Deep Learning In Internet Finance Practice

In this era of rapid development of the Internet and big data,the technology of artificial intelligence has played an increasingly important role in the decision-making of intelligent mar-keting,helping companies to improve user experience and enhance corporate competitiveness.Two aspects of decision-making for the intelligent marketing are crucial.The one is to choose the right time,and the other one is to choose the right marketing plan.This paper focuses on these two important issues in the marketing scenario.For the problem of choosing the right time for marketing,a direct method is to judge customers' demand.In order to determine the customers' demand,we need to predict what the customer might do in the recent future.In this paper,we develop a novel deep learning model to solve the problem of personalized sequential prediction.Our proposed model has an encoder-to-decoder structure to reasonably preserve the inferred relationship between the personalized features and the historical behavioral sequences.The model could make personalized sequential predictions in a more reasonable way and the structure could enhance the flexibility and interpretability of the model.We also modify the forget gate structure of the long-short term memory(LSTM)and propose a variant of the long-short-term memory,called the attentioned long-short-term memory,to expand the attention of the structure.This attentioned long-short-term memory converges quicker and better while still maintain the similar performance in open dataset IMDB.In addition,in dealing with person-alized prediction problems in real-world datasets provided by our cooperative company,our proposed model with the attentioned long-short-term memory achieves an accuracy about 10%higher than the model with the vanilla long-short-term memory.The personalized sequential prediction model with an encoder-to-decoder structure proposed in this paper has also achieved good results in the online experiments in the real-world business scenario.Besides,because deep reinforcement learning is suitable for modeling the real world environment,we model the personalized decision problem in intelligent marketing with the reinforcement learning.In general,agents of the reinforcement learning make a fresh start in each task,ignoring related prior knowledge on other tasks,which results in a large amount of repetitive work on similar scenes.However,the amount of data for a task is very limited,and the information that can be obtained is limited.There may be also some new products appearing in each new task.There is usually an item-based cold start problem in intelligent marketing scenarios.It could be helpful for these problems to share prior knowledge across different tasks.Meta-learning is a key mechanism to share knowledge across different tasks to solve specific tasks with only few samples or zero sample.Therefore,we need to integrate the meta-learning method and make full use of the prior knowledge to make the model maintain a stable performance in the new task,even though the new task lacks training data and has some products that never appear.Here,in order to solve the problem of zero-shot learning and item-based cold start in marketing scenarios,we propose a personalized recommendation decision model based on deep reinforcement learning and meta-learning.It can be seen from the experiment results that the model can maintain a stable performance in the absence of training data.At the same time,it can quickly adapt to changes of the action space for different marketing activities.The model has improved by 14.6%compared to the comparison model in the real-world business scenario and it can achieve a good result.In addition,non-stationary environments has a great influence on the model's training and working.It is pointed out here that the data preprocessing method with time-division normalization can effectively and the method can provide a relatively stationary environment for the model.It can be seen from the experimental results that this method can make the environment relatively stable,so that the performance of the model is guaranteed to a certain extent.Therefore,equipped with our proposed methods,decision-making in marketing scenarios can become more convenient and more effective.