Given a list of offers and some associated metadata around the retailers and brands that are sponsoring the offer, we must provide a tool that allows users to search for offers via text input from the user.
Open the jupyter notebook on colab and follow the instructions.
Skip to the Offer Recommendation to see the final result if you are not interested in the details of the approach.
My approach is to use a simple tf-idf model to recommend offers to the users. I augment the offers with additional data using an LLM, Llama 2 70b q4 in particular. The LLM is used to generate additional keywords for each offer. This is because the offers are very short and do not contain much information. Then, I train a tf-idf model on the original data and the augmented data. There is a single function that handles giving a top-5 recommendation to the user. Tf-idf is a simple model that can be easily trained and deployed. Other word embeddings such as GloVe or BERT models can replace the existing model with minimal changes to the code.
Ultimately, our approach is to encode the search, and finding the most similar offers in the vector space.
Note that the search query itself can be augmented as well to include additional information. The usage of LLMs could be replaced by pre-defined rules or knowledge graphs. This function can become and API endpoint that can be used by the front-end to recommend offers to the user. The performance and the results are not quantified in this work, but it can be easily done by having access to labeled data. I prefer to finalize the full pipeline including the evaluation prior to moving forward with a more complex model. Furthermore, the expectation of the model also must be defined, and we cannot use "good" as a descriptor for the model. Future development can utilize this code as baseline and build on top of it. The visualization is handled by streamlit, which is a simple and easy to use tool for building quick prototypes. The UI is not the focus of this work, but it can be easily improved by using a more sophisticated UI framework, but that is outside my expertise.
Build the docker image and run the container. The container will be used to run the code. The code is written in Python
docker build -t offer-recommender -f docker/slim.dockerfile .Run it inside docker container, map ports for streamlit.
docker run -it -p 8501:8501 --rm -v $(pwd):/app offer-recommender streamlit run src/offer_recommender.pyAll the scripts assume the environment is properly setup.
We will merge the offers and the metadata into a single dataframe. The metadata does not seem to contain much information, therefore, we will use all the available information for each offer and pass it to an LLM to augment our input with additional keywords.
Note that this step cannot be performed without direct access to an LLM. Currently, the code users a privately ran instance of Llama 2 70b q4 using text generation inference from Huggingface. The endpoint APIs are similar to ChatGPT, therefore, the code can be easily modified to use ChatGPT instead. Using Llama 2 with this level of quantization is not ideal and will result in suboptimal results, nonetheless, it is sufficient for the purpose of building a proof of concept.
Run the below command to generate the augmented input. The augmented input will be saved in data/offers_augmented.csv.
You must be inside the docker container to run the code.
python src/llm_augmentation.pyWe will use a simple tf-idf model to recommend offers to the user. The tf-idf model will be trained on the augmented input. The simplicity and low computational cost of the tf-idf model makes it a good candidate for this task. In the later stages, we can utilize cosine similarity to recommend offers to the user.
python src/recommender_model.pyWe will use the trained tf-idf model to recommend offers to the user. The values are fixed at the moment, but again, this is a proof of concept. The code uses tfidf model trained on offers and the metadata first. If the similarity score is below a certain threshold, it will use the tfidf model trained on the augmented input.
python streamlit run src/offer_recommender.pyThe file output.txt contains the outputs of the model. We will explore the results to ensure that the model is working as expected. Ultimately, we need a way to evaluate this work properly to determine the best approach. The recommendations are also provided in results.csv.
The results look good, expanded keywords looks useful, but the expanded model has not been used due to high score for the original model.
0.55 Arber, at Target
0.42 L'Oréal Paris Makeup, spend $30 at Target
0.42 L'Oréal Paris Makeup, spend $35 at Target
0.41 L'Oréal Paris True Match Foundation at Target
0.40 L'Oreal Paris True Match Foundation at Target
This time, we are using the augmented keywords. The results are average, I looked for potential matches in the offers and I couldn't find any relevant offers at first glance. I suppose relevant offers targeting baby product are relevant, this needs to be evaluated further.
0.56 Whole pizza at Casey's
0.56 Whole Pizza at Casey's
0.56 Whole Pizza at Casey's
0.54 Whole Pizza Pie at Casey's
0.51 2 Pack OR 2 Liter AND Whole Pizza at Casey's