An ordered regression method and system for processing tabular data

Abstract

This invention belongs to the field of tabular data prediction and provides a method and system for ordered regression of tabular data. The method includes: encoding query samples and a tabular data training set using a multimodal encoder to obtain initial and final query sample features, training features, and a candidate label set; calculating the cumulative inter-class distance of each type of feature in the training feature candidate set and retrieving sample features and corresponding labels similar to the final query sample features from the training features and label candidate set; calculating contribution weights and the similarity between similar sample features and the final query sample features using similar sample labels and cumulative inter-class distances; integrating the similarity and contribution weights with the initial query sample features, and then inputting the results into a predictor to obtain ordered labels. Because the cumulative inter-class distance can reflect the hierarchy, unequal distance, and inclusiveness of ordered knowledge, it promotes classification prediction and can predict the ordered labels of query samples relatively accurately.

Description

Technical Field

[0001] This invention belongs to the field of tabular data prediction technology, and in particular relates to an ordered regression method and system for processing tabular data. Background Technology

[0002] Ordinal regression is a classic problem in machine learning, specifically designed for predicting data with ordered labels. It has wide applications in real life, such as movie ratings, product ratings, and lesion severity assessment. A table is a structured dataset composed of rows and columns.

[0003] Currently, research on ordered regression mainly focuses on fields such as computer vision and natural language processing. The inventors found ordered regression methods for image data and 3D point cloud data, but no methods for tabular data. For example, Patent Document 1: CN117333485A discloses a WSI survival prediction method based on a weakly supervised deep ordinal regression network, which inputs image data into an ordinal regression network to perform survival prediction and obtain the prediction result. Patent Document 2: CN106250818A discloses a face age estimation method based on total order-preserving projection, which performs total order projection on feature data extracted from face images to obtain low-dimensional features, inputs them into an ordered regression classification model, and then estimates the age. Patent Document 3: CN105975916A discloses an age estimation method based on a multi-output convolutional neural network and ordered regression, which uses face images and a binary classification ordered regression method for age estimation. Patent document 4, CN111428619A, discloses a 3D point cloud head pose estimation system and method based on ordered regression and soft labeling. It processes 3D point cloud data of the head generated from depth images to estimate head pose. While patent documents 1 to 3 all use images as input for level estimation, patent document 4 uses 3D point cloud data as input for head pose estimation.

[0004] Due to the unique discreteness and heterogeneous multimodal nature of tabular data, ordered regression methods outside the field of tabular data prediction cannot be used for tabular data. Therefore, there is an urgent need for an ordered regression method for processing tabular data. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and to propose an ordered regression method for processing tabular data. This invention also discloses an ordered regression system for processing tabular data.

[0006] On one hand, the present invention provides an ordered regression method for processing tabular data, comprising:

[0007] Step 1: Use a multimodal encoder to encode the query sample twice to obtain the initial query sample features and the final query sample features. Use a multimodal encoder to encode the tabular data training set twice to obtain the training feature candidate set and the label candidate set. Each sample in the tabular data training set is tabular data including entity features and corresponding ordered labels, and the query sample is tabular data including entity features.

[0008] Step 2: Calculate the cumulative inter-class distance of each type of feature obtained by classifying it according to ordered labels in the training feature candidate set, and use the retrieval tool to retrieve multiple similar sample features and corresponding multiple similar sample labels that are similar to the features of the final query sample from the training feature candidate set and the label candidate set.

[0009] Step 3: Calculate the contribution weight of multiple similar sample features by using the cumulative inter-class distance values ​​of multiple similar sample labels and various features, and calculate the similarity between the retrieved multiple similar sample features and the final query sample features.

[0010] Step 4: Referring to the attention mechanism, similarity and contribution weights are fused to obtain contextual features;

[0011] Step 5: Integrate the initial query sample features and context features to obtain the final feature representation of the query sample;

[0012] Step 6: Use the predictor to predict the final feature representation to obtain the ordered labels of the query samples.

[0013] In an improved ordinal regression method for processing tabular data, step 2, which involves calculating the cumulative inter-class distances of features obtained after classifying the training feature candidate set according to ordinal labels, includes:

[0014] Step 2-1: Calculate the inter-class distance between each category of features obtained by classifying them according to ordered labels in the training feature candidate set and the baseline category features;

[0015] Step 2-2: For each type of feature, sum the inter-class distances between this type of feature and the baseline feature to obtain the summed inter-class distance value of this type of feature.

[0016] In an improved ordinal regression method for processing tabular data, where the baseline category is a first class, step 2-1 includes:

[0017] Calculate the mean features of each category obtained after classifying the training feature candidate set according to ordered labels;

[0018] Calculate the distance from the mean feature of each type of feature to the first type of feature to obtain the inter-class distance of each type of feature.

[0019] In an improved ordinal regression method for processing tabular data, step 3, which calculates the contribution weights of multiple similar sample features using the sum of inter-class distances of multiple similar sample labels and features, includes:

[0020] Using the candidate set of labels and the cumulative inter-class distance values ​​of various features, a label inter-class weight dictionary is constructed, which includes the label of each category and the corresponding cumulative inter-class distance value.

[0021] Using multiple similar sample labels, the corresponding cumulative inter-class distance value is found from the inter-class weight dictionary of the labels;

[0022] The accumulated inter-class distances are encoded to obtain the contribution weights of features from multiple similar samples.

[0023] In an improved ordinal regression method for processing tabular data, step 4 includes:

[0024] The context features are obtained by multiplying the similarity and contribution weights together.

[0025] In an improved ordinal regression method for processing tabular data, step 5 includes:

[0026] The initial query sample features and context features are added together to obtain the final feature representation of the query sample.

[0027] In an improved ordinal regression method for processing tabular data, during training, the query sample is one sample in the training set; during validation, the query sample is one sample in the validation set; during testing, the query sample is one sample in the test set; and in actual use, the query sample is one piece of data with an ordinal label to be determined.

[0028] In an improved ordinal regression method for processing tabular data, the data type of the tabular data includes any one, two, or all of the following three types:

[0029] Numeric, character, and binary types.

[0030] In an improved ordinal regression method for processing tabular data, the ordinal regression method is applied in product rating, age prediction, soil utility rating, and retrieval-related rating scenarios.

[0031] On the other hand, the present invention provides an ordered regression system based on the above-described ordered regression method for processing tabular data, comprising:

[0032] The query encoding module is used to encode the query sample twice using a multimodal encoder to obtain the initial query sample features and the final query sample features. The query sample is tabular data that includes entity features.

[0033] The training set encoding module is used to encode the tabular data training set twice using a multimodal encoder to obtain a training feature candidate set and a label candidate set. Each sample in the tabular data training set is tabular data that includes entity features and corresponding ordered labels.

[0034] The ordered knowledge enhancement module is used to calculate the cumulative inter-class distance of each type of feature obtained by classifying the training feature candidate set according to ordered labels;

[0035] The retrieval module is used to retrieve multiple similar sample features and corresponding multiple similar sample labels that are similar to the features of the final query sample from the training feature candidate set and label candidate set using a retrieval tool.

[0036] The contribution weight calculation module is used to calculate the contribution weight of multiple similar sample features by using the cumulative inter-class distance values ​​of the labels of multiple similar samples and various features.

[0037] The similarity calculation module is used to calculate the similarity between the features of multiple similar samples retrieved and the features of the final query sample.

[0038] The fusion module is used to fuse similarity and contribution weights, referencing the attention mechanism, to obtain contextual features;

[0039] The integration module is used to integrate the initial query sample features and context features to obtain the final feature representation of the query sample; and

[0040] The prediction module is used to predict the final feature representation using a predictor to obtain the ordered labels of the query samples.

[0041] Compared with the prior art, the advantages of the present invention are:

[0042] The solution of this invention combines ordered regression and deep learning methods. It uses inter-class distance to represent the knowledge intervals embodied by ordered data of different categories, that is, the hierarchical and unequal distance between classes. By accumulating the distances, it reflects the inclusiveness of ordered knowledge and uses it as the basis for calculating contribution weights. This process uncovers the hierarchical relationship between data features, promotes better classification and prediction of ordered tabular data, and thus can predict the ordered labels of query samples more accurately. Attached Figure Description

[0043] Figure 1 This is a flowchart of the ordered regression method for processing tabular data in Embodiment 1 of the present invention;

[0044] Figure 2 This is a schematic diagram illustrating the process of the deep learning model for tables processing tabular data in Embodiment 2 of the present invention.

[0045] Figure 3 This is a schematic diagram of an ordered regression system for processing tabular data in Embodiment 3 of the present invention. Detailed Implementation

[0046] The ordered regression method for processing tabular data of the present invention first encodes the multimodal content of the tabular data training set and the query samples using a multimodal encoder. Then, based on the features of the query samples, it retrieves multiple similar features (e.g., top-k similar features) and similar labels from the training feature candidate set, and calculates the cumulative inter-class distance for each class in the training feature candidate set. Next, it uses the cumulative inter-class distance and similar labels to calculate the contribution weight V of n similar features in the attention mechanism, and calculates the similarity K*Q value in the attention mechanism using the top-k similar features. Finally, it fuses label knowledge and similar features using K*Q and V, integrates the query features and similar features to obtain the final feature representation of the query features, and inputs it into the predictor to obtain the ordered regression prediction value.

[0047] The ordered regression method of this invention takes tabular data as input and uses the TabR model, a retrieval-based tabular deep learning model, as the backbone network. An ordered knowledge enhancement module is added to this backbone network to form a new tabular deep learning model, aiming to predict ordered labels. The TabR model is detailed in the paper Yury Gorishniy, Ivan Rubachev, Nikolay Kartasev, Daniil Shlenskii, Akim Kotelnikov, and Artem Babenko. Tabr: Unlocking the power of retrieval-augmented tabular deep learning. arXiv preprint arXiv:2307.14338,2023.

[0048] The method and system of the present invention are used to process ordered tabular data, and the application scenarios include, but are not limited to, product rating, age prediction, soil utility rating and search relevance rating.

[0049] Product ratings, such as wine quality ratings, can be found in the open-source dataset at https: / / openml.org / d / 287.

[0050] Age prediction, such as the abalone age prediction dataset, can be found at https: / / archive.ics.uci.edu / dataset / 1 / abalone.

[0051] The soil utility rating scale eucalyptus, and its open-source dataset can be found at https: / / www.openml.org / d / 188.

[0052] For details on the open-source dataset for retrieval relevance prediction, please see https: / / www.microsoft.com / en-us / research / project / mslr / .

[0053] The datasets generated in these scenarios are all presented in Excel format, with each cell containing a number or string, and ordered labels as level labels (such as level meaning labels like 1, 2, 3, 4, 5).

[0054] The prediction goal is to predict the level of the entity represented by a row of data. The meaning of each column may differ across datasets, but they all share the common characteristic of being presented in tabular form. Each cell is formatted as a numerical value or a single string, with ordered labels representing the level.

[0055] Taking the wine quality dataset as an example, each row represents a sample, and there are 12 columns in total. Columns 1 to 11 are entity features, and columns 12 are the corresponding ordered labels. Their meanings are shown in Table 1 below.

[0056] Table 1

[0057]

[0058] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0059] Example 1

[0060] like Figure 1 As shown, this is an ordered regression method for processing tabular data according to Embodiment 1 of the present invention, comprising:

[0061] Step 101: Encode the query sample twice using a multimodal encoder to obtain the initial query sample features and the final query sample features. Then, encode the tabular data training set using a multimodal encoder to obtain the training feature candidate set and the label candidate set.

[0062] In the training set of tabular data, each sample is tabular data including entity features and corresponding ordered labels. The query sample is tabular data including entity features. The multimodal encoder performs the first encoding on the query sample to obtain the initial query sample features. The multimodal encoder performs the second encoding on the initial query sample features to obtain the final query sample features.

[0063] The query sample is simply the original data of a row in the table. The first encoding encodes features of different modalities (types) to represent them numerically, paying attention to important feature values ​​(because the multimodal encoder is learnable). The second encoding simply adds a normalization layer to scale the features to a similar range, reducing the bias of simply retrieving similar features and improving retrieval generalization performance. The features encoded in the first encoding are ultimately used when merging with context features because they contain more accurate semantic information; the more multimodal encoders a feature passes through, the greater the possibility of semantic bias.

[0064] The table data contains numbers, characters, and may also be in different forms such as sentences or images. These different forms of data are called multimodal data. Therefore, in this embodiment of the invention, a multimodal encoder is used to process the table data.

[0065] Here, different table datasets are needed for different query samples. The training set in these datasets is used to construct the corresponding training feature candidate set.

[0066] For example, when querying samples containing 11 columns of feature data reflecting wine quality, it is necessary to construct a candidate set of wine quality training features using a wine quality table data training set. This training set contains many samples, each with 12 columns of data, where the first 11 columns are feature data reflecting wine quality and the 12th column is wine quality grade data.

[0067] When querying samples that are soil utility feature data, it is necessary to use the soil utility table data training set to construct a soil utility training feature candidate set.

[0068] Here, the query sample is one sample in the training set during training, one sample in the validation set during validation, and one sample in the test set during testing. In actual applications, it is the query data.

[0069] In step 1, the candidate feature set contains features, and the candidate label set contains ordered labels. The features in the candidate feature set still correspond to the labels in the candidate label set.

[0070] In this embodiment of the invention, the novel table deep learning model is trained using a training set. It continuously learns and optimizes parameters through loss function feedback until the function converges, at which point training stops. After training, the encoder, retrieval unit, predictor, etc., all contain the trained parameters. In actual use, ordered label prediction can be performed directly using query data and the training set.

[0071] Step 102: Calculate the cumulative inter-class distance of each type of feature obtained after classifying by ordered labels in the training feature candidate set, and use the retrieval tool to retrieve multiple similar sample features and corresponding multiple similar sample labels that are similar to the features of the final query sample from the training feature candidate set and the label candidate set.

[0072] Since the categories are ordered and the distances between these ordered categories are not equal, and the inclusion property of ordered knowledge is found, that is, high-level knowledge contains low-level knowledge, step 102 uses the sum of inter-class distances to represent this. This makes the regression method of the present invention not only consider the order relationship between different categories, but also enhance the ordered knowledge between categories, making the subsequent ordered label prediction more accurate.

[0073] The above calculation of the cumulative inter-class distances of features obtained by classifying the candidate training features according to ordered labels may include steps 102-1 and 102-2:

[0074] Step 102-1: Calculate the inter-class distance between each category of features obtained after classifying by ordered labels in the training feature candidate set and the baseline category features;

[0075] Specifically, assuming the ordered labels can be 1, 2, and 3, entity features with ordered labels of 1 in the training feature candidate set can be classified into the first category, entity features with ordered labels of 2 in the training feature candidate set can be classified into the second category, entity features with ordered labels of 2 in the training feature candidate set can be classified into the third category.

[0076] The inter-class distance here can be Euclidean distance, or it can be Manhattan distance, Chebyshev distance, Minkowski distance, cosine distance, Hamming distance, etc. Any distance that can be calculated between features can be used. In application, it is necessary to try different formulas one by one to find the optimal distance formula based on different data characteristics.

[0077] The baseline category here can be selected based on the total number of categories, such as selecting category one and category two.

[0078] Specifically, the benchmark category is the first category, and step 102-1 includes:

[0079] Calculate the mean features of each category of features obtained after classifying the candidate training features by ordered labels; calculate the distance between the mean features of each category and the first category feature to obtain the inter-class distance of each category of features.

[0080] Here, we use unequal-distance class mean features to approximate the unequal distance of ordered knowledge between classes.

[0081] Step 102-2: For each type of feature, sum the inter-class distances between this type of feature and the baseline feature to obtain the summed inter-class distance value of this type of feature.

[0082] The accumulated distance values ​​here can be combined with ordered labels to form a dictionary, which is used to determine the contribution weight.

[0083] Step 103: Calculate the contribution weight of multiple similar sample features by using the cumulative inter-class distance values ​​of multiple similar sample labels and various features, and calculate the similarity between the retrieved multiple similar sample features and the query sample features.

[0084] Specifically, step 103 includes the following steps:

[0085] Step 1: Construct a label inter-class weight dictionary using the candidate label set and the cumulative inter-class distance values ​​of various features. The label inter-class weight dictionary includes the label of each category and the corresponding cumulative inter-class distance value.

[0086] Step 2: Using multiple similar sample labels, find the corresponding cumulative inter-class distance value from the label inter-class weight dictionary;

[0087] Step 3: Encode the accumulated inter-class distance values ​​to obtain the contribution weights of features of multiple similar samples.

[0088] For example, based on the retrieved n similar sample feature rows, there are corresponding n ordered labels representing categories. For each of these n ordered labels representing categories, the cumulative inter-class distance value corresponding to the label with the same label is found in the dictionary, and finally n cumulative inter-class distance values ​​are obtained. However, these cumulative inter-class distance values ​​cannot be directly used as the contribution weights of the n features, or in other words, directly using the cumulative inter-class distance values ​​as contribution values ​​does not perform well. Therefore, an encoder (containing nn.linear, nn.Embedding layers, etc.) is used here, so that the contribution weights of the n similar features can be learned during model training.

[0089] Step 104: Referring to the attention mechanism, similarity and contribution weights are fused to obtain contextual features.

[0090] Specifically, the similarity and contribution weights can be multiplied together to obtain contextual features.

[0091] Step 105: Integrate the initial query sample features and context features to obtain the final feature representation of the query sample.

[0092] Specifically, the initial query sample features and context features are added together to obtain the final feature representation of the query sample.

[0093] Step 105 integrates the contribution weights obtained based on the cumulative inter-class distance into the query sample features through the features described below, thus enriching the feature representation.

[0094] Step 106: Use the predictor to predict the final feature representation to obtain the ordered labels of the query samples.

[0095] In the scheme of Embodiment 1 of this invention, ordered regression and deep learning are combined. For different datasets, corresponding training set features are extracted to construct a candidate set of training features. Then, the distance between various features is calculated, which represents the distance between ordered levels, and then combined with ordered labels to form a dictionary. Multiple samples similar to the query sample are retrieved through similarity, and then the level distance (weight) and similarity features are fused using an attention mechanism and integrated into the query sample features to enrich the feature expression. Finally, the data is input into the predictor for prediction. This not only provides an ordered regression method for tabular data, but also accurately predicts the ordered labels of query samples.

[0096] Example 2

[0097] This embodiment 2 provides a detailed description of the solution in embodiment 1 of the present invention, including the following Step 1 to Step 12.

[0098] Step 1: First, refer to the TabR model data format and divide the original table data into numeric type Xnum, character type Xcat, and binary type Xbin according to the data type.

[0099] Assume the dataset eucalyptus (https: / / www.openml.org / d / 188) has one row per sample, 515 rows for the training set, 147 rows for the validation set, and 74 rows for the test set, with a total of 19 columns, including 13 numerical columns, 5 character columns, and 1 binary column. The ordered labels represent 5 categories.

[0100] The process of the tabular deep learning model of this invention for processing tabular data is described in [reference needed]. Figure 2 For comparison Figure 2 To understand the solution of Embodiment 2 of the present invention.

[0101] Step 2: Use the multimodal encoder in the TabR model to encode the training set sample set X twice to obtain the sample training feature candidate set F.

[0102] X = cat(X) num ,X cat ,X bin ), where cat is a concatenation operation.

[0103] F = W k W x (X), where Wx and Wk are multimodal encoders.

[0104] Continuing with the previous example, assuming that the feature dimension of each sample is 128 after training feature encoding, then F is a tensor of (515, 128).

[0105] Step 3: Using the training feature candidate set F from the training set, calculate the mean feature Fi for each type of feature:

[0106] Fi = M(F), where i = 1, ..., c; c is the number of categories, and M is the mean algorithm.

[0107] Continuing with the previous example, since the dataset has 5 categories, Fi is a tensor of (5, 128).

[0108] Step 4: Calculate the distance from each type of feature to the first type of feature, and obtain the inter-class distance D. i .

[0109] D i =F i -F1, where i = 1, ..., c; c is the number of categories.

[0110] Continuing with the previous example, if Euclidean distance is used, the final tensor obtained is Di = (1, 5).

[0111] Step 5: Calculate the cumulative inter-class distance S for each class. i And construct the dictionary Dict.

[0112] The cumulative distance is used to indicate that a higher-level feature value contains a lower-level feature value.

[0113] Where i = 1, ..., c, j are the category ordinal numbers.

[0114] Dict = {key = y i :value=S i}, where y i It corresponds to the label of the i-th class.

[0115] Following the previous example, S i It is D i The sum of the preceding terms, therefore S iIt is also a tensor of (1, 5); Dict has 5 keys and 5 values.

[0116] Step 6: Use the multimodal encoder W in the TabR model x and multimodal encoder W k For query x q Perform encoding twice.

[0117] f x =W x (x q ), f k =W k (f x )

[0118] Query sample features f x This refers to the initial query sample features, and the query sample features f. k This refers to the final query sample features.

[0119] Continuing with the previous example, here x q During training, x represents one sample from the training set; during validation, it represents one sample from the validation set; and during testing, it represents one sample from the test set. q It is the shape of (1, 19), f k and f x All are tesnsors of (1, 128).

[0120] Step 7: Use the retrieval tool R in the TabR model to find n features f that match the query sample. k The set of similar features Fs.

[0121] Fs = R(f k →F)

[0122] Continuing with the previous example, if n is 96, then Fs is the tesnor of (1, 96, 128).

[0123] Step 8: Calculate the similarity KQ between the retrieved features and the query features:

[0124] KQ = softmax(-||f k -Fs|| 2 ), where softmax is the normalization function.

[0125] Continuing with the previous example, KQ is a tensor of (1, 96).

[0126] Step 9: Calculate the contribution weights of the n similar features based on the ordered labels Y of the retrieved similar features:

[0127] Value = W y(Dict[y=Y α ]), where α=1,...,n,W y It is composed of nn.linear and nn.Embedding layers in a neural network, and its purpose is to encode integer or floating-point values ​​to obtain the overall contribution weights. Physically, it means obtaining n different contribution weights for n similar features.

[0128] Using the previous example, the value is a tensor of (1, 96, 128).

[0129] Step 10: Following the attention mechanism, fuse similar features to obtain the context feature context_x.

[0130] context_x = KQ*Value, where * represents matrix multiplication, which is the multiplication between the KQ matrix and the Value matrix.

[0131] Using the previous example, context_x is a tensor of (1, 128).

[0132] Step 11: Integrate query sample features f x The final feature representation context_fx is obtained from context_x.

[0133] context_fx = f x +context_x

[0134] Continuing with the previous example, context_fx is a tensor of (1, 128).

[0135] Step 12: Input the final features into the predictor P for prediction.

[0136]

[0137] Following the example above, The maximum value is selected from the tenor prediction probabilities of (1, 5), which is the ordered label of the prediction.

[0138] The embodiments of the present invention use inter-class distance to represent the interval of ordered knowledge, and the inclusion of ordered knowledge is reflected by accumulation. Experimental verification shows that it improves wine quality rating by 1.13%, eucalyptus by 1.39%, and abalone by 0.15%.

[0139] Example 3

[0140] This embodiment 3 proposes an ordered regression system based on the ordered regression method for processing tabular data described in embodiments 1 and 2, such as... Figure 3As shown, it includes:

[0141] The query encoding module 31 is used to encode the query sample twice using a multimodal encoder to obtain the initial query sample features and the final query sample features, wherein the query sample is tabular data including entity features;

[0142] The training set encoding module 32 is used to encode the tabular data training set twice using a multimodal encoder to obtain a training feature candidate set and a label candidate set. Each sample in the tabular data training set is tabular data including entity features and corresponding ordered labels.

[0143] The ordered knowledge enhancement module 33 is used to calculate the cumulative inter-class distance of each type of feature obtained by classifying the training feature candidate set according to ordered labels;

[0144] The retrieval module 34 is used to retrieve multiple similar sample features and corresponding multiple similar sample labels that are similar to the features of the final query sample from the training feature candidate set and the label candidate set using a retrieval tool.

[0145] The contribution weight calculation module 35 is used to calculate the contribution weight of multiple similar sample features by using the cumulative inter-class distance values ​​of multiple similar sample labels and various features.

[0146] The similarity calculation module 36 is used to calculate the similarity between the features of multiple similar samples retrieved and the features of the final query sample;

[0147] Fusion module 37 is used to fuse similarity and contribution weights with reference to the attention mechanism to obtain contextual features;

[0148] Integration module 38 is used to integrate the initial query sample features and context features to obtain the final feature representation of the query sample; and

[0149] The prediction module 39 is used to predict the final feature representation using a predictor to obtain the ordered labels of the query samples.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An ordered regression method for processing tabular data, comprising: Step 1: Use a multimodal encoder to encode the query sample twice to obtain the initial query sample features and the final query sample features. Use a multimodal encoder to encode the tabular data training set twice to obtain the training feature candidate set and the label candidate set. Each sample in the tabular data training set is tabular data including entity features and corresponding ordered labels, and the query sample is tabular data including entity features. Step 2: Calculate the cumulative inter-class distance of each type of feature obtained by classifying it according to ordered labels in the training feature candidate set, and use the retrieval tool to retrieve multiple similar sample features and corresponding multiple similar sample labels that are similar to the features of the final query sample from the training feature candidate set and the label candidate set. Step 3: Calculate the contribution weight of multiple similar sample features by using the cumulative inter-class distance values ​​of multiple similar sample labels and various features, and calculate the similarity between the retrieved multiple similar sample features and the final query sample features. Step 4: Referring to the attention mechanism, similarity and contribution weights are fused to obtain contextual features; Step 5: Integrate the initial query sample features and context features to obtain the final feature representation of the query sample; Step 6: Use the predictor to predict the final feature representation to obtain the ordered labels of the query samples.

2. The ordered regression method for processing tabular data as described in claim 1, characterized in that, Step 2, which involves calculating the cumulative inter-class distances of features obtained after classifying the training feature candidate set according to ordered labels, includes: Step 2-1: Calculate the inter-class distance between each category of features obtained by classifying them according to ordered labels in the training feature candidate set and the baseline category features; Step 2-2: For each type of feature, sum the inter-class distances between this type of feature and the baseline feature to obtain the summed inter-class distance value of this type of feature.

3. The ordered regression method for processing tabular data as described in claim 2, characterized in that, The benchmark category is the first category, and step 2-1 includes: Calculate the mean features of each category obtained after classifying the training feature candidate set according to ordered labels; Calculate the distance from the mean feature of each type of feature to the first type of feature to obtain the inter-class distance of each type of feature.

4. The ordered regression method for processing tabular data as described in any one of claims 1-3, characterized in that, Step 3, which involves calculating the contribution weights of multiple similar sample features using the cumulative inter-class distances of multiple similar sample labels and various features, includes: Using the candidate set of labels and the cumulative inter-class distance values ​​of various features, a label inter-class weight dictionary is constructed, which includes the label of each category and the corresponding cumulative inter-class distance value. Using multiple similar sample labels, the corresponding cumulative inter-class distance value is found from the inter-class weight dictionary of the labels; The accumulated inter-class distances are encoded to obtain the contribution weights of features from multiple similar samples.

5. The ordered regression method for processing tabular data as described in any one of claims 1-3, characterized in that, Step 4 includes: The context features are obtained by multiplying the similarity and contribution weights together.

6. The ordered regression method for processing tabular data as described in any one of claims 1-3, characterized in that, Step 5 includes: The initial query sample features and context features are added together to obtain the final feature representation of the query sample.

7. The ordered regression method for processing tabular data as described in any one of claims 1-3, characterized in that, During training, the query sample is one sample in the training set; during validation, the query sample is one sample in the validation set; during testing, the query sample is one sample in the test set; and in actual use, the query sample is one piece of data with an ordered label to be determined.

8. The ordered regression method for processing tabular data as described in any one of claims 1-3, characterized in that, The data type of the table data includes any one, any two, or all of the following three types: Numeric, character, and binary types.

9. The ordered regression method for processing tabular data as described in any one of claims 1-3, characterized in that, The ordered regression method is applied in product rating, age prediction, soil utility rating, and retrieval-related rating scenarios.

10. An ordered regression system based on the ordered regression method for processing tabular data as described in claim 1, comprising: The query encoding module is used to encode the query sample twice using a multimodal encoder to obtain the initial query sample features and the final query sample features. The query sample is tabular data that includes entity features. The training set encoding module is used to encode the tabular data training set twice using a multimodal encoder to obtain a training feature candidate set and a label candidate set. Each sample in the tabular data training set is tabular data that includes entity features and corresponding ordered labels. The ordered knowledge enhancement module is used to calculate the cumulative inter-class distance of each type of feature obtained by classifying the training feature candidate set according to ordered labels; The retrieval module is used to retrieve multiple similar sample features and corresponding multiple similar sample labels that are similar to the features of the final query sample from the training feature candidate set and label candidate set using a retrieval tool. The contribution weight calculation module is used to calculate the contribution weight of multiple similar sample features by using the cumulative inter-class distance values ​​of the labels of multiple similar samples and various features. The similarity calculation module is used to calculate the similarity between the features of multiple similar samples retrieved and the features of the final query sample. The fusion module is used to fuse similarity and contribution weights, referencing the attention mechanism, to obtain contextual features; The integration module is used to integrate the initial query sample features and context features to obtain the final feature representation of the query sample; and The prediction module is used to predict the final feature representation using a predictor to obtain the ordered labels of the query samples.

Images