An automated feature engineering method and device for collaborative feature set search

Abstract

The application discloses a kind of automation feature engineering method and device for collaborative feature set search, the method comprises: obtaining table data to be processed;The automation feature engineering of table data is modeled as a Markov decision process, the Markov decision process is processed based on hierarchical reinforcement learning, then collaborates all feature searches to obtain the new feature set corresponding to table data.The method of the application can reduce the human consumption of feature engineering, is superior to the prior art, and has the advantages of high efficiency, high performance, good generalization, etc.

Description

Technical Field

[0001] This invention relates to the field of computer-automated feature engineering, and in particular to an automated feature engineering method and apparatus for collaborative feature set search. Background Technology

[0002] In modern computing infrastructure, most web-based applications involve the analysis and mining of tabular data (also known as structured data), largely thanks to the extensive deployment of database systems. Over the past few decades, research tasks surrounding tabular data have proliferated. These research lines have significantly advanced several domain-specific tasks, such as risk management, fraud detection, advertising recommendation, and marketing technology. A common feature among them is the involvement of a complete machine learning pipeline built upon tabular data. Arguably, the most critical step in building a tabular data mining system is feature engineering. This can be considered similar to other data forms, such as images, text, or video clips, where representation learning is crucial. While feature engineering is typically derived from highly experienced data scientists or engineers, the rapid development of reinforcement learning and automated machine learning offers an automated alternative. In fact, reinforcement learning-driven automated engines share many similarities with human experts. Notably, the feature engineering process can include feature selection, combination, and arithmetic operations. The manual process undertaken by human experts, encompassing the exploration of these potential plans, can clearly be viewed as a trial-and-error process. For example, a data expert would first propose a feature engineering plan, train a model, and then check the model's results through a specific testing protocol. Based on their extensive experience, data scientists will modify the feature plan and repeat the above process until they converge.

[0003] In this regard, automated feature engineering is known for its powerful exploration capabilities, as well as its detailed exploration schemes, good industrial friendliness, and extremely low labor costs.

[0004] This invention considers all previous AutoFE frameworks, including Microsoft's most influential Neural Feature Search (NFS), the current state-of-the-art (SOTA) method DIFER, and well-known deep feature synthesis-based functional tools. We believe that the shortcomings of these existing technologies can be categorized from three main perspectives: overall system performance, efficiency, and practicality. In terms of performance, existing methods lack high-order complex features and ignore the coupling between features; in terms of efficiency, most techniques have high sample complexity and suffer from feature explosion; in terms of practicality, these techniques cannot process extremely large datasets and have poor generalization performance on unseen datasets. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an automated feature engineering method and apparatus for collaborative feature set search.

[0006] According to a first aspect of the present invention, an automated feature engineering method for collaborative feature set search is provided, the method specifically including the following steps:

[0007] Retrieve the table data to be processed;

[0008] The automated feature engineering of tabular data is modeled as a Markov decision process. The Markov decision process is processed based on hierarchical reinforcement learning, and then all features are searched collaboratively to obtain a new feature set corresponding to the tabular data.

[0009] According to a second aspect of the present invention, an automated feature engineering apparatus for collaborative feature set search is provided, comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described automated feature engineering method for collaborative feature set search.

[0010] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described automated feature engineering method for cooperative feature set search.

[0011] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention models the automated feature engineering of tabular data as a Markov decision process. By constructing a collaborative feature set search framework based on a hierarchical reinforcement learning scheme, it can efficiently generate various high-order complex features from structured data, ultimately forming the most expressive feature set. This effectively reduces computational overhead and improves efficiency when performing downstream tasks. The method of this invention has the advantages of strong versatility and ease of use. Attached Figure Description

[0012] Figure 1 This is an architecture diagram of automated feature engineering for collaborative feature set search provided in an embodiment of the present invention;

[0013] Figure 2 This is a schematic diagram of the strategy network structure provided in an embodiment of the present invention;

[0014] Figure 3 This is a graph showing the results of testing the method of this invention on 26 datasets;

[0015] Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0018] like Figure 1 As shown, this embodiment of the invention provides an automated feature engineering method for collaborative feature set search. Feature engineering plays a crucial role in building tabular data mining systems and traditionally requires experienced human experts to perform. The method proposed in this embodiment searches structured data for classification and regression tasks at the feature set granularity and learns a hierarchical strategy through reinforcement learning. The method specifically includes the following steps:

[0019] (1) Obtain the table data to be processed: C = (c1, ... c2) n ), where c i Let represent the i-th initial feature column, and n represent the number of feature columns.

[0020] It should be noted that the tabular data in this example was obtained from publicly available structured data source websites such as OpenML, Kaggle, and UCI.

[0021] (2) The automated feature engineering of the tabular data is modeled as a Markov decision process. Based on hierarchical reinforcement learning, the Markov decision process is processed to obtain a new feature set corresponding to the tabular data. Specifically, step (2) includes the following steps:

[0022] (2.1) The automated feature engineering of tabular data is modeled as a Markov decision process;

[0023] Specifically, in this example, a Markov decision process is defined as a tuple.

[0024] Among them, agents The trainee interacts with the environment ξ, which receives the previous state matrix S. i-1 Select action set a i Calculate the reward r and return the next state matrix S. i Among them, action set a i It is the set of all characteristic operations.

[0025] (A) Action set a iDefinition of space: The core of automated feature engineering is to perform feature transformation operations and feature execution operations on feature columns.

[0026] On the one hand, feature transformation operations are implemented based on the following feature transformation functions:

[0027] Unary operations include: tanh, square, inverse, log, sqrt, sigmoid, none, and other value conversion functions;

[0028] Binary operations include combinational arithmetic functions such as addition, subtraction, multiplication, division, and combination.

[0029] In this example, the discrete features can be further integrated into the search space by adding a combine transformation operation.

[0030] On the other hand, the feature execution operations include: replace and / or append.

[0031] In other words, the hierarchical strategy obtained by the method of the present invention consists of a set of actions—feature transformation operations and feature execution operations.

[0032] (B) Definition of state S: State S i Let a represent the action sequence in the i-th round. i In other words, this invention uses the action sequence of all feature columns as the state. S0 is initialized as an action sequence without any transitions.

[0033] (C) Agency Definition of policy network: agent From the policy network π: Modeling, where P is the agent The probability of the sampling action.

[0034] (D) Definition of Environment ξ: Environment ξ consists of a machine learning algorithm L with preset hyperparameters and an evaluation method E. In this example, feature engineering is performed on the corresponding feature columns based on the predicted action sequence, and the results are evaluated using the machine learning algorithm L and the evaluation method E.

[0035] For classification tasks, the machine learning algorithm L is selected from Linear Support Vector Machine (LinearSVC), Logistic Regression, Random Forest, and Gradient Boosting Tree (XGBOOST / LightGBM); for regression tasks, the machine learning algorithm L is selected from Linear Support Vector Machine, Linear Regression, Random Forest, and Gradient Boosting Tree. The evaluation method E for classification tasks is specified by the F1-score, and the evaluation method E for regression tasks is specified by 1-RAE.

[0036] (E) Reward Function Definition: Reward Function It is done by calculating the new feature set corresponding to the tabular data. The k-fold cross-validation performance was used to establish the method.

[0037] (2.2) The Markov decision process based on hierarchical reinforcement learning includes: constructing a policy network, sampling multiple complete feature engineering plans P on the policy network, and applying the feature engineering plans P to the original features C to obtain a new feature set. Calculate the reward r, and update the policy network according to the reward r until convergence. Specifically, step (2.2) includes the following steps:

[0038] (2.2.1) Construct a policy network based on the definition of Markov decision process and collaborative search strategy;

[0039] like Figure 2 As shown, the policy network includes an embedding layer, which is used to encode the action set a. i Encode a continuous vector; input this encoding into the first Long Short-Term Memory (LSTM) network. M This is used to predict feature execution decisions, resulting in a continuous feature execution vector. The continuous feature execution vector is then decoded to obtain the feature execution operation M (i.e., the feature execution policy). Simultaneously, the continuous feature execution vector is input into the second Long Short-Term Memory (LSTM) network. T This is used to predict feature transformation decisions, resulting in a continuous vector of feature transformations. Decoding this continuous vector yields the feature transformation operation T (i.e., the feature transformation strategy).

[0040] It should be noted that in this example, the first Long Short-Term Memory (LSTM) network is designed. M Second Long Short-Term Memory Network (LSTM) T Predict the feature transformation decision and the execution decision respectively, where T and M represent the feature transformation sequence and the decision sequence. The first Long Short-Term Memory (LSTM) network... M Second Long Short-Term Memory Network (LSTM)T Both rely on a collaborative search strategy, meaning the policy network π considers the complementarity and relationships between features, then determines the operations for all feature columns at once. To obtain the feature transformation operation T and the feature execution operation M, double mirror path prediction is required. It's important to note that, based on the different parameter convergence and operational complexity of the two LSTMs, this example chooses to place them together, using the second Long Short-Term Memory (LSTM) network. T First Long Short-Term Memory Network (LSTM) M Following this, each Long Short-Term Memory (LSTM) network is followed by a decoder module, which can consist of several fully connected layers and a non-linear activation function. After the decoder's output, a softmax function is used as the head to generate a probability distribution in the action space.

[0041] (2.2.2) Next, feature engineering plans are sampled on the constructed policy network to obtain feature engineering plans P. Feature engineering plans P are then applied to the original features C to obtain a new feature set.

[0042] The expression for feature engineering plan sampling of the constructed policy network is as follows:

[0043] t i ~p(t|S i ),m i ~p(m|S i ),

[0044] Where i refers to the index wheel, and m is the number of samples. i and t i Sequences of feature execution operations M and feature transformation operations T are formed separately. Finally, these sequences are combined to obtain a complete action sequence S. i+1 =M∪T. Simultaneously, multiple action sequences S are obtained through a K-round continuous sampling strategy network. i Ultimately, this leads to a complete feature engineering plan P = (S1,…S…) K To find the optimal feature engineering plan P, the policy network is required to maximize its expected reward. This is because the reward signal... Since it is not differentiable, the Proximal Policy Optimization (PPO) algorithm was used to train the policy network.

[0045] Specifically, during the sampling process, the Long Short-Term Memory (LSTM) network relies on the backbone of the collaborative feature set search. T A feature transformation strategy T is generated for the entire feature set at once, utilizing information from all features. In specific implementation, the initial dataset C = (c1, ... c2) nBased on their feature types, features are categorized into discrete features (classification features) and continuous features (numerical features), and specific preprocessing is performed, including normalization, missing value imputation, and discrete-to-continuous conversion. For the i-th round, a sampled action sequence S is obtained. i =(s i1 ,...s in ), where s ij This represents an action with a two-layer policy, including a feature transformation policy and a feature execution policy. In the i ∈ [1, K], the algorithm will use the state S generated by the policy network. i-1 Generate the current state S i It considers the interaction information of all feature columns and directly generates the action sequence corresponding to each column once. Ultimately, all action sequences constitute an action plan P. This action plan P is applied to the current dataset, and by implementing corresponding hierarchical strategies, each feature column is modified to obtain a new feature set. Then directly evaluate the entire feature set. This invention also establishes several parallel information samplers, which record the action information (including the complete feature engineering plan, calculated reward value, formed feature set, etc.) obtained by each thread in each round of iterative search.

[0046] At a higher level, the Long Short-Term Memory (LSTM) network... M The system generates post-processing (replacement or append decision M) decisions for new features formed after feature transformation operation T, enabling autonomous management of feature quality and quantity. It predicts the execution policy for each feature transformation associated with each feature column; that is, the policy network predicts an additional policy at a higher level to decide whether to replace the current feature column or append it. For replacement decisions, the generated transformed feature is replaced with the feature at the index position of the current decision. For append decisions, the generated feature is appended after the positions of the first n feature columns, without further exploration.

[0047] (2.2.3) For the new feature set Perform k-fold cross-validation to obtain the reward r, and update the policy network based on the reward r until convergence.

[0048] Specifically, for the new feature set After performing k-fold cross-validation and obtaining the reward r, since the reward r is non-differentiable, the Proximal Policy Optimization (PPO) algorithm is chosen to update the policy network in this example. For the new feature set... The k-fold cross-validation evaluation includes: feature engineering on the corresponding feature columns based on the predicted action sequence, and evaluation using a machine learning algorithm L and an evaluation method E. For classification tasks, the machine learning algorithm L is selected from Linear Support Vector Machine (LinearSVC), Logistic Regression (LogisticRegression), Random Forest (RandomForest), and Gradient Boosting Tree (XGBOOST / LightGBM); for regression tasks, the machine learning algorithm L is selected from Linear Support Vector Machine, Linear Regression (LinearRegression), Random Forest, and Gradient Boosting Tree. The evaluation method (metric) E for classification tasks is specified by the F1-score, and the evaluation method for regression tasks is specified by 1-RAE (Relative Absolute Error).

[0049] According to the PPO algorithm, the importance sampling ratio is I. θ The following can be calculated:

[0050]

[0051] In the formula, a t Let s be an action sequence with step size t. t This is the state with a step size of t.

[0052] Set a baseline value, calculate the difference between the current reward r and the baseline, and then standardize it into the advantage function A. θ (a t s t The value of ). Then, using the clip method, the advantage function C′ is defined, with the expression:

[0053] C′(a t s t ) = clip(I θ ,1-ε,1+ε)A θ (a t s t ),

[0054] Where ε is a hyperparameter (usually set to 0.2), clip can change the first parameter I θ Cut into the interval consisting of two parameters, 1-ε and 1+ε.

[0055] Therefore, the loss function of the policy network can be written as follows.

[0056]

[0057] Finally, the loss function can be defined as:

[0058] J Final(θ)=J PPO (θ)+E loss (θ)

[0059] Among them, E loss (θ) is an entropy reward to ensure sufficient exploration.

[0060] The update process of the policy network also includes setting the initial hyperparameters of the policy network: In this embodiment, an action sequence S0 without any transformations is initialized, each action is embedded as an 8-size vector, and the hidden layer size of the two unidirectional Long Short Memory networks is 300. The method sets the maximum number of search iterations to 400 and uses the Adam optimizer to train the agent. The number of feature engineering plans sampled per epoch is 24, and the learning rate is 0.001.

[0061] (2.3) Input the initial target dataset and the initially set hyperparameters into the policy network after optimization iteration, and perform another inference to obtain an action plan P′ with good scores and high generalization. Then, apply P′ to the initial target dataset C to form a new feature set after feature engineering. This feature set is then used as input for downstream tasks (including classification and regression tasks).

[0062] Example 1

[0063] This invention employs the Airfoil dataset for feature engineering. Airfoil is a regression task containing an initial feature set of five columns: <frequency, angle, length, velocity, and thickness>, with the target being the sound pressure of an aircraft. In this embodiment, the following steps are performed:

[0064] (A) First, the automated feature engineering model of the table data C = (c1, ... c5) is modeled as a Markov decision process, for example, c1 is the frequency feature column.

[0065] (B) Then, initialize the parameters of the policy network π according to step (2.2.3) and perform sampling multiple times for K consecutive rounds to obtain multiple complete feature engineering plans P. Apply the feature engineering plans P to the original feature set C to obtain a new feature set. Calculate the reward r, and update the policy network based on the reward r until convergence.

[0066] (C) In the policy network after iteration, perform another inference to obtain an action plan P′ with good scores and high generalization. Apply P′ to the initial target dataset C to form a new feature set after feature engineering. The new feature set formed in this embodiment A subset of features is The constructed feature subset can improve the performance accuracy of this regression task by 26%.

[0067] Example 2

[0068] This invention uses the Fertility dataset for feature engineering. Airfoil is a classification task that includes an initial feature set of features such as Age, Childish diseases, Accident, and Sedentary. The prediction target is sperm concentration. The specific steps of this invention are described in Example 1, resulting in a new feature set. A new feature is (in This represents the combine operation (i.e., feature crossing), which can improve the performance accuracy of this classification task by 5%.

[0069] Example 3

[0070] On a machine equipped with an Intel(R) Xeon(R) CPU (E5-2680v4@2.40GHz, 28 cores), 503GB of RAM, and an 8TB hard drive, the automated feature engineering method for collaborative feature set search provided in this invention was implemented simultaneously on several publicly available machine learning datasets (OpenML, UCI repository, Kaggle). This example compares the optimization results with other existing automated feature engineering methods for structured data, parallel automated machine learning methods (AutoML), and feature modeling methods. Figure 3 As shown. It can be seen that the Catch and Catch of this invention... - (Catch - This is a variant of the invention without the combine operation, which offers a significant performance improvement over the base column of the original data; and it achieves optimal performance across all datasets compared to competing methods such as DIFFER, NFS, AutoFEAT, and DFS. Furthermore, for other parallel methods (AutoML and Feature modling), this invention achieves optimal performance on over 80% of datasets.

[0071] Corresponding to the aforementioned embodiments of the automated feature engineering method for collaborative feature set search, the present invention also provides embodiments of an automated feature engineering apparatus for collaborative feature set search.

[0072] See Figure 4 The present invention provides an automated feature engineering apparatus for collaborative feature set search, comprising one or more processors for implementing the automated feature engineering method for collaborative feature set search in the above embodiments.

[0073] The embodiments of the automated feature engineering device for collaborative feature set search of this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 4 The diagram shown is a hardware structure diagram of any device with data processing capabilities, where the automated feature engineering device for collaborative feature set search of the present invention is located. (Except for...) Figure 4 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0074] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0075] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0076] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the automated feature engineering method for cooperative feature set search described in the above embodiments.

[0077] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0078] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. An automated feature engineering method for collaborative feature set search, characterized in that, The method specifically comprises the following steps: Obtaining table data to be processed; Modeling the automated feature engineering of the table data as a Markov decision process, processing the Markov decision process based on hierarchical reinforcement learning, and then searching for a new feature set corresponding to the table data in cooperation with all features; The processing of the Markov decision process based on hierarchical reinforcement learning and the searching for the new feature set corresponding to the table data in cooperation with all features comprise: According to the definition of Markov decision process and the cooperative search strategy, a strategy network is constructed, a plurality of complete feature engineering plans P are sampled on the strategy network, the feature engineering plans P are applied to the original features C, and an updated feature set is obtained A reward r is calculated, and the strategy network is optimized according to the reward r; The table data to be processed is input into the optimized strategy network, a reasoning feature engineering plan P' is obtained, and then P' is applied to the table data to be processed C to form a new feature set after feature engineering The policy network comprises an encoding layer configured to encode the action set a i into a continuous vector; input the encoded vector into a first long short-term memory network (LSTM) M to predict a feature execution decision and obtain a feature execution continuous vector; decode the feature execution continuous vector to obtain a feature execution operation M; simultaneously input the feature execution continuous vector into a second long short-term memory network (LSTM) T to predict a feature conversion decision and obtain a feature conversion continuous vector; and decode the feature conversion continuous vector to obtain a feature conversion operation T.

2. The automated feature engineering method for collaborative feature set search according to claim 1, characterized in that, Modeling the automated feature engineering of the table data as a Markov decision process comprises: A Markov decision process is defined as a tuple where the agent is trained to interact with an environment ξ that receives a previous state matrix S i-1 , selects a set of actions a i , computes a reward r, and returns a next state matrix S i .

3. The automated feature engineering method for collaborative feature set search of claim 2, wherein, Action set a i comprises a feature conversion operation and a feature execution operation; The feature conversion operation is implemented based on a feature conversion function, and the feature conversion function is a value conversion function including tanh, square, inverse, log, sqrt, sigmoid, and none, or a combined arithmetic function including addition, subtraction, multiplication, division, and combine; The feature execution operation comprises a replacement operation and an addition operation.

4. The automated feature engineering method for collaborative feature set search of claim 2, wherein, State matrix S i Let a represent the action sequence in the i-th round. i ;acting From the policy network π: Modeling, where P is the agent The probability of the sampled action; the environment ξ, composed of a machine learning algorithm and evaluation method with preset hyperparameters; the reward function. By calculating the new feature set corresponding to the tabular data The performance of k-fold cross-validation was obtained.

5. The automation feature engineering method for collaborative feature set search according to claim 1, wherein, Sampling a plurality of complete feature engineering plans P on the policy network comprises: The expression of the sampling of the feature engineering plan on the constructed policy network is as follows: t i ~ p(t | S i ), m i ~ p(m | S i ), where i denotes the index wheel, m i and t i form the sequence of feature execution operations M and feature transformation operations T, respectively; a complete sequence of actions is denoted by S i+1 = M U T; A plurality of action sequences S are obtained by a recurrent K-round continuous sampling strategy network i , and a feature engineering plan P = (S1,... S K ) is obtained.

6. The automation feature engineering method for collaborative feature set search according to claim 1, characterized in that, Optimizing the policy network according to the reward r comprises: on the new feature set perform k-fold cross-validation to obtain rewards r, perform feature engineering on the corresponding feature columns according to the predicted action sequence, and evaluate by a machine learning algorithm L and an evaluation method E, update the policy network according to the rewards r until convergence; For a classification task, the machine learning algorithm L is selected from linear support vector machines, logistic regression, random forests, and gradient boosting trees; and the evaluation method E is specified by F1-score. For a regression task, the machine learning algorithm L is selected from linear support vector machines, linear regression, random forests, and gradient boosting trees; and the evaluation method E is specified by 1-RAE.

7. An apparatus for automated feature engineering oriented to collaborative feature set search, comprising a memory and a processor, characterized in that, The memory is coupled with the processor; wherein the memory is configured to store program data, and the processor is configured to execute the program data to implement the automated feature engineering method for cooperative feature set search according to any one of claims 1-6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the automated feature engineering method for cooperative feature set search according to any one of claims 1-6.

Images