Ranking prediction for time-series data using learning to rank (Case Study: Top mobile games prediction)

Learning to rank as one of machine learning technique becomes emerging topic for ranking problems. However, this technique has not been well applied in time-series data. This paper proposes ranking model for time-series data using learning to rank which uses top mobile games as our case study to develop top mobile games prediction system. We employ learning to rank algorithm using mobile games and time-series attributes such as complete date, date index per year, and day index and year in our prediction system, MGPrediction+. Experiment has been performed to identify the best learning model for top mobile games prediction. Result shows that LambdaMART has the best NDCG@100 at 0.6985 using original dataset. Our best feature set contains original attributes with additions of day index and year which give NDCG@100 score at 0.7375. To sum up, learning to rank for time-series data can be performed better by using time attribute, especially for top mobile games prediction.