Automated industry classification with deep learning
An automated predictive analytics system disclosed herein provides a novel technique for industry classification. Leveraging specific API to construct a database of companies labeled with the industries to which they belong, the automated predictive analytics system trains a deep neural network to predict the industries of novel companies. The automated predictive analytics system examines the capacity of the model to predict six-digit NAICS codes, as well as the ability of the model architecture to adapt to other industry segmentation schemas. Additionally, the automated predictive analytics system investigates the ability of the model to generalize despite the presence of noise in the labels in the training set. Finally, the automated predictive analytics system explores the possibility of increasing predictive precision by thresholding based on the confidence scores that the model outputs along with its predictions. The automated predictive analytics system finds that the approach yields six-digit NAICS code predictions that surpass the precision of gold-standard databases.
1. A method of generating industry classifications, the method comprising:
constructing a training dataset of using a database, wherein the database includes an index tying a plurality of companies with a set of attributes;
representing the plurality of companies using sparse feature vectors;
training a classifier on mini batches of data from the dataset;
extracting relevant keywords or keyword phrases for one or more of the plurality of companies from one or more of webpages, titles, meta descriptions, meta keywords, hyperlinks, and visible text;
determining a vocabulary of keywords and keyword phrases relevant in describing the services and product offerings of one or more of the plurality of companies;
generating categories for sparse features of the dataset;
generating labels for the plurality of companies in the dataset; and
training a neural-network model using the labeled dataset.
2. The method of claim 1 , further comprising performing natural language processing on the one or more of the webpages, titles, meta descriptions, meta keywords, hyperlinks, and visible text to extract relevant keywords.
3. The method of claim 1 , further comprising web crawling internet data to acquire the one or more of the webpages, titles, meta descriptions, meta keywords, hyperlinks, and visible text.
4. The method of claim 1 , further comprising determining a frequency at which each keyword or keyword phrase appears within the one of more of the webpages, titles, meta descriptions, meta keywords, hyperlinks, and visible text.
5. The method of claim 4 , further comprising performing a bag-of-words analysis on the keyword or keyword phrases to determine the frequency of each keyword or keyword phrase.
6. The method of claim 1 , wherein the training dataset includes a weight for each class based on a ratio of a total number of training examples to a number of training examples for that class.
7. The method of claim 6 , wherein the weight causes the training examples of one class to be up-weighted.
8. The method of claim 6 , wherein the weight causes the training examples of one to be down-weighted.
9. The method of claim 6 , wherein the weight is also based on verified and unverified training examples.
10. The method of claim 9 , further comprising up-weighting the verified training examples.
11. The method of claim 10 , further comprising down-weighting the unverified training examples.
12. The method of claim 1 , further comprising updating parameters of the neural-network model with an Adam optimizer with Nesterov momentum.
13. The method of claim 1 , further comprising:
inputting a second dataset associated with a company into the neural-network model; and
causing the neural-network model to output a classification code associated with the company based on the second dataset.
14. The method of claim 13 , wherein the neural-network model generates a confidence score for the classification code based on the second dataset.
15. The method of claim 14 , further comprising thresholding the confidence score to remove the classification code when the confidence score is below a given value.
16. The method of claim 1 , further comprising generating a term frequency inverse document frequency (tf-idf) value, the tf-idf value being a ratio of a frequency with which the keywords and keyword phrases occur in a document or collected information about the one or more companies to a frequency with which the keywords and keyword phrases occur in a plurality of documents or collected information about the one or more companies.
17. The method of claim 16 , wherein one of the sparse features is the tf-idf value calculated from a pooled set of texts across all sources for the one or more companies.
18. The method of claim 16 , wherein one of the sparse features is the tf-idf value calculated from homepage title, meta description, meta keywords, or combinations thereof associated with the one or more companies.
19. The method of claim 16 , wherein one of the sparse features is the tf-idf value calculated from a homepage hyperlink associated with the one or more companies.