Introduction & CF

• recommender systems help match users with items

#

Long form:

Recommender systems are software agents that elicit the interests and preferences of individual consumers […] and make recommendations accordingly. They have the potential to support

and improve the quality of the decisions consumers make while searching for and selecting products online.

• explosion of choice - information overload

• Tools to guide through decision process

• increacesed number of options

• exploratory search

• User: finds new things

  • narrow down set of choices

  • help explore space of options

  • entertainment

• Targeted exposure

• Create brand awareness

• Increase trust and loyalty

• Increase sales, click trough rates, conversion etc

popularity distributions have a long tail

• Pareto principle 80/20 rule: 80% of plays comes from

20% of items

• demographic-based

• collaborative filtering

  • show me what people with similar taste like

• content-based

• knowledge-based

• hybrid recommenders

  • composition of various recommenders

  • 
    

No it is different to content-based recommenders, domain agnostic: ideas apply to tother types of items

rantings

Users give ratings to items

• users’ behavior (i.e., ratings) is guided by their preferences

• users’ preferences remain stable and consistent over time

• thus, we can expect similar users to prefer similar items

Memory-based methods remeber the entire history of user-item interaction

e.g Neighborhood methods

model-based methods build a model that describes the history and make recommendatiosn from that.

e.g. Matrix factorization

Item-Item-CF empfiehlt ähnliche Artikel basierend auf der Ähnlichkeit zwischen den Bewertungen von Artikeln durch verschiedene Benutzer. Die Methode geht davon aus, dass Benutzer, die ein bestimmtes Produkt mögen, ähnliche Produkte auch mögen könnten.

Beispiel:

Wenn ein Benutzer beispielsweise ein Buch über Geschichte bewertet hat, wird Item-Item-CF ähnliche Bücher aus dem Bereich Geschichte empfehlen. Das Modell berechnet Ähnlichkeiten zwischen den einzelnen Items und gibt dann eine Empfehlung basierend auf der Bewertung des Benutzers und den Ähnlichkeiten zwischen den einzelnen Items.

ChatGPT

User-User Memory-based CF: User-User-CF empfiehlt ähnliche Artikel basierend auf der Ähnlichkeit zwischen den Bewertungen von Benutzern. Das Modell geht davon aus, dass Benutzer, die ähnliche Dinge mögen, wahrscheinlich auch in Zukunft ähnliche Dinge mögen werden.

Beispiel

Wenn ein Benutzer beispielsweise mehrere Artikel im Bereich Sport bewertet hat, wird User-User-CF anderen Benutzern empfehlen, die ähnliche Artikel im Bereich Sport bewertet haben. Das Modell berechnet die Ähnlichkeit zwischen den Bewertungen von Benutzern und gibt dann Empfehlungen basierend auf der Bewertung des Benutzers und den Ähnlichkeiten zwischen den einzelnen Benutzern.

ChatGPT

The rating of a target user to a target item

Take the average rating -> score = all ratings that exist for that item / number of users

we calculate the weighted average rating instead of average

score = (similartity weight between users u,v * all ratings that exist for that item) / normalized by similartity weight between users u,v

Filtering to predict the rating that a user u would give to an item i, based on the ratings of similar users. (chatGPT)

• For some is good 3/5 for some 5/5

1. Central tendency bias: This occurs when users tend to rate items in the middle of the rating scale, rather than at the extremes. This can make it difficult to distinguish between items that are truly mediocre and those that are exceptional or terrible.

2. Leniency bias: This occurs when users tend to rate items higher than they actually deserve. This can be due to a desire to be positive or a lack of critical evaluation of the item.

3. Severity bias: This occurs when users tend to rate items lower than they actually deserve. This can be due to a desire to be critical or a negative impression of the item.

4. Item popularity bias: This occurs when users rate popular items more favorably than less popular items, regardless of their actual quality. This can lead to a "rich get richer" effect, where popular items become even more popular due to their high ratings.

5. User preference bias: This occurs when users rate items based on their personal preferences, rather than the quality of the item itself. This can lead to recommendations that reflect the user's existing preferences, rather than exposing them to new or diverse items.

   (Chat GPT)

Instead of using ratings -> use deviations of how far off ratings are from mean ratings

-> center around the mean

r(v) = source user

r(u) = target user

Yes if my mean rating is 4/5 and all users give +2 above theier average, my predicted rating is 6

-> in practicse, restrict the predicted rating at a final step

Pearson correlation (PCC)

• with some small adaptation

• similar to the mean-centered cosine similarity

We want the similarity of the users u,v the formular is:

u = target user

w(u,v) = items rated by both users and deviations of agreement / idividual rating by earch user normalized to [-1,1]

Normal PC means (x̄ and ȳ) are calculated over all items in the dataset, regardless of whether they have been rated by both users or not. This means that the means can be affected by differences in the rating scales or rating behavior between users.

adjusted pearson means (x̄ and ȳ)are only calculated over the items that have been rated by both users. This helps to eliminate the influence of differences in rating scales or rating behavior between users, as only the common items are used in the calculation.

By using the adjusted Pearson correlation coefficient, the similarity between users is based only on their rating patterns for the common items, which can improve the accuracy and reliability of the similarity metric in Collaborative Filtering.

(ChatGPT)

it measures how closely aligned the two vectors representing the rating patterns are.

• u and v are the vectors representing the rating patterns of two users (or items)

• u ⋅ v is the dot product of the two vectors

• ||u|| and ||v|| are the magnitudes (or lengths) of the two vectors

neighborhood contains users with the highest similarity to the target user

• in practice neighborhood contains only users that have rated the target item

application/domain dependent

best practise: 25-100 most similar users - neighborhood

Very high, that is why there are many optimisation techniques.

Neigborhood creation

• look among users that have rated at least one of those items

• fehlt noch was

Item recommendation

• only itmes that have been rated by at least one neighbor -> non rated items cannto be recommended

Much. more users than items

few ratings compared to product of items and users ( rating per user per product)

pairwise user similarities -> cost = number of user^2