Graded multi-label classification (GMLC) is the task of assigning to each data a set of relevant labels with corresponding membership grades. This paper is interested in GMLC for large and evolving datasets where data are collected from a possibly infinite stream. Many commercial and non-commercial websites acquire such data by giving users the opportunity to rank items any time using an ordinal scale like one-to-five star rating. Typically these collected data are sparse because users rank only a small subset of items. Websites rely on recommender systems to dynamically adapt the recommended item set for each user. Hence, the applied recommender system should remain scalable and efficient when dealing with sparse data. State-of-the-art methods related to GMLC were tested only in batch mode. Their performance in an incremental mode is not investigated, especially in presence of sparse data and concept drifts. This paper presents our proposed incremental GMLC method which answers the above challenges and can be applied to build a recommender system. This method is tested on the well-known MovieLens and Jester datasets, and it is able to adapt to concept drifts and maintain the Hamming loss at a low level.