--- kind: concept slug: collaborative-filtering title: Collaborative filtering summary: >- Predicts a user's rating of an item from past ratings by learning a vector per user and per item. Rating-only methods struggle when users and items have few ratings. Adding review text helps: DeepCoNN reports 7.6% to 8.7% lower MSE than the best baseline on Yelp, Amazon and Beer. tags: [recommender-systems, rating-prediction, matrix-factorization] papers: ["1701.04783"] related: [] status: draft updated_at: 2026-10-18 --- ## Overview A recommender estimates how a user would rate an item they have not rated yet. Collaborative filtering does this from the ratings matrix alone, the observed ratings of many users on many items [arxiv:1701.04783 §1]. The matrix is very sparse, and users or items with few ratings get poor predictions [arxiv:1701.04783 §1]. **Shared structure.** Every variant scores a (user, item) pair from learned representations of the user and the item. Prediction quality is compared by test MSE on an 80/10/10 split [arxiv:1701.04783 §3.3] [arxiv:1701.04783 eq. 11]. ## Variants The variants below differ in **what the representations are learned from**: ratings only, or ratings plus review text. ### Rating-only matrix factorization Uses only the rating matrix. In [arxiv:1701.04783] it is a baseline: MF is fit with ALS and tuned over latent-factor sizes {25, 50, 100, 150, 200} and regularisation {0.001, 0.01, 0.1, 1.0} [arxiv:1701.04783 §3.2] [arxiv:1701.04783 §3.3]. Reported test MSE is 1.792 (Yelp), 1.471 (Amazon) and 0.612 (Beer). PMF reaches 1.783, 1.460 and 0.527 [arxiv:1701.04783 tab. 3]. Card: [Joint Deep Modeling of Users and Items Using Reviews for Recommendation](/papers/1701.04783). ### Review-augmented two-tower model (DeepCoNN) Learns each user's and each item's representation from text. All reviews a user wrote form one document, and all reviews of an item form another. Each document goes through its own CNN: pretrained word vectors, 1-D convolution, max-over-time pooling and a fully connected layer [arxiv:1701.04783 §2.3] [arxiv:1701.04783 §2.4]. A factorization machine scores the concatenated user and item vectors [arxiv:1701.04783 §2.5]. Both networks are trained together [arxiv:1701.04783 §2.6]. Reported test MSE is 1.441 (Yelp), 1.268 (Amazon) and 0.273 (Beer), against 1.574, 1.372 and 0.299 for the best baseline, CDL [arxiv:1701.04783 tab. 3]. Ablations on Yelp show that a learned item-embedding table in place of the item CNN gives 1.577, a dot product in place of the FM gives 1.491, and random word vectors give 1.799 [arxiv:1701.04783 tab. 4]. Card: [Joint Deep Modeling of Users and Items Using Reviews for Recommendation](/papers/1701.04783). ## When to use which | Situation | Use | Why | |---|---|---| | Only star ratings exist | Rating-only MF | There is no text to learn from. The paper reports MF and PMF as its rating-only references [arxiv:1701.04783 tab. 3]. | | Reviews exist and many users or items have few ratings | Review-augmented model | The gain is largest for groups with the fewest reviews [arxiv:1701.04783 §3.6] [arxiv:1701.04783 fig. 3]. Overall MSE is 7.6% to 8.7% below the best baseline [arxiv:1701.04783 tab. 3]. | | Memory or latency budget is tight | Review-augmented model, with a length cap you choose | Documents hold all of a user's reviews, and the paper does not state a maximum length, padding or truncation [arxiv:1701.04783 §2.3]. Memory grows with document length. | | A user has no reviews | Not covered | The cold-start analysis covers few training reviews. The paper does not say how a user with zero reviews is handled [arxiv:1701.04783 §3.6]. | | New users or items arrive often | Review-augmented model, with caution | The paper argues a new review document can be pushed through without refitting, but no experiment tests this [arxiv:1701.04783 §2.7.2]. | | Comparing against the reported numbers | Remove the scored review first | The paper does not say whether the review of the pair being predicted is excluded from its user and item documents [arxiv:1701.04783 §2.3]. If it is included, the reported gains may be inflated. | ## Implementation sketch Review-augmented variant, following [arxiv:1701.04783 §2.2–§2.6]. Our own pseudocode; the paper gives no code. ```pseudocode # Build one document per user and per item from training-split reviews # (drop the review of the pair being scored; see Papers) [§2.3] for each user u: D_u = all training reviews written by u for each item i: D_i = all training reviews written for i # Encode one document with one tower (Net_u or Net_i) def encode(D, params): # one pretrained vector per word, order kept (dimension 300) [eq. 1] V = [embed(w) for w in D] # n x 300 # 100 kernels, window 3 words, bias, ReLU [eq. 2] Z = [relu(conv1d(V, K_j) + b_j) for j in 1..100] # keep the largest value of each feature map [eq. 4] O = [max(Z_j) for Z_j in Z] # fully connected layer to a 50-dim vector, ReLU, dropout [eq. 6] return dropout(relu(W · O + g)) # Score one (user, item) pair x = encode(D_u, user_params) # 50-dim y = encode(D_i, item_params) # 50-dim pred = factorization_machine(concat(x, y)) # eq. 7 # Train: RMSprop, batch 100, learning rate 0.002 [§3.3] # Loss is not stated in the paper; squared error is the likely choice ``` ## Contradictions Only one paper covers this concept, so no disagreement between papers is known. Within [arxiv:1701.04783] the reported percentages do not agree. The conclusion gives 8.5% for Yelp and 7.6% for Beer, and 8.7% for Amazon [arxiv:1701.04783 §5]. The table gives Yelp 8.5%, Amazon 7.6% and Beer 8.7% [arxiv:1701.04783 tab. 3]. The "8.3% average" is the mean of the three percentages. The relative gain on the averaged MSE row is about 8% [arxiv:1701.04783 tab. 3]. ## Papers - [Joint Deep Modeling of Users and Items Using Reviews for Recommendation](/papers/1701.04783): DeepCoNN, a two-tower CNN over review text with a factorization machine that predicts the rating.