Mobile Market Recommend System Research And Development

Most of the people download the app、electronic-book、cartoon through the mobile store. With the fast growth of the app、electronic-book、cartoon types, how to help user to find his interesting item is one of the most important function of mobile store. When the user know what he needs exactly, he may find them easily by search. When the user can not sure what he need or difficulty to describe it with key words, recommendation can help them in some way.Our work is to make a related recommendation system for the mobile store managed by a communication operator, it includes app、electronic-book、cartoon. When the user views the item’s detail, it will recommend four related items.According to the user interaction data and item meta data, we use different recommend algorithm for different types item. We use logic regression to combine content-based and item collaborative filtering, the training data is from the online user behavior. Before combining the content-based and collaborative filtering, we use the user viewing app’s type to filter recommend result those are not the same type to increase the function similar between the recommend results and user viewing app. For those type of items that have similar function, we add the app value model to adjust the recommendation result. The app value model can be used when the viewing app and recommend app has similar function, it can be learn from user behavior. The click and convert rate is higher than pure item collaborative filtering with 30% and 20%. For e-book and cartoon, because of a few user behavior, we use content-based algorithm. The Feature we used in content-based algorithm is TF-IDF.The data of mobile store including user behavior and items, we use the Hadoop to build out recommendation system. Some recommendation algorithm implementation must be optimized in Hadoop with high efficent. When implementing the content-based algorithm in Hadoop, we cut the hot data and recombine them, we also decrease the redundancy of data. By cutting and recombining the hot data, the data can be distributed in cluster really. Decreasing redundancy in Hadoop means decrease disk and network IO. After this optimization, the run time of e-book content-based similar decrease from 15 hours into 6 hours.The recommended results are stored in Hbase, we also provide recommendation web service to be called by mobile client to get the recommendation result. The mobile store recommendation had been on line in Nov 2014. The number of page view is more than 1 million one day. It has running several month. The click and convert rate can make the communication operator satisfied.