VDB Comparison
Generate custom embeddings with ALL your structured and unstructured data.Try Superlinked

How to Build a Recommender System From Clickstream Data

#Recommendations
Publication Date: May 6, 2024

If you have a website with browsable products, your visitors generate a virtual flood of clickstream data. Say, for example, your site is for hotel bookings, and a site visitor browses hotels, starting with the Tranquil Haven Resort, and ending at the Evergreen Retreat Lodge:

graph LR; id1[Tranquil Haven Resort]-->id2[Summit Vista Inn]; id2[Summit Vista Inn]-->id3[Azure Sands Hotel]; id3[Azure Sands Hotel]-->id4[Evergreen Retreat Lodge];

Why did they follow this route in particular? Maybe they wanted to check hotels with certain amenities, in the same location, with a certain style. We can't tell exactly what took the user down a particular path (clickstream). But you can take the clickstreams of all your users, use it to create embeddings for the hotels in the same vector space. In this space, the more "similar" the hotels are, the closer together they will be.

An embedding is a numerical representation of categories that carry more meaning than a simple one-hot encoding. It is a vector of a float, and the length of the vector is called the embedding dimension.

In this article, we'll show you how - with only a few lines of code, using techniques from NLP (natural language processing) - you can create embeddings that embody clickstream data, and use these embeddings to set up a simple recommendation system.

All you need to do is:

  1. collect the clickstream data (usually very simple)
  2. clean it, if necessary
  3. use libraries such as Gensim to train an algorithm (e.g., Word2Vec) and generate our hotel embeddings on our clickstream data

Before we begin building our recommendation system (RecSys), we need to install some requisite libraries: pip install gensim umap-learn pandas matplotlib.

Now that those are in place, let's get started!

Gather and clean your data

To set up our example, we'll use this artificial dataset. The first few lines out of 10,000 look like this:

554>11>956>952>108>174>587>243>153>863>935>87>841 521>621>978>678>379>703>738 457>730>162>42>126>542>972>168>878 325>819>597>158>379>407>960>912>247 194>185>189>183>180>173>994>199>197>192>187 882>719>500>171>692>685>336>616>86>470>908>349>124>621>774>698>159>833 54>495>78 592>780>792 ...

The numbers are hotel IDs (but this will work for numbers representing anything). To train our model and create embeddings for the hotels, we need to preprocess this data slightly.

with open("clickstreams.txt", "r") as f: raw = f.read() clickstream = [line.split(">") for line in raw.split()]

After this preprocessing step, your data looks like this:

[['554', '11', '956', '952', '108', '174', '587', '243', '153', '863', '935', '87', '841'], ['521', '621', '978', '678', '379', '703', '738'], ['457', '730', '162', '42', '126', '542', '972', '168', '878'], ['325', '819', '597', '158', '379', '407', '960', '912', '247']] ...

Train the model, create embeddings

Once you have your data in this format, you can start the model training and create your embeddings. You just type the following:

from gensim.models import Word2Vec model = Word2Vec( clickstream, vector_size=16, sg=1, epochs=10, )

Note that Gensim's Word2Vec model has many more parameters to check out.

Once the training has created the learned embeddings, you can access them like this:

import pandas as pd embeddings = pd.DataFrame( model.wv.vectors, index=model.wv.index_to_key )

The first five rows of the resulting index of embeddings look like this:

0123456789101112131415
991.480.520.270.670.11-0.94-1.16-0.29-0.050.85-0.11.040.8-0.46-0.130.75
9001.730.24-0.290.610.340.94-0.43-0.730.771-0.5-0.60.40.390.92-0.66
7000.21.180.060-0.58-1.05-1.360.321.080.04-0.57-0.73-0.680.020.4-0.12
8000.931.040.70.050.75-0.230.310.720.790.370.051.320.31.030.19-1.18
1991.5-0.75-0.07-1.02-1.290.23-0.60.061.04-0.150.770.53-0.470.620.38-0.18

The index's first column is the hotel ID. The next 16 columns (numbered 0-15) represent the learned embedding. There are 16 numbers because we set vector_size=16.

Now that you have an embedding for each hotel, you're ready to implement your hotel RecSys!

Recommendation system: suggesting other hotels

Starting from one hotel, now embedded in a shared vector space with embeddings for all your other hotels, you can simply ask for its closest embeddings. Querying hotel 123, for example, will make your model generate a list of similar hotels.

Using Gensim, you can do this easily. Query

model.wv.similar_by_word("123")

and you receive

[('131', 0.9089741706848145), ('127', 0.8981987237930298), ('122', 0.8846656084060669), ('118', 0.8609521985054016), ('115', 0.8502810597419739), ('243', 0.8356520533561707), ('454', 0.8332993984222412), ('106', 0.8226854801177979), ('103', 0.8177634477615356), ('765', 0.8113372921943665)]

(Your numbers may differ.)

Here's how to read your output (above): Hotel 131 is most similar to Hotel 123, with a cosine similarity of about 0.91. The second most similar is Hotel 127, with a cosine similarity of about 0.9, and so on.

So, starting with just the clickstream data of your site visitors, you've successfully created a RecSys!

Cosine similarity is a measure of how similar two vectors are. It can range from -1 to 1 (similar to the Pearson correlation coefficient). If both vectors point in the same direction, cosine similarity is at the high end (closer to 1).

The formula is

Cosine Similarity(A,B)=ABAB=i=1nAiBii=1nAi2i=1nBi2\text{Cosine Similarity}(\mathbf{A}, \mathbf{B}) = \frac{\mathbf{A} \cdot \mathbf{B}}{\|\mathbf{A}\|\cdot \|\mathbf{B}\|}=\frac{\sum_{i=1}^{n} A_i \cdot B_i}{\sqrt{\sum_{i=1}^{n} A_i^2} \cdot \sqrt{\sum_{i=1}^{n} B_i^2}}

where A=(A1,,An)\mathbf{A}=(A_1, \ldots, A_n) and B=(B1,,Bn)\mathbf{B}=(B_1, \ldots, B_n) are two vectors.

Cosine Similarity

Under the hood

Gensim's Word2Vec model is, basically, a fast implementation of ideas in Mikolov et al.'s groundbreaking 2013 NLP paper, Efficient Estimation of Word Representations in Vector Space.

To train our model and create our embeddings, we use the Skip-gram version sg=1 (rather than CBOW version) of Mikolov et al.'s Word2Vec model. Skip-gram Word2Vec works as follows.

Starting with the clickstream from our introduction (above):

graph LR; id1[Tranquil Haven Resort]-->id2[Summit Vista Inn]; id2[Summit Vista Inn]-->id3[Azure Sands Hotel]; id3[Azure Sands Hotel]-->id4[Evergreen Retreat Lodge];

We are trying to make our model understand (and embed) the relationships between the clickstream entities (in our case, hotels). These relationships can be represented as skip-gram pairs. Skip-grams are a popular NLP method for training a model to predict the context, or neighboring words, of a given target word. We want our model to be able to predict, for a given hotel ("focus"), who its immediate neighbors in a clickstream will be - i.e., the hotel right before it, and the hotel right after it. We can represent the clickstream data from our example above as a table of skip-gram pairs (below):

FocusPrediction
Tranquil Haven ResortSummit Vista Inn
Summit Vista InnTranquil Haven Resort
Summit Vista InnAzure Sands Hotel
Azure Sands HotelSummit Vista Inn
Azure Sands HotelEvergreen Retreat Lodge
Evergreen Retreat LodgeAzure Sands Hotel

For predicting user behavior on websites we want our model to learn a probability score that a user having clicked on one given hotel (e.g., Tranquil Haven Resort) will click on another given hotel (e.g., Summit Vista Inn), and not on another hotel (e.g., Evergreen Retreat Lodge). This is a binary classification problem, so our model should be trained on binary data. This allows the model to directly learn a probability score of transitions between a specific hotel and another specific hotel, and makes the model architecture easier and more efficient. Here's how the clickstream data looks as binary training data for our model.

FocusOther HotelNeighbor?
Tranquil Haven ResortSummit Vista InnYes
Summit Vista InnTranquil Haven ResortYes
Summit Vista InnAzure Sands HotelYes
Azure Sands HotelSummit Vista InnYes
Azure Sands HotelEvergreen Retreat LodgeYes
Evergreen Retreat LodgeAzure Sands HotelYes

But before we train our binary classifier on this data, we need to address a limitation of our training data. It's constructed only of positive pairs - i.e., hotels that were clicked sequentially in the clickstream. That is, the target column Neighbor? contains only a single label: "yes". Our model needs to learn more than just what constitutes "good" output - i.e., relevant hotel recommendations. It also needs to learn "bad" output - which hotels are irrelevant hotel recommendations. Mikolov et al.'s proposed solution for this was negative sampling. It needs to learn some "no"s.

Negative sampling

If your clickstream training data consists only positive rows of neighbors (skip-grams), you can add some random negative pairs of hotels - ones that were not adjacent in the clickstreams - to ensure that the model learns how to avoid recommending hotels that are irrelevant.

In our case, such a table of training data with added negative sampling might look like:

FocusOther HotelNeighbor?
Tranquil Haven ResortSummit Vista InnYes
Tranquil Haven ResortSome Other Hotel 1No
Summit Vista InnTranquil Haven ResortYes
Summit Vista InnSome Other Hotel 2No
Summit Vista InnAzure Sands HotelYes
Summit Vista InnSome Other Hotel 1No
Azure Sands HotelSummit Vista InnYes
Azure Sands HotelSome Other Hotel 3No
Azure Sands HotelEvergreen Retreat LodgeYes
Azure Sands HotelSome Other Hotel 3No
Evergreen Retreat LodgeAzure Sands HotelYes
Evergreen Retreat LodgeSome Other Hotel 1No

By training our model on this data, we can create embeddings that encode both negative and positive relationships between hotels. In other words, we'll have a binary classifier that learns to predict whether a given pair of hotels are likely to be clicked successively, or not, and on this basis provide hotel recommendations.

And this, in a nutshell, is what happens under the hood.

Hyperparameters

The ratio of positive to negative samples in the training data is a variable hyperparameter (a parameter whose value is set before the learning process begins) that can be optimized to fine-tune to get better embeddings - i.e., better performance. There are many other hyperparameters we can use to fine-tune - for example, embedding size - called vector_size in Gensim, which we've seen already. Another important hyperparameter is window_size, which enlarges the focus of the model, allowing us to take account of not just the immediate clickstream hotel neighbors but also those two or three steps away from the "focus" hotel.

Conclusion

We've discussed how to take advantage of your website's clickstream data to easily create a recommendation system. You need only record your clickstream data, do a little preprocessing, and write about ten lines of code to train a model (using Gensim, for example) and create embeddings. Our example use case is hotels, but you can apply this approach to any clickstream data.

You may also just want to visualize your embeddings to gain more insight into your users' behavior. For this, you can use, for example, PCA, t-SNE, or UMAP to project your embeddings in two dimensions:

import matplotlib.pyplot as plt from umap import UMAP t = UMAP(random_state=0).fit(model.wv.vectors) plt.scatter(t.embedding_[:, 0], t.embedding_[:, 1])

Here's just such a 2D visualization of our example hotel embeddings, generated with UMAP.

Projected Embeddings

As you can see, certain hotels form clusters. Though we can't know just from clustering what about the hotels is similar, we can further investigate the clusters for shared characteristics, and thereby gain insights into our data.

If you want to take advantage of the value latent in your clickstream data, whether it's to glean insights, or even build a working recommendation system, we encourage you to give our code a try!

Don't be a stranger. Let us know how it goes!

Contributors

Stay updated with VectorHub

Continue Reading