Sentiment-Based Deep Auto Encoder Recommender System (DAERS)

Sentiment analysis, an eminent domain of data mining, refers to elucidation and classification of sentimental contents exhibited on social media platform. Twitter is a popular micro blogging platform where many user tweet to exhibit their thoughts regarding different domains. Immense research has been carried out in the field of sentiment analysis varying from lexicon-based approach to intricate deep learning models. Information from user tweet-feeds can be used to make cross domain recommendations. User generated content (reviews and comments) crawled from web requires analysis to discover the latent thematic structure in the corpora of both source and target domains, with no user-item overlapping between the given domains. Our proposed approach incorporates the power of deep learning to exploit user sentiments that helps to infer ratings based on user reviews and thereby resolving data sparsity issue in target domain. Also, to address Cold Start Recommendations (CSR) for new users we put forward a low-rank Deep Auto-Encoder (DAE), trained back-to-back with no prior training of individual layers. Encoder in DAE maps the behaviour space to attributes space and Decoder reconstructs the behaviour space from attribute space. Three cruxes challenged in this paper include; domain-shift, CSR and spurious correlations. We hypothetically exhibit that: a) DAE models generalize exceptionally than the linear models, and b) Use of normalization technique such as Dropout at the representation layer is necessary to prevent over fitting.