Resource intelligent recommendation method and device based on improved frequent growth mode algorithm

Abstract

The invention provides an intelligent test resource recommendation method and device based on an improved frequent growth mode algorithm, and the method comprises the steps: carrying out the preprocessing of original test resource data, and outputting a preprocessed data set; performing feature extraction on the preprocessed test resource data; receiving a test demand text input by a user, performing semantic understanding, and then outputting a test demand semantic representation; based on the test resource data representation of feature vectorization, executing an improved frequent growth mode algorithm, and outputting a frequent mode tree; screening out a frequent association rule set according to a preset index; combining the screened frequent association rule set with the test demand semantic representation, applying the combined frequent association rule set and the test demand semantic representation to a collaborative filtering recommendation model, and generating a personalized test resource recommendation list for a user; according to the method, intelligent recommendation can help the simulation test to utilize the resources more efficiently, and intelligent recommendation of the resources is realized.

Description

Technical Field

[0001] The present invention relates to the fields of data analysis and mining and machine learning, and in particular to a resource intelligent recommendation method and device based on an improved frequent growth pattern algorithm. Background Art

[0002] Recommendation system technology can tap into the hidden value of data, capture customer needs by collaborating with user data and project attributes, and provide personalized information services, so that users can obtain the information they need, thereby improving the effective utilization of data. Recommendation systems play an important role in alleviating the problem of data overload, helping users discover potential interests and alleviating the problem of users being unable to find the information they need due to excessive data.

[0003] Recommendation systems have become the core of many e-commerce and multimedia platforms. Personalized recommendation services can help platforms attract users' attention and increase user visits. Recommendation systems provide a continuous driving force for the development of network platforms, and their commercial value has also attracted the attention of industry and academia. As a popular technology, deep learning has shown unlimited potential in many fields such as computer vision and natural language processing, and has also provided new methods for recommendation systems. With the powerful representation ability of deep learning technology, it learns the latent vector representation of users and items, mines users' historical behavior data, diversified product data, and contextual scene information, captures users' potential preferences, and generates more accurate personalized recommendation lists for users.

[0004] Traditional recommendation algorithms mainly include collaborative filtering recommendation algorithm, matrix decomposition, logistic regression model, factorization machine model, gradient boosting tree + logistic regression (GBDT + LR) model. Traditional recommendation algorithms usually do not require a lot of domain knowledge, are prone to loss of contextual information, have poor expression ability, and are prone to overfitting.

[0005] With the development of deep learning technology, the combination of deep learning and recommendation systems can alleviate the problem of insufficient expression ability of traditional recommendation models. Deep learning has stronger representation ability and requires a large amount of data to train the model, which can alleviate the problems of large data scale and data sparsity. Common structures of deep learning include: multi-layer perceptron, convolutional neural network, recurrent neural network, attention mechanism, etc.

[0006] The multilayer perceptron is a neural network with a feedforward structure. Data passes through the input layer, multiple hidden layers, and the output layer to calculate the final result. The BP back-propagation algorithm is used to supervise the training of the neural network, adjust the weights of each layer of neurons, fit nonlinear functions, and reduce the error between the predicted value and the true value. Multilayer perceptrons are often used in recommendation systems to mine high-order feature interactions [8] and learn potential data patterns.

[0007] Convolutional neural network (CNN) is a network structure that imitates the biological visual system[9]. It uses convolution operations to process two-dimensional data features and is widely used in the field of computer vision. The parameter sharing of convolution operations in CNN reduces the number of parameters that need to be learned in the model and is more computationally efficient than fully connected neural networks. In recommendation systems, convolutional neural networks are mainly used to extract visual features, text features, and integrate user portrait features to capture user preferences from more aspects. They are often used in scenarios such as image recommendation, news recommendation, and multimodal recommendation.

[0008] In order to solve the problem of information loss and gradient disappearance and explosion caused by long time intervals, new variants of long short-term memory network (LSTM) and gated recurrent unit (GRU) are constructed. In the recommendation system, recurrent neural network can be used for session-based recommendation, based on the user's current session behavior, to learn the user's interest migration process and predict the next item that the user may interact with.

[0009] The attention mechanism is a local signal processing mechanism that imitates human vision. When observing things, people usually focus on some important information and reduce their attention to irrelevant information, so as to make judgments quickly. The attention mechanism helps the recommendation model select more effective features, allowing the model to focus on more important information and reduce the impact of data noise on the results. The training process of deep learning is often regarded as a "black box". The entire training process is unpredictable and the output results cannot provide good explanations. Summary of the invention

[0010] In view of this, the present invention provides a test resource intelligent recommendation method based on an improved frequent growth pattern algorithm, comprising the following steps:

[0011] Step 1: preprocess the original test resource data and output the preprocessed data set;

[0012] Step 2, extracting features from the preprocessed test resource data, mapping the structured data and the unstructured data into a unified feature vector representation, and outputting a feature vectorized test resource data representation;

[0013] Step 3, receiving the test requirement text input by the user, performing semantic understanding and outputting the semantic representation of the test requirement;

[0014] Step 4, based on the feature vectorized representation of the experimental resource data, executing an improved frequent growing pattern algorithm to output a frequent pattern tree;

[0015] Step 5, generating frequent patterns according to the frequent pattern tree, and filtering out frequent association rule sets according to preset indicators;

[0016] Step 6: Combine the filtered frequent association rule set with the test requirement semantic representation, apply them to a collaborative filtering recommendation model, and generate a personalized test resource recommendation list for the user.

[0017] In particular, the preprocessing in step 1 includes cleaning, standardization and missing value filling.

[0018] In particular, in step 2, a deep learning model is used to extract features and vectorize the representation of the test resource data.

[0019] In particular, the step 3 performs the following sub-steps:

[0020] Step 3.1 uses a pre-trained language model to perform word segmentation and requirement object labeling on the test requirement text;

[0021] Step 3.2: Divide the paragraphs according to the context of the required object;

[0022] Step 3.3: Input the requirement object and the paragraph it is in into the semantic understanding model, extract and annotate the related attribute words.

[0023] In particular, the step 3.3 is specifically as follows:

[0024] Step 3.3.1 Use the RoBERTa-wwm-ext pre-trained model to segment the extracted text and annotate the required objects, and divide and annotate the paragraphs according to the context of the required objects;

[0025] Step 3.3.2 uses the requirement object annotation results in step 3.3.1 and the paragraph where the object is located as the input of the RoBERTa-wwm-ext model, extracts the attribute words associated with the requirement object in the paragraph, and uses the CRF algorithm for labeling.

[0026] In particular, in the CRF algorithm, for a given sequence X(x1, x2, ..., x n ), whose corresponding label is Y(y1, y2, …, y n ), the following conditions are met:

[0027] P(y i |X,y1,…,y i-1 ,y i+1 , …, y n )=P(y i |X,y i-1 ,y i+1 )

[0028] P represents the conditional probability, that is, given the observation sequence X and except y i The i-th label y is the condition of other label sequences y. iThe conditional probability of

[0029] Let U be the weight matrix output by the decoding layer, and then we can get the evaluation score S(X, Y), that is:

[0030]

[0031] Among them, A is the transfer matrix, n is the sequence length, and the maximum probability of each word corresponding to the sequence label Y can be calculated using the Softmax function. During training, the maximum likelihood method is generally used to solve the maximum posterior probability of P(Y|X):

[0032]

[0033] In particular, the improved frequent growth pattern algorithm in step 4 includes:

[0034] Step 4.1 Scan the data set and count the support of each item Scan the data set to count the support of each item;

[0035] Step 4.2: cluster each item using the K-means algorithm, select clusters whose support is not less than the threshold, and form a single item set whose support reaches a preset threshold, that is, a frequent 1-item set F1;

[0036] Step 4.3 Repeat the above steps until no new candidate item sets can be generated, and a list of high-order frequent item sets is obtained;

[0037] Step 4.4: sort the frequent item sets in descending order according to support, insert the items in the sorted frequent item set list into the red-black tree according to support, and store an item and its support information in each node;

[0038] Step 4.5 constructs and outputs the frequent pattern tree based on the red-black tree.

[0039] In particular, the step 5 comprises:

[0040] Step 5.1, constructing a corresponding conditional frequent pattern tree for each frequent item set according to the frequent pattern tree;

[0041] Step 5.2, mining the conditional frequent pattern tree to generate a frequent pattern list, where each frequent pattern corresponds to a conditional frequent pattern tree;

[0042] Step 5.3, calculate the support and confidence of each frequent pattern.

[0043] Step 5.4, according to the set minimum support threshold and minimum confidence threshold, filter out the frequent association rule set that meets the conditions.

[0044] In particular, the step 6 specifically includes:

[0045] Step 6.1, matching the test requirement semantic representation with the frequent association rule set to obtain a matching result;

[0046] Step 6.2, according to the matching results, find the rule subset most relevant to the user's needs;

[0047] Step 6.3, integrating the relevant rule subset as auxiliary information into the content-based collaborative filtering recommendation model; the collaborative filtering model generates a personalized candidate list of experimental resources for the target user based on the user's historical behavior data, the content characteristics of the experimental resources and the auxiliary association rules;

[0048] Step 6.4: Based on user feedback, continuously optimize and iterate the parameters of the collaborative filtering model to improve the accuracy of recommendations.

[0049] The present invention also discloses a test resource intelligent recommendation device based on an improved frequent growth mode algorithm, comprising:

[0050] The data preprocessing module is used to preprocess the original test resource data and output the preprocessed data set;

[0051] The feature extraction and representation learning module is used to extract features from the preprocessed test resource data, map structured data and unstructured data into a unified feature vector representation, and output a feature vectorized test resource data representation;

[0052] The test requirement text analysis and semantic understanding module is used to receive the test requirement text input by the user, perform semantic understanding and output the semantic representation of the test requirement;

[0053] An improved frequent growth pattern algorithm module, used for executing an improved frequent growth pattern algorithm based on the experimental resource data representation of feature vectorization, and outputting a frequent pattern tree;

[0054] The association rule generation module is used to generate frequent patterns according to the frequent pattern tree and filter out frequent association rule sets according to indicators.

[0055] Beneficial effects:

[0056] The present invention discloses an intelligent test resource recommendation technology based on an improved frequent growth mode algorithm, which is applied to military simulation tests to adapt to the characteristics of a wide variety of resources, high information complexity, and large real-time changes in simulation training in simulation tests, thereby ensuring the authenticity and practicality of test training and improving resource utilization efficiency.

[0057] The present invention relates to an intelligent test resource recommendation technology based on an improved frequent growth pattern algorithm, and the technical field mainly includes data analysis and mining, machine learning, decision support and optimization, multimodal information processing, etc. The present invention has important significance for realizing an intelligent test resource recommendation technology based on an improved frequent growth pattern algorithm.

[0058] Improve resource utilization efficiency. Military simulation test resources are usually limited. Intelligent recommendation can help simulation tests use these resources more efficiently. Through intelligent resource recommendation, more accurate matching of test requirements and resource characteristics can be achieved, thus avoiding waste and duplication of resources and improving resource utilization efficiency.

[0059] To improve training effects, intelligent resource recommendations can recommend the most suitable simulation test resources based on the characteristics of test needs and training objectives, which helps to provide a simulation environment that is closer to actual combat scenarios, enhance the authenticity and practicality of training, and improve training effects.

[0060] Reduce training costs. Through intelligent resource recommendations, existing simulation test resources can be better utilized, avoiding unnecessary investment and purchase, thereby reducing test costs. Intelligent recommendations can reasonably arrange the order and time of resource use based on resource characteristics and usage, maximize the use of existing resources, and reduce idleness and waste of resources.

[0061] Strengthening decision support can provide reference opinions and recommended solutions for decision makers, help them make wise decisions, and improve the decision-making effect of simulation training.

[0062] Improve training innovation capabilities. Intelligent resource recommendation can analyze resource characteristics and requirements, discover potential problems and room for improvement in resource utilization, discover new training models and methods, improve training innovation capabilities, and promote the development of simulation training. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the process architecture of the method proposed in the present invention;

[0064] Figure 2 It is a schematic diagram of the pretreatment process proposed in the present invention;

[0065] Figure 3 It is a schematic diagram of the frequent growth model algorithm flow proposed in the present invention. DETAILED DESCRIPTION

[0066] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0067] The present invention provides a method for intelligently recommending experimental resources based on an improved frequent growth model algorithm, and its process architecture diagram is as follows: Figure 1As shown, the method includes:

[0068] Step 1: preprocess the original test resource data and output the preprocessed data set;

[0069] The preprocessing in step 1 includes cleaning, standardization and missing value filling.

[0070] Before capturing feature information, the military simulation test resource data needs to be preprocessed, including data cleaning, data standardization, missing value processing and other steps to ensure data dimension unification and data consistency. Figure 2 As shown. The data preprocessing uses an interactive module to provide an input interface to input the data files that need to be cleaned and the specific requirements for cleaning the data. The result display module provides the download of the clean data compression package and the comparison of the data before and after cleaning. The cleaning process mainly includes missing value filling, entity recognition and true value discovery, and inconsistency repair. Among them, entity recognition clusters entity tuples, and true value discovery can find possible true values ​​in data conflicts. Inconsistency detection is used to find erroneous data and make possible repairs under the logical rules of data dependence. For example, simulation resource A depends on simulation resources B and C, but the dependent data source in A is null, and it is necessary to traverse the data for trial repair. The value filling part detects the missing part of the data and fills it.

[0071] Step 2, extract features from the preprocessed test resource data, map structured data and unstructured data into a unified feature vector representation, and output a feature vectorized test resource data representation; in step 2, a deep learning model is used to extract features and vectorize the representation of the test resource data. After data cleaning, feature information needs to be captured to obtain the feature source for intelligent recommendation of test resources. Whether feature information is captured appropriately is an important factor affecting the subsequent recommendation effect. For feature information capture, most of the elements involved in the simulation test should be covered, including various characteristics of the test equipment, environmental conditions, geographical terrain, meteorological information characteristics, etc. Specifically including feature extraction steps:

[0072] For structured data (such as numerical data, categorical data, etc.):

[0073] a) Use the Embedding layer to vectorize discrete features;

[0074] b) For continuous numerical features, normalization and standardization can be used for processing.

[0075] For unstructured data:

[0076] a) Use pre-trained models such as BERT / RoBERTa to encode and obtain the vector representation of the text;

[0077] b) The attention mechanism can be used to capture the key information in the text.

[0078] Feature fusion steps:

[0079] The encoded vectors of structured data and unstructured data are concatenated to construct a vector representation of the overall characteristics of the experimental resources.

[0080] Feature interaction steps:

[0081] A deep learning model (such as DNN, FM, etc.) is selected to cross-combine the previously obtained overall feature vectors to automatically learn the high-order feature interaction relationship.

[0082] Vector output steps:

[0083] The deep learning model ultimately maps the experimental resources into a unified and vectorized data representation as the input of step 4.

[0084] Therefore, step 2 uses a deep learning model to automatically extract features from the original test resource data, which can fully mine useful patterns in structured and unstructured data and fuse them into a unified vector form to form a data representation that the algorithm can directly process. This end-to-end vectorization method improves the quality and generalization ability of feature expression.

[0085] Step 3, receiving the test requirement text input by the user, and outputting the test requirement semantic representation after semantic understanding; said step 3 performs the following sub-steps:

[0086] Step 3.1 uses a pre-trained language model to perform word segmentation and requirement object labeling on the test requirement text;

[0087] Step 3.2: Divide the paragraphs according to the context of the required object;

[0088] Step 3.3: Input the requirement object and the paragraph it is in into the semantic understanding model, extract and annotate the related attribute words.

[0089] The step 3.3 is specifically as follows:

[0090] Step 3.3.1 Use the RoBERTa-wwm-ext pre-trained model to segment the extracted text and annotate the required objects, and divide and annotate the paragraphs according to the context of the required objects;

[0091] Step 3.3.2 uses the requirement object annotation results in step 3.3.1 and the paragraph where the object is located as the input of the RoBERTa-wwm-ext model, extracts the attribute words associated with the requirement object in the paragraph, and uses the CRF algorithm for labeling.

[0092] In the CRF algorithm, for a given sequence X(x1, x2, ..., x n ), whose corresponding label is Y(y1, y2, …, y n ), the following conditions are met:

[0093] P(y i |X,y1,…,y i-1 ,y i+1 , …, y n )=P(y i |X,y i-1 ,y i+1 )

[0094] P represents the conditional probability, that is, given the observation sequence X and except y i The i-th label y is the condition of other label sequences y. i The conditional probability of

[0095] Let U be the weight matrix output by the decoding layer, and then we can get the evaluation score S(X, Y), that is:

[0096]

[0097] Among them, A is the transfer matrix, n is the sequence length, and the maximum probability of each word corresponding to the sequence label Y can be calculated using the Softmax function. During training, the maximum likelihood method is generally used to solve the maximum posterior probability of P(Y|X):

[0098]

[0099] Step 4: Based on the feature vectorized representation of the experimental resource data, an improved frequent growth pattern algorithm is executed to output a frequent pattern tree; the frequent pattern tree is also called FP-tree, which is a special data structure used in the frequent pattern growth (FP-Growth) algorithm. e , namely "frequent pattern tree", its process is as follows Figure 3 shown.

[0100] The present invention uses data mining and association rule mining technology to discover the potential patterns and association rules between resource features and test requirements. By analyzing the association rules of the frequently associated item set between resource feature data and requirements, a more accurate match between features and requirements is obtained, thereby providing a more accurate recommendation effect.

[0101] Dimension reduction is performed on the features obtained by feature selection and extraction, thereby reducing the feature dimension and improving mining efficiency. Association rule mining is a common data mining technology used to discover frequent item sets and association rules in data sets. The frequent growth model algorithm (FP-growth algorithm) is used to establish frequent item sets and association rules.

[0102] The key step of mining in the traditional FP-growth algorithm is to construct FP-tree nodes. The resource data set is traversed for the first time, and the minimum support count min_sup is set to 3. All frequent 1-item sets can be found through one mining; then all non-frequent 1-item sets are deleted and sorted in descending order according to the support count of the data items. The data set is traversed for the second time, the item header table is created and the FP-tree node is constructed. The root node of the tree is initially set to null, and a branch is created for each transaction in turn. The mining method of FP-tree is bottom-up. According to the data items in the item header table, the conditional pattern base is traversed in turn for mining, and the minimum support count is set to 2. Mining starts from node p until all nodes are mined, and all frequent item sets that can be mined in the transaction database are obtained.

[0103] The FP-growth algorithm adopts a divide-and-conquer search concept to mine frequent itemsets. However, the amount of simulation data is huge. In order to discover the association rules between experimental resources and target requirements, the traditional frequent growth model algorithm requires too much calculation and easily causes meaningless data to interfere with the mining of association rules, resulting in low execution efficiency. Therefore, the present invention proposes an improved FP-growth algorithm by (1) using the K-means algorithm to construct a frequent itemset list; (2) using a red-black FP-tree instead of a traditional unordered FP-tree.

[0104] In the improved frequent growth pattern algorithm step of the recommendation method, the whole process of FP-tree is: count the support of each item in the data set; use K-means clustering algorithm to cluster items based on their support, screen out clusters whose support is not less than the set threshold, and form a single item set whose support reaches the preset threshold, that is, frequent 1 item set F1; iteratively generate a list of high-order frequent item sets; arrange the frequent item sets in order from large to small according to the support, and insert them into the red-black tree data structure; construct FP-tree according to the frequent item set sequence inserted into the red-black tree. The constructed FP-tree can efficiently store the frequent pattern information in the data set. In the subsequent association rule generation step, a conditional FP tree will be constructed based on the FP-tree to discover frequent patterns and generate association rules. In general, the whole process of FP-tree includes a series of processes such as statistical item support, cluster screening, generation of high-order item sets, sorting based on support, and finally construction of FP-tree, which is the core data structure of the improved frequent growth pattern algorithm in this method.

[0105] The improved frequent growth pattern algorithm in step 4 includes:

[0106] Step 4.1 Scan the data set and count the support of each item Scan the data set to count the support of each item;

[0107] Step 4.2: cluster each item using K-means algorithm to select clusters whose support is not less than a threshold, and the clusters are unified to form a frequent 1-item set F1;

[0108] Step 4.3 Repeat the above steps until no new candidate item sets can be generated, and a list of high-order frequent item sets is obtained;

[0109] Step 4.4: sort the frequent item sets in descending order according to support, insert the items in the sorted frequent item set list into the red-black tree according to support, and store an item and its support information in each node;

[0110] Step 4.5 constructs and outputs the frequent pattern tree based on the red-black tree, and constructs an FP-tree according to the red-black tree:

[0111] Initialize an empty FP tree and head pointer table; for each transaction data, sort the item set in descending order of support and filter out infrequent items; for each sorted item, start from the root node and find out whether the child node of the item exists in the FP tree. If it does, update the support count of the node; if it does not exist, insert the item into the FP tree and update the corresponding head pointer table; after processing all transaction data, a complete FP tree is obtained. Construct a conditional FP tree. For each frequent item set, obtain the prefix path of the item based on the path and support information of the item set in the FP tree to construct a conditional pattern base;

[0112] A conditional FP tree is constructed based on the conditional pattern base; if the conditional FP tree meets the evaluation indicators, including support, execution, imbalance ratio and other indicators, the frequent association rule is generated, otherwise the association rule is eliminated.

[0113] Step 5, generating frequent patterns according to the frequent pattern tree, and filtering out frequent association rule sets according to preset indicators; Step 5.1, constructing a corresponding conditional frequent pattern tree for each frequent item set according to the frequent pattern tree;

[0114] Step 5.2, mining the conditional frequent pattern tree to generate a frequent pattern list, where each frequent pattern corresponds to a conditional frequent pattern tree;

[0115] Step 5.3, calculate the support and confidence of each frequent pattern.

[0116] Step 5.4, according to the set minimum support threshold and minimum confidence threshold, filter out the frequent association rule set that meets the conditions.

[0117] Calculate the support of each frequent pattern, support = pattern count / total number of transactions;

[0118] For each frequent pattern generated, use the rule L->R to calculate its confidence; confidence = support (L->R) / support (L);

[0119] L->R represents the association rule from the left item set (L) to the right item set (R) in association rule mining. Specifically, an item set is an unordered set of elements, which can be products, services, or any objects that can be combined. An association rule is a logical expression in the form of L->R, where L and R are disjoint item sets. L is called the antecedent of the rule, and R is called the consequent of the rule.

[0120] The rule L->R means that if a transaction contains all items in the item set L, then the transaction is likely to also contain items in the item set R.

[0121] For example, an association rule could be: {bread, butter} -> {milk}

[0122] This means that if a shopping basket contains bread and butter, then there is a high probability that the shopping basket also contains milk. For each frequent pattern L->R found, we calculate:

[0123] Support = Support (L, R) / Total number of transactions, indicating the frequency of the pattern in the data set

[0124] Confidence = Support (L, R) / Support (L), which indicates the probability that R will occur when L occurs.

[0125] Finally, by setting the support and confidence thresholds, a subset of frequent association rules with a high degree of confidence can be screened out for subsequent recommendation decisions. Therefore, the meaning of L->R is the representation of association rules in association rule mining.

[0126] According to the set minimum support threshold and minimum confidence threshold, the frequent association rule sets that meet the conditions are screened out.

[0127] Step 6: Combine the selected frequent association rule set with the semantic representation of the test requirements and apply it to the collaborative filtering recommendation model to generate a personalized test resource recommendation list for the user. Step 6 specifically includes:

[0128] Step 6.1, matching the test requirement semantic representation with the frequent association rule set to obtain a matching result;

[0129] Step 6.2, according to the matching results, find the rule subset most relevant to the user's needs;

[0130] Step 6.3, integrating the relevant rule subset as auxiliary information into the content-based collaborative filtering recommendation model; the collaborative filtering model generates a personalized candidate list of experimental resources for the target user based on the user's historical behavior data, the content characteristics of the experimental resources and the auxiliary association rules;

[0131] Step 6.4: Based on user feedback, continuously optimize and iterate the parameters of the collaborative filtering model to improve the accuracy of recommendations.

[0132] In this invention, the association rules obtained by the improved FP-growth algorithm are applied to the collaborative filtering model. By mining the usage data of simulation resources, decision support is provided in a series of application scenarios, such as the test resources that are often matched in different environments, the coordination and mutual exclusion relationship between test resources, the popularity and scarcity of test resources, and the priority application scenarios of limited test resources. Based on the collaborative filtering model, test resources are recommended to users, the efficiency, quality and training effect of the test are improved, and the authenticity and practicality are enhanced.

[0133] The present invention also discloses a test resource intelligent recommendation device based on an improved frequent growth mode algorithm, comprising:

[0134] The data preprocessing module is used to preprocess the original test resource data and output the preprocessed data set;

[0135] The feature extraction and representation learning module is used to extract features from the preprocessed test resource data, map structured data and unstructured data into a unified feature vector representation, and output a feature vectorized test resource data representation;

[0136] The test requirement text analysis and semantic understanding module is used to receive the test requirement text input by the user, perform semantic understanding and output the semantic representation of the test requirement;

[0137] An improved frequent growth pattern algorithm module, used for executing an improved frequent growth pattern algorithm based on the experimental resource data representation of feature vectorization, and outputting a frequent pattern tree;

[0138] The association rule generation module is used to generate frequent patterns according to the frequent pattern tree and filter out frequent association rule sets according to indicators.

[0139] This embodiment corresponds one-to-one with the previous method embodiment in terms of features, so it will not be described in detail.

[0140] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0141] It is obvious to those skilled in the art that the embodiments of the present invention are not limited to the details of the above exemplary embodiments, and that the embodiments of the present invention can be implemented in other specific forms without departing from the spirit or basic features of the embodiments of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the embodiments of the present invention is limited by the attached claims rather than the above description, so it is intended to include all changes that fall within the meaning and scope of the equivalent elements of the claims in the embodiments of the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved. In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units, modules or devices stated in the system, device or terminal claims can also be implemented by the same unit, module or device through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

[0142] Finally, it should be noted that the above implementation modes are only used to illustrate the technical solutions of the embodiments of the present invention and are not intended to limit them. Although the embodiments of the present invention have been described in detail with reference to the above preferred implementation modes, those skilled in the art should understand that the technical solutions of the embodiments of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm, characterized in that: The following steps are involved: Step 1: preprocess the original test resource data and output the preprocessed data set; Step 2, extracting features from the preprocessed test resource data, mapping the structured data and the unstructured data into a unified feature vector representation, and outputting a feature vectorized test resource data representation; Step 3, receiving the test requirement text input by the user, performing semantic understanding and outputting the semantic representation of the test requirement; Step 4, based on the feature vectorized representation of the experimental resource data, executing an improved frequent growing pattern algorithm to output a frequent pattern tree; Step 5, generating frequent patterns according to the frequent pattern tree, and filtering out frequent association rule sets according to preset indicators; Step 6: Combine the filtered frequent association rule set with the test requirement semantic representation, apply them to a collaborative filtering recommendation model, and generate a personalized test resource recommendation list for the user.

2. The method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm as claimed in claim 1, characterized in that: The preprocessing in step 1 includes cleaning, standardization and missing value filling.

3. The method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm according to claim 2, characterized in that: In the step 2, a deep learning model is used to extract features and vectorize the representation of the test resource data.

4. A method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm according to claim 3, characterized in that: Step 3 performs the following sub-steps: Step 3.1 uses a pre-trained language model to perform word segmentation and requirement object labeling on the test requirement text; Step 3.2: Divide the paragraphs according to the context of the required object; Step 3.3: Input the requirement object and the paragraph it is in into the semantic understanding model, extract and annotate the related attribute words.

5. The method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm according to claim 4, characterized in that: The step 3.3 is specifically as follows: Step 3.3.1 Use the RoBERTa-wwm-ext pre-trained model to segment the extracted text and annotate the required objects, and divide and annotate the paragraphs according to the context of the required objects; Step 3.3.2 uses the requirement object annotation results in step 3.3.1 and the paragraph where the object is located as the input of the RoBERTa-wwm-ext model, extracts the attribute words associated with the requirement object in the paragraph, and uses the CRF algorithm for labeling.

6. The method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm according to claim 5, characterized in that: In the CRF algorithm, for a given sequence X(x1, x2, ..., x n ), whose corresponding labels are Y(y1, y2, ..., y n ), the following conditions are met: P(y i |X,y1,...,y i-1 ,y i+1 ,...,y n )=P(y i |X,y i-1 ,y i+1 ) P represents the conditional probability, that is, given the observation sequence X and except y i The i-th label y is the condition of other label sequences y. i The conditional probability of Let U be the weight matrix output by the decoding layer, and then we can get the evaluation score S(X, Y), that is: Among them, A is the transfer matrix, n is the sequence length, and the maximum probability of each word corresponding to the sequence label Y can be calculated using the Softmax function. During training, the maximum likelihood method is generally used to solve the maximum posterior probability of P(Y|X):

7. The method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm as claimed in claim 6, characterized in that: The improved frequent growth pattern algorithm in step 4 includes: Step 4.1 Scan the data set and count the support of each item Scan the data set to count the support of each item; Step 4.2: cluster each item using the K-means algorithm, select clusters whose support is not less than the threshold, and form a single item set whose support reaches a preset threshold, that is, a frequent 1-item set F1; Step 4.3 Repeat the above steps until no new candidate item sets can be generated, and a list of high-order frequent item sets is obtained; Step 4.4: sort the frequent item sets in descending order according to support, insert the items in the sorted frequent item set list into the red-black tree according to support, and store an item and its support information in each node; Step 4.5 constructs and outputs the frequent pattern tree based on the red-black tree.

8. The method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm according to claim 7, characterized in that: The step 5 comprises: Step 5.1, constructing a corresponding conditional frequent pattern tree for each frequent item set according to the frequent pattern tree; Step 5.2, mining the conditional frequent pattern tree to generate a frequent pattern list, where each frequent pattern corresponds to a conditional frequent pattern tree; Step 5.3, calculate the support and confidence of each frequent pattern. Step 5.4, according to the set minimum support threshold and minimum confidence threshold, filter out the frequent association rule set that meets the conditions.

9. The method for intelligently recommending experimental resources based on an improved frequent growth pattern algorithm according to claim 8, characterized in that: The step 6 specifically includes: Step 6.1, matching the test requirement semantic representation with the frequent association rule set to obtain a matching result; Step 6.2, according to the matching results, find the rule subset most relevant to the user's needs; Step 6.3, integrating the relevant rule subset as auxiliary information into the content-based collaborative filtering recommendation model; the collaborative filtering model generates a personalized candidate list of experimental resources for the target user based on the user's historical behavior data, the content characteristics of the experimental resources and the auxiliary association rules; Step 6.4: Based on user feedback, continuously optimize and iterate the parameters of the collaborative filtering model to improve the accuracy of recommendations.

10. An intelligent test resource recommendation device based on an improved frequent growth model algorithm, characterized in that: include: The data preprocessing module is used to preprocess the original test resource data and output the preprocessed data set; The feature extraction and representation learning module is used to extract features from the preprocessed test resource data, map structured data and unstructured data into a unified feature vector representation, and output a feature vectorized test resource data representation; The test requirement text analysis and semantic understanding module is used to receive the test requirement text input by the user, perform semantic understanding and output the semantic representation of the test requirement; An improved frequent growth pattern algorithm module, used for executing an improved frequent growth pattern algorithm based on the experimental resource data representation of feature vectorization, and outputting a frequent pattern tree; The association rule generation module is used to generate frequent patterns according to the frequent pattern tree and filter out frequent association rule sets according to indicators.

Images