Data Science

So far, we've assumed that certain context isn't relevant in making recommendations about a given item for a given user
Specifically, we've ignored certain context like:
- Time
- User location
- Marketing channel
- Etc.
Realistically, this information is typically very relevant for most recommenders, since users make decision based on this context
There are three general methods for including context in our recommendations:
1. Contextual Prefiltering
2. Contextual Postfiltering
3. Contextual Modeling
For contextual modeling, our rating matrix could look like the following with time included as context:

ratingmatrixrec

Location: recommendations made by a store in Hawaii might be different for a store in Alaska
- Customers who receive recommendations restaurant in one area will receive different recommendations in another area
Time: Movie recommendations received today may not be relevant $10$ years later
- Apparel recommendations will be different depending on the season
Intent: Restaurant recommendations may change if the user is dining for one or for two
- Apparel recommendations may be different if users are shopping for a gift
Channel: Email recommendations can be different than recommendations made on the site
Conditions: Recpmmendations by a department store may include rain boots or a poncho depending on the weather

A non-contextual recommender is trained on $U \times I \times R$ data
- Specifically, its a data set containing a given rating for each user and item
- This is also referred to as the rating matrix
Then, this non-contextual recommender can predict a rating for a given user and item
Typically, we'll input a user into the model and receive a list of item recommendations

\overset{\text{Data}}{\boxed{U \times I \times R}} \to \overset{\text{Model}}{\boxed{u \to R}} \to \overset{\text{Recommendations}}{\boxed{i_{1}, i_{2}, ...}}

A contextual recommender with contextual prefiltering is one way of including contextual information in our recommender
Essentially, contextual prefiltering refers to building a recommender filtered on each value of a context feature
For example, we may want to build a separate recommender for each day of the week
First, the recommender is filtered on each value of our contextual feature
Then, it's trained on $U \times I \times R$ data
- Specifically, its a data set containing a given rating for each user and item
- This is also referred to as the rating matrix
Then, this contextual recommender can predict a rating for a given user and item for a given contextual value
Typically, we'll input a user into the model and receive a list of item recommendations

\overset{\text{Data}}{\boxed{U \times I \times R \times C}} \to \text{Filter on each } C \to

\to \overset{\text{Context Data}}{\boxed{U \times I \times R}} \to \overset{\text{Model}}{\boxed{u \to R}} \to \overset{\text{Recommendations}}{\boxed{i_{1}, i_{2}, ...}}

Before prefiltering, we must first consider the trade-off between sparsity and accuracy from contextualization
Specifically, adding context will increase the size of the rating matrix, which will make the data more sparse
On one hand, we can increase the accuracy of recommendations with user-item interactions
However, contextualization decreases the quantity of data available for the recommender
- Since the data sparsity grows
- This can impact the quality of recommendations

A contextual recommender with contextual postfiltering is another way of including contextual information in our recommender
Essentially, contextual postfiltering refers to filtering on recommendations based on values from a content feature after we make recommendations
For example, we can filter (or re-rank) items based on an item's designated gender once receiving an initial list of non-contextual recommendations from our recommender
First, the recommender is trained on $U \times I \times R$ data
- Specifically, its a data set containing a given rating for each user and item
Then, this recommender can predict a rating for a given user and item
Lastly, we'll filter on items satisfying a value from our context feature only

\overset{\text{Data}}{\boxed{U \times I \times R}} \to \overset{\text{Model}}{\boxed{u \to R}} \to \overset{\text{Recommendations}}{\boxed{i_{1}, i_{2}, ...}} \to

\to \text{Filter on each } C \to \overset{\text{Contextual Recommendations}}{\boxed{i^{c}_{1}, i^{c}_{2}, ...}}

The most intuitive solution for contextual recommendation is baking context in our actual recommendations
To do this, we'll need to create a hybrid model that predicts ratings as a function of item, users, and contexts
- Essentially, we extend matrix factorization from ordinary collaborative filtering to create embeddings for context/content features as well
- This is called a hybrid model because we combine:
  - The ability for content-based models to include content features
  - The ability for collaborative filtering to make recommendations about items without interactions from users
In general, the two types of contextual hybrid modelings are:
- Nearest neighbor modeling
- Latent factor modeling
An example of a hybrid model using contextual information is LightFM

\overset{\text{Data}}{\boxed{U \times I \times C \times R}} \to \overset{\text{Model}}{\boxed{u, c \to R}} \to \overset{\text{Recommendations}}{\boxed{i_{1}, i_{2}, ...}}

A temporal feature is one of the most important contextual features we can include in a context-aware recommender
For example, including a temoporal feature will capture variability in an item's preference over time
Specificially, it could capture fashion changes for clothes over time
Or, it could capture drifts in user preferences over time caused by changes in tastes
To do this, we can use contextual hybrid modeling (using nearest neighbor or latent factor modeling)

Hybrid Filtering Methods

Non-Personalized Methods