TEXTUAL DATA AUGMENTATION FOR NEURAL NETWORKS

The cold-start scenario is a critical problem for recommendation systems, especially in dynamically changing domains such as online news services. In this research, we aim at addressing the cold-start situation by adapting an unsupervised neural User2Vec method to represent new users and articles in a multidimensional space. Toward this goal, we propose an extension of the Doc2Vec model that is capable of representing users with unknown history by building embeddings of their metadata labels along with item representations. We evaluate our proposed approach with respect to different parameter configurations on three real-world recommendation datasets with different characteristics. Our results show that this approach may be applied as an efficient alternative to the factorization machine-based method when the user and item metadata are used and hence can be applied in the cold-start scenario for both new users and new items. Additionally, as our solution represents the user and item labels in the same vector space, we can analyze the spatial relations among these labels to reveal latent interest features of the audience groups as well as possible data biases and disparities.

Complementing information about particular points, places, or institutions, i.e., so-called Points of Interest (POIs) can be achieved by matching data from the growing number of geospatial databases; these include Foursquare, OpenStreetMap, Yelp, and Facebook Places. Doing this potentially allows for the acquisition of more accurate and more complete information about POIs than would be possible by merely extracting the information from each of the systems alone. Problem: The task of Points of Interest matching, and the development of an algorithm to perform this automatically, are quite challenging problems due to the prevalence of different data structures, data incompleteness, conflicting information, naming differences, data inaccuracy, and cultural and language differences; in short, the difficulties experienced in the process of obtaining (complementary) information about the POI from different sources are due, in part, to the lack of standardization among Points of Interest descriptions; a further difficulty stems from the vast and rapidly growing amount of data to be assessed on each occasion. Research design and contributions: To propose an efficient algorithm for automatic Points of Interest matching, we: (1) analyzed available data sources—their structures, models, attributes, number of objects, the quality of data (number of missing attributes), etc.—and defined a unified POI model; (2) prepared a fairly large experimental dataset consisting of 50,000 matching and 50,000 non-matching points, taken from different geographical, cultural, and language areas; (3) comprehensively reviewed metrics that can be used for assessing the similarity between Points of Interest; (4) proposed and verified different strategies for dealing with missing or incomplete attributes; (5) reviewed and analyzed six different classifiers for Points of Interest matching, conducting experiments and follow-up comparisons to determine the most effective combination of similarity metric, strategy for dealing with missing data, and POIs matching classifier; and (6) presented an algorithm for automatic Points of Interest matching, detailing its accuracy and carrying out a complexity analysis. Results and conclusions: The main results of the research are: (1) comprehensive experimental verification and numerical comparisons of the crucial Points of Interest matching components (similarity metrics, approaches for dealing with missing data, and classifiers), indicating that the best Points of Interest matching classifier is a combination of random forest algorithm coupled with marking of missing data and mixing different similarity metrics for different POI attributes; and (2) an efficient greedy algorithm for automatic POI matching. At a cost of just 3.5% in terms of accuracy, it allows for reducing POI matching time complexity by two orders of magnitude in comparison to the exact algorithm.