Weighted Matrix Factorization

Weighted matrix factorization is the standard implicit feedback adaptation of factor models: an interaction says “some preference evidence exists,” while its count or strength says how confident the system should be. It avoids treating every missing pair as a strong negative.

Preference and confidence

Hu, Koren, and Volinsky separate binary preference $p_{ui}$ from confidence $c_{ui}$:

$$p_{ui}=1\{r_{ui}>0\},c_{ui}=1+\alpha r_{ui}.$$

Here $r_{ui}$ is the observed implicit interaction count or strength for user $u$ and item $i$. The binary preference $p_{ui}$ records whether any positive evidence exists, while confidence $c_{ui}$ increases with interaction strength; $\alpha$ controls how fast confidence grows.

The objective is

$$\underset{X,Y}{\min }\sum _{u,i}{c}_{ui}(p_{ui}-x_u^{\top }{y}_i{)}^2+\lambda (\Vert x_u{\Vert }^2+\Vert y_i{\Vert }^2).$$

The user factor $x_u$ and item factor $y_i$ produce a preference score by dot product. The confidence term makes errors on observed, repeated interactions count more than errors on missing entries, and $\lambda$ regularizes factor sizes.

For fixed item factors, each user update is a weighted ridge solve, closely related to ALS.

Worked example

This snippet solves a confidence-weighted user-factor update and scores all items against the learned user vector.

import numpy as np
r = np.array([3., 0., 1.])
C = np.diag(1 + 4 * r)
Y = np.array([[.8, .1], [.2, .7], [.6, .3]])
p = (r > 0).astype(float)
x = np.linalg.solve(Y.T @ C @ Y + 0.1*np.eye(2), Y.T @ C @ p)
print("user_factor", np.round(x, 3).tolist())
print("scores", np.round(Y @ x, 3).tolist())

Observed output:

user_factor [1.283, 0.111]
scores [1.038, 0.335, 0.804]

The item with three interactions has higher confidence and pulls the user factor most strongly. Bayesian personalized ranking instead trains pairwise orderings from positive and sampled negative items.

Item	Interaction count $r$	Preference $p$	Confidence $c=1+4r$	Effect
0	3	1	13	Strong pull toward item 0’s factor.
1	0	0	1	Weak evidence; not a confident dislike.
2	1	1	5	Positive pull, but less than item 0.

Caveats

The confidence formula is a modeling choice, not a fact about preference. Counts can reflect exposure, autoplay, bots, or interface placement. Tune $\alpha$, regularization, and negative treatment against top-k evaluation, and inspect long-tail coverage so popularity does not dominate every factor.

References

• Hu, Koren, and Volinsky, 2008, Collaborative Filtering for Implicit Feedback Datasets
• Koren, Bell, and Volinsky, 2009, Matrix Factorization Techniques for Recommender Systems

Nav

Section — Recommendation Systems and Personalization

← Alternating Least Squares Implicit Feedback Recommendation →