Ranking Learning Methods, Systems, and Applications Based on Evolutionary Conditional Generative Adversarial Networks

Abstract

This invention relates to a ranking learning method, system, and application based on evolutionary conditional generative adversarial networks (GANs). The method includes the following steps: Step 1: Initialize relevant parameters and input a ranking learning dataset; Step 2: Initialize a generator population and a discriminator; Step 3: Iteratively train the generator population and discriminator in an adversarial manner to generate an optimal model, wherein the generator population is trained using an evolutionary strategy; Step 4: Predict and rank information retrieval targets based on the selected optimal model. Compared with existing technologies, this invention incorporates an evolutionary strategy into a conditional generative adversarial network to design a Listwise ranking learning method, employing multiple loss functions for optimization, leveraging their respective strengths to compensate for weaknesses, thereby evolving and finding better solutions to produce a high-performing model.

Description

Technical Field

[0001] This invention relates to the fields of information retrieval and deep learning, and in particular to a ranking learning method based on evolutionary conditional generative adversarial networks. Background Technology

[0002] Learning to rank (LTR) is a machine learning technique that combines a large number of ranking features to automatically train a ranking model to solve ranking problems in fields such as information retrieval. It plays a crucial role in practical applications such as search engines, e-commerce systems, recommendation systems, and question-answering systems, and increasingly, LTR methods are being widely applied in these scenarios. Many well-known internet companies, such as Baidu, Sogou, Yahoo, Google, Yandex, and Bing search engines, rely on LTR methods to provide users with high-quality search results. For example, the LTR algorithm RankNet has been applied to Microsoft's Bing search engine to optimize the ranking of web search results, and the LTR algorithm LogisticRank has been applied to Yahoo's search engine to optimize the ranking of search results.

[0003] Based on the different input data examples used to train the model, ranking learning methods can be divided into three main types: Pointwise (single document level), Pairwise (document pair level), and Listwise (document list level). Among them, Listwise ranking learning methods have a more natural model design for ranking problems. They use the list of all search document results corresponding to each query as a training instance to directly optimize the ranking list, which is more in line with the essence of ranking. In recent years, Listwise ranking learning methods have become increasingly popular.

[0004] Based on the different machine learning techniques used in training the model, ranking learning methods can be categorized into: perceptron-based, neural network-based, support vector machine-based, extreme learning machine-based, Bayesian-based, boosting-based, tree-based, evolutionary algorithm-based, deep learning-based, and other ranking learning methods. With the explosive development of deep learning technology, deep learning-based ranking learning methods have gradually become the mainstream approach in academia and industry. For example, using recurrent neural networks, convolutional neural networks, deep neural networks, deep Q-networks, deep reinforcement learning, and generative adversarial networks to design ranking learning methods to solve ranking problems in information retrieval has become a hot research topic.

[0005] The first pioneering work using generative adversarial networks (GANs) to solve information retrieval problems was the Information Retrieval Generative Adversarial Networks (IRGAN) proposed by Wang et al. in 2017. It borrows the idea of ​​a generator and discriminator adversarially competing in GANs, employing the minimax algorithm from game theory to integrate the generative and discriminative retrieval models into a unified framework through adversarial training. This allows the two models to improve each other, ultimately leading to more accurate document retrieval. IRGAN only designed Pointwise IRGAN and Pairwise IRGAN ranking learning methods, but not Listwise ranking learning methods. In 2022, Hai-Tao Yu et al. proposed Listwise adversarial ranking learning based on IRGAN, which we will name IRGAN-Listwise here, filling the gap in IRGAN's lack of a Listwise IRGAN. IRGAN-Listwise uses a fixed and static single adversarial target defined in traditional generative adversarial networks to iteratively train the generator and discriminator, which is not conducive to improving the generation performance from multiple perspectives, and its effectiveness needs to be further improved. Summary of the Invention

[0006] To overcome the shortcomings of the existing technologies, this patent innovatively integrates generative adversarial network technology and evolutionary strategies to provide a listwise ranking learning method (Evolutionary Conditional Generative Adversarial Networks-Learning ToRank, ECGAN-LTR) based on evolutionary conditional generative adversarial networks for the ranking learning problem.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] As a first aspect of the present invention, a ranking learning method based on evolutionary conditional generative adversarial networks is provided, comprising the following steps:

[0009] Step 1: Initialize relevant parameters and input the sorting learning dataset;

[0010] Step 2: Initialize the generator population and discriminator;

[0011] Step 3: Alternately iterate and adversarially train the generator population and discriminator to produce the optimal model, wherein the generator population is trained using an evolutionary strategy;

[0012] Step 4: Based on the selected optimal model, predict and rank the information retrieval targets.

[0013] As a preferred technical solution, the initialization-related parameters include: the initialization generator population size μ, and the number of generator mutation operators N. m The total number of training iterations is epochs, and the discriminator update step size N in each epoch is... d The top-K batch size K and the hyperparameters of the Adam optimizer.

[0014] As a preferred technical solution, the initialization of the generator population and discriminator specifically involves using random weights. Initialize the generator population Initialize the discriminator D with random weights φ. φ .

[0015] As a preferred technical solution, the alternating iterative adversarial training generator population and discriminator specifically include:

[0016] For each epoch, the generator population is trained once, and the discriminator is trained N times. d This method involves iteratively training the generator population and discriminator by minimizing and maximizing the overall optimization objective function, until the algorithm converges or reaches a given number of epochs.

[0017] The overall optimization objective function is as follows:

[0018]

[0019] Where θ and φ represent the parameter sets of the generator G and discriminator D, respectively, r represents the relevance label between the document and the query, N represents the total number of queries, P represents the probability distribution, and E represents the expected value; P true (list|q n (r) represents the probability distribution of positive samples, i.e., given a query q n The probability distribution of the ground-truth document sorting list obtained by sorting the documents according to their relevance tags; P θ (list|q n (r) represents the generative retrieval model, which is generated by the generator G. θ From a given query q n The probability distribution of the top-K document ranking list obtained by downsampling is the probability distribution of the negative samples obtained by sampling; f φ (list|q n (r) represents the discriminative retrieval model, which is composed of a discriminator D. φEstimate the probability that a sorted list is either a ground-truth sorted list or a generator-generated document sorted list; PL(f φ (list|q n ,r)) represents the sorting probability distribution calculated using the Plackett-Luce model; P represents the distribution of ground-truth documents in a sorted learning dataset. true (list|q n The mathematical expectation of r). P represents the document permutation distribution of the negative samples obtained from sampling. θ (list|q n The mathematical expectation of r);

[0020] Generator G θ Considered an evolutionary population, discriminator D φ As the environment, for the evolutionary operation in each epoch, the generator G θ Update different mutation operators, including likelihood, cosine, and cross-entropy functions, to adapt to the current environment D. φ Choose a high-performance offspring generator. Participate in subsequent adversarial iterative game training.

[0021] As a preferred technical solution, the training of the generator population specifically includes:

[0022] Step a.1. For the generator population G θ Each generator in For different mutation operators, the generator is iteratively trained according to the following steps. Where 1≤j≤μ, 1≤h≤N m ;

[0023] Step a.1.1. Using a generator Sample top-K data from the ranking training set based on scores.

[0024] Step a.1.2. Optimize the generator objective function using policy gradient, according to... Calculate the gradient of the selected mutation operator to update the gradient. The generator objective function is as follows:

[0025]

[0026] In the formula, θ is the parameter set of the generator, N is the total number of queries, and r is the relevance label between the document and the query; P θ (list|qn (r) represents the generative retrieval model, which is generated by the generator G. θ From a given query q n The probability distribution of the top-K document ranking list obtained by downsampling is the probability distribution of the negative samples obtained by sampling; f φ (list|q n (r) represents the discriminative retrieval model, which is composed of a discriminator D. φ Estimate the probability that a sorted list is either a ground-truth sorted list or a generator-generated document sorted list; PL(f φ (list|q n ,r)) represents the sorting probability distribution calculated using the Plackett-Luce model; P represents the document permutation distribution of the negative samples obtained from sampling. θ (list|q n The mathematical expectation of r);

[0027] The mutation operator h includes the Likelihood, Cosine, and CrossEntropy operators;

[0028] Step a.1.3. Update the sub-network parameters The generator network is optimized using the Adam optimizer to update its parameters, generating the offspring generator. Right now Where, θ j Let α, β1, and β2 be the network parameters of the j-th generator, and let α, β2 be the hyperparameters of the Adam optimizer.

[0029] Step a.1.4. According to the fitness function The calculation method is used to evaluate the fitness function value F of the offspring generator. j,h , where f φ (d z ,q n ) represents the scoring function of the discriminator, PL(f φ (d z ,q n )) represents the ranking probability calculated using the Plackett-Luce model, where K is the total number of documents;

[0030] Step a.2. Generate the offspring generator population G. θ The fitness function values ​​F j,h Sort the data and generate a sorted list. Right now

[0031] Step a.3. Sort the list of each offspring generator by fitness function value. Select the optimal generator network and update the network parameters of each child generation. Generate a new generator population.

[0032] As a preferred technical solution, the likelihood operator is as follows:

[0033]

[0034] in, Represents generator p θ Given a query q, obtain documents d from the top-K document sorted list list. z The rating values, among which, This refers to the document d at position z in the sorted list list. z ; This represents the ranking probability calculated using the Plackett-Luce model, and its calculation method is as follows: It is determined by the discriminator D φ The reward function value used for policy gradient-based reinforcement learning is calculated as follows:

[0035]

[0036] in, This represents the ranking probability calculated using the Plackett-Luce model;

[0037] The likelihood ranking probability of the generator is maximized by minimizing the likelihood mutation operator to confuse the discriminator's discrimination behavior.

[0038] As a preferred technical solution, the cosine variation operator is as follows:

[0039]

[0040] in, Represents generator p θ Given a query q, obtain documents d from the top-K document sorted list list. z The rating value, It is determined by the discriminator D φ The reward function value used to evaluate the policy gradient in policy gradient-based reinforcement learning; Let ψ(·) represent a mapping function that preserves the documents d in a sorted list of documents in a given query q. z The rating values, this list is based on the rating values ​​of each document.z The ideal sorted list, i.e., the ground-truth sorted list, is formed by sorting the relevance label values ​​of the query q from largest to smallest.

[0041] As a preferred technical solution, the CrossEntropy operator is as follows:

[0042]

[0043] in, Represents generator p θ Given a query q, obtain documents d from the top-K document sorted list list. z The rating value, It is determined by the discriminator D φ The feedback is the reward function value used for the policy gradient in policy gradient-based reinforcement learning. Let ψ(·) represent a mapping function that preserves the documents d in a most ideally sorted list associated with a given query q. z The rating value, Represents a most idealized sorted list of documents d in the list. z The ordering probability, Document d represents a sorted list of predictions from a model. z The ordering probabilities are all calculated using the softmax function, and their calculation methods are as follows:

[0044] and

[0045] As a preferred technical solution, the training of the discriminator specifically includes:

[0046] Perform N on the discriminator d Iterative training with operations of ×μ times or less:

[0047] Step b.1. Given query q n Use the current generator Sample the top-K negative samples from the ranking learning training dataset based on their scores. And compared with the top-K positive sample data sorted by relevance labels. Combine;

[0048] Step b.2. Train the discriminator using the discriminator objective function based on the permutations generated by the current generator Gθj and the permutations generated by the ground-truth labels. The discriminator objective function is as follows:

[0049]

[0050] In the formula, φ represents the parameter set of the discriminator D, r represents the relevance label between the document and the query, N represents the total number of queries, and P true (list|q n (r) represents the probability distribution of positive samples, i.e., given a query q n The probability distribution of the ground-truth document sorting list obtained by sorting the documents according to their relevance tags; This represents a generative retrieval model, which is generated by a generator. From a given query q n The probability distribution of the top-K document ranking list obtained by downsampling is the probability distribution of the negative samples obtained by sampling; f φ (list|q n (r) represents the discriminative retrieval model, which is composed of a discriminator D. φ Estimate the probability that a sorted list is either a ground-truth sorted list or a generator-generated document sorted list; PL(f φ (list|q n ,r)) represents the sorting probability distribution calculated using the Plackett-Luce model; P represents the ground-truth document permutation distribution P of the sorting training dataset. true (list|q n The mathematical expectation of r). This represents the document arrangement distribution of the negative samples obtained from sampling. The mathematical expectation;

[0051] Calculate the gradient g of the discriminator's objective function. φ ,Right now

[0052]

[0053] In the formula, This represents the scoring function of the discriminator. This represents the ranking probability calculated using the Plackett-Luce model, where K is the total number of documents;

[0054] Step b.3. Optimize the discriminator network using the Adam optimizer to update the discriminator network parameters φ, i.e.

[0055] φ←Adam(g φ ,φ,α,β1,β2)

[0056] In the formula, α, β1, and β2 are the hyperparameters of the Adam optimizer.

[0057] As a preferred technical solution, the step of predicting and ranking information retrieval targets based on the selected optimal model specifically includes: testing the ranking learning test dataset based on the selected optimal model to compare the effects of different ranking learning methods, and also applying it to the ranking learning system to predict the ranking of information retrieval targets, and then applying it to the information retrieval system to solve practical application problems.

[0058] As a second aspect of the present invention, a system is provided that employs any of the above-described ranking learning methods based on evolutionary conditional generative adversarial networks, the system comprising:

[0059] The model training module utilizes a ranking learning method based on evolutionary conditional generative adversarial networks to train the model. It includes a generator model training submodule and a discriminator model training submodule, and its functions include evaluating and saving the trained model. The model training module performs the following operations: loading the ranking learning training dataset; setting the relevant parameters of the ranking learning method; and conducting iterative game training on the selected ranking learning training dataset using the ranking learning method to obtain the generator population model and the discriminator model.

[0060] Model Validation Module: This module validates, selects, and saves the optimal model using the model trained in the model training module. It includes a generator model validation submodule and a discriminator model validation submodule. Its functions include evaluating and selecting / saving the validated model. The model validation module performs the following operations: loading the ranking learning validation dataset; validating the model trained in the model training module on the selected ranking learning validation dataset to obtain the model with the best performance evaluation; and selecting and saving the optimal model as the test model for the model prediction module.

[0061] The model testing module utilizes the optimal performance model selected in the model validation module to test the model. It includes a generator model testing submodule and a discriminator model testing submodule, and its functions include performance evaluation and prediction result saving. The performance evaluation metrics include effectiveness evaluation metrics in the information retrieval domain. The model testing module performs the following operations: loading the ranking learning test dataset; testing the optimal performance model saved in the model validation module on the selected ranking learning test dataset to obtain relevance scores and some performance metric evaluation values ​​in information retrieval; determining whether the relevance score is greater than a set threshold: if so, returning documents that are relevant to the query; otherwise, returning documents that are irrelevant to the query.

[0062] As a third aspect of the present invention, an application of any of the above-described ranking learning methods based on evolutionary conditional generative adversarial networks is provided.

[0063] The ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks was applied to search engines including Baidu, Google, Bing, Yahoo, Sogou, and Yandex. The specific application steps are as follows:

[0064] Step 1: Embed the ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks into the search engine;

[0065] First, obtain the web pages from the search engine's web page index database to obtain initial data for ranking learning;

[0066] Secondly, data preprocessing is performed on a portion of the web pages in the search engine's web page index database, ranking features are extracted and relevance is labeled on the web pages, and a ranking learning dataset for the search engine is constructed in the format of a standard ranking learning dataset.

[0067] Then, on the constructed ranking learning dataset, the ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks is used to alternately iterate game adversarial optimization training, validation and selection of models to produce the optimal generator model and discriminator model.

[0068] Finally, the resulting optimal model is incorporated into the search engine's ranking system;

[0069] Step 2: Execute query - web search, evaluate the performance of the best model generated by the embedded Evolutionary Conditional Generative Adversarial Network-based ranking learning method ECGAN-LTR, predict and present the ranking results;

[0070] In search engines that incorporate the ranking learning method ECGAN-LTR, users can repeatedly execute queries—web page searches.

[0071] First, the user enters the query term they want to search in the search box of the search engine and clicks search to perform a web search;

[0072] Secondly, the search engine's ranking system finds all web pages containing the query term from the search engine's web page index database, evaluates the performance of the optimal model generated by the embedded Evolutionary Conditional Generative Adversarial Network-based ranking learning method ECGAN-LTR, calculates the web page ranking score, and predicts the ranking result of the web page search.

[0073] Finally, the web search ranking results are returned to the search page in a certain way to be presented to the search user.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] This invention provides a novel and more effective listwise ranking learning method based on evolutionary conditional generative adversarial networks (ECGAN-LTR). ECGAN-LTR addresses the ranking learning problem by incorporating evolutionary strategies such as mutation, evaluation, and selection within the conditional generative adversarial network framework to evolve the generator population. It then engages in adversarial gameplay with the discriminator, iteratively training the optimal generator and discriminator models alternately. ECGAN-LTR integrates evolutionary strategies into the conditional generative adversarial network. In each iteration, the generator adapts to the current environment (i.e., the discriminator) based on multiple loss functions, eliminating individuals with low fitness. Furthermore, compared to a single loss function, this invention employs multiple loss functions (likelihood mutation, cosine mutation, and cross-entropy mutation) for optimization, allowing them to complement each other and evolve to find better solutions, resulting in a superior model. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating the ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks of the present invention.

[0077] Figure 2 This is a flowchart illustrating the ranking learning system based on evolutionary conditional generative adversarial networks of the present invention.

[0078] Figure 3 This is a functional structure diagram of the ranking learning system based on evolutionary conditional generative adversarial networks of the present invention;

[0079] Figure 4 This is an interface diagram of the ranking learning system based on evolutionary conditional generative adversarial networks in an embodiment of the present invention;

[0080] Figure 5 A comparison of the P@K performance metrics of the ECGAN-LTR method and the IRGAN-Listwise method on the ranking learning dataset OHSUMED;

[0081] Figure 6 A comparison of the NDCG@K performance metrics of the ECGAN-LTR method and the IRGAN-Listwise method on the ranking learning dataset OHSUMED;

[0082] Figure 7 A comparison of the P@K performance metrics of the ECGAN-LTR method and the IRGAN-Listwise method on the MQ2008 sorting learning dataset;

[0083] Figure 8A comparison of the NDCG@K performance metrics of the ECGAN-LTR method and the IRGAN-Listwise method on the MQ2008 sorting learning dataset;

[0084] Figure 9 A comparison of the P@K performance metrics of the ECGAN-LTR method and the IRGAN-Listwise method on the MSLR-WEB10K ranking learning dataset;

[0085] Figure 10 A comparison of the NDCG@K performance metrics of the ECGAN-LTR method and the IRGAN-Listwise method on the MSLR-WEB10K ranking learning dataset;

[0086] Figure 11 This is a schematic diagram illustrating the application of ECGAN-LTR, a ranking learning method based on evolutionary conditional generative adversarial networks. Detailed Implementation

[0087] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0088] Example 1

[0089] As a first aspect of the present invention, an evolutionary conditional generative adversarial network-based ranking learning method, ECGAN-LTR, is provided. This method is a novel approach that incorporates evolutionary operations such as mutation, evaluation, and selection into the Conditional Generative Adversarial Network (CGAN) framework to evolve the generator population in CGAN and engage in an iterative adversarial game with the discriminator to generate a high-performance model. The method models the ranking learning process as an iterative game between two opponents (i.e., a generator population and a discriminator). The generator aims to generate (or select) permutations that resemble ground-truth rankings to deceive the discriminator, while the discriminator aims to clearly distinguish between ground-truth permutations and permutations generated by its opponent's generator. In the ECGAN-LTR method, the optimal parameters θ and φ of the generator and discriminator can be iteratively learned by minimizing and maximizing the same overall optimization objective function (1) to train and adversarially optimize their respective models.

[0090]

[0091] Among them, generator G θ The aim is to minimize the objective function, that is, to minimize the objective function expression (2):

[0092]

[0093] The generator fits all possible permutations of list ~ p true (list|q n The true distribution on r) is obtained, and random permutations are generated to deceive the discriminator. The discriminator evaluates the probability that a permutation is a ground-truth permutation or a permutation generated by the generator. Its goal is to maximize the log-likelihood of correctly distinguishing between ground-truth permutations and permutations generated by the generator, that is, to maximize the objective function (3):

[0094]

[0095] Where θ and φ represent the parameter sets of the generator G and discriminator D, respectively, r represents the relevance label between the document and the query, N represents the total number of queries, P represents the probability distribution, and E represents the expected value; P true (list|q n (r) represents the probability distribution of positive samples, i.e., given a query q n The probability distribution of the ground-truth list of documents obtained by sorting them according to their relevance labels; P θ (list|q n (r) represents the generative retrieval model, which is generated by the generator G. θ From a given query q n The probability distribution of the top-K document ranking list obtained by downsampling is the probability distribution of the negative samples obtained by sampling; f φ (list|q n (r) represents the discriminative retrieval model, which is composed of a discriminator D. φ Estimate the probability that a sorted list is either a ground-truth sorted list or a generator-generated document sorted list; PL(f φ (list|q n ,r)) represents the sorting probability distribution calculated using the Plackett-Luce model; P represents the ground-truth document permutation distribution P of the sorting training dataset. true (list|q n The mathematical expectation of r). P represents the document permutation distribution of the negative samples obtained from sampling. θ (list|q n The mathematical expectation of ,r).

[0096] Generator G θ Considered an evolutionary population, discriminator D φ As the environment. For each round of evolutionary operations, the generator G... θ Update different mutation functions to adapt to the current environment D φ The mutation function mainly includes the likelihood function, cosine function, and cross-entropy function. According to the "survival of the fittest" principle, only the best-performing offspring generators will be selected. It survives and participates in subsequent adversarial iterative game training. Unlike the two-player minimax game with a single fixed and static adversarial training objective in the Information Retrieval Generative Adversarial Network (IRGAN), the ECGAN-LTR ranking learning method evolves and seeks better solutions by incorporating the advantages of multiple dynamic adversarial objectives with different mutation functions to produce a high-performance model.

[0097] The workflow framework of ECGAN-LTR, a ranking learning method based on evolutionary conditional generative adversarial networks, is as follows: Figure 1 As shown, the process mainly includes four steps: initializing relevant parameters and input ranking learning dataset, initializing generator population and discriminator, alternating iterative adversarial training of generator population and discriminator to generate optimal model, and predicting and ranking information retrieval targets based on the selected optimal model.

[0098] The specific steps of ECGAN-LTR, a ranking learning method based on evolutionary conditional generative adversarial networks, are as follows:

[0099] Step 1: Initialize the relevant parameters and input the sorting learning dataset.

[0100] Initialize the generator population size μ and the number of generator mutation operators N. m The total number of training iterations is epochs, and the discriminator update step size N in each epoch is... d The top-K batch size K, the hyperparameters of the Adam optimizer such as α, β1 and β2, and the input sorting training dataset are given.

[0101] Step 2: Initialize the generator population and discriminator.

[0102] Use random weights Initialize the generator population Initialize the discriminator D with random weights φ. φ .

[0103] Step 3: Alternately iterate and adversarially train the generator population and discriminator to produce the optimal model.

[0104] For each epoch, the generator population is trained once, and the discriminator is trained N times. dThis method involves iteratively training the generator population and discriminator by minimizing and maximizing the same overall optimization objective function (1) alternately until the algorithm converges or reaches a given number of epochs. The specific training methods for the generator population and discriminator are as follows:

[0105] (a) Training the generator population: Training the generator population by implementing evolutionary strategies such as mutation, evaluation, and selection.

[0106] Step a.1 For the generator population G θ Each generator in For different mutation operators, the generator is iteratively trained according to the following steps. Where 1≤j≤μ, 1≤h≤N m :

[0107] Step a.1.1 Using a generator Sample top-K data from the ranking training set based on scores.

[0108] Step a.1.2 Optimize the objective function (2) using policy gradient, according to... Calculate the gradient of the selected mutation operator to update the gradient. The mutation operator h can include operators such as likelihood mutation, cosine mutation, and cross-entropy mutation. The likelihood mutation operator, cosine mutation operator, and cross-entropy mutation operator here are variant mutation operators designed based on the ideas of Likelihood Loss in the ranking learning method ListMLE, Cosine Loss in the ranking learning method RankCosine, and Cross Entropy Loss in the ranking learning method ListNet. Their calculation methods are designed as shown in formulas (4) to (6):

[0109] (1) Likelihood operator:

[0110]

[0111] Here, Represents generator p θ Given a query q, obtain documents d from the top-K document sorted list list. z The rating values, among which, This refers to the document d at position z in the sorted list list. z ; This represents the ranking probability calculated using the Plackett-Luce model, and its calculation method is as follows: It is determined by the discriminator D φ The reward function value used in policy gradient-based reinforcement learning is calculated as follows: The likelihood ranking probability of the generator is maximized by minimizing the likelihood mutation operator to confuse the discriminator's discrimination behavior.

[0112] (2) Cosine variation operator:

[0113]

[0114] Here, and The meaning of the likelihood mutation operator and The meanings of these correspond to the same thing; Let ψ(·) represent a mapping function that preserves the documents d in a most ideally sorted list associated with a given query q. z The rating values, this list is based on the rating values ​​of each document. z The ideal sorted list, i.e., the ground-truth sorted list, is formed by sorting the relevance label values ​​of the query q from largest to smallest.

[0115] (3) CrossEntropy operator:

[0116]

[0117] Here, and The meaning of the likelihood mutation operator and The meanings correspond to the same thing. The meaning of corresponds to the meaning in the cosine mutation operator; Represents a most idealized sorted list of documents d in the list. z The ordering probability, Document d represents a sorted list of predictions from a model. z The ordering probabilities are all calculated using the softmax function, and their calculation methods are as follows:

[0118] and

[0119] Step a.1.3 Update the child network parameters The generator network is optimized using the Adam optimizer to update its parameters, generating the offspring generator. Right now

[0120] Step a.1.4 According to the fitness function The calculation method is used to evaluate the fitness function value F of the offspring generator. j,h .

[0121] Step a.2 Generate the offspring generator population G. θ The fitness function values ​​F j,h Sort the data and generate a sorted list. Right now

[0122] Step a.3 generates a list of each offspring generator sorted by its fitness function value. Select the optimal generator network and update the network parameters of each child generation. Generate a new generator population.

[0123] (II) Training the discriminator.

[0124] Perform N on the discriminator d Iterative training with operations of ×μ times or less:

[0125] Step b.1 Given query q n Use the current generator Sample the top-K negative samples from the ranking learning training dataset based on their scores. And compared with the given top-K positive sample data sorted by relevance labels. Combine;

[0126] Step b.2 Based on the current generator The generated permutations and the permutations generated by the ground-truth labels are used to train the discriminator using function (3), that is, to compute the gradient g of the discriminator objective function (3). φ , that is

[0127]

[0128] Step b.3 optimizes the discriminator network using the Adam optimizer to update the discriminator network parameters φ, i.e., φ←Adam(g φ ,φ,α,β1,β2).

[0129] Step 4: Based on the selected optimal model, predict and rank the information retrieval targets.

[0130] The selected optimal model can be tested on a ranking learning test dataset to compare the effectiveness of different ranking learning methods. It can also be applied to ranking learning systems to predict the ranking of information retrieval targets, and then applied to information retrieval systems to solve practical application problems.

[0131] This invention provides a novel and more effective listwise ranking learning method, ECGAN-LTR. This method addresses the ranking learning problem by incorporating evolutionary strategies such as mutation, evaluation, and selection into the Conditional Generative Adversarial Network framework to evolve the discriminator population. It also engages in adversarial game with the generator to iteratively train the optimal discriminator model and generator model alternately.

[0132] 1) An innovative sorting learning method of the Listwise type was proposed.

[0133] The ranking learning method based on evolutionary conditional generative adversarial networks (GANs) invented in this patent is a novel listwise ranking learning method. This method provides a new type of listwise ranking learning approach based on evolutionary GANs for information retrieval, establishing a new paradigm for the research of listwise ranking learning that integrates GANs and evolutionary algorithms. It pioneers the application of evolutionary GANs in designing listwise ranking learning methods, opening up a new perspective for the research of ranking learning methods and exploring a new approach for colleagues dedicated to ranking learning research.

[0134] 2) The technical effect achieved by this invention is as follows: The Listwise ranking learning method ECGAN-LTR integrates an evolutionary strategy into the conditional generative adversarial network. In each evolution, the generator adapts to the current environment (i.e., the discriminator) based on multiple different loss functions, removing individuals with low fitness. Compared to a single loss function, using multiple loss functions (likelihood mutation, cosine mutation, cross-entropy mutation) for optimization allows for mutual optimization, thereby evolving and finding better solutions to produce a high-performing model.

[0135] Example 2

[0136] As a second aspect of the present invention, in order to verify the effectiveness of the invented listwise ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks, this embodiment implements a system incorporating the listwise ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks using Python language under the PyCharm 2019 integrated development environment and the deep learning framework PyTorch. The basic process framework of the system is as follows: Figure 2As shown, this method trains a model on a ranking learning training dataset using the ECGAN-LTR (Evolutionary Conditional Generative Adversarial Network) listwise ranking learning method described in the previous embodiment. The trained model is then validated on a ranking learning validation dataset to select and save the optimal model. Finally, the saved optimal model is used to predict results on a ranking learning test dataset to obtain the test outcome. For model training, the generator population is evolved based on the ranking learning training dataset, and it is trained adversarially with the discriminator to optimize the model. For model validation, based on the ranking learning validation dataset, the performance metrics of each generation of models are calculated, and the optimal generator and discriminator models are selected and saved according to the principle of survival of the fittest. For model testing, based on the ranking learning test dataset, the saved optimal model is used to test and predict the ranking results, and the predicted ranking results are output.

[0137] The ranking learning system based on evolutionary conditional generative adversarial networks mainly includes a model training module, a model validation module, and a model testing module. Its functional structure is as follows: Figure 3 As shown in the diagram, the model training module utilizes the Listwise Ranking Learning Method (ECGAN-LTR) based on Evolutionary Conditional Generative Adversarial Networks designed in this invention to train the model. It mainly includes a generator model training submodule and a discriminator model training submodule, which primarily implement functions such as evaluating and saving the trained model. The main operations within the model training module include: loading the ranking learning training dataset; setting the relevant parameters of the ECGAN-LTR ranking learning method; and performing alternating iterative game training on the selected ranking learning training dataset using the ECGAN-LTR ranking learning method to obtain the generator population model and the discriminator model.

[0138] The model validation module uses the model trained in the model training module to validate, select, and save the optimal model. It mainly includes a generator model validation submodule and a discriminator model validation submodule, which mainly implement functions such as evaluating the validation model and selecting and saving the validation model. The main operations in the model validation module are: loading the ranking learning validation dataset; validating the model trained in the model training module on the selected ranking learning validation dataset to obtain the model with the best performance evaluation (including the optimal discriminator model and the optimal generator model), and selecting and saving the optimal model as the test model for the model testing module.

[0139] The model testing module uses the optimal model ultimately selected and saved in the model validation module to test the model. It mainly includes a generator model testing submodule and a discriminator model testing submodule, which primarily implement performance evaluation and prediction result saving functions. Performance evaluation metrics mainly include effectiveness evaluation metrics in the information retrieval field, such as Normalized Discounted Cumulative Gain (NDCG) and Precision (P). The main operations in the model testing module are: loading the ranking learning test dataset; testing the optimal performance model saved in the model validation module on the selected ranking learning test dataset to obtain relevance scores and some performance metric evaluation values ​​in information retrieval; and determining whether the relevance score is greater than a set threshold: if yes, return documents that are relevant to the query (represented by 1); otherwise, return documents that are irrelevant to the query (represented by 0).

[0140] After installing Python 3.8 and PyCharm 2019 or later, along with PyTorch 1.12.1, PrettyTable 3.4.1, and other software, you can run a ranking learning system based on evolutionary conditional generative adversarial networks. The system provides a GUI interface, such as... Figure 4 As shown, its usage and operation methods are as follows:

[0141] First, the user clicks the "Ranking Learning Dataset Path Selection" button to select the path to the ranking learning dataset. Second, the user clicks the "Generator and Discriminator Model Output Path Selection" button to select the output paths for the generator and discriminator models. Next, the user selects the desired ranking learning dataset by clicking the down arrow (drop-down list) in the ranking learning dataset selection box, such as OHSUMED, MQ2008, and MSLR-WEB10K. Then, the user can modify the default settings of some parameters, such as the size of k in top-k, the generator population size, the generator learning rate, the discriminator learning rate, the discriminator update step size, and the epochs of the generative adversarial network. Finally, the user clicks the "Train Generator Population and Discriminator Model" button to run the Listwise ranking learning method ECGAN-LTR to train the generator population and discriminator models through alternating iterative game theory, and waits for the results. Finally, the user clicks the "Test Optimal Generator and Discriminator Models" button to test the optimal generator and discriminator models generated by running the Listwise sorting learning method ECGAN-LTR, and waits for the results.

[0142] For the ECGAN-LTR method, the adjustable parameters include: the selection of the ranking training dataset, the setting of the top-k size (default value is 2, adjustable range is an integer greater than 1; when the top-k value is greater than the number of ranked documents under the query, then top-k is the total number of documents corresponding to the query), the discriminator learning rate and the generator learning rate (default value is 0.001, adjustable range must be a floating-point value greater than 0 and less than 1), the number of discriminator epochs (default value is 8, adjustable range must be an integer greater than 1), the number of generator populations (default value is 1, adjustable range must be an integer greater than 1), and the total number of epochs for generating adversarial network games (default value is 1, adjustable range must be an integer greater than 1), etc.

[0143] The experiment was conducted on three publicly available standard ranking learning datasets: OHSUMED from LETOR 3.0, MQ2008 from LETOR 4.0, and the Microsoft ranking learning dataset MSLR-WEB10K. The performance of the dataset was compared with that of the IRGAN-listwise ranking learning method.

[0144] Figures 5 to 10 The figures show a comparison of the ECGAN-LTR and IRGAN-listwise methods on the ranking learning datasets OHSUMED, MQ2008, and MSLR-WEB10K, focusing on performance metrics P@K and NDCG@K. ECGAN-LTR-D and ECGAN-LTR-G represent the discriminator and generator of ECGAN-LTR, respectively, while IRGAN-Listwise-D and IRGAN-Listwise-G represent the discriminator and generator of IRGAN-Listwise. The experimental results demonstrate that the listwise ranking learning method ECGAN-LTR achieves competitive performance and outperforms the IRGAN-Listwise ranking learning method overall.

[0145] Example 3

[0146] According to a third aspect of the present invention, an application of the ranking learning method based on evolutionary conditional generative adversarial networks is provided. A specific application example is as follows:

[0147] The ranking learning method ECGAN-LTR, based on evolutionary conditional generative adversarial networks, can be applied to ranking applications requiring real-world applications such as search engines, e-commerce systems, recommendation systems, advertising systems, and question-answering systems. By embedding the ranking model trained using this method into the ranking system of information retrieval applications, this model can predict the ranking of web pages for the user's search query, thereby optimizing the search result ranking, improving the effectiveness of information retrieval, and enhancing the user experience.

[0148] The ranking learning method ECGAN-LTR, based on evolutionary conditional generative adversarial networks, is applied to search engines such as Baidu, Google, Bing, Yahoo, Sogou, and Yandex as an example. The operational implementation process of applying the ECGAN-LTR ranking learning method to search engines is as follows: Figure 11 As shown, the implementation steps are as follows:

[0149] Step 1. Embed the ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks into the search engine.

[0150] First, obtain web pages from the search engine's web page index database to obtain initial data for ranking learning.

[0151] Secondly, data preprocessing is performed on a portion of the web pages in the search engine's web page index database. Ranking features are extracted and relevance labels are added to the web pages, and a ranking learning dataset for the search engine is constructed in the format of a standard ranking learning dataset.

[0152] Then, on the constructed ranking learning dataset, the ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks is used to alternately iterate game adversarial optimization training, validation and selection of models to produce the optimal generator model and discriminator model.

[0153] Finally, the resulting optimal model is incorporated into the search engine's ranking system.

[0154] Step 2. Perform a query-web search, evaluate the performance of the best model generated by the embedded ECGAN-LTR ranking learning method, and predict and present the ranking results.

[0155] In search engines that incorporate the ECGAN-LTR ranking learning method, users can repeatedly execute queries—web page searches.

[0156] First, the user enters the query term they want to search in the search box of the search engine and clicks search to perform a web search.

[0157] Secondly, the search engine's ranking system retrieves all web pages containing the query term from the search engine's web page index database. It then evaluates the performance using the optimal model generated by the embedded ECGAN-LTR ranking learning method, calculating a page ranking score to predict which pages should rank higher and which should rank lower. This yields the web page search ranking results.

[0158] Finally, the web search results are returned to the "Search" page in a specific manner to be presented to the search user.

[0159] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A ranking learning method based on evolutionary conditional generative adversarial networks, characterized in that, Includes the following steps: Step 1: Initialize relevant parameters and input the sorting learning dataset; Step 2: Initialize the generator population and discriminator; Step 3: Alternately iterate adversarially to train the generator population and discriminator to produce the optimal model. The generator population is trained using an evolutionary strategy, as follows: For each epoch, the generator population is trained once, and the discriminator is trained once. This method involves iteratively training the generator population and discriminator by minimizing and maximizing the overall optimization objective function, until the algorithm converges or reaches a given number of epochs. The overall optimization objective function is as follows: in, and Representing generators respectively and discriminator The parameter set, Tags representing the relevance of the document to the query. Represents the total number of queries. Represents a probability distribution. Represents mathematical expectation; This represents the probability distribution of positive samples, i.e., given a query. The following is a sorted list of ground-truth documents, ordered by relevance tags. The probability distribution; This represents a generative retrieval model, which is generated by a generator. From a given query Downsampling to obtain top- Document sorting list The probability distribution of the negative samples obtained from sampling; This represents a discriminative retrieval model, which consists of a discriminator. Estimate a sorted list It is a ground-truth sorted list Or a sorted list of documents generated by the generator. The probability of; This represents the order probability distribution calculated using the Plackett-Luce model; This represents the distribution of ground-truth documents in a sorted learning dataset. The mathematical expectation, This represents the document arrangement distribution of the negative samples obtained from sampling. The mathematical expectation; Generator Considered an evolutionary population, discriminator As an environment, for the evolutionary operation in each epoch, the generator Update different mutation operators, including likelihood, cosine, and cross-entropy functions, to adapt to the current environment. Choose a high-performance offspring generator. Participate in subsequent adversarial iterative game training; Step 4: Based on the selected optimal model, predict and rank the information retrieval targets.

2. The ranking learning method based on evolutionary conditional generative adversarial networks according to claim 1, characterized in that, The initialization-related parameters include: initializing the generator population size. Number of generator mutation operators The total number of training iterations (epochs), and the discriminator update step size in each epoch. ,top- Batch size And the hyperparameters of the Adam optimizer.

3. The ranking learning method based on evolutionary conditional generative adversarial networks according to claim 1, characterized in that, The training of the generator population specifically includes: Step a.

1. Test the generator population Each generator in For different mutation operators, the generator is iteratively trained according to the following steps. ,in, , ; Step a.1.

1. Using a generator Sample top-ranked data from the training set based on scores. Sample data ; Step a.1.

2. Optimize the generator objective function using policy gradient, according to... Calculate the gradient of the selected mutation operator to update the gradient. The objective function of the generator is as follows: In the formula, For the generator's parameter set, To query the total number, These are tags indicating the relevance of the document to the query. This represents a generative retrieval model, which is generated by a generator. From a given query Downsampling to obtain top- Document sorting list The probability distribution of the negative samples obtained from sampling; This represents a discriminative retrieval model, which consists of a discriminator. Estimate a sorted list It is a ground-truth sorted list Or a sorted list of documents generated by the generator. The probability of; This represents the order probability distribution calculated using the Plackett-Luce model; This represents the document arrangement distribution of the negative samples obtained from sampling. The mathematical expectation; The mutation operator This includes the Likelihood, Cosine, and CrossEntropy operators; Step a.1.

3. Update the sub-network parameters The generator network is optimized using the Adam optimizer to update its parameters and generate offspring generators. ,Right now ,in, For the first j Network parameters of a generator population , , These are the hyperparameters of the Adam optimizer; Step a.1.

4. According to the fitness function The calculation method is used to evaluate the fitness function value of the offspring generator. ,in, This represents the scoring function of the discriminator. This represents the ranking probability calculated using the Plackett-Luce model. K Total number of documents; Step a.

2. Process the resulting offspring generator population. fitness function values Sort the data and generate a sorted list. ,Right now ; Step a.

3. Sort the list of each offspring generator by fitness function value. Select the optimal generator network and update the network parameters of each child generation. This generates a new generator population.

4. The ranking learning method based on evolutionary conditional generative adversarial networks according to claim 3, characterized in that, The likelihood mutation operator is as follows: in, Represents generator From a given query Downsampling to obtain top- Document sorting list Documents in The rating values, among which, Represents a sorted list Middle row in Location document ; This represents the ranking probability calculated using the Plackett-Luce model, and its calculation method is as follows: ; It is determined by the discriminator The reward function value used for policy gradient-based reinforcement learning is calculated as follows: in, This represents the ranking probability calculated using the Plackett-Luce model; The likelihood ranking probability of the generator is maximized by minimizing the likelihood mutation operator to confuse the discriminator's discrimination behavior.

5. The ranking learning method based on evolutionary conditional generative adversarial networks according to claim 3, characterized in that, The cosine variation operator is as follows: in, Represents generator From a given query Downsampling to obtain top- Document sorting list Documents in The rating value, It is determined by the discriminator The reward function value used to evaluate the policy gradient in policy gradient-based reinforcement learning; Represent a mapping function Preserves the association with a given query An ideal sorted list Chinese documents The rating value, According to each document Rather than query The ideal sorted list is formed by sorting the relevance label values ​​from largest to smallest, which is the ground-truth sorted list.

6. The ranking learning method based on evolutionary conditional generative adversarial networks according to claim 3, characterized in that, The CrossEntropy operator is as follows: in, Represents generator From a given query Downsampling to obtain top- Document sorting list Documents in The rating value, It is determined by the discriminator The feedback is the reward function value used for the policy gradient in policy gradient-based reinforcement learning. Represent a mapping function It preserves the association with a given query. An ideal sorted list Chinese documents The rating value, Represents a most idealized sorted list Chinese document The ordering probability, Represents a sorted list of model predictions Chinese document The ordering probabilities are all calculated using the softmax function, and their calculation methods are as follows: and .

7. The ranking learning method based on evolutionary conditional generative adversarial networks according to claim 1, characterized in that, The training of the discriminator specifically includes: Execute on the discriminator Iterative training with the following operations: Step b.

1. Given a query Use the current generator Sample top-ranked data from the training dataset based on their scores. negative sample data and compared with top-ranked by relevance labels Positive sample data Combine; Step b.

2. Based on the current generator The generated permutations and the permutations generated by the ground-truth labels are trained using a discriminator objective function, which is shown below: In the formula, Representative discriminator The parameter set, Tags representing the relevance of the document to the query. Represents the total number of queries. This represents the probability distribution of positive samples, i.e., given a query. The following is a sorted list of ground-truth documents, ordered by relevance tags. The probability distribution; This represents a generative retrieval model, which is generated by a generator. From a given query Downsampling to obtain top- Document sorting list The probability distribution of the negative samples obtained from sampling; This represents a discriminative retrieval model, which consists of a discriminator. Estimate a sorted list It is a ground-truth sorted list Or a sorted list of documents generated by the generator. The probability of; This represents the order probability distribution calculated using the Plackett-Luce model; This represents the ground-truth document permutation distribution of the ordered learning dataset. The mathematical expectation, This represents the document arrangement distribution of the negative samples obtained from sampling. The mathematical expectation; Calculate the gradient of the discriminator's objective function. ,Right now In the formula, This represents the scoring function of the discriminator. This represents the ranking probability calculated using the Plackett-Luce model. K Total number of documents; Step b.

3. Optimize the discriminator network using the Adam optimizer to update the discriminator network parameters. ,Right now In the formula, , , These are the hyperparameters of the Adam optimizer.

8. A system employing the ranking learning method based on evolutionary conditional generative adversarial networks as described in any one of claims 1-7, characterized in that, The system includes Model training module: The model is trained using a ranking learning method based on evolutionary conditional generative adversarial networks. It includes a generator model training submodule and a discriminator model training submodule. The functions include evaluating the trained model and saving the trained model. The model training module performs the following operations: loading the ranking learning training dataset; setting the relevant parameters of the ranking learning method; and using the ranking learning method to perform alternating iterative game training on the selected ranking learning training dataset to obtain the generator population model and the discriminator model. Model Validation Module: This module validates, selects, and saves the optimal model using the model trained in the model training module. It includes a generator model validation submodule and a discriminator model validation submodule. Its functions include evaluating and selecting / saving the validated model. The module performs the following operations: loading the ranking learning validation dataset; validating the model trained in the model training module on the selected ranking learning validation dataset to obtain the model with the best performance evaluation; and selecting and saving the optimal model as the test model for the model testing module. The model testing module utilizes the optimal performance model selected in the model validation module to test the model. It includes a generator model testing submodule and a discriminator model testing submodule, and its functions include performance evaluation and prediction result saving. The performance evaluation metrics include effectiveness evaluation metrics in the information retrieval domain. The model testing module performs the following operations: loading the ranking learning test dataset; testing the optimal performance model saved in the model validation module on the selected ranking learning test dataset to obtain relevance scores and some performance metric evaluation values ​​in information retrieval; determining whether the relevance score is greater than a set threshold: if so, returning documents that are relevant to the query; otherwise, returning documents that are irrelevant to the query.

9. An application of the ranking learning method based on evolutionary conditional generative adversarial networks as described in any one of claims 1-7, characterized in that, The ranking learning method based on evolutionary conditional generative adversarial networks, named ECGAN-LTR, has been applied to search engines including Baidu, Google, Bing, Yahoo, Sogou, and Yandex. The specific application steps are as follows: Step 1: Embed the ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks into the search engine; First, obtain the web pages from the search engine's web page index database to obtain initial data for ranking learning; Secondly, data preprocessing is performed on a portion of the web pages in the search engine's web page index database, ranking features are extracted and relevance is labeled on the web pages, and a ranking learning dataset for the search engine is constructed in the format of a standard ranking learning dataset. Then, on the constructed ranking learning dataset, the ranking learning method ECGAN-LTR based on evolutionary conditional generative adversarial networks is used to alternately iterate game adversarial optimization training, validation and selection of models to produce the optimal generator model and discriminator model. Finally, the resulting optimal model is incorporated into the search engine's ranking system; Step 2: Execute query - web search, evaluate the performance of the best model generated by the embedded Evolutionary Conditional Generative Adversarial Network-based ranking learning method ECGAN-LTR, predict and present the ranking results; In search engines that incorporate the ranking learning method ECGAN-LTR, users can repeatedly execute queries—web page searches. First, the user enters the query term they want to search in the search box of the search engine and clicks search to perform a web search; Secondly, the search engine's ranking system finds all web pages containing the query term from the search engine's web page index database, evaluates the performance of the optimal model generated by the embedded Evolutionary Conditional Generative Adversarial Network-based ranking learning method ECGAN-LTR, calculates the web page ranking score, and predicts the ranking result of the web page search. Finally, the web search ranking results are returned to the search page in a certain way to be presented to the search user.

Images