Use reviews to create a recommendation system that works

If you’ve ever purchased a product online and marveled at the futility and non-applicability of “related items” that haunt the purchase and after-sales process, you already understand that systems popular and traditional recommendations tend to lack understanding. relationships between potential purchases.

If you buy an unlikely and infrequent item, like an oven, recommendations for other ovens will probably be superfluous, although the worst recommender systems won’t recognize this. In the 2000s, for example, TiVO’s recommendation system created a first controversy in this sector by reassigning the perceived sexuality of a user, who then sought to “re-masculinize” his user profile by selecting war movies – a crude approach to algorithm review.

Worse still, you don’t have to buy anything from (say) Amazon, or start watching a movie you’re browsing the description of on any major streaming platform, for the information-starved recommendation algorithms happily start along the wrong path; searches, searches and clicks within the “details” pages are sufficient, and this sparse (and likely incorrect) information is likely to carry over into future browsing sessions on the platform.

Trying to forget a recommendation system

Sometimes it is possible to intervene: Netflix provides a “thumbs up/down” system which should in theory help its machine learning algorithms remove certain embedded concepts and words from your recommendation profile (although its effectiveness has been questioned and it remains much easier to scale a personalized recommendation algorithm from scratch than to remove unwanted ontologies), while Amazon allows you to remove titles from your customer history, which should demote any unwanted domains that have infiltrated your recommendations.

Hulu has a similar feature, while HBO Max has partially opted out of purely algorithm-based recommendation systems, given their current shortcomings.

None of these strictly consumer-level experiments even address the widespread and growing criticism of recommendation systems from “passive” advertising platforms (where noticeable changes occur due to public anger), or the incendiary topic of AI recommendations on social media, where sites such as YouTube, Twitter and Facebook continue to face criticism for irrelevant or even harmful recommendations.

The machine doesn’t seem to know what we want, unless we want adjacent element that came up in our search – even though this item is essentially a duplicate or alternative to the main item we may have just purchased, rather than a potential add-on or accessory purchase.

Accurate recommendations with review data

A new research collaboration between China and Australia offers a new method to address these irrelevant recommendations, using external user reviews to better understand the actual relationships between items during a shopping session. In testing, the architecture outperformed all current state-of-the-art methods, offering hope to recommender systems that have a better internal map of item dependencies:

RI-GNN outperforms leading competitors in item relationship accuracy, best performing on sessions with more than five items. The system was tested against Pet Supplies and Movies and TV datasets from Amazon Review Data (2018). Source: https://arxiv.org/pdf/2201.12532.pdf

To boot, the project takes on the notable challenge of creating recommendations even in anonymous sessions, where the recommender system doesn’t have access to details provided by the user, such as purchase history, or own reviews. user’s line on past purchases.

The new newspaper is called Rethink Adjacent Dependency in Session-Based Recommendationsand comes from researchers at Qilu University of Technology and Beijing Institute of Technology in China, RMIT University in Melbourne, and the Australian Institute of Artificial Intelligence at the University of Technology Sydney.

And after?

The primary task of session-based recommendations (SBR) is to determine the “next” item from the current item, based on its calculated relationship to the current item. Concretely, this could manifest as a listing of “related items” in an items page for a birdcage on an e-commerce website.

If you’re buying a birdcage, what else are you likely to need? Well, at the very least, you’ll need a bird to put in it – it’s a true addiction. However, the birdcage is in the ontology pet items, where the birds are not sold. Perversely, cat food is in the same ontology, although the addition of a cat feeding bowl as an associated recommendation for a birdcage product is a false addiction – an erroneous and erroneous association.

Excerpt from the article: true and false relationships between several items, visualized on the right as an inter-item graph.

Excerpt from the article: true and false relationships between several items, visualized on the right as an inter-item graph.

As is often the case in machine learning architectures, it is difficult to persuade a recommender system that a “remote” entity (bird does not appear at all in pet products) can have an intrinsic and important relationship with an element, whereas elements that belong to the same category and are very close in terms of function and central concept (such as cat bowl), may be orthogonal or directly opposed to the intended purchase.

The only way to create these mappings between “non-adjacent” entities is to crowdsource the problem, since the relationships in question are a facet of human experience, cannot be guessed programmatically, and are likely beyond the scope affordable from conventional approaches to labeling datasets, such as Amazon Mechanical Turk.

Therefore, the researchers used natural language processing (NLP) mechanisms to extract salient words from reviews of a product, and used the frequencies of these analyzes to create incorporations capable of “matching” elements apparently distant.

The architecture of the Inter-Item Graph Neural Network (RI-GNN) refined by revision.

The architecture of the Inter-Item Graph Neural Network (RI-GNN) refined by revision.

Architecture and data

As the new article notes, previous work of a similar nature has leveraged a logged-in user’s revision history to provide rudimentary mappings. Both DeepCONN and RNS have used this approach. However, this does not take into account that a user may have written no reviews, or no relevant reviews for a particular item that is “out of range” of their usual shopping habits. Additionally, it is somewhat of a “white box” approach, as it assumes the user has already engaged enough with the POS to create an account and log in.

The extended Graph Neural Network (GNN) proposed by the researchers takes a more oracle-like approach, deriving real dependencies a prioriso that, presumably, the anonymous, disconnected user can benefit from more relevant recommendations with minimal input required.

The augmented examination system is titled Refined inter-element graph neural network (RI-GNN). The researchers tested it against two datasets from Amazon, Pet supplies and Movies and TV. While this solves the issue of review availability quite well, an implementation in the wild would need to locate and scrape a proper review database. Such a dataset source could, in theory, be anything from social media posts to Quora responses.

High-level relationship mappings of this nature would, furthermore, be valuable for a range of machine learning applications beyond recommender systems. Many ongoing projects are crippled by the lack of inter- and intra-domain mapping due to limited funds and scope, while the commercial momentum of a truly informed and participatory e-commerce recommendation system could potentially fill this gap.

Metrics and tests

The authors tested RI-GNN against two versions of each dataset, each including a user’s purchase history and general product reviews. Items appearing less than five times have been removed and the user’s history split into units of one week. The first version of the dataset had all sessions with more than one item, and the second all sessions with more than five items.

The project used P@K (accuracy) and MRR@K (Mean Reciprocal Rank) for its evaluation measures. The rival architectures tested were: S-KNN; GRU4Rec; S-POP; STAMP; BERT4Rec; DHCN; GCE-GNN; SR-GNN; and NARM.

The frame was trained in batches of 100 on Adam at a learning rate of 0.001, with the number of subjects set to 24 and 20, respectively, for Pet supplies and Movies and TV.

First published February 1, 2022.


Source link

Comments are closed.