ProductLink - product explorer with recommendation

Abstract

Comparison shopping is natural for a human. Nowadays, shopping online is one of the methods for the consumer to reduce costs for searching products and its information. However, existing platforms provide limited features for comparison shopping includes the abilities to explore, search and choose health product. The project aims to create a platform namely "ProductLink" that focus on product (packaged food) exploration, provide features which are price comparison, product recommendation, image-based product search and information that help to decide a health product. Four recommenders are implemented and sixteen users were extracted from the MovieLens dataset and assume the movie is similar to product. The one-features approach achieved around 66% accuracy, which was the highest among the four recommenders. It uses the random forest for prediction, the major feature was the product name and it has been converted to number by sum the ASCII of each characters. The two-features approach has achieved around 20% of accuracy, it uses product name and product tags as features. The pre-processing of the product name was same as the one-features approach, while the product tags uses the one-hot encoding to convert to a set of binary number. The reason of a poor result compares with the one-features approach maybe the data sparsity after including the product tags. Ensemble approach had achieved around 20% of accuracy, it combine the prediction results of both the slope one and SVM by getting the average/maximum/minimum of the rating value. The reason of a poos result may be due to the difference of the algorithms and it consume much more resource for computation. Slope one approach aims to predict product that a user may interest based on other user browser history on a product. This approach uses item-based recommendation to predict the amount of a user view a product. Image-based search with crop function was implemented, it extracts the image fingerprint into a hash value, then compare their similarity with hamming distance, achieved around 60% of accuracy in the correctness test. PACE labelling was aimed to give information for a user to determine how healthy a product is, instead of only giving the nutrition facts, the amount of time to exercise to burn calories of a product was presented, the time was calculated using the calories of a product and a person detail. Two JavaScript components was implemented and were reused to create another component. The two-components output an HTML DOM object with event handler attached, it was inspired by the component-based library, Reactjs. My implementation encapsulates the output and the event handler, and it can be invoked in HTML and JavaScript. This design enhances maintainability and reusability. Promise was used to overcome the drawbacks from the callback approach when handling process after asynchronous calls. Using callback may appear into nested callbacks and this reduces code readability, harder to debug and problematic when there are multiple asynchronous calls. Promise can be chained together in a linear way and ability to wait for multiple asynchronous calls. As a result, a responsive web app was successfully implemented includes product recommenders, image-based search, PACE label and Web API.