RecSys’11: Utilizing related products for post-purchase recommendation in e-commerce
Recommendation system paper challenge (8/50)
Why this paper?
RecSys’11 Best Short Paper.
What problem do they solve?
Designing a top-n recommender system for the post-purchase stage
What are the challenges?
- relevance of recommendation
- recommendation coverage
- time sensitivity
What others solve this problem?
Traditional approaches: Content-Based and Collaborative Filtering
Content-Based: It is good for col-start problem
Collaborative Filtering: there are memory-based and model-based.
How do they tackle the temporal problem?
They apply time-decay function to adjust the similarity of the items.
However they usually do similar product recommendation becomes less meaningful for the user because they already purchased.
What is their model?
Category-level Relevance Factor
It is MLE (maximum likelihood estimation) for the likelihood of user purchase category k given user purchased category a. For computing MLE, it calculated all over user and historical data to get the estimation.
Product-level Relevance Factor
In order to alleviate the cold-start problem, they consider content-based information.
Product-level Behavioral Relevance
It indicates that pk was likely to be purchased after pa in history
They only consider those records >threshold_user, which controls the balance between reliability and accuracy.
If more users purchase pk after they purchase pa, S(pk, pa) will be higher.
r_ik is the ( 1 / p_ik), if the user purchase 5 times for p_k, but only one time from p_i, so r_ik will be (1/5) = 0.2
Product-level Content Relevance
In this case, even if users never purchased product pk after the active product pa, pk can still be recommended if its content is related to pa’s content. They extract content value from the product title.
Temporal Factor
Here’s example:
The system should recommend camera and then camera lens and then filter.
But if it recommend filter directly after camera, it is very bad.
It suggests that the timing is equally important for recommendation as relevance.
They utilize 3 time windows:
- within 1 day
- from 1 day to 1 week
- from 1 week to 1 month
Category Level Behavioral Relevance with t_w
Define k_a,k,m,t as the value of k_a,k,m in the time window t
Product Level Behavioral Relevance with t_w
Estimating the Recommendation Score (combining all of them)
What is the Dataset? eBay
A sample of 3 million unique eBay users from Jan 1, 2010 to June 30, 2010. These users purchased around 70,000 unique products in 200 product categories.
What is the metrics?
Recommendation Coverage
Purchase Rate:
Baseline mode?
item-based collaborative filtering recommender system
The result
Other related blogs:
Trust-aware recommender systems
Performance of recommender algorithms on top-n recommendation tasks
A Matrix Factorization Technique with Trust Propagation for Recommendation in Social Networks
Hidden Factors and Hidden Topics: Understanding Rating Dimensions with Review Text
Beyond Clicks: Dwell Time for Personalization
RecSys’15: Context-Aware Event Recommendation in Event-based Social Networks
Best paper in RecSys:
https://recsys.acm.org/best-papers/
My Website: