Skip to content

Personalized Recommendation Engine: Guided by Behavior

Notifications You must be signed in to change notification settings

rahul99/PRE-Guide

Repository files navigation

PRE-Guide

Personalized Recommendation Engine: Guided by Behavior

Summary

Recommendation Engine application guided by behavior. The behavior of the person is inferred using his/her temporal activites in the Social Media platforms such as Facebook/Twitter/linkedIn. In the inference phase, the social media activity of the user is recorded and a trained topic model predicts the topic that a user has been talking about or is showing interest in. This behavior can be used as a guide to recommend products to the user. Within the scope of this project, we show a prototype to demonstrate book recommendation based on Social behavior. In the offline phase, we train topic-models to predict the behavior of the person and export these models for evaluation. The models can re-trained and updated periodically.

Backend

[1] We work on Twitter 20 Newsgroups Dataset to do topic-modelling. Specifically, Given several tweets of a user in a particular time window, we predict the probability of a topic in each tweet and pick the top-k categories on a ranked list.

[2] We work on Goodbooks-10k Dataset to find out find out the ranked list of books in each category.

[3] A mapping is required to create between categories in [1] and [2].

[4] Optional: Optionally, we use Book Dataset, that contains a huge collection of book names in each category.

Frontend

[1] A web server will be setup, that takes in input as the username of the person.

[2] We query this username in the Twitter Scrapper and crawl the last K tweets of this person as a list of JSON objects.

[3] This list is sent to backend where recommendation computations happen as follows:

  • [a] Collect the JSON list, parse it and create a sequence of feature vectors, one for each json

  • [b] Predict a topic for each feature vector. Add the probabilities and create a cumulative list of top-k categories.

  • [c] Query a top recommended book in each category and create a list of this category.

[4] The client API collects this list and displays in the browser in the order.

Releases

No releases published

Packages

No packages published