Forecasting with time series imaging
1. Model Selection: extract time series features with computer vision method, select the best forecasting method for each time series, and our selection method produces comparable result with a single best forecasting method in M4 dataset.
2. Model Averaging: design the objective function and train the best weight value for every method based on automatic feature with XGBoost, the results are comparable with top 10 results in M4 Competition.
Personalized Recommendation and Sales Forecasting Based on Offline Book Retail Data
1. Hybrid Recommendation: design a feature-based hybrid recommendation algorithm, extract multi-source features with text representation
methods(word2vec, LDA) and combine them with matrix factorization
method LFM, the results are much better than a single LFM matrix
factorization method.
2. Old Book Sales Forecasting: sales forecasting for old books with the
sliding window and GBDT to predict the sales of old books and the
average weekly sales error is less than 5 books.
3. New Book Sales Forecasting: develop a hybrid method, which is a
combination of feature-based forecasting and text clustering.
Sentiment Analysis Based on Weibo
1. Transfer Learning For Sentiment Analysis: propose a model-based framework for sentiment analysis with the pre-trained model from Weibo corpus and calculate the probability value of 4 kinds of emotion.
2. Aspect Sentiment Analysis: use a ranking algorithm based on a bipartite graph to rank the candidates of sentiment expressions and use a refining algorithm according to semantic similarity to extract some expressions from the low-rank set.
Word Prediction and Sentence Expression Standard Detection Based on Business Corpus
1. Hybrid Word Segmentation Algorithm: a combination of multiple word
segmentation algorithms according to corpus characteristics and the
accuracy rate reaches 75%.
2. Word Prediction: use RNN and N-gram model to do word prediction
and the accuracy rate reaches 70%.
3. Sentence Expression Standard Detection: word-level Convolutional
Networks for standard classification and the accuracy rate reaches 90%.