The invention discloses a feature engineering recommendation method based on spark streaming real-time streams. The method comprises the steps that expanded logs and clicked logs of a client are acquired, washed and then added into a distributed message queue; spark streaming is used for subscribing log streams of the expanded logs and the clicked logs, and stream data in the two log streams are combined in engineering; operation is performed on the stream data to generate labels to identify expanded and clicked stream data and expanded and non-clicked stream data; multidimensional features are constructed for the expanded logs and the clicked logs according to basic features, and meanwhile basic time features are combined; the stream data with new features is subjected to off-line training and on-line training to generate recommended stream data. A feature extraction mode widely applicable to most fields is provided, the problem of a small application range of feature engineering is solved, timeliness of feature engineering is realized by adopting the mode of on-line domination and off-line correction, and the effectiveness and accuracy of the features are achieved through a series of feature combination transformation.