Multi-Objective Recommender System 2022

Build a recommender system based on real-world e-commerce sessions

Exploratory Data Analysis:

• We have divided dataset from three parts, including "clicks", "add2cart", and "orders" and we observe that "clicks" happen more often than the others.
• We figure out the correlation among three types.

Experiment:

• Intuitive method
• Original: The first idea is to recommend "In-cart items" after "orders", "Clicked items" after "carts".
• Sort by time length: Add a new column "minutes" to Train-data. Sorting by descending "minutes".
• Recommend Clicked items: Recommend click items from "best-sold item" to "clicked item".
• Recommend Cart-items: Recommending clicked-items instead of "pair-sold items" adding to cart-in items.
• Word2vec
• We use the gensim library to apply the Word2vec model, which offers extremely fast training on the CPU, and using the AnnoyIndex library of Spotify to compute distance(similarity) and find the top 20 nearest neighbors as candidates.
• Word2vec+LightGBM
• To improve the performance, we try to build a system with a learning to rank model to rerank candidates. We chose LightGBM as rerank model. 
• Co-Visitation method
• Most recent previously visited items.
• Items previously visited multiple times.
• Co-visitation.
• Type Weight – Assign type weights to action pairs click/add2cart/order which is shorter than one day. Choose top candidates for each session.
• Frequency Weight – Assign frequency weights to add2cart/order pairs which is shorter than 7 days. Choose top candidates for each session.
• Interval Weight - Co-visitation matrix of click to click by time weighting (for clicks). Choose top candidates for each session.
• Current popular items.
• Combine these four rules choose top twenty to the submission.

Recommendations for Future Work:

• Voting ensemble.
• Ensemble Learning: we design our algorithm to improve our model.

Conclusion:

Firstly, we success to finish data visualizations for data analysis. And then, we successfully use the vector space model named word2vec to do word embedding and retrieve information, and the score about 0.52 to 0.53. In order to improve our performance, we add the deep learning techniques named LGBM to implement our experiments. On the other hand, we also try the candidate Rerank model, and the score is 0.575. We combine the LGBM and candidate Rerank, and the score is

increase 0.577. Last but not least, we use the voting ensemble to up 0.578. Besides, we propose the two methods on recommendations for future work due to hoping increase score.