----------------------- REVIEW 1 ---------------------
SUBMISSION: 118
TITLE: Jointly modeling products and resource pages for task-oriented recommendation
AUTHORS: Brendan Duncan, Surya Kallumadi, Taylor Berg-Kirkpatrick and Julian McAuley

----------- Overall evaluation -----------
SCORE: 0 (borderline paper)
----- TEXT:
Main Contributions:
This paper investigates a novel approach to obtaining user intent signals by leveraging resource pages associated with particular home improvement tasks. The work demonstrates how this can be incorporated into a recommender system to improve product recommendations, as well as successfully recommend DIY project pages.


Strengths:
- The problem is well motivated with real-world business impact of the work.
- Experiments well conducted also comparing the SASRec-based approach with multi-task training approach with insights on performance comparison.
- Experimental results compared with other approaches to show the benefit of SASRec with resource page interaction data.


Limitations:
- The approach is mainly based on SASRec with modification of leveraging resource pages interactions in addition to user-product interactions, hence a limited incremental contribution.
- Results are limited to offline results on NDCG@k and HR@k, with no online results indicating user impact in production environment. The production deployment plan is also limited to only recommending project pages (and not products).
    - The results are compared where the system recommends low-level product categories (referred to as products), and not individual items (mentioned in Section 2).



----------------------- REVIEW 2 ---------------------
SUBMISSION: 118
TITLE: Jointly modeling products and resource pages for task-oriented recommendation
AUTHORS: Brendan Duncan, Surya Kallumadi, Taylor Berg-Kirkpatrick and Julian McAuley

----------- Overall evaluation -----------
SCORE: 2 (accept)
----- TEXT:
This paper seems well-written and clearly fulfils many of the criteria for the industrial stream.
The idea is generalisable to other web content ecommerce situations and is described in appropriate detail.



----------------------- REVIEW 3 ---------------------
SUBMISSION: 118
TITLE: Jointly modeling products and resource pages for task-oriented recommendation
AUTHORS: Brendan Duncan, Surya Kallumadi, Taylor Berg-Kirkpatrick and Julian McAuley

----------- Overall evaluation -----------
SCORE: 2 (accept)
----- TEXT:
Paper focuses developing a model for capturing high-level or task-oriented user intent to be used in a sequential recommender system. The particular application context is home improvement shopping where users often have a task context beyond just shopping for a specific tool. The authors argue that being able to capture such a task-oriented or usage context will allow for better quality recommendations about other related products or services to recommend based on their previous actions. Their approach interpretes user-clicks on both product pages as well as resource pages e.g. How-To or DIY pages as intent signals in a usage context.

The proposed sequential recommendation the model is based on the probability of the next item in a given sequence where the sequence is a sequence of user interactions with either product page or a resource page. The approach uses on auto-regressive transformer model in which word embeddings were replaced with interactions and a positional embedding matrix was learned. An optional embedding matrix was also learn which represents whether the interaction was with the project page or the product page. Both embeddings were combined to get a final embedding and fed to the lowest layer of the Transformer model.
Evaluation was done using a two layer transformer model and 100 dimensional embeddings. The optional embedding matrx didnt seem to have a significant impact so they were excluded from final experiments.The system was able to be adjusted to be used for either product or project recommendation by conditioning the model on some parameter functions.

Evaluation: Experiments were done to assess the quality of product recommendations and project recommendation. The experiments compared both a baseline approach with the sequential recommendation without the intent model and other modeling strategies. The results showed a good improvement in both product and project predictions over all methods tried. In particular, it showed that good quality product predictions are still possible even in the absence of user interaction with project pages (which represents a good percentage of user interactions on the website).
Deployment: The deployment context is to use to support email marketing campaigns beginning this year. To this end, the model was trained on a full year of user interaction data from a large home improvement store and sequential action data was considered as users that had activity in consecutive months which was incorporated in prediction code. The times needed for both training were in the order of minutes, which is quite reasonable for the planned use case.

Initial steps for deployment have been taken and plans for timeline and strategy for full deployment were given. However, full scale deployment is still pending.