AI Get Local

Using A.I. to solve the challenge of product categorisation and organisation – how we classified and organised over 2 million products from 5,000 online Retailers across Ireland to create an annotated, organised, and searchable product data set. The first challenge when developing our GetLocal.ie platform, was the ingestion and categorisation of products from thousands of pre-existing Irish online Retailers into a predefined single data structure. The Getlocal.ie platform would require each product to follow a structured categorisation taxonomy to enable search, refine, and related products results. Identifying Tensorflow A.I. learning models as the primary option to develop a categorisation model, we commenced a manual categorisation project to develop a training data set that would be used to train our initial A.I. model.

First, we needed training data: Building an annotation dashboard enabled us to query 30,000 broad-based products. We undertook a human-led manual triple verification project to classify these 30,000 products across 5k classes. We further expanded the training data in QA’d steps up to 300,000 broad-based products.

This project provided us with extremely high confidence categorisation and formed the training data / seed for our first A.I. categorisation model. This proved very successful, however, not without challenges, we found limitations in the ability of the model to decipher between abstract product images, as explained by our Director of Platforms and Services – Alan Linnane.

Alan Linnane, Director of Platforms & Services “One classification that initially challenged the model was music records. For example:‘‘The Velvet Underground and Nico’ album cover features an image of ‘Banana 10’ by Andy Warhol. This abstract product image initially misled the A.I. model to categorise the product into a suggested category of Fruits and Vegetables. It’s a perfect example of how A.I. models often create unexpected inferences that require resolution”.

We followed with three further large-scale iterations of our A.I. categorisation model. We used a transformer model design incorporating Image Classification and Google’s BERT (for NLP understanding) to make a release ready version of our A.I. categorisation model. This has enabled us to successfully achieve a high confidence categorisation score on 1.4 million products via A.I. technologies. The scale of this operation and platform is impossible for a small team without using A.I..

Consumer text-based search queries are AI driven on getlocal.ie. We provide lightning fast, accurate search results to users in less than 100ms total response time. Following our successful categorisation of 1.4 million products, we turned our attention to further developing our search technology.

We evaluated many existing A.I. Open Source projects to deliver at scale within budget. Early efforts with Spotify’s ANNOY Project showed promise but also limitations. We finally choose Apache Solr search platform, specifically Solr Indexing with Approximate Nearest Neighbour A.I. embedding Search combined with a hybrid Solr cloud cluster. Implementing a mix of In-Ram and Disk-Backed Replicas ensured high-quality results with cost-effective speed, resilience, and redundancy.

Initially, our search model focused on breadcrumb and keywords search. This enabled complex keyword search against physical location to render multi-attribute results and dynamically produced related product views.

Early releases were successful but not without limitations, as explained by our Business Planning Analyst – Eoin Dunne. “colloquial or natural language would lead a user to search for ‘Tommy-Hilfiger-Perfume’ however, we found in reality products were not defined as ‘Tommy-Hilfiger-Perfume’ but rather ‘Tommy-Hilfiger-Eau-de-Parfum’. By interrogating our product results over time, we could spotanomalies like these which led us to evolve our search technology to includeimage-recognition data in our search model.”

This led us to develop and evolve our A.I. search experience using search technologies from OpenA.I. CLIP Model, GPT3 and ImageNet to combine embeddings from keyword, breadcrumb, and image-inferenced data. We solved the additional technology hurdle of performance at scale across the new model by utilising server-less A.I. inferencing and a mix of In-Ram and Disk backed database replica clusters along with intelligent cloud-based load balancing to give us sub 100ms total. The cumulative effect was; that our result quality and speed soared. This approach to infrastructure also offers near limitless scaling potential within the platform context.

Derek Conniffe – Lead Developer “inspired by Roblox, a gaming ecosystem, we built out our own real-time A.I. inference platform operating in a server-less environment. This offers unlimited scaling potential for the platform.”

We further developed a percentile-based algorithm to determine statistically high confidence result sets. When fed back into our A.I. search model, we refined our result set further into direct-match and related-match result sets, driving relevance for the consumers’ query and reducing search noise.

Innovation lies at the heart of getlocal.ie platform from inception. We continue to evolve our A.I. search experience. Our recent beta launch of A.I.- driven Visual Search in August 2022 is a vanguard feature for the platform and shows significant promise in terms of customer experience and result quality.

Jade O’Connor VP of Product & Marketing “A.I. technology allows us to enhance the user’s experience in new and innovative ways. At the core of getlocal.ie is the question: How do we enable Irish consumers to buy online from local businesses? Visual Search is an incredibly useful way to generate meaningful consumer search results across a broad-base product platform. This makes it an exciting and useful development for our audience.”

Our approach to Visual Search is pure A.I. based and incorporates a hybrid of traditional multidimensional distance search algorithms and newer approximate nearest neighbour (ANN) search technologies. For image searching a combination of batch and real-time A.I. inferencing is applied in order to identify and deliver the best ranking results. When married with user experience focussed design, it opens a gateway for consumers to discover products incredibly intuitively.

Key features are:
• Search similar products by clicking on the image search icon in any search result
set.
• Taking or uploading your own image to be used as the search source.
• Follow search examples from the desktop prompt.
• Visual Search Chrome Extension (in development)
• This allows a user to perform a search based on an uploaded photo or by querying
product images.