基于时间动态性的场感知分解机模型

Field-aware factorization machine model based on time dynamics

场感知分解机模型FFM能够有效解决高维数据特征组合的稀疏问题且具有较高的预测准确度和计算效率,广泛应用于推荐系统领域.FFM在建模时没有考虑时间动态性因素,而真实场景中部分特征值会随着时间发生变化,并在不同时间段对预测影响程度不同.鉴于此,提出一种基于时间动态性的场感知分解机模型tFFM.该模型考虑两类时间动态性:偏置动态性和特征动态性.前者从用户行为和物品流行趋势变化角度分别进行动态建模,并基于时间窗口技术设置不同粒度的时间因子;后者将特征细分为随时间变化的动态特征和保持稳定的静态特征,采用ReLU激活函数建立时间函数.采用统一特征编码方式,并设计一种样本数据生成和存取策略,能够大幅降低模型的训练和预测时间复杂度.利用随机优化方法Adam对目标进行优化,实验结果表明,tFFM比目前广泛应用的FM和FFM相关方法具有更高的预测准确度.

The field-aware factorization machine model(FFM)is widely used in the field of recommender systems since it can effectively solve the sparse problem of high dimensional feature combination with high prediction accuracy and computation efficiency.However,the FFM does not consider time dynamics in the modeling phase.In real scene,some feature values change with time and they will have different effects on prediction at different time.A field-aware factorization machine model based on the time dynamics FFM(tFFM)is proposed.The model takes into account two kinds of time dynamics,bias dynamics and feature dynamics.The former is modelled dynamically from user behavior changes and popularity of items respectively,and different time granularities based on time-window technique are used.The latter subdivides features into static features that maintain stability and dynamic features that change with time,and the time function is established by using the ReLU activation function.The unified feature encoding method is used and a sample data representation and access strategy is designed to greatly reduce the time complexity of model training and prediction.The stochastic optimization method Adam is used to optimize the target.The experimental results show that the tFFM can obtain higher prediction accuracy compared with the state-of-the-art methods related to factorization machines such as the FM and the FFM.