Study On User Interest Drift Mining Algorithm Based On Association Analysis

In the advent of "big data", people have been submerged in the mess of data and can't find the real demanding information. Recommender system mines people's potential needs from massive consumption history become an effective way to alleviate "data overload”.But the traditional recommender system has three common problems: the first one is the data sparsity, users in a real website always has less consumption history, so it is inaccurate to predict user's real interest; The second one is the difficulty to catch the user's interest drift with the context transforming, the traditional recommender system assume the user's interest is fixed so can not capture user's interest drift over time; The third one is over fitting,recommender system tends to recommend similar products to users which make users feel boring. This article aim to solve the above three problems in recommendation system and made the following research.First, we build the data imputation model to solve the data sparsity problem. We mine user association group from the whole comsumption history, then define user association group occupation and correlation to filter the groups which has weak correlation. After the consumption history is divided with fixed time window, we fill the sessions which has rare items or even no with the items consumed by raleted user association groups.Second, we build interest drift model to solve the problem of difficulty to catch the user's interest drift. We use a Latent Dirichlet Allocation(LDA) to model user's interest in each session. We get user's interest distribution in each session then to predict user's interest next time with a loss function to model user's interest drift over time.Third, we build the interest extend model to solve the problem of over fitting. We build a graph with time sequence and get the most related users by the random walk algorithm, and then get the relation between items in the same way. We integrate the result of traditional user based collaborative filtering and the relation between items to alleviate the problem of over fitting in recommender system.Experiments show that: the algorithm we proposed can alleviate the above three problems and recommend to users the information and goods they really need.