A cue recommendation method and system based on entity embedding and dynamically adjusting deep neural networks

Abstract

This invention proposes a clue recommendation method and system based on a dynamically adjusted deep neural network with entity embedding. The method includes the following steps: S1, acquiring sample data, establishing a sample database, and storing it in a structured manner; S2, performing feature encoding and numerical transformation on sample features and business entity features; S3, converting features into dense embeddings of uniform dimension, and distinguishing between business entity feature embeddings and sample feature embeddings; S4, constructing a recommendation model formed by fusing a deep neural network structure based on conditionalized deep neural network with business entity feature embedding using a factorization machine; S5, outputting recommendation data using the recommendation model. This invention can more accurately characterize the clue sample conversion preferences under different business entities, resulting in better clue recommendation quality.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence recommendation technology, and in particular to a cue recommendation method and system based on entity embedding dynamically adjusted deep neural networks. Background Technology

[0002] In the car sales process, dealerships (business entities) need to sift through a large number of sales leads. Machine learning methods such as logistic regression and gradient boosting trees can model lead features to predict lead conversion probabilities. Ranking leads based on probability helps dealerships prioritize customers with higher conversion probabilities. Traditional machine learning models rely on manual feature engineering and experience-based feature design and construction by domain experts. These methods (traditional machine learning models) are costly to develop and limited in large-scale applications.

[0003] In the field of recommender systems, there are many models that automatically learn feature representations, such as Factorization Machines, Wide&Deep, DeepFM, and their variants. These models map discrete features into dense embedding vectors, and deep neural networks model high-order feature interactions end-to-end, eliminating or reducing reliance on manual feature engineering and achieving significant results in many recommendation and ranking tasks.

[0004] Existing standard recommendation models typically assume that different users follow similar decision-making logic. Therefore, standard recommendation models are generally static, globally consistent user decision functions, employing fixed network structures and sharing model parameters. However, this assumption is insufficient for recommendation ranking tasks at the business entity level. Our empirical observations show that data from different business entities not only exhibit shifts in preference distribution but also reveal differences in the structural diversity of the business entities' decision-making logic. Standard recommendation models only distinguish different business entities through business entity embedding, essentially adjusting only the connection weights in the feature interaction graph, and cannot change the topological structure of feature interactions.

[0005] Representative methods in conditional computation frameworks attempt to achieve the above operations by dynamically adjusting the model structure or parameters based on the input. MoE, HyperNetwork, and Meta-learning are representative methods in conditional computation frameworks, but they have significant limitations in industrial-scale entity scenarios. MoE relies on discrete expert selection, making it difficult to scale to a large number of entities; HyperNetwork requires per-entity parameter generation, which is costly and unstable for long-tail business entities; Meta-learning requires per-entity adaptation, making industrial deployment difficult. Therefore, we need a lightweight solution that strikes a balance between parameter efficiency and expressive power. Models such as GateNet have demonstrated that gating mechanisms can implicitly select feature interaction paths by suppressing or amplifying specific signals. However, their gates operate on the embeddings or hidden layer outputs of each feature, using their respective feature embeddings as their respective feature gating signals, essentially remaining standard recommendation models.

[0006] In summary, we need a lightweight solution that strikes a balance between parameter efficiency and model expressive power. Summary of the Invention

[0007] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a clue recommendation method and system based on entity embedding and dynamically adjusting deep neural networks.

[0008] To achieve the above-mentioned objectives of this invention, this invention provides a cue recommendation method based on entity embedding dynamically adjusted deep neural networks, comprising the following steps:

[0009] Step 1: The data parsing module acquires sample data, establishes a sample database, and stores it in a structured format. The characteristics of the samples should include "basic attribute features" and "business statistical features," i.e., the attributes of the samples themselves; they should also include corresponding "business entity features" (e.g., dealer stores, regions). The set of sample feature fields is denoted as... The set of business entity feature fields is denoted as , , ,

[0010] in, Represents the set of feature domains of a sample;

[0011] Distinguishing the union symbol;

[0012] Represents the set of feature domains of a business entity;

[0013] It represents the set of all feature fields (the complete set);

[0014] Intersection symbol;

[0015] Indicates the empty set;

[0016] Basic attributes and characteristics include, but are not limited to: city, desired car model, gender, age, budget range, and source channel;

[0017] Business statistical features include, but are not limited to: historical request count, transaction popularity of vehicle models in different cities, and conversion statistics of leads in different time windows;

[0018] The business entity feature corresponding to the sample is at least one unique identifier code, which belongs to discrete features;

[0019] Step 2, the feature processing module processes the sample features. and business entity characteristics Perform feature encoding and numerical conversion of features.

[0020] Discrete features Encode the discrete values ​​to unique consecutive integer indices;

[0021] ,

[0022] in, Indicates the first Discrete features The set of all possible indexes;

[0023] Indicates the first Discrete features (i.e., the first) The index of the first value in each feature domain;

[0024] Representing discrete features The index of the second value;

[0025] It is a discrete feature No. The index of each value;

[0026] It is a discrete feature Number of categories (values) possessed;

[0027] For continuous features Perform numerical transformation to make it conform to an approximate normal distribution;

[0028] ,

[0029] in, This represents a continuous eigenvalue that, after data transformation, follows an approximately normal distribution.

[0030] It is the cumulative distribution function of the standard normal distribution;

[0031] yes The inverse function;

[0032] It is the empirical distribution function;

[0033] Step 3: The feature embedding module converts the features into dense embeddings of a uniform dimension and distinguishes between business entity feature embeddings and sample feature embeddings.

[0034] Discrete features are input into the embedding layer using integer indices to construct a discrete feature embedding matrix. The embedding matrix is ​​initialized using Xavier uniform initialization.

[0035] ,

[0036] in, It is the first Discrete features The embedding matrix;

[0037] It is the first Discrete features No. A specific value in The corresponding embedding in;

[0038] It is a discrete feature The Value index (e.g., "Chongqing" is coded as 57, at this time) );

[0039] Indicates that it belongs to the symbol;

[0040] Represent the space of real numbers;

[0041] It is a discrete feature The number of categories it possesses;

[0042] It is the embedding dimension (e.g., 8, 16, 32 are hyperparameters).

[0043] Continuous features By using a fully connected linear mapping to an embedding space of the same dimension as the discrete features:

[0044] ,

[0045] in, Representing continuous features The embedding vector;

[0046] Represents the weight vector of a linear mapping;

[0047] This represents the continuous feature input value after numerical conversion;

[0048] Indicates the bias term;

[0049] Indicates that it belongs to the symbol;

[0050] Represent the space of real numbers;

[0051] It is the embedded dimension;

[0052] The embedded features are grouped according to business semantics, dividing the embeddings into a sample feature embedding set and a business entity feature embedding set. (Sample feature embedding set) Depend on Embedded mapping Composition, denoted as Business entity feature embedding set ;in It is the first Discrete features The embedding matrix; This is the embedding matrix of the feature domains of the business entities.

[0053] Step 4: Construct a recommendation model consisting of an "entity-embedded dynamically adjustable deep neural network" formed by fusing a "factorization machine structure" with a "conditional dynamic entity gating network".

[0054] The described "Conditionalized Dynamic Entity Gating Network" is an asymmetric deep neural network structure based on the decoupling of "data flow" and "control flow." The "data flow" is responsible for extracting sample features. The general high-order representation serves as the "feature extraction backbone"; the "control flow" incorporates the features of business entities. The embedded input is directly fed into the "conditional gating layer", which belongs to the "dynamic entity gating path".

[0055] The "conditional gating layer" can be applied to the first layer, any intermediate layer, or multiple layers of a deep neural network. The "conditional gating layer" can include residual connections or be simplified to a structure that does not include residual connections.

[0056] Decoupling the "control flow" and "data flow" requires imposing asymmetric constraints on the input of the "conditional dynamic entity gating network": sample feature embedding. Direct input transformation, conditional feature embedding This signal is then independently converted into a gating signal, used to dynamically modulate the activation of the hidden layer. Both signals influence the output through different function paths.

[0057] 4.1 Constructing the Factorization Machine Module:

[0058] ,

[0059] in, This represents the output scalar of the factorization machine module;

[0060] Indicates the total number of feature fields;

[0061] Indicates the first Linear weight parameters for each feature domain;

[0062] Indicates the first Input values ​​for each feature domain;

[0063] Represents the dot product operation of vectors;

[0064] They represent the first The first feature domain and the first Embedding vectors corresponding to each feature domain;

[0065] 4.2 Constructing a conditional dynamic entity gating network structure module:

[0066] 4.2.1, Define asymmetric input:

[0067] Feature extraction backbone input:

[0068]

[0069] in, This represents the concatenated sample feature embedding vector;

[0070] This represents field-wise concatenation of vectors arranged in the order of their feature domains. ;

[0071] Represents the set of sample feature embeddings The Middle Feature embedding vectors corresponding to each feature domain;

[0072] Indicates that it belongs to the symbol;

[0073] Represents the set of sample feature embeddings;

[0074] Input to the conditional gating layer:

[0075]

[0076] in, This represents the input to the conditional gating layer, serving as the modulation "condition";

[0077] This indicates the embedding of business entity features corresponding to the sample. ;

[0078] This represents the set of features embedded in the business entity;

[0079] 4.2.2, Constructing a residual condition gating mechanism:

[0080] Gating signal generation mechanism:

[0081] ,

[0082] Residual connection gating mechanism:

[0083] ,

[0084] ,

[0085] in, It is a dynamically generated gated modulation vector (or scalar). The gated vector can be in scalar, vector, or broadcastable tensor form, as long as it can perform element-wise modulation operations with the corresponding hidden layer output.

[0086] Forming residual connections, Indicates and Broadcast of all-one vectors or scalars of the same dimension;

[0087] It is the sigmoid activation function;

[0088] This represents the weight matrix of the gated layer;

[0089] express transpose;

[0090] The input to the conditional gating layer is the embedding of business entity features. ;

[0091] This represents the bias vector in the gating layer;

[0092] Represent the space of real numbers;

[0093] This represents the dimension of the first hidden layer in a deep neural network.

[0094] The dimension representing the embedding of business entity features;

[0095] This represents the output vector of the first hidden layer after being modulated by the gated layer;

[0096] Represents the ReLU activation function;

[0097] This represents the weight matrix of the first hidden layer of a deep neural network.

[0098] This represents the concatenated sample feature embedding vector;

[0099] This represents the bias vector of the first hidden layer of a deep neural network;

[0100] Represents the Hadamard product symbol;

[0101] express The dimension of the embedding vector;

[0102] This represents a deep neural network submodule consisting of at least one layer of nonlinear transformations and linear mappings;

[0103] 4.3 Model Fusion and Prediction:

[0104] ,

[0105] in, It is a predicted value;

[0106] It is the sigmoid activation function;

[0107] This represents the output scalar of the factorization machine module;

[0108] This represents the output scalar of a deep neural network.

[0109] Step 5: Use the recommendation model to recommend data.

[0110] The present invention also discloses a computer system, comprising:

[0111] processor;

[0112] Memory used to store processor-executable instructions;

[0113] The processor is configured to implement a cue recommendation method based on entity embedding dynamically adjusted deep neural networks when executing the executable instructions.

[0114] The present invention also discloses a computer-readable storage medium, comprising:

[0115] A memory on which computer programs are stored;

[0116] A processor for executing the program in the memory to implement a cue recommendation method based on entity embedding dynamically adjusted deep neural networks.

[0117] In summary, by adopting the above technical solution, the present invention is able to:

[0118] (1) Compared to modeling each business entity independently, this invention only requires adding low-dimensional embedding vectors to the entities, and the core parameters of the model are completely shared. This can maintain the lightweight nature of the model and reduce the training cost and industrial deployment difficulty in large-scale business scenarios.

[0119] (2) A simple and effective conditional gating mechanism based on business entities is proposed, which can achieve adaptive structural adjustment without sacrificing scalability.

[0120] (3) Unlike the standard recommendation model that uses feature-indiscriminate concatenation input, the innovative decoupling architecture of “control flow” and “data flow” proposed in this invention can effectively improve the modeling effect of long-tail business entities and solve the technical problem that small sample entities cannot converge.

[0121] (4) Experiments show that this method has a higher AUC level than standard DeepFM and other models, and can more accurately characterize the conversion preferences of clue samples under different business entities, resulting in better clue recommendation quality.

[0122] The effectiveness of the proposed method was verified through large-scale experiments on real-world transformation tasks.

[0123] Additional aspects and advantages of the invention 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 the invention. Attached Figure Description

[0124] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0125] Figure 1 This is a schematic flowchart of the overall process of the present invention.

[0126] Figure 2 This is a schematic block diagram of the standard DeepFM model in the existing technology.

[0127] Figure 3 This is a schematic block diagram of the overall structure of the improved model one of the present invention.

[0128] Figure 4 This is a schematic block diagram of the overall structure of the improved model two provided by the present invention.

[0129] Figure 5 This is a schematic diagram of the model training process in a specific application scenario (clue recommendation) of the present invention.

[0130] Figure 6 This is a schematic diagram of the training process of Model 2 under a specific application scenario (clue recommendation) of the present invention.

[0131] Figure 7 This is a schematic block diagram of a standard gating structure in existing technology.

[0132] Figure 8 This is a schematic block diagram of the dynamic entity condition gating structure of the present invention.

[0133] Figure 9 This is a schematic block diagram of the dynamic entity condition residual gating structure of the present invention. Detailed Implementation

[0134] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown 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 are only used to explain the present invention, and should not be construed as limiting the present invention.

[0135] The overall processing flow of this invention, from raw data parsing and feature engineering to the final generation of prediction probabilities through a dynamic entity gating network, is illustrated in the schematic diagram below. Figure 1 .

[0136] S1, Data parsing module acquisition clueSample data, establish clue The sample database is stored in a structured manner. clue The characteristics of the sample data should include "basic attribute characteristics" and "business statistical characteristics," that is... clue The sample itself should possess attributes; it should also include the "business entity characteristics" (e.g., dealer stores, regions) corresponding to the lead sample. The set of lead feature fields is denoted as... The set of business entity feature fields is denoted as , , ;

[0137] S1-1, basic attribute features include but are not limited to: city, intended car model, gender, age, budget range, and source channel; business statistical features include but are not limited to: historical request count, transaction popularity of car model in city dimension, and conversion statistics of leads in different time windows;

[0138] S1-2, the business entity feature corresponding to the clue sample is at least a unique identifier code, which serves as a discrete feature;

[0139] S2, Feature processing module processes clue sample features and business entity characteristics Perform encoding and feature transformation.

[0140] S2-1, for discrete features Encode the discrete values ​​to unique consecutive integer indices;

[0141]

[0142] yes Number of categories (values) possessed;

[0143] yes No. The index of each possible value;

[0144] S2-2, for continuous features Perform numerical transformation to make it conform to an approximately normal distribution.

[0145]

[0146] It is the cumulative distribution function of the standard normal distribution. It is its inverse function;

[0147] It is the empirical distribution function;

[0148] S2-2-1, specifically, the values ​​of continuous features. Sort and calculate the rank. The rank value represents its relative position within the entire sample. For example: if ,but ;

[0149] S2-2-2, Rank normalization:

[0150]

[0151] in This represents the total number of samples. At this point, the minimum value corresponds to 0, and the maximum value corresponds to 1.

[0152] S2-2-3, truncation:

[0153] ,

[0154] in, This represents the truncated value;

[0155] This indicates taking the smaller value;

[0156] This indicates taking the larger value;

[0157] This represents the normalized value;

[0158] This represents the preset cutoff threshold, whose value can be taken as follows: ;

[0159] S2-2-4, mapped to the standard normal distribution:

[0160]

[0161] The entire process can be viewed as:

[0162]

[0163] S3, the feature embedding module converts features into dense embeddings of a uniform dimension and distinguishes between business entity feature embeddings and clue sample feature embeddings.

[0164] S3-1, transforming features into uniform-dimensional dense embeddings.

[0165] S3-1-1, Discrete features are input into the embedding layer using integer indices to construct the discrete feature embedding matrix. The embedding matrix is ​​initialized using the Xavier uniform initialization method.

[0166]

[0167] It is the first Discrete features The embedding matrix; yes The Middle Value embedding;

[0168] It is a discrete feature The number of categories it possesses;

[0169] It is a discrete feature The Value index (e.g., "Chongqing" is coded as 57, at this time) );

[0170] It is the embedding dimension (e.g., 8, 16, 32 are hyperparameters).

[0171] S3-1-2, Continuous Features By using a fully connected linear mapping to an embedding space of the same dimension as the discrete features:

[0172]

[0173] For continuous characteristic values The numerical conversion result;

[0174] S3-2, group the embedded features according to business semantics, dividing the embeddings into a clue sample feature embedding set and a business entity feature embedding set. (Clues sample feature embedding set) Depend on Embedded mapping Composition, denoted as Business entity feature embedding set ;in Let i be the embedding matrix corresponding to the i-th feature domain; This is the embedding matrix of the feature domains of the business entities.

[0175] S4. Construct an "Entity Embedding Dynamic Adjustment Deep Neural Network" model formed by fusing a "Factor Factorization Machine Structure" with a "Conditional Dynamic Entity Gating Network". The gating modulation layer can be applied to the first layer, any intermediate layer, or multiple layers of the deep neural network. End-to-end joint optimization of the entire model is performed using the sample feature embeddings of clues and the feature embeddings of business entities. The input to the Factor Factorization Machine module consists of a set of clue sample feature embeddings, which are independent of the store feature embedding set. In the Factor Factorization Machine module, feature cross-calculation is performed only between the clue feature embeddings. Figure 2 , Figure 3 , Figure 4 This visually demonstrates the differences between the present invention and existing technologies. (See reference) Figure 2 This invention uses a standard DeepFM model structure and globally shared parameters; the core architecture of this invention is based on... Figure 3 (Improved Model 1) or Figure 4 (Improved Model 2) By introducing business entity feature embedding as a conditional signal, dynamic modulation of the output of the hidden layer of the neural network is achieved.

[0176] S4-1, Building the Model Architecture

[0177] S4-1-1, Constructing the Factorization Machine Module

[0178]

[0179] S4-1-2, Constructing a Conditional Dynamic Entity Gated Network Structure Module

[0180] S4-1-2-1, Define asymmetric input:

[0181] Feature extraction backbone input:

[0182]

[0183] in This represents field-wise concatenation of vectors in the order of their feature domains.

[0184] Input to the conditional gating layer:

[0185]

[0186] This indicates the embedding of business entity features corresponding to the clue sample.

[0187] S4-1-2-2, Constructing the residual gating network:

[0188] Gating signal generation mechanism:

[0189] ,

[0190] Residual connection gating mechanism:

[0191] ,

[0192] ,

[0193] in, It is a gated modulation vector (or scalar). The gated vector can be in scalar, vector, or broadcastable tensor form, as long as it can perform element-wise modulation operations with the corresponding hidden layer output.

[0194] Forming residual connections, Indicates and Broadcast of all-one vectors or scalars of the same dimension;

[0195] Indicates conditional gating layer input;

[0196] It is the sigmoid activation function;

[0197] It is a gated modulation vector (or scalar). The gated vector can be in scalar, vector, or broadcastable tensor form, as long as it can perform element-wise modulation operations with the corresponding hidden layer output.

[0198] This represents a deep neural network submodule consisting of at least one layer of nonlinear transformations and linear mappings.

[0199] In this embodiment, the gating layer is key to achieving a decoupled architecture between the "control flow" and the "data flow". In existing standard gating network structures, the gating signal originates from the feature itself, as referenced... Figure 7 The dynamic entity conditional gating network structure proposed in this invention generates gating vectors through business entity embedding. Perform element-wise multiplication with the hidden layer output, refer to... Figure 8 In another preferred embodiment, the present invention employs... Figure 9 The dynamic entity conditional residual gated network structure shown above has residual connections of the original feature representation superimposed on the output, which can enhance the stability of model training.

[0200] S4-1-3, Model Fusion and Prediction:

[0201]

[0202] It is a predicted value. It is the sigmoid activation function;

[0203] S4-2, Model Training and Parameter Optimization

[0204] This embodiment uses the following training process for joint parameter optimization. Figure 5 and Figure 6 The diagrams show the end-to-end training process of the improved model 1 and improved model 2 provided by this invention in a real-world scenario.

[0205] S4-2-1, Dataset Partitioning and Preprocessing: The data in the sample database is divided into training, validation, and test sets in an 8:1:1 ratio. The validation set is used during model training to monitor model performance and adjust the learning rate and early stopping strategy. The test set is only used to evaluate the model's capabilities and does not participate in any model training process. Strict time blocking is used for dataset partitioning. , , Let these represent the time spans of the training set, validation set, and test set, respectively, and they need to satisfy... and .

[0206] S4-2-2, Batch processing, set the batch size to 2048.

[0207] S4-2-3, Loss Function Construction: The binary cross-entropy loss function is adopted. In implementation, the sigmoid activation function is integrated into the loss function for calculation.

[0208] ,

[0209] in, Represents the loss function;

[0210] Indicates the total sample size;

[0211] It is the first The true label value of each sample;

[0212] This represents the unnormalized log odds (Logits) of the model output.

[0213] It is the sigmoid activation function;

[0214] S4-2-4, Optimizer and Parameter Update: The AdamW optimizer (Adam with Decoupled Weight Decay) is used to adaptively adjust the learning rate. L2 regularization is applied, and the decay coefficient can be... .

[0215] S4-2-5, regularization and early stopping mechanism, sets the dropout rate of the Dropout layer to 0.3, evaluates the AUC metric using the validation set, and terminates training when the validation set shows no improvement within 10 consecutive epochs.

[0216] Compared with the prior art, the present invention has the following beneficial effects:

[0217] This invention achieves decoupling of "control flow" and "data flow" in deep neural networks through a dynamic conditional gating mechanism based on business entity embedding. Compared to modeling each entity independently, this method can achieve adaptive structural adjustment by using low-dimensional embedding vectors of business entities while sharing the backbone parameters of the model. Simultaneously, this mechanism can solve the problem of convergence difficulty in modeling long-tail business entities due to sample sparsity. Experiments show that this invention significantly improves the AUC level of clue sample conversion prediction and has better personalized recommendation performance.

[0218] Table 1. Comparison of AUC performance of different recommendation models on the car sales lead dataset.

[0219]

[0220] Note: N is the number of business entities, P is the number of main model parameters, and E is the number of entity embedding parameters. )

[0221] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A cue recommendation method based on entity embedding dynamically adjusted deep neural networks, characterized in that, Includes the following steps: S1. Obtain sample data, establish a sample database, and perform structured storage; S2, for sample features and business entity characteristics Perform feature encoding and numerical conversion of features; S3 converts features into dense embeddings of a uniform dimension and distinguishes between business entity feature embeddings and sample feature embeddings; S4. Construct an entity embedding dynamically adjustable deep neural network recommendation model formed by fusing a factorization machine structure with a conditional dynamic entity gating network structure. S5 uses the recommendation model to output recommendation data.

2. The cue recommendation method based on entity embedding dynamically adjusted deep neural network according to claim 1, characterized in that, Step S1 includes: The characteristics of the sample include basic attribute characteristics and business statistical characteristics, that is, the attributes of the sample itself; it also includes the corresponding business entity characteristics. The set of sample feature domains is denoted as The set of feature domains of business entities is denoted as , , , in, Represents the set of feature domains of a sample; The union symbol represents the set of sets. Represents the set of feature domains of a business entity; Represents the set of all feature domains; Intersection symbol; This represents the empty set.

3. The cue recommendation method based on entity embedding dynamically adjusted deep neural network according to claim 1, characterized in that, Step S2 includes: Discrete features Encode the discrete values ​​by mapping them to unique consecutive integer indices: ， in, Indicates the first Discrete features The set of all possible indexes; Indicates the first Discrete features The index of the first value; Representing discrete features The index of the second value; It is a discrete feature No. The index of each value; It is a discrete feature The number of categories it possesses; For continuous features Perform a numerical transformation to make it conform to an approximately normal distribution: ， in, This represents a continuous eigenvalue that, after data transformation, follows an approximately normal distribution. It is the cumulative distribution function of the standard normal distribution; yes The inverse function; This is the empirical distribution function.

4. The cue recommendation method based on entity embedding dynamically adjusted deep neural network according to claim 1, characterized in that, Step S3 includes: Discrete features are input into the embedding layer using integer indices to construct a discrete feature embedding matrix; the embedding matrix is ​​initialized using the Xavier uniform initialization method. ， in, It is the first Discrete features The embedding matrix; It is the first Discrete features No. A specific value in The corresponding embedding in; It is a discrete feature The Value index; Indicates that it belongs to the symbol; Represents the space of real numbers; It is a discrete feature The number of categories it possesses; It is the embedded dimension; Continuous features By using a fully connected linear mapping to an embedding space of the same dimension as the discrete features: ， in, Representing continuous features The embedding vector; Represents the weight vector of a linear mapping; This represents the continuous feature input value after numerical conversion; Indicates the bias term; Indicates that it belongs to the symbol; Represents the space of real numbers; It is the embedded dimension; The embedded features are grouped according to business semantics, dividing the embeddings into a sample feature embedding set and a business entity feature embedding set; the sample feature embedding set... Depend on Embedded mapping Composition, denoted as Business entity feature embedding set ;in It is the first Discrete features The embedding matrix; This is the embedding matrix of the feature domains of the business entities.

5. The cue recommendation method based on entity embedding dynamically adjusted deep neural network according to claim 1, characterized in that, The factorization machine in step S4 includes: ， in, This represents the output scalar of the factorization machine module; Indicates the total number of feature domains; Indicates the first Linear weight parameters for each feature domain; Indicates the first Input values ​​for each feature domain; Represents the dot product operation of vectors; They represent the first The first feature domain and the first Embedding vectors corresponding to each feature domain.

6. The cue recommendation method based on entity embedding dynamically adjusted deep neural network according to claim 1, characterized in that, The main input for feature extraction in step S4 is: ， in, This represents the concatenated sample feature embedding vector; This represents the concatenation of vectors in the order of their feature domains. Represents the set of sample feature embeddings The Middle Feature embedding vectors corresponding to each feature domain; Indicates that it belongs to the symbol; This represents the set of sample feature embeddings.

7. The cue recommendation method based on entity embedding dynamically adjusted deep neural network according to claim 1, characterized in that, The gating signal generation in step S4 includes: ， Residual connection gating includes: ， ， in, It is a dynamically generated gated modulation vector. Forming residual connections, Indicates and Broadcast of all-one vectors or scalars of the same dimension; It is the sigmoid activation function; This represents the weight matrix of the gated layer; express transpose; Indicates conditional gating layer input; This represents the bias vector in the gating layer; Represents the space of real numbers; This represents the dimension of the first hidden layer in a deep neural network. Indicates the dimension of the embedding of business entity features; This represents the output vector of the first hidden layer after being modulated by the gated layer; Represents the ReLU activation function; This represents the weight matrix of the first hidden layer of a deep neural network. This represents the concatenated sample feature embedding vector; This represents the bias vector of the first hidden layer of a deep neural network; Represents the Hadamard product symbol; express The dimension of the embedding vector; This represents a deep neural network submodule consisting of at least one layer of nonlinear transformations and linear mappings.

8. The cue recommendation method based on entity embedding dynamically adjusted deep neural network according to claim 1, characterized in that, The model fusion and prediction in step S4 include: ， in, It is a predicted value; It is the sigmoid activation function; This represents the output scalar of the factorization machine module; This represents the output scalar of a deep neural network.

9. A computer system, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the cue recommendation method based on entity embedding dynamically adjustable deep neural network as described in any one of claims 1 to 8 when executing the executable instructions.

10. A computer-readable storage medium, characterized in that, include: A memory on which computer programs are stored; A processor for executing the program in the memory to implement the cue recommendation method based on entity embedding dynamically adjusted deep neural networks as described in any one of claims 1 to 8.

Images