the initial coarse-ranking model includes: A SENET layer is added to the dual-tower model to obtain the initial dual-tower model for training. Positive and negative samples are collected and concatenated to form a sample stream. The positive samples are material samples in the final recommendation results output by the initial recommendation system and which the user has already followed. The negative samples are material samples outside the final recommendation results output by the initial recommendation system and which the user has already followed. The sample stream is used to train the initial training dual-tower model and the initial deep structure model. After training, the SeNet layer in the trained dual-tower model is removed to obtain the improved dual-tower model. The improved dual-tower model and the trained deep structure model are used as the trained coarse-ranked model.

3. The material handling method as described in claim 2, characterized in that, During the training of the initial training dual-tower model and the initial deep structure model using the sample stream, the SeNet layer is used to determine the input feature category, which is used to determine the category of user features and the category of material features extracted from the recalled materials.

4. The material handling method as described in claim 2, characterized in that, The process of collecting positive and negative samples and concatenating them into a sample stream specifically includes: The first negative sample is collected from the output of the recall model of the initial recommendation system; A second negative sample is collected from the output of the initial coarse-ranking model of the initial recommendation system. A third negative sample is collected from the output of the fine-ranking model of the initial recommendation system; A fourth negative sample is selected from the final recommendation results output by the initial recommendation system that the user has not paid attention to. Positive samples are selected from the final recommendation results output by the initial recommendation system that the user has already followed. The first negative sample, the second negative sample, the third negative sample, the fourth negative sample, and the positive sample are concatenated into a sample stream.

5. A material handling system, characterized in that, include: The recalled material acquisition module is used to acquire recalled materials obtained through the recall model and input the recalled materials into the trained coarse-ranking model; wherein, the trained coarse-ranking model includes an improved dual-tower model and a trained deep structure model, and the trained coarse-ranking model is obtained by training the initial coarse-ranking model; The improved dual-tower model is used to extract user features and material features of recalled materials from the recalled materials based on pre-determined input features that define categories; input the user features of the recalled materials into the user tower of the improved dual-tower model to obtain a user output vector; input the material features of the recalled materials into the material tower of the improved dual-tower model to obtain a material output vector; and based on the user output vector and the material output vector, obtain a first preset number of output results of the improved dual-tower model as initial screening and coarse discharge materials. The deep structure model is used to extract user features and material features of the primary screening and coarsely discharged materials from the primary screening and coarsely discharged materials; and to perform splicing processing on the user features and material features of the primary screening and coarsely discharged materials; based on the splicing processing result, a second preset number of output results of the deep structure model are obtained as coarsely discharged materials, the second preset number being less than the first preset number; the coarsely discharged materials are used to input into the fine-sorting model to obtain the final recommendation result; wherein, the deep structure model is used to perform a second coarse-sorting scoring and ranking on the primary screening and coarsely discharged materials.

6. The material handling system as described in claim 5, characterized in that, The initial recommendation system includes the recall model, the initial coarse-ranking model, and the fine-ranking model, wherein the initial coarse-ranking model adopts a dual-tower model structure; The material handling system also includes a coarse-sorting model training module, used for: A SENET layer is added to the dual-tower model to obtain the initial training dual-tower model. Positive and negative samples are collected and concatenated to form a sample stream. The positive samples are material samples from the final recommendation results output by the initial recommendation system and followed by the user, and the negative samples are material samples from the final recommendation results output by the initial recommendation system and not followed by the user. The sample stream is used to train the initial training dual-tower model and the initial deep structure model. After training, the SENET layer in the trained training dual-tower model is removed to obtain the improved dual-tower model. The improved dual-tower model and the trained deep structure model are used as the trained coarse-ranking model.

7. The material handling system as described in claim 6, characterized in that, The coarse-ranking model training module also includes a sample stream splicing submodule, used for: A first negative sample is collected from the output of the recall model of the initial recommendation system; a second negative sample is collected from the output of the initial coarse-ranking model of the initial recommendation system; a third negative sample is collected from the output of the fine-ranking model of the initial recommendation system; a fourth negative sample is selected from the final recommendation results output by the initial recommendation system that the user has not followed; and a positive sample is selected from the final recommendation results output by the initial recommendation system that the user has followed. The first negative sample, the second negative sample, the third negative sample, the fourth negative sample, and the positive sample are concatenated into a sample stream.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the material handling method as described in any one of claims 1-4.

9. A computer device, characterized in that, It includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the material handling method as described in any one of claims 1-4.

Images