- Parth Srivastava, UG Sophomore, IIT Kanpur
- Gurbaaz Singh Nandra, UG Sophomore, IIT Kanpur
- Shubh Agarwal, UG Sophomore, IIT Kanpur
The persuit of personalization has long been carried out in shopping websites. From personal dashboards to recommender systems, we've come a long way to give the user more personalized experience. In this hackathon, we've developed the following functionalities and integrated them into our demo website.
- Personalized Discount Coupon Generator
- Personalized Product Recommendation using Multi-Armed Contextual Bandits
The Personalized Discount Coupon Generator gives unique discounts to the users, based on the user's activities like time duration, page views, etc. The Product Recommndation System using Multi-Armed Contextual Bandits is a unique approach to recommedation. As opposed to Collaborative Filtering, this approach improves over time, and hence provides more utility as compared to the older approach. Furthur details regarding the 2 functionalities can be found below.
- Bandits : Contains the model, encoders and the training code for the Multi-Armed Contextual Bandits.
- Classifier : Contains the data, model and the encoders for the Personalized Discount Coupons.
- Frontend : Contains the HTML, JS and CSS files for the frontend of our project.
- Heroku-APIs : Contains the code of the APIs for both the functionalitites, and the deployment code for Heroku.
We deployed our project using Firebase Hosting. The deployed verion of the website is available HERE. The steps to test the functionalities are -
- The Discount Coupon Value is automatically shown on entering the site, based on the User type. One can easily view the value of the dummy variables generated by the code in the console.
- The recommended books can be viewed in the Recommended Books section. Clicking on the book will trigger the learn function of the model, and hence generate new recommendations.
NOTE : On visiting the website for the first time, the functionalities may not respond immediately. This is because the dynos in Heroku take time to start, and therefore refreshing the page once or twice may be required.
We used an open source dataset for the training of the model.The data can be found here The dataset contained 24 unique features like avg_session_duration, hits,etc. and contained 3 labels - 'buyer','window' and 'hesitant'. The description of the labels is given below :
- Buyer - The customers who often purchase products from the website fall in this category.
- Window - This category includes customers who need a little push over the edge to buy the products.
- Hesitant - This category includes customers who seldom buy products from our website, and therefore need a large incentive to get them over the edge.
We used an lgbm classifier for the problem. We used a total of 20 features for our model, with a 80:20 train/dev split. The model gave an accuracy of 99.345% on the dev set.
We deployed the model using Flask API and Heroku. For every user that opens our website, we use javascript to generate dummy values for the features that we use in our model ( In actual use case, this can simply be done by pulling the values of the user from the database). We then post those values(after encoding the non-numerical values using the respective encoders) to the /predict endpoint of our model, and fetch the classification result from our model. Based on the result, we provide the user with a personalized discount coupon, a coupon which can help us (a shopping firm) maximize our utility.
We used an open source dataset from Kaggle for this task. The dataset can be found here. One can clearly see that the dataset has many outliers, null values, etc, and hence required some good amount of pre-processing. Also, in order to train the bandits on the dataset, we had to combine the 3 files into one. Our pre-processing code can be found in pre-processing.ipynb. The dataset obtained after the final pre-processing can be found here
The finally prepared dataset contains fields like location, country, age, book title, etc, with a total of 383,842 data points.
We used Vowpal Wabbit to train our Multi-Armed Contextual Bandits. It is simple to use and easily integrable with python. We use cb_explore algorithm, with an epsilon value of 0.2, so as to focus on both, exploration as well as exploitation. Since our final dataset had a total of 135567 unique books, therefore, our final algorithm for training was - cb_explore 135567 --epsilon 0.2. We used 10 - book_rating as the cost of choosing the particular datapoint, as cost is the opposite of reward. We used book_rating/10.0 as the probability of choosing the datapoint. We use the clicks as the reward function in the model.
We again used Flask API and Heroku for deploying our model. For every user that visits our website, we generate dummy features like location,age, etc, and push them to the /predict endpoint of our API, which in turn provides us with the top 6 recommendations for the user. These 6 recommendations are then displayed on our homepage in the recommendations section. Once a user selects a particular option, the values previously generated, along with the book_name, are pushed to the /learn endpoint of the API, which learns from the pushed datapoint, and again generates 6 new recommendations after learning. As we know, from the basics of Reinforcement Learning, that by learning, the model tries to maximise its utility by minimising the error, and hence generates perfect recommendations after some learning iterations.