I. Project Source Code
Project source code URL: (This project is not open source at present.)
II. Introduction
The patterns of user are different from each other, but almost of them have general regular patterns, which reflect personalized features of users, that is, user profile. By calculating user’s features in different dimensions, it can do personalized recommendation for users with these data effectively.
The user profile includes long-term user profile and real-time user profile.
III. Long-term User Profile
The long-term user profile, whose data derive from user logs during at least several months, that is multidimensional user features, and it can be stored into offline databases. It includes gender, age, area, interest of commodity categories, preference of product word, preference of brand word, extended attribute, purchasing power, preference of shop, preference of discount, gender and age of their children, and so on. The user logs are processed, and some crucial information can be extracted from original SDK event tracking logs, in that way, the session logs are gotten completely. All the steps above are to upgrade data granularity of user profile.
User’s deep behavior : dwell time, browse comments.
User’s initiative feedback : search, write comments, complaints.
User’s cross-platform behavior : the accounts of different devices correspond to the same user. By uniforming behaviors of them, we can establish a vivid user profile.
IV. Real-time User Profile
The real-time user profile includes real-time feedback, current behaviors, whose granularity is in seconds level, e.g.30 seconds, 1 hour, 6 hours, 1day, 3 days, 1 week.
The flow of user’s real-time behavior can be calculated as real-time features, and Storm is used for the calculation.
V. Reference
[1]…
[2]…
[3]…