My goal was to start a career I was passionate about, despite my lack of funds.

Because of this choice I have accomplished a lot over the last few months. I published my own website, was posted in a major online data science publication, and was given scholarships to a competitive computer science graduate program.

In the following article, I give guidelines and advice so you can make your own data science curriculum. I hope to give others the tools to begin their own educational journey. So they can begin to work towards a more passionate career in data science.

A Quick Note

When I say “data science”, I am referring to the collection of tools that turn data into real-world actions. These include machine learning, database technologies, statistics, programming, and domain-specific technologies.

A few resources to start out your journey.

The internet is a chaotic mess. Learning from it can often feel like drinking from the fun end of a fire-hose.

There are simpler alternatives that offer to sort the mess for you.

Sites like Dataquest, DataCamp, and Udacity all offer to teach you data science skills. Each creating an education program that shepherds you from topic to topic. Each requires little course-planning on your part.

The problem? They cost too much, they don’t teach you how to apply concepts in a job setting, and they prevent you from exploring your own interests and passions.

There are free alternatives like edX and coursera which offer one-off courses diving into specific topics. If you learn well from videos or a classroom setting, these are excellent ways to learn data science.

As I discussed elsewhere, I recently started my third Custom GPT research project. My first one was less successful than my last two. This article is an in-depth look at a successful project, the one named after a trail, a race, a place: the Middlesex Fells Trail Analyzer.

This past December (2023), I finished an ultramarathon (32 miles) on the Skyline trail in the Middlesex Fells in Stoneham MA. It is an annual race and this one was my ninth time. Most of my races are GPS recorded on Strava. Figure 1 shows the course on a Strava display. Stava is a sports social media platform for recording and sharing sports activities and is the source of the GPS data used in the experiment described in this article.

I started this experiment by asking ChatGPT to analyze GPS data from this year’s race. The experiment grew into a larger investigation that included race data from the last five races. During my research, I found that bundling my race data and instructions into a GPT to be helpful. Thus. the Middlesex Fells Trail Analyzer Custom GPT was born and is helping to support my work with ChatGPT/GPTs. Topics of interest include:

• What prompt/instruction patterns to use?

• How to design effective knowledge structures that ChatGPT can use?

• How to improve ChatGPT code generation?

In this ongoing experiment, I used a customized GPT to simplify the set-up and preparation of experiments. This means that I don’t have to reload and unnecessarily reprocess data, or update configuration and prompts with each ChatGPT session.

This is a common occurrence, particularly with off-the-shelf RAG implementations. Documents may lack complete answers to the query, contain redundant information, or include irrelevant details. Furthermore, the order in which these documents are presented may not consistently match the user’s intent.

In this post, we will explore three effective techniques to enhance document retrieval in RAG-based applications:

1. Query expansion

2. Cross-encoder re-ranking

3. Embedding adaptors

By incorporating these techniques, you can retrieve more pertinent documents that closely match the user’s query, thereby increasing the impact of the generated answer.

Let’s have a look 👇.

If you’re interested in ML content, detailed tutorials and practical tips from the industry, follow my newsletter. It’s called The Tech Buffet.

At Picnic, we’re not just an online supermarket; we’re the modern milkman. This means that we want to make the shopping experience of our customers as easy as possible while delivering the best personal service.

To do this we couldn’t do without recommender systems. Recommender systems are used in many places within Picnic. From ranking customers’ search results and previously bought items, to showing the most relevant recipes for each customer. Our goal is to personalise every aspect of the grocery shopping experience using machine learning. However, recommendations aren’t just about algorithms; it’s about helping our customers save time, find the right things, and curate the shopping experience they deserve.

What do we have, what do we want?

We currently have multiple recommendation models in operation. One example is a model focussed on customers’ repeat purchases, called Customer Article Rebuy Prediction (or CARP in short 🐟). CARP is used in many places in the Picnic app. Most notably in the previous purchases page where all of your previously bought articles are ranked by CARP to create your personal shopping list. CARP is a relatively simple machine learning model (*Under the hood CARP is powered by XGBoost) *with handcrafted features tailored to predicting repeat purchases.

These features include things like the average rebuying frequency of an item or the periodicity of a customer doing their shopping. Grocery shopping actually contains many of these patterns and they are often very clear! Indeed a large part of customers shopping is repeat behaviour. That makes repeat item recommendation not only quite straightforward but also very useful!

However, to realise our goal of personalising every aspect of the shopping experience, we also want to serve recommendations in general, i.e. for any article in our store, including the ones a customer hasn’t bought before.

Beyond your weekly shopping list: repeat vs explore

Making the distinction between unbought, or explored articles, and repeat articles, is non-trivial. As noted in the literature [1, 2, 3, 4, 5], this element plays a vital role in grocery shopping recommendation performance.

As mentioned above, our CARP model is already running in production and serving millions of repeat recommendations on a daily basis. Developing a more general model that can additionally handle explore recommendations is however quite a difficult task.

The challenges of Explore Recommendations

Customers in grocery shopping have highly personal preferences that discriminate between sometimes very similar products (think of competing brands or subtle differences in flavours) while the assortment is huge and the distribution of customer-article interactions has a long tail. Ok, that’s a bit of a mouthful. Let’s break that down.

#creating classes: blueprint class Car: pass

Honda=Car() #it will inherit all properties from class Car Suzuki=Car() Let’s build something, as I mentioned before: OOPs connect programming with the real world, and we’ll see how. Class, a blueprint, can contain:

1. Variables (defined for whole class),

2. Attributes (personality traits or information commonly known to all objects). These are called upon creation of objects.

3. Methods (functions or procedures) associated with a class. We have to call these manually, unlike attributes.