Combined wide and deep machine learning models for automated database element processing systems, methods and apparatuses
A method of automated database element processing includes training a wide machine learning model with historical feature vector inputs to generate a wide ranked element output. The method includes training a deep machine learning model with the historical feature vector inputs to generate a deep ranked element output. The method includes generating a set of inputs specific to an individual entity, obtaining a set of current article database elements, and creating a feature vector input according to the set of inputs and the set of current article database elements. The method includes processing the feature vector input with the wide machine learning model to generate a wide ranked element list, processing the feature vector input with the deep machine learning model to generate a deep ranked element list, and merging database elements of the wide and deep ranked element lists to generate a ranked element recommendation output.
1 . A computer-implemented method of dynamic machine learning model retraining in a content recommendation system, the method comprising:
receiving, by at least one processor, a plurality of article elements published by at least one content provider over a time period;
determining, by the at least one processor, an article publication count associated with the plurality of article elements published by the at least one content provider during the time period;
comparing, by the at least one processor, the article publication count to a training criteria;
in response to determining that the article publication count satisfies the training criteria based on the comparison, automatically adjusting a model retraining frequency for at least one machine learning model used for article recommendations;
obtaining, by the at least one processor, updated article elements published since a last model training;
obtaining, by the at least one processor, updated user interaction data associated with the updated article elements;
retraining, by the at least one processor, the at least one machine learning model using the updated article elements and the updated user interaction data;
generating, by the at least one processor and using the retrained machine learning model, personalized article recommendations for at least one user based on user-specific input data; and
providing the personalized article recommendations to a user device for display.
2 . The method of claim 1 , wherein the time period comprises a sliding time window.
3 . The method of claim 2 , wherein the sliding time window has a duration of at least one week.
4 . The method of claim 1 , wherein the article publication count is calculated based on a number of article elements published in the time period.
5 . The method of claim 1 , wherein the training criteria comprises an elapsed time threshold.
6 . The method of claim 5 , wherein automatically adjusting the model retraining frequency comprises increasing the model retraining frequency in response to the elapsed time threshold being exceeded.
7 . The method of claim 1 , further comprising obtaining a new click history since the last model training.
8 . The method of claim 7 , further comprising retraining the at least one machine learning model using the new click history.
9 . The method of claim 1 , wherein the at least one machine learning model comprises a wide machine learning model and a deep machine learning model.
10 . The method of claim 9 , wherein retraining comprises jointly retraining both the wide machine learning model and the deep machine learning model.
11 . The method of claim 9 , wherein the wide machine learning model comprises a term frequency inverse document frequency (TFIDF) model or a best matching (BM25) model.
12 . The method of claim 9 , wherein the deep machine learning model comprises a Doc2Vec model or a deep contrastive learning (DECLUTR) model.
13 . The method of claim 1 , wherein the at least one machine learning model comprises a supervised trained model.
14 . The method of claim 1 , wherein the at least one machine learning model comprises an unsupervised model.
15 . The method of claim 1 , wherein the updated article elements comprise articles published within at least the last six months.
16 . The method of claim 1 , wherein the updated user interaction data comprises user click history from at least the last forty-five days.
17 . The method of claim 1 , wherein the article elements comprise news articles having temporal relevance windows of a week.
18 . The method of claim 1 , wherein determining the article publication count comprises analyzing publication timestamps of the plurality of article elements.
19 . The method of claim 1 , wherein the model retraining frequency is adjusted to a frequency based on a time scale selected from the group consisting of: an hour, a five minute period, a day, a week, and a month.
20 . The method of claim 1 , wherein the method is implemented using a scalable serverless compute engine.