| |
| exec(open('/tmp/dig/lat.py').read()) |
| df=ALL.unique(['u','ix']); R=sp.csr_matrix((np.ones(df.height,np.float32),(df['u'].to_numpy(),df['ix'].to_numpy())),shape=(NU,NI)) |
| def side(Rm,Y,alpha,lam): |
| d=Y.shape[1]; Rc=Rm.tocoo(); YtY=Y.T@Y |
| O=(Y[Rc.col][:,:,None]*Y[Rc.col][:,None,:]).reshape(len(Rc.col),d*d) |
| S=sp.csr_matrix((np.ones(len(Rc.row),np.float32),(Rc.row,np.arange(len(Rc.row)))),shape=(Rm.shape[0],len(Rc.row))) |
| A=(S@O).reshape(-1,d,d)*alpha+YtY[None]+lam*np.eye(d)[None] |
| b=(1+alpha)*(Rm@Y) |
| return np.linalg.solve(A,b[...,None])[...,0] |
| rng=np.random.default_rng(0) |
| for d in (8,12,16): |
| for alpha in (5.,20.): |
| Y=rng.normal(0,0.1,(NI,d)).astype(np.float64) |
| for it in range(10): |
| X=side(R,Y,alpha,1.0); Y=side(R.T.tocsr(),X,alpha,1.0) |
| ev(cosf(Y),f'ALS d{d} a{alpha} item-cos',res=(d==12)) |
| |
| T=np.load('/tmp/dig/tr_ev.npz'); Vt=np.load('/tmp/dig/val_t.npz') |
| W=nrm(np.load('V_ab55.npy')) |
| su=np.unique(Vt['evuser']); m=np.isin(T['user'],su) |
| us=np.repeat(T['user'][m],10); ci=T['cand'][m].ravel(); yy=(np.arange(10)[None,:]==T['lab'][m][:,None]).ravel().astype(float) |
| order=np.argsort(us,kind='stable'); us,ci,yy=us[order],ci[order],yy[order] |
| uu,st=np.unique(us,return_index=True); en=np.r_[st[1:],len(us)] |
| ur={} |
| for lam in (1.,10.): |
| for u0,a0,b0 in zip(uu,st,en): |
| X=W[ci[a0:b0]]; t=yy[a0:b0]-yy[a0:b0].mean() |
| ur[u0]=np.linalg.solve(X.T@X+lam*np.eye(X.shape[1]),X.T@t) |
| Uv=np.array([ur.get(u,np.zeros(W.shape[1])) for u in Vt['evuser']]) |
| xr=(Uv[Vt['evix']]*W[Vt['cix']]).sum(1) |
| ev(xr,f'per-user ridge lam{lam} (imps+clicks)',res=False) |
| xc=cosf(np.load('V_ab55.npy')); ev(xc+0*xr,'history centroid cos (same V)',res=False) |
| print('stack centroid+ridge',resid(np.c_[xc,xr],2),' centroid alone',resid(xc,2)) |
|
|
