Research On Latent Fctor Model And Its Optimization Algorithms Of Machine Learning

With gradually increasing of various data sources and data collection methods,machine learning and data mining technology have more abundant applicationscenarios. The nature of machine learning is to address the representation and modelwhich are two major problems of intelligent learning. The former requires thecomputer to process the data for feature representation, this representation hasgeneralization and abstraction storage characteristics; the latter need to learn unifiedmodel to represent source data for achieving intelligent understanding, prediction,reasoning and other functions. Traditional supervised learning models are facing thescalability and performance bottlenecks of artificial feature engineering, whiletraditional unsupervised learning confronting the inadequate representation ofhidden structure and pattern in data and difficulties in evaluation of results. Thispaper studies the latent factor model which is an effective method in machinelearning to solve the representation and model problems. The nature of latent factormodel is automatically learn the high-level abstract representation of input data toimprove the performance of predictive models and hidden structure representationof data.First, for the dyadic data and general data types in traditional supervisedlearning tasks, we study the factorization models which use the latent factor vectorsto model the interaction between input features to solve the dependency of artificialfeature engineering and data sparsity problems. By setting the prediction accuracyas the objective function of supervised learning model, this method can blend thehigh-order interactions between features as new features to improve the accuracy ofpredictive models. Although this supervised factorization model can achievehigh-level feature extension learning, it still needs some help and optimization ofartificial feature engineering. Since single layer model such as factor vector orassumption of linear model has limited capabilities of data representation, We studythe hierarchical latent factor model which emphasizes the importance of automaticfeature learning to achive more expressive power of raw data. By utilizing thetraditional structure of neural network, this model can achieve bottom-uprepresentation learning and step by step feature learning of the data. For supervisedhierarchical model, by optimizing the structure of the network and sharing theconnection weights, multilayer neural network can be trained effectively to learnlocal features of data efficiently. To solve the limitation of available annotation datain supervised learning models, we can train the hierarchical factor model to realizeunsupervised self-learning to get the structural parameters based on lots of available unlabeled data. These structural parameters got by layer-wise training can be used toimprove the abstract representation of labeled data, so it can improve the predictionaccuracy and generalization of supervised learning models.