System and method for machine learning architecture for electronic field autofill
Approaches are proposed herein for machine learning based electronic field autofill. In particular, machine learning approaches are utilized to estimate inputs for entry into fields. The machine learning approach described herein includes a proposed feature engineering approach where collected data obtained from various webpages (e.g., checkout pages) is encoded for training and inference. The approach is particularly useful in respect of providing a machine learning mechanism that is resilient in spite of inconsistent and varying approaches of implementing form fields.
1 . A computer implemented system for machine learning architecture for electronic field autofill with positional pattern recognition, the system comprising:
a processor operating in conjunction with computer memory and non-transitory computer readable storage, the processor configured to:
extract raw input data sets from one or more webpages, the raw input data sets representing raw markup language machine instructions used in rendering the one or more webpages;
identify one or more input fields in the raw input data sets, and for each input data field, extract label information including a label, a previous label corresponding to a label of an immediately preceding input field, and a next label corresponding to a label of an immediately succeeding input field;
extend the label information using feature engineering approaches including at least word embedding of label text using a sentence transformer model to generate dense vector representations of the label, the previous label, and the next label, and clustering approaches to associate each label with at least a type encoding, wherein the clustering approaches comprise:
applying a clustering algorithm to the dense vector representations of the previous labels to generate a previous label cluster number for each input field, wherein semantically similar previous labels are clustered; and
applying the clustering algorithm to the dense vector representations of the next labels to generate a next label cluster number for each input field, wherein semantically similar next labels are clustered;
train a machine learning model architecture using the extended label information including the previous label cluster numbers and the next label cluster numbers to establish a trained machine learning model architecture;
receive a new webpage having at least one input data field for inference; and
for the at least one input data field, process the new webpage and associated markup language using the trained machine learning model architecture to identify a predicted label, and based on the predicted label, conduct an electronic field autofill of the at least one input data field.
2 . The system of claim 1 , wherein the extracted label information includes at least one of a maxlength of the label, a previous label, or a next label.
3 . The system of claim 1 , wherein the type encoding includes at least a cluster type encoding for the label, a previous cluster label, or a next cluster label.
4 . The system of claim 1 , wherein the feature engineering approaches include processing the raw input data sets using a sentence transformer model to conduct a mapping to a dense vector.
5 . The system of claim 4 , wherein the dense vector is utilized to determine the type encoding.
6 . The system of claim 5 , wherein determining the type encoding utilizes one-hot encoding.
7 . The system of claim 6 , wherein the extended label information is represented as a concatenated data frame.
8 . The system of claim 1 , wherein the machine learning model architecture is an XGBoost classification model, and the machine learning model architecture is trained until at least a threshold accuracy level is reached.
9 . The system of claim 8 , wherein the machine learning model architecture generates logit values representing predictions of labels for each of the at least one input data field, and a highest logit value is used to determine what information is used for the electronic field autofill.
10 . The system of claim 1 , wherein the machine learning model architecture resides on a backend server and copies of the trained machine learning model architecture are deployed to reside on local memory on user devices for conducting local electronic field autofill.
11 . A computer implemented method using machine learning architecture for electronic field autofill with positional pattern recognition, the method comprising:
extracting raw input data sets from one or more webpages, the raw input data sets representing raw markup language machine instructions used in rendering the one or more webpages;
identifying one or more input fields in the raw input data sets, and for each input data field, extract label information including a label, a previous label corresponding to a label of an immediately preceding input field, and a next label corresponding to a label of an immediately succeeding input field;
extending the label information using feature engineering approaches including at least word embedding of label text using a sentence transformer model to generate dense vector representations of the label, the previous label, and the next label, and clustering approaches to associate each label with at least a type encoding, wherein the clustering approaches comprise:
applying a K-means clustering algorithm to the dense vector representations of the previous labels to generate a previous label cluster number for each input field, wherein semantically similar previous labels are clustered; and
applying the K-means clustering algorithm to the dense vector representations of the next labels to generate a next label cluster number for each input field, wherein semantically similar next labels are clustered;
training a machine learning model architecture using the extended label information to establish a trained machine learning model architecture;
receiving a new webpage having at least one input data field for inference; and
for the at least one input data field, processing the new webpage and associated markup language using the trained machine learning model architecture to identify a predicted label, and based on the predicted label, conducting an electronic field autofill of the at least one input data field.
12 . The method of claim 11 , wherein the extracted label information includes at least one of a maxlength of the label, a previous label, or a next label.
13 . The method of claim 11 , wherein the cluster type encoding includes at least a cluster type encoding for the label, a previous cluster label, or a next cluster label.
14 . The method of claim 11 , wherein the feature engineering approaches include processing the raw input data sets using a sentence transformer model to conduct a mapping to a dense vector.
15 . The method of claim 14 , wherein the dense vector is utilized to determine the type encoding.
16 . The method of claim 15 , wherein determining the type encoding utilizes one-hot encoding.
17 . The method of claim 16 , wherein the extended label information is represented as a concatenated data frame.
18 . The method of claim 11 , wherein the machine learning model architecture is an XGBoost classification model, and the machine learning model architecture is trained until at least a threshold accuracy level is reached.
19 . The method of claim 18 , wherein the machine learning model architecture generates logit values representing predictions of labels for each of the at least one input data field, and a highest logit value is used to determine what information is used for the electronic field autofill.
20 . A non-transitory computer readable medium including machine-interpretable instructions, which when executed, cause a processor to perform steps of a method for using machine learning architecture for electronic field autofill with positional pattern recognition, the method comprising:
extracting raw input data sets from one or more webpages, the raw input data sets representing raw markup language machine instructions used in rendering the one or more webpages;
identifying one or more input fields in the raw input data sets, and for each input data field, extract label information including a label, a previous label corresponding to a label of an immediately preceding input field, and a next label corresponding to a label of an immediately succeeding input field;
extending the label information using feature engineering approaches including at least word embedding of label text using a sentence transformer model to generate dense vector representations of the label, the previous label, and the next label, and clustering approaches to associate each label with at least a type encoding, wherein the clustering approaches comprise:
applying a K-means clustering algorithm to the dense vector representations of the previous labels to generate a previous label cluster number for each input field, wherein semantically similar previous labels are clustered; and
applying the K-means clustering algorithm to the dense vector representations of the next labels to generate a next label cluster number for each input field, wherein semantically similar next labels are clustered;
training a machine learning model architecture using the extended label information to establish a trained machine learning model architecture;
receiving a new webpage having at least one input data field for inference; and
for the at least one input data field, processing the new webpage and associated markup language using the trained machine learning model architecture to identify a predicted label, and based on the predicted label, conducting an electronic field autofill of the at least one input data field.