This is the code for a hackathon problem given by GreyAtom
Hackathon Problem Statement:
Predict the Segment - Hotstar
- Determining the demographics of customers is one of the most key tasks in the advertising domain. Advertisers usually want to target customers based on demographic attributes. However, it is difficult to get demographic data from all the customers since that can add friction to the user experience.
- At Hotstar, we have detailed information on all the content that customers watch, let’s call it “watch patterns” and we’d like to use this signal to fine tune demo-targeting.
- We are seeking a machine learning based solution using which we can learn patterns from customers whose watch patterns are already known. In this competition, the task is to generate predictive models that can best capture the behavior. Participants are free to use any open source external data. Data Set Information : A zipped file containing train, test and sample submission files are given. The training dataset consists of data corresponding to 200,000 customers and the test dataset consists of 100,000 customers. Both training and test data is in the form of json dict, where key is masked user ID and value is an aggregation of all records corresponding to the user as described below. Description ID: a unique identifier titles: titles of the shows watched by the user and watch_time on different titles in the format “title:watch_time” separated by comma, e.g. “JOLLY LLB:23, Ishqbaaz:40”. watch_time is in seconds genres: same format as titles cities: same format as titles tod: total watch time of the user spreaded across different time of days (24 hours format) in the format “time_of_day:watch_time” separated by comma, e.g. “1:454, “17”:5444”. dow: total watch time of the user spreaded across different days of week (7 days format) in the format “day_of_week:watch_time” separated by comma, e.g. “1:454, “6”:5444” segment target variable. consider them as interest segments. For modeling, encode pos = 1, neg = 0
My approach: Convert all cities, genres, days of week and time of day into individual variables with the seconds as the value- so all variables have the same unit of comparison - seconds. Once this is done, we can apply the ML models to derive the final solution.